summary refs log tree commit diff
diff options
context:
space:
mode:
-rw-r--r--README.rst4
-rw-r--r--docs/postgres.rst6
-rw-r--r--docs/tcp_replication.rst223
-rw-r--r--docs/turn-howto.rst38
-rwxr-xr-xscripts-dev/nuke-room-from-db.sh43
-rwxr-xr-xscripts/synapse_port_db4
-rw-r--r--synapse/app/appservice.py50
-rw-r--r--synapse/app/client_reader.py22
-rw-r--r--synapse/app/federation_reader.py22
-rw-r--r--synapse/app/federation_sender.py190
-rwxr-xr-xsynapse/app/homeserver.py11
-rw-r--r--synapse/app/media_repository.py22
-rw-r--r--synapse/app/pusher.py122
-rw-r--r--synapse/app/synchrotron.py346
-rwxr-xr-xsynapse/app/synctl.py5
-rw-r--r--synapse/appservice/__init__.py38
-rw-r--r--synapse/config/emailconfig.py16
-rw-r--r--synapse/config/registration.py1
-rw-r--r--synapse/config/voip.py8
-rw-r--r--synapse/config/workers.py4
-rw-r--r--synapse/federation/send_queue.py308
-rw-r--r--synapse/handlers/federation.py43
-rw-r--r--synapse/handlers/message.py2
-rw-r--r--synapse/handlers/presence.py73
-rw-r--r--synapse/handlers/typing.py7
-rw-r--r--synapse/notifier.py34
-rw-r--r--synapse/push/mailer.py6
-rw-r--r--synapse/push/push_tools.py7
-rw-r--r--synapse/replication/resource.py4
-rw-r--r--synapse/replication/slave/storage/_base.py31
-rw-r--r--synapse/replication/slave/storage/account_data.py49
-rw-r--r--synapse/replication/slave/storage/deviceinbox.py23
-rw-r--r--synapse/replication/slave/storage/devices.py24
-rw-r--r--synapse/replication/slave/storage/events.py57
-rw-r--r--synapse/replication/slave/storage/presence.py19
-rw-r--r--synapse/replication/slave/storage/push_rule.py23
-rw-r--r--synapse/replication/slave/storage/pushers.py16
-rw-r--r--synapse/replication/slave/storage/receipts.py24
-rw-r--r--synapse/replication/slave/storage/room.py11
-rw-r--r--synapse/replication/tcp/__init__.py30
-rw-r--r--synapse/replication/tcp/client.py196
-rw-r--r--synapse/replication/tcp/commands.py346
-rw-r--r--synapse/replication/tcp/protocol.py640
-rw-r--r--synapse/replication/tcp/resource.py290
-rw-r--r--synapse/replication/tcp/streams.py464
-rw-r--r--synapse/rest/client/v1/voip.py5
-rw-r--r--synapse/rest/client/v2_alpha/thirdparty.py8
-rw-r--r--synapse/server.py4
-rw-r--r--synapse/storage/deviceinbox.py11
-rw-r--r--synapse/storage/devices.py6
-rw-r--r--synapse/storage/events.py88
-rw-r--r--synapse/storage/pusher.py42
-rw-r--r--synapse/storage/state.py36
-rw-r--r--synapse/types.py4
-rw-r--r--synapse/util/async.py7
-rw-r--r--synapse/util/caches/descriptors.py42
-rw-r--r--synapse/util/logcontext.py6
-rw-r--r--synapse/visibility.py3
-rw-r--r--tests/appservice/test_appservice.py4
-rw-r--r--tests/replication/slave/storage/_base.py30
-rw-r--r--tests/storage/test__base.py2
-rw-r--r--tests/util/caches/test_descriptors.py38
-rw-r--r--tests/util/test_snapshot_cache.py4
63 files changed, 3419 insertions, 823 deletions
diff --git a/README.rst b/README.rst
index b9c854ad48..759197f5ff 100644
--- a/README.rst
+++ b/README.rst
@@ -108,10 +108,10 @@ Installing prerequisites on ArchLinux::
     sudo pacman -S base-devel python2 python-pip \
                    python-setuptools python-virtualenv sqlite3
 
-Installing prerequisites on CentOS 7::
+Installing prerequisites on CentOS 7 or Fedora 25::
 
     sudo yum install libtiff-devel libjpeg-devel libzip-devel freetype-devel \
-                     lcms2-devel libwebp-devel tcl-devel tk-devel \
+                     lcms2-devel libwebp-devel tcl-devel tk-devel redhat-rpm-config \
                      python-virtualenv libffi-devel openssl-devel
     sudo yum groupinstall "Development Tools"
 
diff --git a/docs/postgres.rst b/docs/postgres.rst
index 402ff9a4de..b592801e93 100644
--- a/docs/postgres.rst
+++ b/docs/postgres.rst
@@ -112,9 +112,9 @@ script one last time, e.g. if the SQLite database is at  ``homeserver.db``
 run::
 
     synapse_port_db --sqlite-database homeserver.db \
-        --postgres-config database_config.yaml
+        --postgres-config homeserver-postgres.yaml
 
 Once that has completed, change the synapse config to point at the PostgreSQL
-database configuration file using the ``database_config`` parameter (see
-`Synapse Config`_) and restart synapse. Synapse should now be running against
+database configuration file ``homeserver-postgres.yaml`` (i.e. rename it to 
+``homeserver.yaml``) and restart synapse. Synapse should now be running against
 PostgreSQL.
diff --git a/docs/tcp_replication.rst b/docs/tcp_replication.rst
new file mode 100644
index 0000000000..62225ba6f4
--- /dev/null
+++ b/docs/tcp_replication.rst
@@ -0,0 +1,223 @@
+TCP Replication
+===============
+
+Motivation
+----------
+
+Previously the workers used an HTTP long poll mechanism to get updates from the
+master, which had the problem of causing a lot of duplicate work on the server.
+This TCP protocol replaces those APIs with the aim of increased efficiency.
+
+
+
+Overview
+--------
+
+The protocol is based on fire and forget, line based commands. An example flow
+would be (where '>' indicates master to worker and '<' worker to master flows)::
+
+    > SERVER example.com
+    < REPLICATE events 53
+    > RDATA events 54 ["$foo1:bar.com", ...]
+    > RDATA events 55 ["$foo4:bar.com", ...]
+
+The example shows the server accepting a new connection and sending its identity
+with the ``SERVER`` command, followed by the client asking to subscribe to the
+``events`` stream from the token ``53``. The server then periodically sends ``RDATA``
+commands which have the format ``RDATA <stream_name> <token> <row>``, where the
+format of ``<row>`` is defined by the individual streams.
+
+Error reporting happens by either the client or server sending an `ERROR`
+command, and usually the connection will be closed.
+
+
+Since the protocol is a simple line based, its possible to manually connect to
+the server using a tool like netcat. A few things should be noted when manually
+using the protocol:
+
+* When subscribing to a stream using ``REPLICATE``, the special token ``NOW`` can
+  be used to get all future updates. The special stream name ``ALL`` can be used
+  with ``NOW`` to subscribe to all available streams.
+* The federation stream is only available if federation sending has been
+  disabled on the main process.
+* The server will only time connections out that have sent a ``PING`` command.
+  If a ping is sent then the connection will be closed if no further commands
+  are receieved within 15s. Both the client and server protocol implementations
+  will send an initial PING on connection and ensure at least one command every
+  5s is sent (not necessarily ``PING``).
+* ``RDATA`` commands *usually* include a numeric token, however if the stream
+  has multiple rows to replicate per token the server will send multiple
+  ``RDATA`` commands, with all but the last having a token of ``batch``. See
+  the documentation on ``commands.RdataCommand`` for further details.
+
+
+Architecture
+------------
+
+The basic structure of the protocol is line based, where the initial word of
+each line specifies the command. The rest of the line is parsed based on the
+command. For example, the `RDATA` command is defined as::
+
+    RDATA <stream_name> <token> <row_json>
+
+(Note that `<row_json>` may contains spaces, but cannot contain newlines.)
+
+Blank lines are ignored.
+
+
+Keep alives
+~~~~~~~~~~~
+
+Both sides are expected to send at least one command every 5s or so, and
+should send a ``PING`` command if necessary. If either side do not receive a
+command within e.g. 15s then the connection should be closed.
+
+Because the server may be connected to manually using e.g. netcat, the timeouts
+aren't enabled until an initial ``PING`` command is seen. Both the client and
+server implementations below send a ``PING`` command immediately on connection to
+ensure the timeouts are enabled.
+
+This ensures that both sides can quickly realize if the tcp connection has gone
+and handle the situation appropriately.
+
+
+Start up
+~~~~~~~~
+
+When a new connection is made, the server:
+
+* Sends a ``SERVER`` command, which includes the identity of the server, allowing
+  the client to detect if its connected to the expected server
+* Sends a ``PING`` command as above, to enable the client to time out connections
+  promptly.
+
+The client:
+
+* Sends a ``NAME`` command, allowing the server to associate a human friendly
+  name with the connection. This is optional.
+* Sends a ``PING`` as above
+* For each stream the client wishes to subscribe to it sends a ``REPLICATE``
+  with the stream_name and token it wants to subscribe from.
+* On receipt of a ``SERVER`` command, checks that the server name matches the
+  expected server name.
+
+
+Error handling
+~~~~~~~~~~~~~~
+
+If either side detects an error it can send an ``ERROR`` command and close the
+connection.
+
+If the client side loses the connection to the server it should reconnect,
+following the steps above.
+
+
+Congestion
+~~~~~~~~~~
+
+If the server sends messages faster than the client can consume them the server
+will first buffer a (fairly large) number of commands and then disconnect the
+client. This ensures that we don't queue up an unbounded number of commands in
+memory and gives us a potential oppurtunity to squawk loudly. When/if the client
+recovers it can reconnect to the server and ask for missed messages.
+
+
+Reliability
+~~~~~~~~~~~
+
+In general the replication stream should be considered an unreliable transport
+since e.g. commands are not resent if the connection disappears.
+
+The exception to that are the replication streams, i.e. RDATA commands, since
+these include tokens which can be used to restart the stream on connection
+errors.
+
+The client should keep track of the token in the last RDATA command received
+for each stream so that on reconneciton it can start streaming from the correct
+place. Note: not all RDATA have valid tokens due to batching. See
+``RdataCommand`` for more details.
+
+
+Example
+~~~~~~~
+
+An example iteraction is shown below. Each line is prefixed with '>' or '<' to
+indicate which side is sending, these are *not* included on the wire::
+
+    * connection established *
+    > SERVER localhost:8823
+    > PING 1490197665618
+    < NAME synapse.app.appservice
+    < PING 1490197665618
+    < REPLICATE events 1
+    < REPLICATE backfill 1
+    < REPLICATE caches 1
+    > POSITION events 1
+    > POSITION backfill 1
+    > POSITION caches 1
+    > RDATA caches 2 ["get_user_by_id",["@01register-user:localhost:8823"],1490197670513]
+    > RDATA events 14 ["$149019767112vOHxz:localhost:8823",
+        "!AFDCvgApUmpdfVjIXm:localhost:8823","m.room.guest_access","",null]
+    < PING 1490197675618
+    > ERROR server stopping
+    * connection closed by server *
+
+The ``POSITION`` command sent by the server is used to set the clients position
+without needing to send data with the ``RDATA`` command.
+
+
+An example of a batched set of ``RDATA`` is::
+
+    > RDATA caches batch ["get_user_by_id",["@test:localhost:8823"],1490197670513]
+    > RDATA caches batch ["get_user_by_id",["@test2:localhost:8823"],1490197670513]
+    > RDATA caches batch ["get_user_by_id",["@test3:localhost:8823"],1490197670513]
+    > RDATA caches 54 ["get_user_by_id",["@test4:localhost:8823"],1490197670513]
+
+In this case the client shouldn't advance their caches token until it sees the
+the last ``RDATA``.
+
+
+List of commands
+~~~~~~~~~~~~~~~~
+
+The list of valid commands, with which side can send it: server (S) or client (C):
+
+SERVER (S)
+    Sent at the start to identify which server the client is talking to
+
+RDATA (S)
+    A single update in a stream
+
+POSITION (S)
+    The position of the stream has been updated
+
+ERROR (S, C)
+    There was an error
+
+PING (S, C)
+    Sent periodically to ensure the connection is still alive
+
+NAME (C)
+    Sent at the start by client to inform the server who they are
+
+REPLICATE (C)
+    Asks the server to replicate a given stream
+
+USER_SYNC (C)
+    A user has started or stopped syncing
+
+FEDERATION_ACK (C)
+    Acknowledge receipt of some federation data
+
+REMOVE_PUSHER (C)
+    Inform the server a pusher should be removed
+
+INVALIDATE_CACHE (C)
+    Inform the server a cache should be invalidated
+
+SYNC (S, C)
+    Used exclusively in tests
+
+
+See ``synapse/replication/tcp/commands.py`` for a detailed description and the
+format of each command.
diff --git a/docs/turn-howto.rst b/docs/turn-howto.rst
index 04c0100715..e48628ce6e 100644
--- a/docs/turn-howto.rst
+++ b/docs/turn-howto.rst
@@ -50,14 +50,37 @@ You may be able to setup coturn via your package manager,  or set it up manually
 
        pwgen -s 64 1
 
- 5. Ensure youe firewall allows traffic into the TURN server on
+ 5. Consider your security settings.  TURN lets users request a relay
+    which will connect to arbitrary IP addresses and ports.  At the least
+    we recommend:
+
+       # VoIP traffic is all UDP. There is no reason to let users connect to arbitrary TCP endpoints via the relay.
+       no-tcp-relay
+
+       # don't let the relay ever try to connect to private IP address ranges within your network (if any)
+       # given the turn server is likely behind your firewall, remember to include any privileged public IPs too.
+       denied-peer-ip=10.0.0.0-10.255.255.255
+       denied-peer-ip=192.168.0.0-192.168.255.255
+       denied-peer-ip=172.16.0.0-172.31.255.255
+
+       # special case the turn server itself so that client->TURN->TURN->client flows work
+       allowed-peer-ip=10.0.0.1
+
+       # consider whether you want to limit the quota of relayed streams per user (or total) to avoid risk of DoS.
+       user-quota=12 # 4 streams per video call, so 12 streams = 3 simultaneous relayed calls per user.
+       total-quota=1200
+
+    Ideally coturn should refuse to relay traffic which isn't SRTP;
+    see https://github.com/matrix-org/synapse/issues/2009
+
+ 6. Ensure your firewall allows traffic into the TURN server on
     the ports you've configured it to listen on (remember to allow
-    both TCP and UDP if you've enabled both).
+    both TCP and UDP TURN traffic)
 
- 6. If you've configured coturn to support TLS/DTLS, generate or
+ 7. If you've configured coturn to support TLS/DTLS, generate or
     import your private key and certificate.
 
- 7. Start the turn server::
+ 8. Start the turn server::
  
        bin/turnserver -o
 
@@ -83,12 +106,19 @@ Your home server configuration file needs the following extra keys:
     to refresh credentials. The TURN REST API specification recommends
     one day (86400000).
 
+  4. "turn_allow_guests": Whether to allow guest users to use the TURN
+    server.  This is enabled by default, as otherwise VoIP will not
+    work reliably for guests.  However, it does introduce a security risk
+    as it lets guests connect to arbitrary endpoints without having gone
+    through a CAPTCHA or similar to register a real account.
+
 As an example, here is the relevant section of the config file for
 matrix.org::
 
     turn_uris: [ "turn:turn.matrix.org:3478?transport=udp", "turn:turn.matrix.org:3478?transport=tcp" ]
     turn_shared_secret: n0t4ctuAllymatr1Xd0TorgSshar3d5ecret4obvIousreAsons
     turn_user_lifetime: 86400000
+    turn_allow_guests: True
 
 Now, restart synapse::
 
diff --git a/scripts-dev/nuke-room-from-db.sh b/scripts-dev/nuke-room-from-db.sh
index 58c036c896..1201d176c2 100755
--- a/scripts-dev/nuke-room-from-db.sh
+++ b/scripts-dev/nuke-room-from-db.sh
@@ -9,16 +9,39 @@
 ROOMID="$1"
 
 sqlite3 homeserver.db <<EOF
-DELETE FROM context_depth WHERE context = '$ROOMID';
-DELETE FROM current_state WHERE context = '$ROOMID';
-DELETE FROM feedback WHERE room_id = '$ROOMID';
-DELETE FROM messages WHERE room_id = '$ROOMID';
-DELETE FROM pdu_backward_extremities WHERE context = '$ROOMID';
-DELETE FROM pdu_edges WHERE context = '$ROOMID';
-DELETE FROM pdu_forward_extremities WHERE context = '$ROOMID';
-DELETE FROM pdus WHERE context = '$ROOMID';
-DELETE FROM room_data WHERE room_id = '$ROOMID';
+DELETE FROM event_forward_extremities WHERE room_id = '$ROOMID';
+DELETE FROM event_backward_extremities WHERE room_id = '$ROOMID';
+DELETE FROM event_edges WHERE room_id = '$ROOMID';
+DELETE FROM room_depth WHERE room_id = '$ROOMID';
+DELETE FROM state_forward_extremities WHERE room_id = '$ROOMID';
+DELETE FROM events WHERE room_id = '$ROOMID';
+DELETE FROM event_json WHERE room_id = '$ROOMID';
+DELETE FROM state_events WHERE room_id = '$ROOMID';
+DELETE FROM current_state_events WHERE room_id = '$ROOMID';
 DELETE FROM room_memberships WHERE room_id = '$ROOMID';
+DELETE FROM feedback WHERE room_id = '$ROOMID';
+DELETE FROM topics WHERE room_id = '$ROOMID';
+DELETE FROM room_names WHERE room_id = '$ROOMID';
 DELETE FROM rooms WHERE room_id = '$ROOMID';
-DELETE FROM state_pdus WHERE context = '$ROOMID';
+DELETE FROM room_hosts WHERE room_id = '$ROOMID';
+DELETE FROM room_aliases WHERE room_id = '$ROOMID';
+DELETE FROM state_groups WHERE room_id = '$ROOMID';
+DELETE FROM state_groups_state WHERE room_id = '$ROOMID';
+DELETE FROM receipts_graph WHERE room_id = '$ROOMID';
+DELETE FROM receipts_linearized WHERE room_id = '$ROOMID';
+DELETE FROM event_search_content WHERE c1room_id = '$ROOMID';
+DELETE FROM guest_access WHERE room_id = '$ROOMID';
+DELETE FROM history_visibility WHERE room_id = '$ROOMID';
+DELETE FROM room_tags WHERE room_id = '$ROOMID';
+DELETE FROM room_tags_revisions WHERE room_id = '$ROOMID';
+DELETE FROM room_account_data WHERE room_id = '$ROOMID';
+DELETE FROM event_push_actions WHERE room_id = '$ROOMID';
+DELETE FROM local_invites WHERE room_id = '$ROOMID';
+DELETE FROM pusher_throttle WHERE room_id = '$ROOMID';
+DELETE FROM event_reports WHERE room_id = '$ROOMID';
+DELETE FROM public_room_list_stream WHERE room_id = '$ROOMID';
+DELETE FROM stream_ordering_to_exterm WHERE room_id = '$ROOMID';
+DELETE FROM event_auth WHERE room_id = '$ROOMID';
+DELETE FROM appservice_room_list WHERE room_id = '$ROOMID';
+VACUUM;
 EOF
diff --git a/scripts/synapse_port_db b/scripts/synapse_port_db
index ea367a1281..2e5d666707 100755
--- a/scripts/synapse_port_db
+++ b/scripts/synapse_port_db
@@ -447,9 +447,7 @@ class Porter(object):
 
             postgres_tables = yield self.postgres_store._simple_select_onecol(
                 table="information_schema.tables",
-                keyvalues={
-                    "table_schema": "public",
-                },
+                keyvalues={},
                 retcol="distinct table_name",
             )
 
diff --git a/synapse/app/appservice.py b/synapse/app/appservice.py
index a6f1e7594e..9a476efa63 100644
--- a/synapse/app/appservice.py
+++ b/synapse/app/appservice.py
@@ -26,17 +26,17 @@ from synapse.replication.slave.storage.directory import DirectoryStore
 from synapse.replication.slave.storage.events import SlavedEventStore
 from synapse.replication.slave.storage.appservice import SlavedApplicationServiceStore
 from synapse.replication.slave.storage.registration import SlavedRegistrationStore
+from synapse.replication.tcp.client import ReplicationClientHandler
 from synapse.storage.engines import create_engine
-from synapse.util.async import sleep
 from synapse.util.httpresourcetree import create_resource_tree
-from synapse.util.logcontext import LoggingContext, PreserveLoggingContext
+from synapse.util.logcontext import LoggingContext, PreserveLoggingContext, preserve_fn
 from synapse.util.manhole import manhole
 from synapse.util.rlimit import change_resource_limit
 from synapse.util.versionstring import get_version_string
 
 from synapse import events
 
-from twisted.internet import reactor, defer
+from twisted.internet import reactor
 from twisted.web.resource import Resource
 
 from daemonize import Daemonize
@@ -120,30 +120,25 @@ class AppserviceServer(HomeServer):
             else:
                 logger.warn("Unrecognized listener type: %s", listener["type"])
 
-    @defer.inlineCallbacks
-    def replicate(self):
-        http_client = self.get_simple_http_client()
-        store = self.get_datastore()
-        replication_url = self.config.worker_replication_url
-        appservice_handler = self.get_application_service_handler()
-
-        @defer.inlineCallbacks
-        def replicate(results):
-            stream = results.get("events")
-            if stream:
-                max_stream_id = stream["position"]
-                yield appservice_handler.notify_interested_services(max_stream_id)
-
-        while True:
-            try:
-                args = store.stream_positions()
-                args["timeout"] = 30000
-                result = yield http_client.get_json(replication_url, args=args)
-                yield store.process_replication(result)
-                replicate(result)
-            except:
-                logger.exception("Error replicating from %r", replication_url)
-                yield sleep(30)
+        self.get_tcp_replication().start_replication(self)
+
+    def build_tcp_replication(self):
+        return ASReplicationHandler(self)
+
+
+class ASReplicationHandler(ReplicationClientHandler):
+    def __init__(self, hs):
+        super(ASReplicationHandler, self).__init__(hs.get_datastore())
+        self.appservice_handler = hs.get_application_service_handler()
+
+    def on_rdata(self, stream_name, token, rows):
+        super(ASReplicationHandler, self).on_rdata(stream_name, token, rows)
+
+        if stream_name == "events":
+            max_stream_id = self.store.get_room_max_stream_ordering()
+            preserve_fn(
+                self.appservice_handler.notify_interested_services
+            )(max_stream_id)
 
 
 def start(config_options):
@@ -199,7 +194,6 @@ def start(config_options):
             reactor.run()
 
     def start():
-        ps.replicate()
         ps.get_datastore().start_profiling()
         ps.get_state_handler().start_caching()
 
diff --git a/synapse/app/client_reader.py b/synapse/app/client_reader.py
index e4ea3ab933..9b72c649ac 100644
--- a/synapse/app/client_reader.py
+++ b/synapse/app/client_reader.py
@@ -30,11 +30,11 @@ from synapse.replication.slave.storage.room import RoomStore
 from synapse.replication.slave.storage.directory import DirectoryStore
 from synapse.replication.slave.storage.registration import SlavedRegistrationStore
 from synapse.replication.slave.storage.transactions import TransactionStore
+from synapse.replication.tcp.client import ReplicationClientHandler
 from synapse.rest.client.v1.room import PublicRoomListRestServlet
 from synapse.server import HomeServer
 from synapse.storage.client_ips import ClientIpStore
 from synapse.storage.engines import create_engine
-from synapse.util.async import sleep
 from synapse.util.httpresourcetree import create_resource_tree
 from synapse.util.logcontext import LoggingContext, PreserveLoggingContext
 from synapse.util.manhole import manhole
@@ -45,7 +45,7 @@ from synapse.crypto import context_factory
 from synapse import events
 
 
-from twisted.internet import reactor, defer
+from twisted.internet import reactor
 from twisted.web.resource import Resource
 
 from daemonize import Daemonize
@@ -145,21 +145,10 @@ class ClientReaderServer(HomeServer):
             else:
                 logger.warn("Unrecognized listener type: %s", listener["type"])
 
-    @defer.inlineCallbacks
-    def replicate(self):
-        http_client = self.get_simple_http_client()
-        store = self.get_datastore()
-        replication_url = self.config.worker_replication_url
+        self.get_tcp_replication().start_replication(self)
 
-        while True:
-            try:
-                args = store.stream_positions()
-                args["timeout"] = 30000
-                result = yield http_client.get_json(replication_url, args=args)
-                yield store.process_replication(result)
-            except:
-                logger.exception("Error replicating from %r", replication_url)
-                yield sleep(5)
+    def build_tcp_replication(self):
+        return ReplicationClientHandler(self.get_datastore())
 
 
 def start(config_options):
@@ -209,7 +198,6 @@ def start(config_options):
     def start():
         ss.get_state_handler().start_caching()
         ss.get_datastore().start_profiling()
-        ss.replicate()
 
     reactor.callWhenRunning(start)
 
diff --git a/synapse/app/federation_reader.py b/synapse/app/federation_reader.py
index e52b0f240d..eb392e1c9d 100644
--- a/synapse/app/federation_reader.py
+++ b/synapse/app/federation_reader.py
@@ -27,9 +27,9 @@ from synapse.replication.slave.storage.keys import SlavedKeyStore
 from synapse.replication.slave.storage.room import RoomStore
 from synapse.replication.slave.storage.transactions import TransactionStore
 from synapse.replication.slave.storage.directory import DirectoryStore
+from synapse.replication.tcp.client import ReplicationClientHandler
 from synapse.server import HomeServer
 from synapse.storage.engines import create_engine
-from synapse.util.async import sleep
 from synapse.util.httpresourcetree import create_resource_tree
 from synapse.util.logcontext import LoggingContext, PreserveLoggingContext
 from synapse.util.manhole import manhole
@@ -42,7 +42,7 @@ from synapse.crypto import context_factory
 from synapse import events
 
 
-from twisted.internet import reactor, defer
+from twisted.internet import reactor
 from twisted.web.resource import Resource
 
 from daemonize import Daemonize
@@ -134,21 +134,10 @@ class FederationReaderServer(HomeServer):
             else:
                 logger.warn("Unrecognized listener type: %s", listener["type"])
 
-    @defer.inlineCallbacks
-    def replicate(self):
-        http_client = self.get_simple_http_client()
-        store = self.get_datastore()
-        replication_url = self.config.worker_replication_url
+        self.get_tcp_replication().start_replication(self)
 
-        while True:
-            try:
-                args = store.stream_positions()
-                args["timeout"] = 30000
-                result = yield http_client.get_json(replication_url, args=args)
-                yield store.process_replication(result)
-            except:
-                logger.exception("Error replicating from %r", replication_url)
-                yield sleep(5)
+    def build_tcp_replication(self):
+        return ReplicationClientHandler(self.get_datastore())
 
 
 def start(config_options):
@@ -198,7 +187,6 @@ def start(config_options):
     def start():
         ss.get_state_handler().start_caching()
         ss.get_datastore().start_profiling()
-        ss.replicate()
 
     reactor.callWhenRunning(start)
 
diff --git a/synapse/app/federation_sender.py b/synapse/app/federation_sender.py
index 76c4cc54d1..477e16e0fa 100644
--- a/synapse/app/federation_sender.py
+++ b/synapse/app/federation_sender.py
@@ -23,7 +23,6 @@ from synapse.config.homeserver import HomeServerConfig
 from synapse.crypto import context_factory
 from synapse.http.site import SynapseSite
 from synapse.federation import send_queue
-from synapse.federation.units import Edu
 from synapse.metrics.resource import MetricsResource, METRICS_PREFIX
 from synapse.replication.slave.storage.deviceinbox import SlavedDeviceInboxStore
 from synapse.replication.slave.storage.events import SlavedEventStore
@@ -31,11 +30,11 @@ from synapse.replication.slave.storage.receipts import SlavedReceiptsStore
 from synapse.replication.slave.storage.registration import SlavedRegistrationStore
 from synapse.replication.slave.storage.transactions import TransactionStore
 from synapse.replication.slave.storage.devices import SlavedDeviceStore
+from synapse.replication.tcp.client import ReplicationClientHandler
 from synapse.storage.engines import create_engine
-from synapse.storage.presence import UserPresenceState
-from synapse.util.async import sleep
+from synapse.util.async import Linearizer
 from synapse.util.httpresourcetree import create_resource_tree
-from synapse.util.logcontext import LoggingContext, PreserveLoggingContext
+from synapse.util.logcontext import LoggingContext, PreserveLoggingContext, preserve_fn
 from synapse.util.manhole import manhole
 from synapse.util.rlimit import change_resource_limit
 from synapse.util.versionstring import get_version_string
@@ -50,7 +49,6 @@ from daemonize import Daemonize
 import sys
 import logging
 import gc
-import ujson as json
 
 logger = logging.getLogger("synapse.app.appservice")
 
@@ -59,7 +57,28 @@ class FederationSenderSlaveStore(
     SlavedDeviceInboxStore, TransactionStore, SlavedReceiptsStore, SlavedEventStore,
     SlavedRegistrationStore, SlavedDeviceStore,
 ):
-    pass
+    def __init__(self, db_conn, hs):
+        super(FederationSenderSlaveStore, self).__init__(db_conn, hs)
+
+        # We pull out the current federation stream position now so that we
+        # always have a known value for the federation position in memory so
+        # that we don't have to bounce via a deferred once when we start the
+        # replication streams.
+        self.federation_out_pos_startup = self._get_federation_out_pos(db_conn)
+
+    def _get_federation_out_pos(self, db_conn):
+        sql = (
+            "SELECT stream_id FROM federation_stream_position"
+            " WHERE type = ?"
+        )
+        sql = self.database_engine.convert_param_style(sql)
+
+        txn = db_conn.cursor()
+        txn.execute(sql, ("federation",))
+        rows = txn.fetchall()
+        txn.close()
+
+        return rows[0][0] if rows else -1
 
 
 class FederationSenderServer(HomeServer):
@@ -127,26 +146,27 @@ class FederationSenderServer(HomeServer):
             else:
                 logger.warn("Unrecognized listener type: %s", listener["type"])
 
-    @defer.inlineCallbacks
-    def replicate(self):
-        http_client = self.get_simple_http_client()
-        store = self.get_datastore()
-        replication_url = self.config.worker_replication_url
-        send_handler = FederationSenderHandler(self)
-
-        send_handler.on_start()
-
-        while True:
-            try:
-                args = store.stream_positions()
-                args.update((yield send_handler.stream_positions()))
-                args["timeout"] = 30000
-                result = yield http_client.get_json(replication_url, args=args)
-                yield store.process_replication(result)
-                yield send_handler.process_replication(result)
-            except:
-                logger.exception("Error replicating from %r", replication_url)
-                yield sleep(30)
+        self.get_tcp_replication().start_replication(self)
+
+    def build_tcp_replication(self):
+        return FederationSenderReplicationHandler(self)
+
+
+class FederationSenderReplicationHandler(ReplicationClientHandler):
+    def __init__(self, hs):
+        super(FederationSenderReplicationHandler, self).__init__(hs.get_datastore())
+        self.send_handler = FederationSenderHandler(hs, self)
+
+    def on_rdata(self, stream_name, token, rows):
+        super(FederationSenderReplicationHandler, self).on_rdata(
+            stream_name, token, rows
+        )
+        self.send_handler.process_replication_rows(stream_name, token, rows)
+
+    def get_streams_to_replicate(self):
+        args = super(FederationSenderReplicationHandler, self).get_streams_to_replicate()
+        args.update(self.send_handler.stream_positions())
+        return args
 
 
 def start(config_options):
@@ -205,7 +225,6 @@ def start(config_options):
             reactor.run()
 
     def start():
-        ps.replicate()
         ps.get_datastore().start_profiling()
         ps.get_state_handler().start_caching()
 
@@ -229,9 +248,15 @@ class FederationSenderHandler(object):
     """Processes the replication stream and forwards the appropriate entries
     to the federation sender.
     """
-    def __init__(self, hs):
+    def __init__(self, hs, replication_client):
         self.store = hs.get_datastore()
         self.federation_sender = hs.get_federation_sender()
+        self.replication_client = replication_client
+
+        self.federation_position = self.store.federation_out_pos_startup
+        self._fed_position_linearizer = Linearizer(name="_fed_position_linearizer")
+
+        self._last_ack = self.federation_position
 
         self._room_serials = {}
         self._room_typing = {}
@@ -243,98 +268,35 @@ class FederationSenderHandler(object):
             self.store.get_room_max_stream_ordering()
         )
 
-    @defer.inlineCallbacks
     def stream_positions(self):
-        stream_id = yield self.store.get_federation_out_pos("federation")
-        defer.returnValue({
-            "federation": stream_id,
+        return {"federation": self.federation_position}
 
-            # Ack stuff we've "processed", this should only be called from
-            # one process.
-            "federation_ack": stream_id,
-        })
-
-    @defer.inlineCallbacks
-    def process_replication(self, result):
+    def process_replication_rows(self, stream_name, token, rows):
         # The federation stream contains things that we want to send out, e.g.
         # presence, typing, etc.
-        fed_stream = result.get("federation")
-        if fed_stream:
-            latest_id = int(fed_stream["position"])
-
-            # The federation stream containis a bunch of different types of
-            # rows that need to be handled differently. We parse the rows, put
-            # them into the appropriate collection and then send them off.
-            presence_to_send = {}
-            keyed_edus = {}
-            edus = {}
-            failures = {}
-            device_destinations = set()
-
-            # Parse the rows in the stream
-            for row in fed_stream["rows"]:
-                position, typ, content_js = row
-                content = json.loads(content_js)
-
-                if typ == send_queue.PRESENCE_TYPE:
-                    destination = content["destination"]
-                    state = UserPresenceState.from_dict(content["state"])
-
-                    presence_to_send.setdefault(destination, []).append(state)
-                elif typ == send_queue.KEYED_EDU_TYPE:
-                    key = content["key"]
-                    edu = Edu(**content["edu"])
-
-                    keyed_edus.setdefault(
-                        edu.destination, {}
-                    )[(edu.destination, tuple(key))] = edu
-                elif typ == send_queue.EDU_TYPE:
-                    edu = Edu(**content)
-
-                    edus.setdefault(edu.destination, []).append(edu)
-                elif typ == send_queue.FAILURE_TYPE:
-                    destination = content["destination"]
-                    failure = content["failure"]
-
-                    failures.setdefault(destination, []).append(failure)
-                elif typ == send_queue.DEVICE_MESSAGE_TYPE:
-                    device_destinations.add(content["destination"])
-                else:
-                    raise Exception("Unrecognised federation type: %r", typ)
-
-            # We've finished collecting, send everything off
-            for destination, states in presence_to_send.items():
-                self.federation_sender.send_presence(destination, states)
-
-            for destination, edu_map in keyed_edus.items():
-                for key, edu in edu_map.items():
-                    self.federation_sender.send_edu(
-                        edu.destination, edu.edu_type, edu.content, key=key,
-                    )
-
-            for destination, edu_list in edus.items():
-                for edu in edu_list:
-                    self.federation_sender.send_edu(
-                        edu.destination, edu.edu_type, edu.content, key=None,
-                    )
-
-            for destination, failure_list in failures.items():
-                for failure in failure_list:
-                    self.federation_sender.send_failure(destination, failure)
-
-            for destination in device_destinations:
-                self.federation_sender.send_device_messages(destination)
-
-            # Record where we are in the stream.
-            yield self.store.update_federation_out_pos(
-                "federation", latest_id
-            )
+        if stream_name == "federation":
+            send_queue.process_rows_for_federation(self.federation_sender, rows)
+            preserve_fn(self.update_token)(token)
 
         # We also need to poke the federation sender when new events happen
-        event_stream = result.get("events")
-        if event_stream:
-            latest_pos = event_stream["position"]
-            self.federation_sender.notify_new_events(latest_pos)
+        elif stream_name == "events":
+            self.federation_sender.notify_new_events(token)
+
+    @defer.inlineCallbacks
+    def update_token(self, token):
+        self.federation_position = token
+
+        # We linearize here to ensure we don't have races updating the token
+        with (yield self._fed_position_linearizer.queue(None)):
+            if self._last_ack < self.federation_position:
+                yield self.store.update_federation_out_pos(
+                    "federation", self.federation_position
+                )
+
+                # We ACK this token over replication so that the master can drop
+                # its in memory queues
+                self.replication_client.send_federation_ack(self.federation_position)
+                self._last_ack = self.federation_position
 
 
 if __name__ == '__main__':
diff --git a/synapse/app/homeserver.py b/synapse/app/homeserver.py
index 2cdd2d39ff..990eb477e5 100755
--- a/synapse/app/homeserver.py
+++ b/synapse/app/homeserver.py
@@ -56,6 +56,7 @@ from synapse.util.logcontext import LoggingContext, PreserveLoggingContext
 from synapse.metrics import register_memory_metrics, get_metrics_for
 from synapse.metrics.resource import MetricsResource, METRICS_PREFIX
 from synapse.replication.resource import ReplicationResource, REPLICATION_PREFIX
+from synapse.replication.tcp.resource import ReplicationStreamProtocolFactory
 from synapse.federation.transport.server import TransportLayerServer
 
 from synapse.util.rlimit import change_resource_limit
@@ -222,6 +223,16 @@ class SynapseHomeServer(HomeServer):
                         ),
                         interface=address
                     )
+            elif listener["type"] == "replication":
+                bind_addresses = listener["bind_addresses"]
+                for address in bind_addresses:
+                    factory = ReplicationStreamProtocolFactory(self)
+                    server_listener = reactor.listenTCP(
+                        listener["port"], factory, interface=address
+                    )
+                    reactor.addSystemEventTrigger(
+                        "before", "shutdown", server_listener.stopListening,
+                    )
             else:
                 logger.warn("Unrecognized listener type: %s", listener["type"])
 
diff --git a/synapse/app/media_repository.py b/synapse/app/media_repository.py
index 1444e69a42..26c4416956 100644
--- a/synapse/app/media_repository.py
+++ b/synapse/app/media_repository.py
@@ -25,13 +25,13 @@ from synapse.replication.slave.storage._base import BaseSlavedStore
 from synapse.replication.slave.storage.appservice import SlavedApplicationServiceStore
 from synapse.replication.slave.storage.registration import SlavedRegistrationStore
 from synapse.replication.slave.storage.transactions import TransactionStore
+from synapse.replication.tcp.client import ReplicationClientHandler
 from synapse.rest.media.v0.content_repository import ContentRepoResource
 from synapse.rest.media.v1.media_repository import MediaRepositoryResource
 from synapse.server import HomeServer
 from synapse.storage.client_ips import ClientIpStore
 from synapse.storage.engines import create_engine
 from synapse.storage.media_repository import MediaRepositoryStore
-from synapse.util.async import sleep
 from synapse.util.httpresourcetree import create_resource_tree
 from synapse.util.logcontext import LoggingContext, PreserveLoggingContext
 from synapse.util.manhole import manhole
@@ -45,7 +45,7 @@ from synapse.crypto import context_factory
 from synapse import events
 
 
-from twisted.internet import reactor, defer
+from twisted.internet import reactor
 from twisted.web.resource import Resource
 
 from daemonize import Daemonize
@@ -142,21 +142,10 @@ class MediaRepositoryServer(HomeServer):
             else:
                 logger.warn("Unrecognized listener type: %s", listener["type"])
 
-    @defer.inlineCallbacks
-    def replicate(self):
-        http_client = self.get_simple_http_client()
-        store = self.get_datastore()
-        replication_url = self.config.worker_replication_url
+        self.get_tcp_replication().start_replication(self)
 
-        while True:
-            try:
-                args = store.stream_positions()
-                args["timeout"] = 30000
-                result = yield http_client.get_json(replication_url, args=args)
-                yield store.process_replication(result)
-            except:
-                logger.exception("Error replicating from %r", replication_url)
-                yield sleep(5)
+    def build_tcp_replication(self):
+        return ReplicationClientHandler(self.get_datastore())
 
 
 def start(config_options):
@@ -206,7 +195,6 @@ def start(config_options):
     def start():
         ss.get_state_handler().start_caching()
         ss.get_datastore().start_profiling()
-        ss.replicate()
 
     reactor.callWhenRunning(start)
 
diff --git a/synapse/app/pusher.py b/synapse/app/pusher.py
index ab682e52ec..f9114acfcb 100644
--- a/synapse/app/pusher.py
+++ b/synapse/app/pusher.py
@@ -27,9 +27,9 @@ from synapse.replication.slave.storage.events import SlavedEventStore
 from synapse.replication.slave.storage.pushers import SlavedPusherStore
 from synapse.replication.slave.storage.receipts import SlavedReceiptsStore
 from synapse.replication.slave.storage.account_data import SlavedAccountDataStore
+from synapse.replication.tcp.client import ReplicationClientHandler
 from synapse.storage.engines import create_engine
 from synapse.storage import DataStore
-from synapse.util.async import sleep
 from synapse.util.httpresourcetree import create_resource_tree
 from synapse.util.logcontext import LoggingContext, preserve_fn, \
     PreserveLoggingContext
@@ -89,7 +89,6 @@ class PusherSlaveStore(
 
 
 class PusherServer(HomeServer):
-
     def get_db_conn(self, run_new_connection=True):
         # Any param beginning with cp_ is a parameter for adbapi, and should
         # not be passed to the database engine.
@@ -109,16 +108,7 @@ class PusherServer(HomeServer):
         logger.info("Finished setting up.")
 
     def remove_pusher(self, app_id, push_key, user_id):
-        http_client = self.get_simple_http_client()
-        replication_url = self.config.worker_replication_url
-        url = replication_url + "/remove_pushers"
-        return http_client.post_json_get_json(url, {
-            "remove": [{
-                "app_id": app_id,
-                "push_key": push_key,
-                "user_id": user_id,
-            }]
-        })
+        self.get_tcp_replication().send_remove_pusher(app_id, push_key, user_id)
 
     def _listen_http(self, listener_config):
         port = listener_config["port"]
@@ -166,73 +156,52 @@ class PusherServer(HomeServer):
             else:
                 logger.warn("Unrecognized listener type: %s", listener["type"])
 
+        self.get_tcp_replication().start_replication(self)
+
+    def build_tcp_replication(self):
+        return PusherReplicationHandler(self)
+
+
+class PusherReplicationHandler(ReplicationClientHandler):
+    def __init__(self, hs):
+        super(PusherReplicationHandler, self).__init__(hs.get_datastore())
+
+        self.pusher_pool = hs.get_pusherpool()
+
+    def on_rdata(self, stream_name, token, rows):
+        super(PusherReplicationHandler, self).on_rdata(stream_name, token, rows)
+        preserve_fn(self.poke_pushers)(stream_name, token, rows)
+
     @defer.inlineCallbacks
-    def replicate(self):
-        http_client = self.get_simple_http_client()
-        store = self.get_datastore()
-        replication_url = self.config.worker_replication_url
-        pusher_pool = self.get_pusherpool()
-
-        def stop_pusher(user_id, app_id, pushkey):
-            key = "%s:%s" % (app_id, pushkey)
-            pushers_for_user = pusher_pool.pushers.get(user_id, {})
-            pusher = pushers_for_user.pop(key, None)
-            if pusher is None:
-                return
-            logger.info("Stopping pusher %r / %r", user_id, key)
-            pusher.on_stop()
-
-        def start_pusher(user_id, app_id, pushkey):
-            key = "%s:%s" % (app_id, pushkey)
-            logger.info("Starting pusher %r / %r", user_id, key)
-            return pusher_pool._refresh_pusher(app_id, pushkey, user_id)
-
-        @defer.inlineCallbacks
-        def poke_pushers(results):
-            pushers_rows = set(
-                map(tuple, results.get("pushers", {}).get("rows", []))
+    def poke_pushers(self, stream_name, token, rows):
+        if stream_name == "pushers":
+            for row in rows:
+                if row.deleted:
+                    yield self.stop_pusher(row.user_id, row.app_id, row.pushkey)
+                else:
+                    yield self.start_pusher(row.user_id, row.app_id, row.pushkey)
+        elif stream_name == "events":
+            yield self.pusher_pool.on_new_notifications(
+                token, token,
             )
-            deleted_pushers_rows = set(
-                map(tuple, results.get("deleted_pushers", {}).get("rows", []))
+        elif stream_name == "receipts":
+            yield self.pusher_pool.on_new_receipts(
+                token, token, set(row.room_id for row in rows)
             )
-            for row in sorted(pushers_rows | deleted_pushers_rows):
-                if row in deleted_pushers_rows:
-                    user_id, app_id, pushkey = row[1:4]
-                    stop_pusher(user_id, app_id, pushkey)
-                elif row in pushers_rows:
-                    user_id = row[1]
-                    app_id = row[5]
-                    pushkey = row[8]
-                    yield start_pusher(user_id, app_id, pushkey)
-
-            stream = results.get("events")
-            if stream and stream["rows"]:
-                min_stream_id = stream["rows"][0][0]
-                max_stream_id = stream["position"]
-                preserve_fn(pusher_pool.on_new_notifications)(
-                    min_stream_id, max_stream_id
-                )
-
-            stream = results.get("receipts")
-            if stream and stream["rows"]:
-                rows = stream["rows"]
-                affected_room_ids = set(row[1] for row in rows)
-                min_stream_id = rows[0][0]
-                max_stream_id = stream["position"]
-                preserve_fn(pusher_pool.on_new_receipts)(
-                    min_stream_id, max_stream_id, affected_room_ids
-                )
-
-        while True:
-            try:
-                args = store.stream_positions()
-                args["timeout"] = 30000
-                result = yield http_client.get_json(replication_url, args=args)
-                yield store.process_replication(result)
-                poke_pushers(result)
-            except:
-                logger.exception("Error replicating from %r", replication_url)
-                yield sleep(30)
+
+    def stop_pusher(self, user_id, app_id, pushkey):
+        key = "%s:%s" % (app_id, pushkey)
+        pushers_for_user = self.pusher_pool.pushers.get(user_id, {})
+        pusher = pushers_for_user.pop(key, None)
+        if pusher is None:
+            return
+        logger.info("Stopping pusher %r / %r", user_id, key)
+        pusher.on_stop()
+
+    def start_pusher(self, user_id, app_id, pushkey):
+        key = "%s:%s" % (app_id, pushkey)
+        logger.info("Starting pusher %r / %r", user_id, key)
+        return self.pusher_pool._refresh_pusher(app_id, pushkey, user_id)
 
 
 def start(config_options):
@@ -288,7 +257,6 @@ def start(config_options):
             reactor.run()
 
     def start():
-        ps.replicate()
         ps.get_pusherpool().start()
         ps.get_datastore().start_profiling()
         ps.get_state_handler().start_caching()
diff --git a/synapse/app/synchrotron.py b/synapse/app/synchrotron.py
index 34e34e5580..d39e3161fe 100644
--- a/synapse/app/synchrotron.py
+++ b/synapse/app/synchrotron.py
@@ -16,7 +16,7 @@
 
 import synapse
 
-from synapse.api.constants import EventTypes, PresenceState
+from synapse.api.constants import EventTypes
 from synapse.config._base import ConfigError
 from synapse.config.homeserver import HomeServerConfig
 from synapse.config.logger import setup_logging
@@ -40,15 +40,14 @@ from synapse.replication.slave.storage.presence import SlavedPresenceStore
 from synapse.replication.slave.storage.deviceinbox import SlavedDeviceInboxStore
 from synapse.replication.slave.storage.devices import SlavedDeviceStore
 from synapse.replication.slave.storage.room import RoomStore
+from synapse.replication.tcp.client import ReplicationClientHandler
 from synapse.server import HomeServer
 from synapse.storage.client_ips import ClientIpStore
 from synapse.storage.engines import create_engine
 from synapse.storage.presence import PresenceStore, UserPresenceState
 from synapse.storage.roommember import RoomMemberStore
-from synapse.util.async import sleep
 from synapse.util.httpresourcetree import create_resource_tree
-from synapse.util.logcontext import LoggingContext, preserve_fn, \
-    PreserveLoggingContext
+from synapse.util.logcontext import LoggingContext, PreserveLoggingContext, preserve_fn
 from synapse.util.manhole import manhole
 from synapse.util.rlimit import change_resource_limit
 from synapse.util.stringutils import random_string
@@ -63,7 +62,6 @@ import sys
 import logging
 import contextlib
 import gc
-import ujson as json
 
 logger = logging.getLogger("synapse.app.synchrotron")
 
@@ -107,11 +105,11 @@ UPDATE_SYNCING_USERS_MS = 10 * 1000
 
 class SynchrotronPresence(object):
     def __init__(self, hs):
+        self.hs = hs
         self.is_mine_id = hs.is_mine_id
         self.http_client = hs.get_simple_http_client()
         self.store = hs.get_datastore()
         self.user_to_num_current_syncs = {}
-        self.syncing_users_url = hs.config.worker_replication_url + "/syncing_users"
         self.clock = hs.get_clock()
         self.notifier = hs.get_notifier()
 
@@ -121,17 +119,52 @@ class SynchrotronPresence(object):
             for state in active_presence
         }
 
-        self.process_id = random_string(16)
-        logger.info("Presence process_id is %r", self.process_id)
+        # user_id -> last_sync_ms. Lists the users that have stopped syncing
+        # but we haven't notified the master of that yet
+        self.users_going_offline = {}
 
-        self._sending_sync = False
-        self._need_to_send_sync = False
-        self.clock.looping_call(
-            self._send_syncing_users_regularly,
-            UPDATE_SYNCING_USERS_MS,
+        self._send_stop_syncing_loop = self.clock.looping_call(
+            self.send_stop_syncing, 10 * 1000
         )
 
-        reactor.addSystemEventTrigger("before", "shutdown", self._on_shutdown)
+        self.process_id = random_string(16)
+        logger.info("Presence process_id is %r", self.process_id)
+
+    def send_user_sync(self, user_id, is_syncing, last_sync_ms):
+        self.hs.get_tcp_replication().send_user_sync(user_id, is_syncing, last_sync_ms)
+
+    def mark_as_coming_online(self, user_id):
+        """A user has started syncing. Send a UserSync to the master, unless they
+        had recently stopped syncing.
+
+        Args:
+            user_id (str)
+        """
+        going_offline = self.users_going_offline.pop(user_id, None)
+        if not going_offline:
+            # Safe to skip because we haven't yet told the master they were offline
+            self.send_user_sync(user_id, True, self.clock.time_msec())
+
+    def mark_as_going_offline(self, user_id):
+        """A user has stopped syncing. We wait before notifying the master as
+        its likely they'll come back soon. This allows us to avoid sending
+        a stopped syncing immediately followed by a started syncing notification
+        to the master
+
+        Args:
+            user_id (str)
+        """
+        self.users_going_offline[user_id] = self.clock.time_msec()
+
+    def send_stop_syncing(self):
+        """Check if there are any users who have stopped syncing a while ago
+        and haven't come back yet. If there are poke the master about them.
+        """
+        now = self.clock.time_msec()
+        for user_id, last_sync_ms in self.users_going_offline.items():
+            if now - last_sync_ms > 10 * 1000:
+                self.users_going_offline.pop(user_id, None)
+                self.send_user_sync(user_id, False, last_sync_ms)
 
     def set_state(self, user, state, ignore_status_msg=False):
         # TODO Hows this supposed to work?
@@ -142,15 +175,14 @@ class SynchrotronPresence(object):
     _get_interested_parties = PresenceHandler._get_interested_parties.__func__
     current_state_for_users = PresenceHandler.current_state_for_users.__func__
 
-    @defer.inlineCallbacks
     def user_syncing(self, user_id, affect_presence):
         if affect_presence:
             curr_sync = self.user_to_num_current_syncs.get(user_id, 0)
             self.user_to_num_current_syncs[user_id] = curr_sync + 1
-            prev_states = yield self.current_state_for_users([user_id])
-            if prev_states[user_id].state == PresenceState.OFFLINE:
-                # TODO: Don't block the sync request on this HTTP hit.
-                yield self._send_syncing_users_now()
+
+            # If we went from no in flight sync to some, notify replication
+            if self.user_to_num_current_syncs[user_id] == 1:
+                self.mark_as_coming_online(user_id)
 
         def _end():
             # We check that the user_id is in user_to_num_current_syncs because
@@ -159,6 +191,10 @@ class SynchrotronPresence(object):
             if affect_presence and user_id in self.user_to_num_current_syncs:
                 self.user_to_num_current_syncs[user_id] -= 1
 
+                # If we went from one in flight sync to non, notify replication
+                if self.user_to_num_current_syncs[user_id] == 0:
+                    self.mark_as_going_offline(user_id)
+
         @contextlib.contextmanager
         def _user_syncing():
             try:
@@ -166,49 +202,7 @@ class SynchrotronPresence(object):
             finally:
                 _end()
 
-        defer.returnValue(_user_syncing())
-
-    @defer.inlineCallbacks
-    def _on_shutdown(self):
-        # When the synchrotron is shutdown tell the master to clear the in
-        # progress syncs for this process
-        self.user_to_num_current_syncs.clear()
-        yield self._send_syncing_users_now()
-
-    def _send_syncing_users_regularly(self):
-        # Only send an update if we aren't in the middle of sending one.
-        if not self._sending_sync:
-            preserve_fn(self._send_syncing_users_now)()
-
-    @defer.inlineCallbacks
-    def _send_syncing_users_now(self):
-        if self._sending_sync:
-            # We don't want to race with sending another update.
-            # Instead we wait for that update to finish and send another
-            # update afterwards.
-            self._need_to_send_sync = True
-            return
-
-        # Flag that we are sending an update.
-        self._sending_sync = True
-
-        yield self.http_client.post_json_get_json(self.syncing_users_url, {
-            "process_id": self.process_id,
-            "syncing_users": [
-                user_id for user_id, count in self.user_to_num_current_syncs.items()
-                if count > 0
-            ],
-        })
-
-        # Unset the flag as we are no longer sending an update.
-        self._sending_sync = False
-        if self._need_to_send_sync:
-            # If something happened while we were sending the update then
-            # we might need to send another update.
-            # TODO: Check if the update that was sent matches the current state
-            # as we only need to send an update if they are different.
-            self._need_to_send_sync = False
-            yield self._send_syncing_users_now()
+        return defer.succeed(_user_syncing())
 
     @defer.inlineCallbacks
     def notify_from_replication(self, states, stream_id):
@@ -223,26 +217,24 @@ class SynchrotronPresence(object):
         )
 
     @defer.inlineCallbacks
-    def process_replication(self, result):
-        stream = result.get("presence", {"rows": []})
-        states = []
-        for row in stream["rows"]:
-            (
-                position, user_id, state, last_active_ts,
-                last_federation_update_ts, last_user_sync_ts, status_msg,
-                currently_active
-            ) = row
-            state = UserPresenceState(
-                user_id, state, last_active_ts,
-                last_federation_update_ts, last_user_sync_ts, status_msg,
-                currently_active
-            )
-            self.user_to_current_state[user_id] = state
-            states.append(state)
+    def process_replication_rows(self, token, rows):
+        states = [UserPresenceState(
+            row.user_id, row.state, row.last_active_ts,
+            row.last_federation_update_ts, row.last_user_sync_ts, row.status_msg,
+            row.currently_active
+        ) for row in rows]
+
+        for state in states:
+            self.user_to_current_state[row.user_id] = state
 
-        if states and "position" in stream:
-            stream_id = int(stream["position"])
-            yield self.notify_from_replication(states, stream_id)
+        stream_id = token
+        yield self.notify_from_replication(states, stream_id)
+
+    def get_currently_syncing_users(self):
+        return [
+            user_id for user_id, count in self.user_to_num_current_syncs.iteritems()
+            if count > 0
+        ]
 
 
 class SynchrotronTyping(object):
@@ -257,16 +249,12 @@ class SynchrotronTyping(object):
         # value which we *must* use for the next replication request.
         return {"typing": self._latest_room_serial}
 
-    def process_replication(self, result):
-        stream = result.get("typing")
-        if stream:
-            self._latest_room_serial = int(stream["position"])
+    def process_replication_rows(self, token, rows):
+        self._latest_room_serial = token
 
-            for row in stream["rows"]:
-                position, room_id, typing_json = row
-                typing = json.loads(typing_json)
-                self._room_serials[room_id] = position
-                self._room_typing[room_id] = typing
+        for row in rows:
+            self._room_serials[row.room_id] = token
+            self._room_typing[row.room_id] = row.user_ids
 
 
 class SynchrotronApplicationService(object):
@@ -351,124 +339,89 @@ class SynchrotronServer(HomeServer):
             else:
                 logger.warn("Unrecognized listener type: %s", listener["type"])
 
-    @defer.inlineCallbacks
-    def replicate(self):
-        http_client = self.get_simple_http_client()
-        store = self.get_datastore()
-        replication_url = self.config.worker_replication_url
-        notifier = self.get_notifier()
-        presence_handler = self.get_presence_handler()
-        typing_handler = self.get_typing_handler()
-
-        def notify_from_stream(
-            result, stream_name, stream_key, room=None, user=None
-        ):
-            stream = result.get(stream_name)
-            if stream:
-                position_index = stream["field_names"].index("position")
-                if room:
-                    room_index = stream["field_names"].index(room)
-                if user:
-                    user_index = stream["field_names"].index(user)
-
-                users = ()
-                rooms = ()
-                for row in stream["rows"]:
-                    position = row[position_index]
-
-                    if user:
-                        users = (row[user_index],)
-
-                    if room:
-                        rooms = (row[room_index],)
-
-                    notifier.on_new_event(
-                        stream_key, position, users=users, rooms=rooms
-                    )
+        self.get_tcp_replication().start_replication(self)
 
-        @defer.inlineCallbacks
-        def notify_device_list_update(result):
-            stream = result.get("device_lists")
-            if not stream:
-                return
+    def build_tcp_replication(self):
+        return SyncReplicationHandler(self)
 
-            position_index = stream["field_names"].index("position")
-            user_index = stream["field_names"].index("user_id")
+    def build_presence_handler(self):
+        return SynchrotronPresence(self)
 
-            for row in stream["rows"]:
-                position = row[position_index]
-                user_id = row[user_index]
+    def build_typing_handler(self):
+        return SynchrotronTyping(self)
 
-                room_ids = yield store.get_rooms_for_user(user_id)
 
-                notifier.on_new_event(
-                    "device_list_key", position, rooms=room_ids,
-                )
+class SyncReplicationHandler(ReplicationClientHandler):
+    def __init__(self, hs):
+        super(SyncReplicationHandler, self).__init__(hs.get_datastore())
 
-        @defer.inlineCallbacks
-        def notify(result):
-            stream = result.get("events")
-            if stream:
-                max_position = stream["position"]
-
-                event_map = yield store.get_events([row[1] for row in stream["rows"]])
-
-                for row in stream["rows"]:
-                    position = row[0]
-                    event_id = row[1]
-                    event = event_map.get(event_id, None)
-                    if not event:
-                        continue
-
-                    extra_users = ()
-                    if event.type == EventTypes.Member:
-                        extra_users = (event.state_key,)
-                    notifier.on_new_room_event(
-                        event, position, max_position, extra_users
-                    )
+        self.store = hs.get_datastore()
+        self.typing_handler = hs.get_typing_handler()
+        self.presence_handler = hs.get_presence_handler()
+        self.notifier = hs.get_notifier()
 
-            notify_from_stream(
-                result, "push_rules", "push_rules_key", user="user_id"
-            )
-            notify_from_stream(
-                result, "user_account_data", "account_data_key", user="user_id"
-            )
-            notify_from_stream(
-                result, "room_account_data", "account_data_key", user="user_id"
+        self.presence_handler.sync_callback = self.send_user_sync
+
+    def on_rdata(self, stream_name, token, rows):
+        super(SyncReplicationHandler, self).on_rdata(stream_name, token, rows)
+
+        preserve_fn(self.process_and_notify)(stream_name, token, rows)
+
+    def get_streams_to_replicate(self):
+        args = super(SyncReplicationHandler, self).get_streams_to_replicate()
+        args.update(self.typing_handler.stream_positions())
+        return args
+
+    def get_currently_syncing_users(self):
+        return self.presence_handler.get_currently_syncing_users()
+
+    @defer.inlineCallbacks
+    def process_and_notify(self, stream_name, token, rows):
+        if stream_name == "events":
+            # We shouldn't get multiple rows per token for events stream, so
+            # we don't need to optimise this for multiple rows.
+            for row in rows:
+                event = yield self.store.get_event(row.event_id)
+                extra_users = ()
+                if event.type == EventTypes.Member:
+                    extra_users = (event.state_key,)
+                max_token = self.store.get_room_max_stream_ordering()
+                self.notifier.on_new_room_event(
+                    event, token, max_token, extra_users
+                )
+        elif stream_name == "push_rules":
+            self.notifier.on_new_event(
+                "push_rules_key", token, users=[row.user_id for row in rows],
             )
-            notify_from_stream(
-                result, "tag_account_data", "account_data_key", user="user_id"
+        elif stream_name in ("account_data", "tag_account_data",):
+            self.notifier.on_new_event(
+                "account_data_key", token, users=[row.user_id for row in rows],
             )
-            notify_from_stream(
-                result, "receipts", "receipt_key", room="room_id"
+        elif stream_name == "receipts":
+            self.notifier.on_new_event(
+                "receipt_key", token, rooms=[row.room_id for row in rows],
             )
-            notify_from_stream(
-                result, "typing", "typing_key", room="room_id"
+        elif stream_name == "typing":
+            self.typing_handler.process_replication_rows(token, rows)
+            self.notifier.on_new_event(
+                "typing_key", token, rooms=[row.room_id for row in rows],
             )
-            notify_from_stream(
-                result, "to_device", "to_device_key", user="user_id"
+        elif stream_name == "to_device":
+            entities = [row.entity for row in rows if row.entity.startswith("@")]
+            if entities:
+                self.notifier.on_new_event(
+                    "to_device_key", token, users=entities,
+                )
+        elif stream_name == "device_lists":
+            all_room_ids = set()
+            for row in rows:
+                room_ids = yield self.store.get_rooms_for_user(row.user_id)
+                all_room_ids.update(room_ids)
+            self.notifier.on_new_event(
+                "device_list_key", token, rooms=all_room_ids,
             )
-            yield notify_device_list_update(result)
-
-        while True:
-            try:
-                args = store.stream_positions()
-                args.update(typing_handler.stream_positions())
-                args["timeout"] = 30000
-                result = yield http_client.get_json(replication_url, args=args)
-                yield store.process_replication(result)
-                typing_handler.process_replication(result)
-                yield presence_handler.process_replication(result)
-                yield notify(result)
-            except:
-                logger.exception("Error replicating from %r", replication_url)
-                yield sleep(5)
-
-    def build_presence_handler(self):
-        return SynchrotronPresence(self)
-
-    def build_typing_handler(self):
-        return SynchrotronTyping(self)
+        elif stream_name == "presence":
+            yield self.presence_handler.process_replication_rows(token, rows)
 
 
 def start(config_options):
@@ -514,7 +467,6 @@ def start(config_options):
 
     def start():
         ss.get_datastore().start_profiling()
-        ss.replicate()
         ss.get_state_handler().start_caching()
 
     reactor.callWhenRunning(start)
diff --git a/synapse/app/synctl.py b/synapse/app/synctl.py
index 23eb6a1ec4..8223734845 100755
--- a/synapse/app/synctl.py
+++ b/synapse/app/synctl.py
@@ -125,7 +125,7 @@ def main():
         "configfile",
         nargs="?",
         default="homeserver.yaml",
-        help="the homeserver config file, defaults to homserver.yaml",
+        help="the homeserver config file, defaults to homeserver.yaml",
     )
     parser.add_argument(
         "-w", "--worker",
@@ -202,7 +202,8 @@ def main():
         worker_app = worker_config["worker_app"]
         worker_pidfile = worker_config["worker_pid_file"]
         worker_daemonize = worker_config["worker_daemonize"]
-        assert worker_daemonize  # TODO print something more user friendly
+        assert worker_daemonize, "In config %r: expected '%s' to be True" % (
+            worker_configfile, "worker_daemonize")
         worker_cache_factor = worker_config.get("synctl_cache_factor")
         workers.append(Worker(
             worker_app, worker_configfile, worker_pidfile, worker_cache_factor,
diff --git a/synapse/appservice/__init__.py b/synapse/appservice/__init__.py
index b0106a3597..7346206bb1 100644
--- a/synapse/appservice/__init__.py
+++ b/synapse/appservice/__init__.py
@@ -13,6 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from synapse.api.constants import EventTypes
+from synapse.util.caches.descriptors import cachedInlineCallbacks
 
 from twisted.internet import defer
 
@@ -124,29 +125,23 @@ class ApplicationService(object):
                     raise ValueError(
                         "Expected bool for 'exclusive' in ns '%s'" % ns
                     )
-                if not isinstance(regex_obj.get("regex"), basestring):
+                regex = regex_obj.get("regex")
+                if isinstance(regex, basestring):
+                    regex_obj["regex"] = re.compile(regex)  # Pre-compile regex
+                else:
                     raise ValueError(
                         "Expected string for 'regex' in ns '%s'" % ns
                     )
         return namespaces
 
-    def _matches_regex(self, test_string, namespace_key, return_obj=False):
-        if not isinstance(test_string, basestring):
-            logger.error(
-                "Expected a string to test regex against, but got %s",
-                test_string
-            )
-            return False
-
+    def _matches_regex(self, test_string, namespace_key):
         for regex_obj in self.namespaces[namespace_key]:
-            if re.match(regex_obj["regex"], test_string):
-                if return_obj:
-                    return regex_obj
-                return True
-        return False
+            if regex_obj["regex"].match(test_string):
+                return regex_obj
+        return None
 
     def _is_exclusive(self, ns_key, test_string):
-        regex_obj = self._matches_regex(test_string, ns_key, return_obj=True)
+        regex_obj = self._matches_regex(test_string, ns_key)
         if regex_obj:
             return regex_obj["exclusive"]
         return False
@@ -166,7 +161,14 @@ class ApplicationService(object):
         if not store:
             defer.returnValue(False)
 
-        member_list = yield store.get_users_in_room(event.room_id)
+        does_match = yield self._matches_user_in_member_list(event.room_id, store)
+        defer.returnValue(does_match)
+
+    @cachedInlineCallbacks(num_args=1, cache_context=True)
+    def _matches_user_in_member_list(self, room_id, store, cache_context):
+        member_list = yield store.get_users_in_room(
+            room_id, on_invalidate=cache_context.invalidate
+        )
 
         # check joined member events
         for user_id in member_list:
@@ -219,10 +221,10 @@ class ApplicationService(object):
         )
 
     def is_interested_in_alias(self, alias):
-        return self._matches_regex(alias, ApplicationService.NS_ALIASES)
+        return bool(self._matches_regex(alias, ApplicationService.NS_ALIASES))
 
     def is_interested_in_room(self, room_id):
-        return self._matches_regex(room_id, ApplicationService.NS_ROOMS)
+        return bool(self._matches_regex(room_id, ApplicationService.NS_ROOMS))
 
     def is_exclusive_user(self, user_id):
         return (
diff --git a/synapse/config/emailconfig.py b/synapse/config/emailconfig.py
index 0030b5db1e..fe156b6930 100644
--- a/synapse/config/emailconfig.py
+++ b/synapse/config/emailconfig.py
@@ -71,6 +71,15 @@ class EmailConfig(Config):
             self.email_riot_base_url = email_config.get(
                 "riot_base_url", None
             )
+            self.email_smtp_user = email_config.get(
+                "smtp_user", None
+            )
+            self.email_smtp_pass = email_config.get(
+                "smtp_pass", None
+            )
+            self.require_transport_security = email_config.get(
+                "require_transport_security", False
+            )
             if "app_name" in email_config:
                 self.email_app_name = email_config["app_name"]
             else:
@@ -91,10 +100,17 @@ class EmailConfig(Config):
         # Defining a custom URL for Riot is only needed if email notifications
         # should contain links to a self-hosted installation of Riot; when set
         # the "app_name" setting is ignored.
+        #
+        # If your SMTP server requires authentication, the optional smtp_user &
+        # smtp_pass variables should be used
+        #
         #email:
         #   enable_notifs: false
         #   smtp_host: "localhost"
         #   smtp_port: 25
+        #   smtp_user: "exampleusername"
+        #   smtp_pass: "examplepassword"
+        #   require_transport_security: False
         #   notif_from: "Your Friendly %(app)s Home Server <noreply@example.com>"
         #   app_name: Matrix
         #   template_dir: res/templates
diff --git a/synapse/config/registration.py b/synapse/config/registration.py
index 87e500c97a..f7e03c4cde 100644
--- a/synapse/config/registration.py
+++ b/synapse/config/registration.py
@@ -69,6 +69,7 @@ class RegistrationConfig(Config):
         trusted_third_party_id_servers:
             - matrix.org
             - vector.im
+            - riot.im
         """ % locals()
 
     def add_arguments(self, parser):
diff --git a/synapse/config/voip.py b/synapse/config/voip.py
index eeb693027b..3a4e16fa96 100644
--- a/synapse/config/voip.py
+++ b/synapse/config/voip.py
@@ -23,6 +23,7 @@ class VoipConfig(Config):
         self.turn_username = config.get("turn_username")
         self.turn_password = config.get("turn_password")
         self.turn_user_lifetime = self.parse_duration(config["turn_user_lifetime"])
+        self.turn_allow_guests = config.get("turn_allow_guests", True)
 
     def default_config(self, **kwargs):
         return """\
@@ -41,4 +42,11 @@ class VoipConfig(Config):
 
         # How long generated TURN credentials last
         turn_user_lifetime: "1h"
+
+        # Whether guests should be allowed to use the TURN server.
+        # This defaults to True, otherwise VoIP will be unreliable for guests.
+        # However, it does introduce a slight security risk as it allows users to
+        # connect to arbitrary endpoints without having first signed up for a
+        # valid account (e.g. by passing a CAPTCHA).
+        turn_allow_guests: True
         """
diff --git a/synapse/config/workers.py b/synapse/config/workers.py
index b165c67ee7..ea48d931a1 100644
--- a/synapse/config/workers.py
+++ b/synapse/config/workers.py
@@ -28,7 +28,9 @@ class WorkerConfig(Config):
         self.worker_pid_file = config.get("worker_pid_file")
         self.worker_log_file = config.get("worker_log_file")
         self.worker_log_config = config.get("worker_log_config")
-        self.worker_replication_url = config.get("worker_replication_url")
+        self.worker_replication_host = config.get("worker_replication_host", None)
+        self.worker_replication_port = config.get("worker_replication_port", None)
+        self.worker_name = config.get("worker_name", self.worker_app)
 
         if self.worker_listeners:
             for listener in self.worker_listeners:
diff --git a/synapse/federation/send_queue.py b/synapse/federation/send_queue.py
index bbb0195228..748548bbe2 100644
--- a/synapse/federation/send_queue.py
+++ b/synapse/federation/send_queue.py
@@ -31,21 +31,19 @@ Events are replicated via a separate events stream.
 
 from .units import Edu
 
+from synapse.storage.presence import UserPresenceState
 from synapse.util.metrics import Measure
 import synapse.metrics
 
 from blist import sorteddict
-import ujson
+from collections import namedtuple
 
+import logging
 
-metrics = synapse.metrics.get_metrics_for(__name__)
+logger = logging.getLogger(__name__)
 
 
-PRESENCE_TYPE = "p"
-KEYED_EDU_TYPE = "k"
-EDU_TYPE = "e"
-FAILURE_TYPE = "f"
-DEVICE_MESSAGE_TYPE = "d"
+metrics = synapse.metrics.get_metrics_for(__name__)
 
 
 class FederationRemoteSendQueue(object):
@@ -220,10 +218,15 @@ class FederationRemoteSendQueue(object):
     def get_current_token(self):
         return self.pos - 1
 
-    def get_replication_rows(self, token, limit, federation_ack=None):
-        """
+    def federation_ack(self, token):
+        self._clear_queue_before_pos(token)
+
+    def get_replication_rows(self, from_token, to_token, limit, federation_ack=None):
+        """Get rows to be sent over federation between the two tokens
+
         Args:
-            token (int)
+            from_token (int)
+            to_token(int)
             limit (int)
             federation_ack (int): Optional. The position where the worker is
                 explicitly acknowledged it has handled. Allows us to drop
@@ -232,9 +235,11 @@ class FederationRemoteSendQueue(object):
         # TODO: Handle limit.
 
         # To handle restarts where we wrap around
-        if token > self.pos:
-            token = -1
+        if from_token > self.pos:
+            from_token = -1
 
+        # list of tuple(int, BaseFederationRow), where the first is the position
+        # of the federation stream.
         rows = []
 
         # There should be only one reader, so lets delete everything its
@@ -244,62 +249,283 @@ class FederationRemoteSendQueue(object):
 
         # Fetch changed presence
         keys = self.presence_changed.keys()
-        i = keys.bisect_right(token)
+        i = keys.bisect_right(from_token)
+        j = keys.bisect_right(to_token) + 1
         dest_user_ids = set(
             (pos, dest_user_id)
-            for pos in keys[i:]
+            for pos in keys[i:j]
             for dest_user_id in self.presence_changed[pos]
         )
 
         for (key, (dest, user_id)) in dest_user_ids:
-            rows.append((key, PRESENCE_TYPE, ujson.dumps({
-                "destination": dest,
-                "state": self.presence_map[user_id].as_dict(),
-            })))
+            rows.append((key, PresenceRow(
+                destination=dest,
+                state=self.presence_map[user_id],
+            )))
 
         # Fetch changes keyed edus
         keys = self.keyed_edu_changed.keys()
-        i = keys.bisect_right(token)
-        keyed_edus = set((k, self.keyed_edu_changed[k]) for k in keys[i:])
+        i = keys.bisect_right(from_token)
+        j = keys.bisect_right(to_token) + 1
+        keyed_edus = set((k, self.keyed_edu_changed[k]) for k in keys[i:j])
 
         for (pos, (destination, edu_key)) in keyed_edus:
-            rows.append(
-                (pos, KEYED_EDU_TYPE, ujson.dumps({
-                    "key": edu_key,
-                    "edu": self.keyed_edu[(destination, edu_key)].get_internal_dict(),
-                }))
-            )
+            rows.append((pos, KeyedEduRow(
+                key=edu_key,
+                edu=self.keyed_edu[(destination, edu_key)],
+            )))
 
         # Fetch changed edus
         keys = self.edus.keys()
-        i = keys.bisect_right(token)
-        edus = set((k, self.edus[k]) for k in keys[i:])
+        i = keys.bisect_right(from_token)
+        j = keys.bisect_right(to_token) + 1
+        edus = set((k, self.edus[k]) for k in keys[i:j])
 
         for (pos, edu) in edus:
-            rows.append((pos, EDU_TYPE, ujson.dumps(edu.get_internal_dict())))
+            rows.append((pos, EduRow(edu)))
 
         # Fetch changed failures
         keys = self.failures.keys()
-        i = keys.bisect_right(token)
-        failures = set((k, self.failures[k]) for k in keys[i:])
+        i = keys.bisect_right(from_token)
+        j = keys.bisect_right(to_token) + 1
+        failures = set((k, self.failures[k]) for k in keys[i:j])
 
         for (pos, (destination, failure)) in failures:
-            rows.append((pos, FAILURE_TYPE, ujson.dumps({
-                "destination": destination,
-                "failure": failure,
-            })))
+            rows.append((pos, FailureRow(
+                destination=destination,
+                failure=failure,
+            )))
 
         # Fetch changed device messages
         keys = self.device_messages.keys()
-        i = keys.bisect_right(token)
-        device_messages = set((k, self.device_messages[k]) for k in keys[i:])
+        i = keys.bisect_right(from_token)
+        j = keys.bisect_right(to_token) + 1
+        device_messages = set((k, self.device_messages[k]) for k in keys[i:j])
 
         for (pos, destination) in device_messages:
-            rows.append((pos, DEVICE_MESSAGE_TYPE, ujson.dumps({
-                "destination": destination,
-            })))
+            rows.append((pos, DeviceRow(
+                destination=destination,
+            )))
 
         # Sort rows based on pos
         rows.sort()
 
-        return rows
+        return [(pos, row.TypeId, row.to_data()) for pos, row in rows]
+
+
+class BaseFederationRow(object):
+    """Base class for rows to be sent in the federation stream.
+
+    Specifies how to identify, serialize and deserialize the different types.
+    """
+
+    TypeId = None  # Unique string that ids the type. Must be overriden in sub classes.
+
+    @staticmethod
+    def from_data(data):
+        """Parse the data from the federation stream into a row.
+
+        Args:
+            data: The value of ``data`` from FederationStreamRow.data, type
+                depends on the type of stream
+        """
+        raise NotImplementedError()
+
+    def to_data(self):
+        """Serialize this row to be sent over the federation stream.
+
+        Returns:
+            The value to be sent in FederationStreamRow.data. The type depends
+            on the type of stream.
+        """
+        raise NotImplementedError()
+
+    def add_to_buffer(self, buff):
+        """Add this row to the appropriate field in the buffer ready for this
+        to be sent over federation.
+
+        We use a buffer so that we can batch up events that have come in at
+        the same time and send them all at once.
+
+        Args:
+            buff (BufferedToSend)
+        """
+        raise NotImplementedError()
+
+
+class PresenceRow(BaseFederationRow, namedtuple("PresenceRow", (
+    "destination",  # str
+    "state",  # UserPresenceState
+))):
+    TypeId = "p"
+
+    @staticmethod
+    def from_data(data):
+        return PresenceRow(
+            destination=data["destination"],
+            state=UserPresenceState.from_dict(data["state"])
+        )
+
+    def to_data(self):
+        return {
+            "destination": self.destination,
+            "state": self.state.as_dict()
+        }
+
+    def add_to_buffer(self, buff):
+        buff.presence.setdefault(self.destination, []).append(self.state)
+
+
+class KeyedEduRow(BaseFederationRow, namedtuple("KeyedEduRow", (
+    "key",  # tuple(str) - the edu key passed to send_edu
+    "edu",  # Edu
+))):
+    TypeId = "k"
+
+    @staticmethod
+    def from_data(data):
+        return KeyedEduRow(
+            key=tuple(data["key"]),
+            edu=Edu(**data["edu"]),
+        )
+
+    def to_data(self):
+        return {
+            "key": self.key,
+            "edu": self.edu.get_internal_dict(),
+        }
+
+    def add_to_buffer(self, buff):
+        buff.keyed_edus.setdefault(
+            self.edu.destination, {}
+        )[self.key] = self.edu
+
+
+class EduRow(BaseFederationRow, namedtuple("EduRow", (
+    "edu",  # Edu
+))):
+    TypeId = "e"
+
+    @staticmethod
+    def from_data(data):
+        return EduRow(Edu(**data))
+
+    def to_data(self):
+        return self.edu.get_internal_dict()
+
+    def add_to_buffer(self, buff):
+        buff.edus.setdefault(self.edu.destination, []).append(self.edu)
+
+
+class FailureRow(BaseFederationRow, namedtuple("FailureRow", (
+    "destination",  # str
+    "failure",
+))):
+    TypeId = "f"
+
+    @staticmethod
+    def from_data(data):
+        return FailureRow(
+            destination=data["destination"],
+            failure=data["failure"],
+        )
+
+    def to_data(self):
+        return {
+            "destination": self.destination,
+            "failure": self.failure,
+        }
+
+    def add_to_buffer(self, buff):
+        buff.failures.setdefault(self.destination, []).append(self.failure)
+
+
+class DeviceRow(BaseFederationRow, namedtuple("DeviceRow", (
+    "destination",  # str
+))):
+    TypeId = "d"
+
+    @staticmethod
+    def from_data(data):
+        return DeviceRow(destination=data["destination"])
+
+    def to_data(self):
+        return {"destination": self.destination}
+
+    def add_to_buffer(self, buff):
+        buff.device_destinations.add(self.destination)
+
+
+TypeToRow = {
+    Row.TypeId: Row
+    for Row in (
+        PresenceRow,
+        KeyedEduRow,
+        EduRow,
+        FailureRow,
+        DeviceRow,
+    )
+}
+
+
+ParsedFederationStreamData = namedtuple("ParsedFederationStreamData", (
+    "presence",  # dict of destination -> [UserPresenceState]
+    "keyed_edus",  # dict of destination -> { key -> Edu }
+    "edus",  # dict of destination -> [Edu]
+    "failures",  # dict of destination -> [failures]
+    "device_destinations",  # set of destinations
+))
+
+
+def process_rows_for_federation(transaction_queue, rows):
+    """Parse a list of rows from the federation stream and put them in the
+    transaction queue ready for sending to the relevant homeservers.
+
+    Args:
+        transaction_queue (TransactionQueue)
+        rows (list(synapse.replication.tcp.streams.FederationStreamRow))
+    """
+
+    # The federation stream contains a bunch of different types of
+    # rows that need to be handled differently. We parse the rows, put
+    # them into the appropriate collection and then send them off.
+
+    buff = ParsedFederationStreamData(
+        presence={},
+        keyed_edus={},
+        edus={},
+        failures={},
+        device_destinations=set(),
+    )
+
+    # Parse the rows in the stream and add to the buffer
+    for row in rows:
+        if row.type not in TypeToRow:
+            logger.error("Unrecognized federation row type %r", row.type)
+            continue
+
+        RowType = TypeToRow[row.type]
+        parsed_row = RowType.from_data(row.data)
+        parsed_row.add_to_buffer(buff)
+
+    for destination, states in buff.presence.iteritems():
+        transaction_queue.send_presence(destination, states)
+
+    for destination, edu_map in buff.keyed_edus.iteritems():
+        for key, edu in edu_map.items():
+            transaction_queue.send_edu(
+                edu.destination, edu.edu_type, edu.content, key=key,
+            )
+
+    for destination, edu_list in buff.edus.iteritems():
+        for edu in edu_list:
+            transaction_queue.send_edu(
+                edu.destination, edu.edu_type, edu.content, key=None,
+            )
+
+    for destination, failure_list in buff.failures.iteritems():
+        for failure in failure_list:
+            transaction_queue.send_failure(destination, failure)
+
+    for destination in buff.device_destinations:
+        transaction_queue.send_device_messages(destination)
diff --git a/synapse/handlers/federation.py b/synapse/handlers/federation.py
index 53f9296399..2d9126dd86 100644
--- a/synapse/handlers/federation.py
+++ b/synapse/handlers/federation.py
@@ -28,7 +28,7 @@ from synapse.api.constants import EventTypes, Membership, RejectedReason
 from synapse.events.validator import EventValidator
 from synapse.util import unwrapFirstError
 from synapse.util.logcontext import (
-    PreserveLoggingContext, preserve_fn, preserve_context_over_deferred
+    preserve_fn, preserve_context_over_deferred
 )
 from synapse.util.metrics import measure_func
 from synapse.util.logutils import log_function
@@ -394,11 +394,10 @@ class FederationHandler(BaseHandler):
             target_user = UserID.from_string(target_user_id)
             extra_users.append(target_user)
 
-        with PreserveLoggingContext():
-            self.notifier.on_new_room_event(
-                event, event_stream_id, max_stream_id,
-                extra_users=extra_users
-            )
+        self.notifier.on_new_room_event(
+            event, event_stream_id, max_stream_id,
+            extra_users=extra_users
+        )
 
         if event.type == EventTypes.Member:
             if event.membership == Membership.JOIN:
@@ -916,11 +915,10 @@ class FederationHandler(BaseHandler):
                 origin, auth_chain, state, event
             )
 
-            with PreserveLoggingContext():
-                self.notifier.on_new_room_event(
-                    event, event_stream_id, max_stream_id,
-                    extra_users=[joinee]
-                )
+            self.notifier.on_new_room_event(
+                event, event_stream_id, max_stream_id,
+                extra_users=[joinee]
+            )
 
             logger.debug("Finished joining %s to %s", joinee, room_id)
         finally:
@@ -1035,10 +1033,9 @@ class FederationHandler(BaseHandler):
             target_user = UserID.from_string(target_user_id)
             extra_users.append(target_user)
 
-        with PreserveLoggingContext():
-            self.notifier.on_new_room_event(
-                event, event_stream_id, max_stream_id, extra_users=extra_users
-            )
+        self.notifier.on_new_room_event(
+            event, event_stream_id, max_stream_id, extra_users=extra_users
+        )
 
         if event.type == EventTypes.Member:
             if event.content["membership"] == Membership.JOIN:
@@ -1084,11 +1081,10 @@ class FederationHandler(BaseHandler):
         )
 
         target_user = UserID.from_string(event.state_key)
-        with PreserveLoggingContext():
-            self.notifier.on_new_room_event(
-                event, event_stream_id, max_stream_id,
-                extra_users=[target_user],
-            )
+        self.notifier.on_new_room_event(
+            event, event_stream_id, max_stream_id,
+            extra_users=[target_user],
+        )
 
         defer.returnValue(event)
 
@@ -1246,10 +1242,9 @@ class FederationHandler(BaseHandler):
             target_user = UserID.from_string(target_user_id)
             extra_users.append(target_user)
 
-        with PreserveLoggingContext():
-            self.notifier.on_new_room_event(
-                event, event_stream_id, max_stream_id, extra_users=extra_users
-            )
+        self.notifier.on_new_room_event(
+            event, event_stream_id, max_stream_id, extra_users=extra_users
+        )
 
         defer.returnValue(None)
 
diff --git a/synapse/handlers/message.py b/synapse/handlers/message.py
index 7a498af5a2..348056add5 100644
--- a/synapse/handlers/message.py
+++ b/synapse/handlers/message.py
@@ -612,7 +612,7 @@ class MessageHandler(BaseHandler):
         @defer.inlineCallbacks
         def _notify():
             yield run_on_reactor()
-            yield self.notifier.on_new_room_event(
+            self.notifier.on_new_room_event(
                 event, event_stream_id, max_stream_id,
                 extra_users=extra_users
             )
diff --git a/synapse/handlers/presence.py b/synapse/handlers/presence.py
index 1ede117c79..9ed5af3cb4 100644
--- a/synapse/handlers/presence.py
+++ b/synapse/handlers/presence.py
@@ -30,6 +30,7 @@ from synapse.api.constants import PresenceState
 from synapse.storage.presence import UserPresenceState
 
 from synapse.util.caches.descriptors import cachedInlineCallbacks
+from synapse.util.async import Linearizer
 from synapse.util.logcontext import preserve_fn
 from synapse.util.logutils import log_function
 from synapse.util.metrics import Measure
@@ -187,6 +188,7 @@ class PresenceHandler(object):
         # process_id to millisecond timestamp last updated.
         self.external_process_to_current_syncs = {}
         self.external_process_last_updated_ms = {}
+        self.external_sync_linearizer = Linearizer(name="external_sync_linearizer")
 
         # Start a LoopingCall in 30s that fires every 5s.
         # The initial delay is to allow disconnected clients a chance to
@@ -509,6 +511,73 @@ class PresenceHandler(object):
         self.external_process_to_current_syncs[process_id] = syncing_user_ids
 
     @defer.inlineCallbacks
+    def update_external_syncs_row(self, process_id, user_id, is_syncing, sync_time_msec):
+        """Update the syncing users for an external process as a delta.
+
+        Args:
+            process_id (str): An identifier for the process the users are
+                syncing against. This allows synapse to process updates
+                as user start and stop syncing against a given process.
+            user_id (str): The user who has started or stopped syncing
+            is_syncing (bool): Whether or not the user is now syncing
+            sync_time_msec(int): Time in ms when the user was last syncing
+        """
+        with (yield self.external_sync_linearizer.queue(process_id)):
+            prev_state = yield self.current_state_for_user(user_id)
+
+            process_presence = self.external_process_to_current_syncs.setdefault(
+                process_id, set()
+            )
+
+            updates = []
+            if is_syncing and user_id not in process_presence:
+                if prev_state.state == PresenceState.OFFLINE:
+                    updates.append(prev_state.copy_and_replace(
+                        state=PresenceState.ONLINE,
+                        last_active_ts=sync_time_msec,
+                        last_user_sync_ts=sync_time_msec,
+                    ))
+                else:
+                    updates.append(prev_state.copy_and_replace(
+                        last_user_sync_ts=sync_time_msec,
+                    ))
+                process_presence.add(user_id)
+            elif user_id in process_presence:
+                updates.append(prev_state.copy_and_replace(
+                    last_user_sync_ts=sync_time_msec,
+                ))
+
+            if not is_syncing:
+                process_presence.discard(user_id)
+
+            if updates:
+                yield self._update_states(updates)
+
+            self.external_process_last_updated_ms[process_id] = self.clock.time_msec()
+
+    @defer.inlineCallbacks
+    def update_external_syncs_clear(self, process_id):
+        """Marks all users that had been marked as syncing by a given process
+        as offline.
+
+        Used when the process has stopped/disappeared.
+        """
+        with (yield self.external_sync_linearizer.queue(process_id)):
+            process_presence = self.external_process_to_current_syncs.pop(
+                process_id, set()
+            )
+            prev_states = yield self.current_state_for_users(process_presence)
+            time_now_ms = self.clock.time_msec()
+
+            yield self._update_states([
+                prev_state.copy_and_replace(
+                    last_user_sync_ts=time_now_ms,
+                )
+                for prev_state in prev_states.itervalues()
+            ])
+            self.external_process_last_updated_ms.pop(process_id, None)
+
+    @defer.inlineCallbacks
     def current_state_for_user(self, user_id):
         """Get the current presence state for a user.
         """
@@ -527,14 +596,14 @@ class PresenceHandler(object):
             for user_id in user_ids
         }
 
-        missing = [user_id for user_id, state in states.items() if not state]
+        missing = [user_id for user_id, state in states.iteritems() if not state]
         if missing:
             # There are things not in our in memory cache. Lets pull them out of
             # the database.
             res = yield self.store.get_presence_for_users(missing)
             states.update(res)
 
-            missing = [user_id for user_id, state in states.items() if not state]
+            missing = [user_id for user_id, state in states.iteritems() if not state]
             if missing:
                 new = {
                     user_id: UserPresenceState.default(user_id)
diff --git a/synapse/handlers/typing.py b/synapse/handlers/typing.py
index 0eea7f8f9c..3b7818af5c 100644
--- a/synapse/handlers/typing.py
+++ b/synapse/handlers/typing.py
@@ -24,7 +24,6 @@ from synapse.types import UserID, get_domain_from_id
 import logging
 
 from collections import namedtuple
-import ujson as json
 
 logger = logging.getLogger(__name__)
 
@@ -288,11 +287,13 @@ class TypingHandler(object):
         for room_id, serial in self._room_serials.items():
             if last_id < serial and serial <= current_id:
                 typing = self._room_typing[room_id]
-                typing_bytes = json.dumps(list(typing), ensure_ascii=False)
-                rows.append((serial, room_id, typing_bytes))
+                rows.append((serial, room_id, list(typing)))
         rows.sort()
         return rows
 
+    def get_current_token(self):
+        return self._latest_room_serial
+
 
 class TypingNotificationEventSource(object):
     def __init__(self, hs):
diff --git a/synapse/notifier.py b/synapse/notifier.py
index 7eeba6d28e..48566187ab 100644
--- a/synapse/notifier.py
+++ b/synapse/notifier.py
@@ -163,6 +163,8 @@ class Notifier(object):
         self.store = hs.get_datastore()
         self.pending_new_room_events = []
 
+        self.replication_callbacks = []
+
         self.clock = hs.get_clock()
         self.appservice_handler = hs.get_application_service_handler()
 
@@ -202,7 +204,12 @@ class Notifier(object):
             lambda: len(self.user_to_user_stream),
         )
 
-    @preserve_fn
+    def add_replication_callback(self, cb):
+        """Add a callback that will be called when some new data is available.
+        Callback is not given any arguments.
+        """
+        self.replication_callbacks.append(cb)
+
     def on_new_room_event(self, event, room_stream_id, max_room_stream_id,
                           extra_users=[]):
         """ Used by handlers to inform the notifier something has happened
@@ -216,15 +223,13 @@ class Notifier(object):
         until all previous events have been persisted before notifying
         the client streams.
         """
-        with PreserveLoggingContext():
-            self.pending_new_room_events.append((
-                room_stream_id, event, extra_users
-            ))
-            self._notify_pending_new_room_events(max_room_stream_id)
+        self.pending_new_room_events.append((
+            room_stream_id, event, extra_users
+        ))
+        self._notify_pending_new_room_events(max_room_stream_id)
 
-            self.notify_replication()
+        self.notify_replication()
 
-    @preserve_fn
     def _notify_pending_new_room_events(self, max_room_stream_id):
         """Notify for the room events that were queued waiting for a previous
         event to be persisted.
@@ -242,14 +247,16 @@ class Notifier(object):
             else:
                 self._on_new_room_event(event, room_stream_id, extra_users)
 
-    @preserve_fn
     def _on_new_room_event(self, event, room_stream_id, extra_users=[]):
         """Notify any user streams that are interested in this room event"""
         # poke any interested application service.
-        self.appservice_handler.notify_interested_services(room_stream_id)
+        preserve_fn(self.appservice_handler.notify_interested_services)(
+            room_stream_id)
 
         if self.federation_sender:
-            self.federation_sender.notify_new_events(room_stream_id)
+            preserve_fn(self.federation_sender.notify_new_events)(
+                room_stream_id
+            )
 
         if event.type == EventTypes.Member and event.membership == Membership.JOIN:
             self._user_joined_room(event.state_key, event.room_id)
@@ -260,7 +267,6 @@ class Notifier(object):
             rooms=[event.room_id],
         )
 
-    @preserve_fn
     def on_new_event(self, stream_key, new_token, users=[], rooms=[]):
         """ Used to inform listeners that something has happend event wise.
 
@@ -287,7 +293,6 @@ class Notifier(object):
 
                 self.notify_replication()
 
-    @preserve_fn
     def on_new_replication_data(self):
         """Used to inform replication listeners that something has happend
         without waking up any of the normal user event streams"""
@@ -510,6 +515,9 @@ class Notifier(object):
             self.replication_deferred = ObservableDeferred(defer.Deferred())
             deferred.callback(None)
 
+        for cb in self.replication_callbacks:
+            preserve_fn(cb)()
+
     @defer.inlineCallbacks
     def wait_for_replication(self, callback, timeout):
         """Wait for an event to happen.
diff --git a/synapse/push/mailer.py b/synapse/push/mailer.py
index 3a50c72e0b..f83aa7625c 100644
--- a/synapse/push/mailer.py
+++ b/synapse/push/mailer.py
@@ -200,7 +200,11 @@ class Mailer(object):
         yield sendmail(
             self.hs.config.email_smtp_host,
             raw_from, raw_to, multipart_msg.as_string(),
-            port=self.hs.config.email_smtp_port
+            port=self.hs.config.email_smtp_port,
+            requireAuthentication=self.hs.config.email_smtp_user is not None,
+            username=self.hs.config.email_smtp_user,
+            password=self.hs.config.email_smtp_pass,
+            requireTransportSecurity=self.hs.config.require_transport_security
         )
 
     @defer.inlineCallbacks
diff --git a/synapse/push/push_tools.py b/synapse/push/push_tools.py
index 287df94b4f..6835f54e97 100644
--- a/synapse/push/push_tools.py
+++ b/synapse/push/push_tools.py
@@ -17,15 +17,12 @@ from twisted.internet import defer
 from synapse.push.presentable_names import (
     calculate_room_name, name_from_member_event
 )
-from synapse.util.logcontext import preserve_fn, preserve_context_over_deferred
 
 
 @defer.inlineCallbacks
 def get_badge_count(store, user_id):
-    invites, joins = yield preserve_context_over_deferred(defer.gatherResults([
-        preserve_fn(store.get_invited_rooms_for_user)(user_id),
-        preserve_fn(store.get_rooms_for_user)(user_id),
-    ], consumeErrors=True))
+    invites = yield store.get_invited_rooms_for_user(user_id)
+    joins = yield store.get_rooms_for_user(user_id)
 
     my_receipts_by_room = yield store.get_receipts_for_user(
         user_id, "m.read",
diff --git a/synapse/replication/resource.py b/synapse/replication/resource.py
index 03930fe958..abd3fe7665 100644
--- a/synapse/replication/resource.py
+++ b/synapse/replication/resource.py
@@ -489,7 +489,7 @@ class ReplicationResource(Resource):
 
         if federation is not None and federation != current_position:
             federation_rows = self.federation_sender.get_replication_rows(
-                federation, limit, federation_ack=federation_ack,
+                federation, current_position, limit, federation_ack=federation_ack,
             )
             upto_token = _position_from_rows(federation_rows, current_position)
             writer.write_header_and_rows("federation", federation_rows, (
@@ -504,7 +504,7 @@ class ReplicationResource(Resource):
 
         if device_lists is not None and device_lists != current_position:
             changes = yield self.store.get_all_device_list_changes_for_remotes(
-                device_lists,
+                device_lists, current_position,
             )
             writer.write_header_and_rows("device_lists", changes, (
                 "position", "user_id", "destination",
diff --git a/synapse/replication/slave/storage/_base.py b/synapse/replication/slave/storage/_base.py
index ab133db872..b962641166 100644
--- a/synapse/replication/slave/storage/_base.py
+++ b/synapse/replication/slave/storage/_base.py
@@ -15,7 +15,6 @@
 
 from synapse.storage._base import SQLBaseStore
 from synapse.storage.engines import PostgresEngine
-from twisted.internet import defer
 
 from ._slaved_id_tracker import SlavedIdTracker
 
@@ -34,8 +33,7 @@ class BaseSlavedStore(SQLBaseStore):
         else:
             self._cache_id_gen = None
 
-        self.expire_cache_url = hs.config.worker_replication_url + "/expire_cache"
-        self.http_client = hs.get_simple_http_client()
+        self.hs = hs
 
     def stream_positions(self):
         pos = {}
@@ -43,35 +41,20 @@ class BaseSlavedStore(SQLBaseStore):
             pos["caches"] = self._cache_id_gen.get_current_token()
         return pos
 
-    def process_replication(self, result):
-        stream = result.get("caches")
-        if stream:
-            for row in stream["rows"]:
-                (
-                    position, cache_func, keys, invalidation_ts,
-                ) = row
-
+    def process_replication_rows(self, stream_name, token, rows):
+        if stream_name == "caches":
+            self._cache_id_gen.advance(token)
+            for row in rows:
                 try:
-                    getattr(self, cache_func).invalidate(tuple(keys))
+                    getattr(self, row.cache_func).invalidate(tuple(row.keys))
                 except AttributeError:
                     # We probably haven't pulled in the cache in this worker,
                     # which is fine.
                     pass
-            self._cache_id_gen.advance(int(stream["position"]))
-        return defer.succeed(None)
 
     def _invalidate_cache_and_stream(self, txn, cache_func, keys):
         txn.call_after(cache_func.invalidate, keys)
         txn.call_after(self._send_invalidation_poke, cache_func, keys)
 
-    @defer.inlineCallbacks
     def _send_invalidation_poke(self, cache_func, keys):
-        try:
-            yield self.http_client.post_json_get_json(self.expire_cache_url, {
-                "invalidate": [{
-                    "name": cache_func.__name__,
-                    "keys": list(keys),
-                }]
-            })
-        except:
-            logger.exception("Failed to poke on expire_cache")
+        self.hs.get_tcp_replication().send_invalidate_cache(cache_func, keys)
diff --git a/synapse/replication/slave/storage/account_data.py b/synapse/replication/slave/storage/account_data.py
index 77c64722c7..efbd87918e 100644
--- a/synapse/replication/slave/storage/account_data.py
+++ b/synapse/replication/slave/storage/account_data.py
@@ -69,38 +69,25 @@ class SlavedAccountDataStore(BaseSlavedStore):
         result["tag_account_data"] = position
         return result
 
-    def process_replication(self, result):
-        stream = result.get("user_account_data")
-        if stream:
-            self._account_data_id_gen.advance(int(stream["position"]))
-            for row in stream["rows"]:
-                position, user_id, data_type = row[:3]
-                self.get_global_account_data_by_type_for_user.invalidate(
-                    (data_type, user_id,)
-                )
-                self.get_account_data_for_user.invalidate((user_id,))
+    def process_replication_rows(self, stream_name, token, rows):
+        if stream_name == "tag_account_data":
+            self._account_data_id_gen.advance(token)
+            for row in rows:
+                self.get_tags_for_user.invalidate((row.user_id,))
                 self._account_data_stream_cache.entity_has_changed(
-                    user_id, position
+                    row.user_id, token
                 )
-
-        stream = result.get("room_account_data")
-        if stream:
-            self._account_data_id_gen.advance(int(stream["position"]))
-            for row in stream["rows"]:
-                position, user_id = row[:2]
-                self.get_account_data_for_user.invalidate((user_id,))
+        elif stream_name == "account_data":
+            self._account_data_id_gen.advance(token)
+            for row in rows:
+                if not row.room_id:
+                    self.get_global_account_data_by_type_for_user.invalidate(
+                        (row.data_type, row.user_id,)
+                    )
+                self.get_account_data_for_user.invalidate((row.user_id,))
                 self._account_data_stream_cache.entity_has_changed(
-                    user_id, position
+                    row.user_id, token
                 )
-
-        stream = result.get("tag_account_data")
-        if stream:
-            self._account_data_id_gen.advance(int(stream["position"]))
-            for row in stream["rows"]:
-                position, user_id = row[:2]
-                self.get_tags_for_user.invalidate((user_id,))
-                self._account_data_stream_cache.entity_has_changed(
-                    user_id, position
-                )
-
-        return super(SlavedAccountDataStore, self).process_replication(result)
+        return super(SlavedAccountDataStore, self).process_replication_rows(
+            stream_name, token, rows
+        )
diff --git a/synapse/replication/slave/storage/deviceinbox.py b/synapse/replication/slave/storage/deviceinbox.py
index f9102e0d89..6f3fb64770 100644
--- a/synapse/replication/slave/storage/deviceinbox.py
+++ b/synapse/replication/slave/storage/deviceinbox.py
@@ -53,21 +53,18 @@ class SlavedDeviceInboxStore(BaseSlavedStore):
         result["to_device"] = self._device_inbox_id_gen.get_current_token()
         return result
 
-    def process_replication(self, result):
-        stream = result.get("to_device")
-        if stream:
-            self._device_inbox_id_gen.advance(int(stream["position"]))
-            for row in stream["rows"]:
-                stream_id = row[0]
-                entity = row[1]
-
-                if entity.startswith("@"):
+    def process_replication_rows(self, stream_name, token, rows):
+        if stream_name == "to_device":
+            self._device_inbox_id_gen.advance(token)
+            for row in rows:
+                if row.entity.startswith("@"):
                     self._device_inbox_stream_cache.entity_has_changed(
-                        entity, stream_id
+                        row.entity, token
                     )
                 else:
                     self._device_federation_outbox_stream_cache.entity_has_changed(
-                        entity, stream_id
+                        row.entity, token
                     )
-
-        return super(SlavedDeviceInboxStore, self).process_replication(result)
+        return super(SlavedDeviceInboxStore, self).process_replication_rows(
+            stream_name, token, rows
+        )
diff --git a/synapse/replication/slave/storage/devices.py b/synapse/replication/slave/storage/devices.py
index ca46aa17b6..4d4a435471 100644
--- a/synapse/replication/slave/storage/devices.py
+++ b/synapse/replication/slave/storage/devices.py
@@ -51,22 +51,18 @@ class SlavedDeviceStore(BaseSlavedStore):
         result["device_lists"] = self._device_list_id_gen.get_current_token()
         return result
 
-    def process_replication(self, result):
-        stream = result.get("device_lists")
-        if stream:
-            self._device_list_id_gen.advance(int(stream["position"]))
-            for row in stream["rows"]:
-                stream_id = row[0]
-                user_id = row[1]
-                destination = row[2]
-
+    def process_replication_rows(self, stream_name, token, rows):
+        if stream_name == "device_lists":
+            self._device_list_id_gen.advance(token)
+            for row in rows:
                 self._device_list_stream_cache.entity_has_changed(
-                    user_id, stream_id
+                    row.user_id, token
                 )
 
-                if destination:
+                if row.destination:
                     self._device_list_federation_stream_cache.entity_has_changed(
-                        destination, stream_id
+                        row.destination, token
                     )
-
-        return super(SlavedDeviceStore, self).process_replication(result)
+        return super(SlavedDeviceStore, self).process_replication_rows(
+            stream_name, token, rows
+        )
diff --git a/synapse/replication/slave/storage/events.py b/synapse/replication/slave/storage/events.py
index d4db1e452e..5fd47706ef 100644
--- a/synapse/replication/slave/storage/events.py
+++ b/synapse/replication/slave/storage/events.py
@@ -201,48 +201,25 @@ class SlavedEventStore(BaseSlavedStore):
         result["backfill"] = -self._backfill_id_gen.get_current_token()
         return result
 
-    def process_replication(self, result):
-        stream = result.get("events")
-        if stream:
-            self._stream_id_gen.advance(int(stream["position"]))
-
-            if stream["rows"]:
-                logger.info("Got %d event rows", len(stream["rows"]))
-
-            for row in stream["rows"]:
-                self._process_replication_row(
-                    row, backfilled=False,
+    def process_replication_rows(self, stream_name, token, rows):
+        if stream_name == "events":
+            self._stream_id_gen.advance(token)
+            for row in rows:
+                self.invalidate_caches_for_event(
+                    token, row.event_id, row.room_id, row.type, row.state_key,
+                    row.redacts,
+                    backfilled=False,
                 )
-
-        stream = result.get("backfill")
-        if stream:
-            self._backfill_id_gen.advance(-int(stream["position"]))
-            for row in stream["rows"]:
-                self._process_replication_row(
-                    row, backfilled=True,
+        elif stream_name == "backfill":
+            self._backfill_id_gen.advance(-token)
+            for row in rows:
+                self.invalidate_caches_for_event(
+                    -token, row.event_id, row.room_id, row.type, row.state_key,
+                    row.redacts,
+                    backfilled=True,
                 )
-
-        stream = result.get("forward_ex_outliers")
-        if stream:
-            self._stream_id_gen.advance(int(stream["position"]))
-            for row in stream["rows"]:
-                event_id = row[1]
-                self._invalidate_get_event_cache(event_id)
-
-        stream = result.get("backward_ex_outliers")
-        if stream:
-            self._backfill_id_gen.advance(-int(stream["position"]))
-            for row in stream["rows"]:
-                event_id = row[1]
-                self._invalidate_get_event_cache(event_id)
-
-        return super(SlavedEventStore, self).process_replication(result)
-
-    def _process_replication_row(self, row, backfilled):
-        stream_ordering = row[0] if not backfilled else -row[0]
-        self.invalidate_caches_for_event(
-            stream_ordering, row[1], row[2], row[3], row[4], row[5],
-            backfilled=backfilled,
+        return super(SlavedEventStore, self).process_replication_rows(
+            stream_name, token, rows
         )
 
     def invalidate_caches_for_event(self, stream_ordering, event_id, room_id,
diff --git a/synapse/replication/slave/storage/presence.py b/synapse/replication/slave/storage/presence.py
index e4a2414d78..dffc80adc3 100644
--- a/synapse/replication/slave/storage/presence.py
+++ b/synapse/replication/slave/storage/presence.py
@@ -48,15 +48,14 @@ class SlavedPresenceStore(BaseSlavedStore):
         result["presence"] = position
         return result
 
-    def process_replication(self, result):
-        stream = result.get("presence")
-        if stream:
-            self._presence_id_gen.advance(int(stream["position"]))
-            for row in stream["rows"]:
-                position, user_id = row[:2]
+    def process_replication_rows(self, stream_name, token, rows):
+        if stream_name == "presence":
+            self._presence_id_gen.advance(token)
+            for row in rows:
                 self.presence_stream_cache.entity_has_changed(
-                    user_id, position
+                    row.user_id, token
                 )
-                self._get_presence_for_user.invalidate((user_id,))
-
-        return super(SlavedPresenceStore, self).process_replication(result)
+                self._get_presence_for_user.invalidate((row.user_id,))
+        return super(SlavedPresenceStore, self).process_replication_rows(
+            stream_name, token, rows
+        )
diff --git a/synapse/replication/slave/storage/push_rule.py b/synapse/replication/slave/storage/push_rule.py
index 21ceb0213a..83e880fdd2 100644
--- a/synapse/replication/slave/storage/push_rule.py
+++ b/synapse/replication/slave/storage/push_rule.py
@@ -50,18 +50,15 @@ class SlavedPushRuleStore(SlavedEventStore):
         result["push_rules"] = self._push_rules_stream_id_gen.get_current_token()
         return result
 
-    def process_replication(self, result):
-        stream = result.get("push_rules")
-        if stream:
-            for row in stream["rows"]:
-                position = row[0]
-                user_id = row[2]
-                self.get_push_rules_for_user.invalidate((user_id,))
-                self.get_push_rules_enabled_for_user.invalidate((user_id,))
+    def process_replication_rows(self, stream_name, token, rows):
+        if stream_name == "push_rules":
+            self._push_rules_stream_id_gen.advance(token)
+            for row in rows:
+                self.get_push_rules_for_user.invalidate((row.user_id,))
+                self.get_push_rules_enabled_for_user.invalidate((row.user_id,))
                 self.push_rules_stream_cache.entity_has_changed(
-                    user_id, position
+                    row.user_id, token
                 )
-
-            self._push_rules_stream_id_gen.advance(int(stream["position"]))
-
-        return super(SlavedPushRuleStore, self).process_replication(result)
+        return super(SlavedPushRuleStore, self).process_replication_rows(
+            stream_name, token, rows
+        )
diff --git a/synapse/replication/slave/storage/pushers.py b/synapse/replication/slave/storage/pushers.py
index d88206b3bb..4e8d68ece9 100644
--- a/synapse/replication/slave/storage/pushers.py
+++ b/synapse/replication/slave/storage/pushers.py
@@ -40,13 +40,9 @@ class SlavedPusherStore(BaseSlavedStore):
         result["pushers"] = self._pushers_id_gen.get_current_token()
         return result
 
-    def process_replication(self, result):
-        stream = result.get("pushers")
-        if stream:
-            self._pushers_id_gen.advance(int(stream["position"]))
-
-        stream = result.get("deleted_pushers")
-        if stream:
-            self._pushers_id_gen.advance(int(stream["position"]))
-
-        return super(SlavedPusherStore, self).process_replication(result)
+    def process_replication_rows(self, stream_name, token, rows):
+        if stream_name == "pushers":
+            self._pushers_id_gen.advance(token)
+        return super(SlavedPusherStore, self).process_replication_rows(
+            stream_name, token, rows
+        )
diff --git a/synapse/replication/slave/storage/receipts.py b/synapse/replication/slave/storage/receipts.py
index ac9662d399..b371574ece 100644
--- a/synapse/replication/slave/storage/receipts.py
+++ b/synapse/replication/slave/storage/receipts.py
@@ -65,20 +65,22 @@ class SlavedReceiptsStore(BaseSlavedStore):
         result["receipts"] = self._receipts_id_gen.get_current_token()
         return result
 
-    def process_replication(self, result):
-        stream = result.get("receipts")
-        if stream:
-            self._receipts_id_gen.advance(int(stream["position"]))
-            for row in stream["rows"]:
-                position, room_id, receipt_type, user_id = row[:4]
-                self.invalidate_caches_for_receipt(room_id, receipt_type, user_id)
-                self._receipts_stream_cache.entity_has_changed(room_id, position)
-
-        return super(SlavedReceiptsStore, self).process_replication(result)
-
     def invalidate_caches_for_receipt(self, room_id, receipt_type, user_id):
         self.get_receipts_for_user.invalidate((user_id, receipt_type))
         self.get_linearized_receipts_for_room.invalidate_many((room_id,))
         self.get_last_receipt_event_id_for_user.invalidate(
             (user_id, room_id, receipt_type)
         )
+
+    def process_replication_rows(self, stream_name, token, rows):
+        if stream_name == "receipts":
+            self._receipts_id_gen.advance(token)
+            for row in rows:
+                self.invalidate_caches_for_receipt(
+                    row.room_id, row.receipt_type, row.user_id
+                )
+                self._receipts_stream_cache.entity_has_changed(row.room_id, token)
+
+        return super(SlavedReceiptsStore, self).process_replication_rows(
+            stream_name, token, rows
+        )
diff --git a/synapse/replication/slave/storage/room.py b/synapse/replication/slave/storage/room.py
index 6df9a25ef3..f510384033 100644
--- a/synapse/replication/slave/storage/room.py
+++ b/synapse/replication/slave/storage/room.py
@@ -46,9 +46,10 @@ class RoomStore(BaseSlavedStore):
         result["public_rooms"] = self._public_room_id_gen.get_current_token()
         return result
 
-    def process_replication(self, result):
-        stream = result.get("public_rooms")
-        if stream:
-            self._public_room_id_gen.advance(int(stream["position"]))
+    def process_replication_rows(self, stream_name, token, rows):
+        if stream_name == "public_rooms":
+            self._public_room_id_gen.advance(token)
 
-        return super(RoomStore, self).process_replication(result)
+        return super(RoomStore, self).process_replication_rows(
+            stream_name, token, rows
+        )
diff --git a/synapse/replication/tcp/__init__.py b/synapse/replication/tcp/__init__.py
new file mode 100644
index 0000000000..81c2ea7ee9
--- /dev/null
+++ b/synapse/replication/tcp/__init__.py
@@ -0,0 +1,30 @@
+# -*- coding: utf-8 -*-
+# Copyright 2017 Vector Creations Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""This module implements the TCP replication protocol used by synapse to
+communicate between the master process and its workers (when they're enabled).
+
+Further details can be found in docs/tcp_replication.rst
+
+
+Structure of the module:
+ * client.py   - the client classes used for workers to connect to master
+ * command.py  - the definitions of all the valid commands
+ * protocol.py - contains bot the client and server protocol implementations,
+                 these should not be used directly
+ * resource.py - the server classes that accepts and handle client connections
+ * streams.py  - the definitons of all the valid streams
+
+"""
diff --git a/synapse/replication/tcp/client.py b/synapse/replication/tcp/client.py
new file mode 100644
index 0000000000..90fb6c1336
--- /dev/null
+++ b/synapse/replication/tcp/client.py
@@ -0,0 +1,196 @@
+# -*- coding: utf-8 -*-
+# Copyright 2017 Vector Creations Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""A replication client for use by synapse workers.
+"""
+
+from twisted.internet import reactor, defer
+from twisted.internet.protocol import ReconnectingClientFactory
+
+from .commands import (
+    FederationAckCommand, UserSyncCommand, RemovePusherCommand, InvalidateCacheCommand,
+)
+from .protocol import ClientReplicationStreamProtocol
+
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class ReplicationClientFactory(ReconnectingClientFactory):
+    """Factory for building connections to the master. Will reconnect if the
+    connection is lost.
+
+    Accepts a handler that will be called when new data is available or data
+    is required.
+    """
+    maxDelay = 5  # Try at least once every N seconds
+
+    def __init__(self, hs, client_name, handler):
+        self.client_name = client_name
+        self.handler = handler
+        self.server_name = hs.config.server_name
+        self._clock = hs.get_clock()  # As self.clock is defined in super class
+
+        reactor.addSystemEventTrigger("before", "shutdown", self.stopTrying)
+
+    def startedConnecting(self, connector):
+        logger.info("Connecting to replication: %r", connector.getDestination())
+
+    def buildProtocol(self, addr):
+        logger.info("Connected to replication: %r", addr)
+        self.resetDelay()
+        return ClientReplicationStreamProtocol(
+            self.client_name, self.server_name, self._clock, self.handler
+        )
+
+    def clientConnectionLost(self, connector, reason):
+        logger.error("Lost replication conn: %r", reason)
+        ReconnectingClientFactory.clientConnectionLost(self, connector, reason)
+
+    def clientConnectionFailed(self, connector, reason):
+        logger.error("Failed to connect to replication: %r", reason)
+        ReconnectingClientFactory.clientConnectionFailed(
+            self, connector, reason
+        )
+
+
+class ReplicationClientHandler(object):
+    """A base handler that can be passed to the ReplicationClientFactory.
+
+    By default proxies incoming replication data to the SlaveStore.
+    """
+    def __init__(self, store):
+        self.store = store
+
+        # The current connection. None if we are currently (re)connecting
+        self.connection = None
+
+        # Any pending commands to be sent once a new connection has been
+        # established
+        self.pending_commands = []
+
+        # Map from string -> deferred, to wake up when receiveing a SYNC with
+        # the given string.
+        # Used for tests.
+        self.awaiting_syncs = {}
+
+    def start_replication(self, hs):
+        """Helper method to start a replication connection to the remote server
+        using TCP.
+        """
+        client_name = hs.config.worker_name
+        factory = ReplicationClientFactory(hs, client_name, self)
+        host = hs.config.worker_replication_host
+        port = hs.config.worker_replication_port
+        reactor.connectTCP(host, port, factory)
+
+    def on_rdata(self, stream_name, token, rows):
+        """Called when we get new replication data. By default this just pokes
+        the slave store.
+
+        Can be overriden in subclasses to handle more.
+        """
+        logger.info("Received rdata %s -> %s", stream_name, token)
+        self.store.process_replication_rows(stream_name, token, rows)
+
+    def on_position(self, stream_name, token):
+        """Called when we get new position data. By default this just pokes
+        the slave store.
+
+        Can be overriden in subclasses to handle more.
+        """
+        self.store.process_replication_rows(stream_name, token, [])
+
+    def on_sync(self, data):
+        """When we received a SYNC we wake up any deferreds that were waiting
+        for the sync with the given data.
+
+        Used by tests.
+        """
+        d = self.awaiting_syncs.pop(data, None)
+        if d:
+            d.callback(data)
+
+    def get_streams_to_replicate(self):
+        """Called when a new connection has been established and we need to
+        subscribe to streams.
+
+        Returns a dictionary of stream name to token.
+        """
+        args = self.store.stream_positions()
+        user_account_data = args.pop("user_account_data", None)
+        room_account_data = args.pop("room_account_data", None)
+        if user_account_data:
+            args["account_data"] = user_account_data
+        elif room_account_data:
+            args["account_data"] = room_account_data
+        return args
+
+    def get_currently_syncing_users(self):
+        """Get the list of currently syncing users (if any). This is called
+        when a connection has been established and we need to send the
+        currently syncing users. (Overriden by the synchrotron's only)
+        """
+        return []
+
+    def send_command(self, cmd):
+        """Send a command to master (when we get establish a connection if we
+        don't have one already.)
+        """
+        if self.connection:
+            self.connection.send_command(cmd)
+        else:
+            logger.warn("Queuing command as not connected: %r", cmd.NAME)
+            self.pending_commands.append(cmd)
+
+    def send_federation_ack(self, token):
+        """Ack data for the federation stream. This allows the master to drop
+        data stored purely in memory.
+        """
+        self.send_command(FederationAckCommand(token))
+
+    def send_user_sync(self, user_id, is_syncing, last_sync_ms):
+        """Poke the master that a user has started/stopped syncing.
+        """
+        self.send_command(UserSyncCommand(user_id, is_syncing, last_sync_ms))
+
+    def send_remove_pusher(self, app_id, push_key, user_id):
+        """Poke the master to remove a pusher for a user
+        """
+        cmd = RemovePusherCommand(app_id, push_key, user_id)
+        self.send_command(cmd)
+
+    def send_invalidate_cache(self, cache_func, keys):
+        """Poke the master to invalidate a cache.
+        """
+        cmd = InvalidateCacheCommand(cache_func.__name__, keys)
+        self.send_command(cmd)
+
+    def await_sync(self, data):
+        """Returns a deferred that is resolved when we receive a SYNC command
+        with given data.
+
+        Used by tests.
+        """
+        return self.awaiting_syncs.setdefault(data, defer.Deferred())
+
+    def update_connection(self, connection):
+        """Called when a connection has been established (or lost with None).
+        """
+        self.connection = connection
+        if connection:
+            for cmd in self.pending_commands:
+                connection.send_command(cmd)
+            self.pending_commands = []
diff --git a/synapse/replication/tcp/commands.py b/synapse/replication/tcp/commands.py
new file mode 100644
index 0000000000..84d2a2272a
--- /dev/null
+++ b/synapse/replication/tcp/commands.py
@@ -0,0 +1,346 @@
+# -*- coding: utf-8 -*-
+# Copyright 2017 Vector Creations Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Defines the various valid commands
+
+The VALID_SERVER_COMMANDS and VALID_CLIENT_COMMANDS define which commands are
+allowed to be sent by which side.
+"""
+
+import logging
+import ujson as json
+
+
+logger = logging.getLogger(__name__)
+
+
+class Command(object):
+    """The base command class.
+
+    All subclasses must set the NAME variable which equates to the name of the
+    command on the wire.
+
+    A full command line on the wire is constructed from `NAME + " " + to_line()`
+
+    The default implementation creates a command of form `<NAME> <data>`
+    """
+    NAME = None
+
+    def __init__(self, data):
+        self.data = data
+
+    @classmethod
+    def from_line(cls, line):
+        """Deserialises a line from the wire into this command. `line` does not
+        include the command.
+        """
+        return cls(line)
+
+    def to_line(self):
+        """Serialises the comamnd for the wire. Does not include the command
+        prefix.
+        """
+        return self.data
+
+
+class ServerCommand(Command):
+    """Sent by the server on new connection and includes the server_name.
+
+    Format::
+
+        SERVER <server_name>
+    """
+    NAME = "SERVER"
+
+
+class RdataCommand(Command):
+    """Sent by server when a subscribed stream has an update.
+
+    Format::
+
+        RDATA <stream_name> <token> <row_json>
+
+    The `<token>` may either be a numeric stream id OR "batch". The latter case
+    is used to support sending multiple updates with the same stream ID. This
+    is done by sending an RDATA for each row, with all but the last RDATA having
+    a token of "batch" and the last having the final stream ID.
+
+    The client should batch all incoming RDATA with a token of "batch" (per
+    stream_name) until it sees an RDATA with a numeric stream ID.
+
+    `<token>` of "batch" maps to the instance variable `token` being None.
+
+    An example of a batched series of RDATA::
+
+        RDATA presence batch ["@foo:example.com", "online", ...]
+        RDATA presence batch ["@bar:example.com", "online", ...]
+        RDATA presence 59 ["@baz:example.com", "online", ...]
+    """
+    NAME = "RDATA"
+
+    def __init__(self, stream_name, token, row):
+        self.stream_name = stream_name
+        self.token = token
+        self.row = row
+
+    @classmethod
+    def from_line(cls, line):
+        stream_name, token, row_json = line.split(" ", 2)
+        return cls(
+            stream_name,
+            None if token == "batch" else int(token),
+            json.loads(row_json)
+        )
+
+    def to_line(self):
+        return " ".join((
+            self.stream_name,
+            str(self.token) if self.token is not None else "batch",
+            json.dumps(self.row),
+        ))
+
+
+class PositionCommand(Command):
+    """Sent by the client to tell the client the stream postition without
+    needing to send an RDATA.
+    """
+    NAME = "POSITION"
+
+    def __init__(self, stream_name, token):
+        self.stream_name = stream_name
+        self.token = token
+
+    @classmethod
+    def from_line(cls, line):
+        stream_name, token = line.split(" ", 1)
+        return cls(stream_name, int(token))
+
+    def to_line(self):
+        return " ".join((self.stream_name, str(self.token),))
+
+
+class ErrorCommand(Command):
+    """Sent by either side if there was an ERROR. The data is a string describing
+    the error.
+    """
+    NAME = "ERROR"
+
+
+class PingCommand(Command):
+    """Sent by either side as a keep alive. The data is arbitary (often timestamp)
+    """
+    NAME = "PING"
+
+
+class NameCommand(Command):
+    """Sent by client to inform the server of the client's identity. The data
+    is the name
+    """
+    NAME = "NAME"
+
+
+class ReplicateCommand(Command):
+    """Sent by the client to subscribe to the stream.
+
+    Format::
+
+        REPLICATE <stream_name> <token>
+
+    Where <token> may be either:
+        * a numeric stream_id to stream updates from
+        * "NOW" to stream all subsequent updates.
+
+    The <stream_name> can be "ALL" to subscribe to all known streams, in which
+    case the <token> must be set to "NOW", i.e.::
+
+        REPLICATE ALL NOW
+    """
+    NAME = "REPLICATE"
+
+    def __init__(self, stream_name, token):
+        self.stream_name = stream_name
+        self.token = token
+
+    @classmethod
+    def from_line(cls, line):
+        stream_name, token = line.split(" ", 1)
+        if token in ("NOW", "now"):
+            token = "NOW"
+        else:
+            token = int(token)
+        return cls(stream_name, token)
+
+    def to_line(self):
+        return " ".join((self.stream_name, str(self.token),))
+
+
+class UserSyncCommand(Command):
+    """Sent by the client to inform the server that a user has started or
+    stopped syncing. Used to calculate presence on the master.
+
+    Includes a timestamp of when the last user sync was.
+
+    Format::
+
+        USER_SYNC <user_id> <state> <last_sync_ms>
+
+    Where <state> is either "start" or "stop"
+    """
+    NAME = "USER_SYNC"
+
+    def __init__(self, user_id, is_syncing, last_sync_ms):
+        self.user_id = user_id
+        self.is_syncing = is_syncing
+        self.last_sync_ms = last_sync_ms
+
+    @classmethod
+    def from_line(cls, line):
+        user_id, state, last_sync_ms = line.split(" ", 2)
+
+        if state not in ("start", "end"):
+            raise Exception("Invalid USER_SYNC state %r" % (state,))
+
+        return cls(user_id, state == "start", int(last_sync_ms))
+
+    def to_line(self):
+        return " ".join((
+            self.user_id, "start" if self.is_syncing else "end", str(self.last_sync_ms),
+        ))
+
+
+class FederationAckCommand(Command):
+    """Sent by the client when it has processed up to a given point in the
+    federation stream. This allows the master to drop in-memory caches of the
+    federation stream.
+
+    This must only be sent from one worker (i.e. the one sending federation)
+
+    Format::
+
+        FEDERATION_ACK <token>
+    """
+    NAME = "FEDERATION_ACK"
+
+    def __init__(self, token):
+        self.token = token
+
+    @classmethod
+    def from_line(cls, line):
+        return cls(int(line))
+
+    def to_line(self):
+        return str(self.token)
+
+
+class SyncCommand(Command):
+    """Used for testing. The client protocol implementation allows waiting
+    on a SYNC command with a specified data.
+    """
+    NAME = "SYNC"
+
+
+class RemovePusherCommand(Command):
+    """Sent by the client to request the master remove the given pusher.
+
+    Format::
+
+        REMOVE_PUSHER <app_id> <push_key> <user_id>
+    """
+    NAME = "REMOVE_PUSHER"
+
+    def __init__(self, app_id, push_key, user_id):
+        self.user_id = user_id
+        self.app_id = app_id
+        self.push_key = push_key
+
+    @classmethod
+    def from_line(cls, line):
+        app_id, push_key, user_id = line.split(" ", 2)
+
+        return cls(app_id, push_key, user_id)
+
+    def to_line(self):
+        return " ".join((self.app_id, self.push_key, self.user_id))
+
+
+class InvalidateCacheCommand(Command):
+    """Sent by the client to invalidate an upstream cache.
+
+    THIS IS NOT RELIABLE, AND SHOULD *NOT* BE USED ACCEPT FOR THINGS THAT ARE
+    NOT DISASTROUS IF WE DROP ON THE FLOOR.
+
+    Mainly used to invalidate destination retry timing caches.
+
+    Format::
+
+        INVALIDATE_CACHE <cache_func> <keys_json>
+
+    Where <keys_json> is a json list.
+    """
+    NAME = "INVALIDATE_CACHE"
+
+    def __init__(self, cache_func, keys):
+        self.cache_func = cache_func
+        self.keys = keys
+
+    @classmethod
+    def from_line(cls, line):
+        cache_func, keys_json = line.split(" ", 1)
+
+        return cls(cache_func, json.loads(keys_json))
+
+    def to_line(self):
+        return " ".join((self.cache_func, json.dumps(self.keys)))
+
+
+# Map of command name to command type.
+COMMAND_MAP = {
+    cmd.NAME: cmd
+    for cmd in (
+        ServerCommand,
+        RdataCommand,
+        PositionCommand,
+        ErrorCommand,
+        PingCommand,
+        NameCommand,
+        ReplicateCommand,
+        UserSyncCommand,
+        FederationAckCommand,
+        SyncCommand,
+        RemovePusherCommand,
+        InvalidateCacheCommand,
+    )
+}
+
+# The commands the server is allowed to send
+VALID_SERVER_COMMANDS = (
+    ServerCommand.NAME,
+    RdataCommand.NAME,
+    PositionCommand.NAME,
+    ErrorCommand.NAME,
+    PingCommand.NAME,
+    SyncCommand.NAME,
+)
+
+# The commands the client is allowed to send
+VALID_CLIENT_COMMANDS = (
+    NameCommand.NAME,
+    ReplicateCommand.NAME,
+    PingCommand.NAME,
+    UserSyncCommand.NAME,
+    FederationAckCommand.NAME,
+    RemovePusherCommand.NAME,
+    InvalidateCacheCommand.NAME,
+    ErrorCommand.NAME,
+)
diff --git a/synapse/replication/tcp/protocol.py b/synapse/replication/tcp/protocol.py
new file mode 100644
index 0000000000..9fee2a484b
--- /dev/null
+++ b/synapse/replication/tcp/protocol.py
@@ -0,0 +1,640 @@
+# -*- coding: utf-8 -*-
+# Copyright 2017 Vector Creations Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""This module contains the implementation of both the client and server
+protocols.
+
+The basic structure of the protocol is line based, where the initial word of
+each line specifies the command. The rest of the line is parsed based on the
+command. For example, the `RDATA` command is defined as::
+
+    RDATA <stream_name> <token> <row_json>
+
+(Note that `<row_json>` may contains spaces, but cannot contain newlines.)
+
+Blank lines are ignored.
+
+# Example
+
+An example iteraction is shown below. Each line is prefixed with '>' or '<' to
+indicate which side is sending, these are *not* included on the wire::
+
+    * connection established *
+    > SERVER localhost:8823
+    > PING 1490197665618
+    < NAME synapse.app.appservice
+    < PING 1490197665618
+    < REPLICATE events 1
+    < REPLICATE backfill 1
+    < REPLICATE caches 1
+    > POSITION events 1
+    > POSITION backfill 1
+    > POSITION caches 1
+    > RDATA caches 2 ["get_user_by_id",["@01register-user:localhost:8823"],1490197670513]
+    > RDATA events 14 ["$149019767112vOHxz:localhost:8823",
+        "!AFDCvgApUmpdfVjIXm:localhost:8823","m.room.guest_access","",null]
+    < PING 1490197675618
+    > ERROR server stopping
+    * connection closed by server *
+"""
+
+from twisted.internet import defer
+from twisted.protocols.basic import LineOnlyReceiver
+from twisted.python.failure import Failure
+
+from commands import (
+    COMMAND_MAP, VALID_CLIENT_COMMANDS, VALID_SERVER_COMMANDS,
+    ErrorCommand, ServerCommand, RdataCommand, PositionCommand, PingCommand,
+    NameCommand, ReplicateCommand, UserSyncCommand, SyncCommand,
+)
+from streams import STREAMS_MAP
+
+from synapse.util.stringutils import random_string
+from synapse.metrics.metric import CounterMetric
+
+import logging
+import synapse.metrics
+import struct
+import fcntl
+
+
+metrics = synapse.metrics.get_metrics_for(__name__)
+
+connection_close_counter = metrics.register_counter(
+    "close_reason", labels=["reason_type"],
+)
+
+
+# A list of all connected protocols. This allows us to send metrics about the
+# connections.
+connected_connections = []
+
+
+logger = logging.getLogger(__name__)
+
+
+PING_TIME = 5000
+PING_TIMEOUT_MULTIPLIER = 5
+PING_TIMEOUT_MS = PING_TIME * PING_TIMEOUT_MULTIPLIER
+
+
+class ConnectionStates(object):
+    CONNECTING = "connecting"
+    ESTABLISHED = "established"
+    PAUSED = "paused"
+    CLOSED = "closed"
+
+
+class BaseReplicationStreamProtocol(LineOnlyReceiver):
+    """Base replication protocol shared between client and server.
+
+    Reads lines (ignoring blank ones) and parses them into command classes,
+    asserting that they are valid for the given direction, i.e. server commands
+    are only sent by the server.
+
+    On receiving a new command it calls `on_<COMMAND_NAME>` with the parsed
+    command.
+
+    It also sends `PING` periodically, and correctly times out remote connections
+    (if they send a `PING` command)
+    """
+    delimiter = b'\n'
+
+    VALID_INBOUND_COMMANDS = []  # Valid commands we expect to receive
+    VALID_OUTBOUND_COMMANDS = []  # Valid commans we can send
+
+    max_line_buffer = 10000
+
+    def __init__(self, clock):
+        self.clock = clock
+
+        self.last_received_command = self.clock.time_msec()
+        self.last_sent_command = 0
+        self.time_we_closed = None  # When we requested the connection be closed
+
+        self.received_ping = False  # Have we reecived a ping from the other side
+
+        self.state = ConnectionStates.CONNECTING
+
+        self.name = "anon"  # The name sent by a client.
+        self.conn_id = random_string(5)  # To dedupe in case of name clashes.
+
+        # List of pending commands to send once we've established the connection
+        self.pending_commands = []
+
+        # The LoopingCall for sending pings.
+        self._send_ping_loop = None
+
+        self.inbound_commands_counter = CounterMetric(
+            "inbound_commands", labels=["command"],
+        )
+        self.outbound_commands_counter = CounterMetric(
+            "outbound_commands", labels=["command"],
+        )
+
+    def connectionMade(self):
+        logger.info("[%s] Connection established", self.id())
+
+        self.state = ConnectionStates.ESTABLISHED
+
+        connected_connections.append(self)  # Register connection for metrics
+
+        self.transport.registerProducer(self, True)  # For the *Producing callbacks
+
+        self._send_pending_commands()
+
+        # Starts sending pings
+        self._send_ping_loop = self.clock.looping_call(self.send_ping, 5000)
+
+        # Always send the initial PING so that the other side knows that they
+        # can time us out.
+        self.send_command(PingCommand(self.clock.time_msec()))
+
+    def send_ping(self):
+        """Periodically sends a ping and checks if we should close the connection
+        due to the other side timing out.
+        """
+        now = self.clock.time_msec()
+
+        if self.time_we_closed:
+            if now - self.time_we_closed > PING_TIMEOUT_MS:
+                logger.info(
+                    "[%s] Failed to close connection gracefully, aborting", self.id()
+                )
+                self.transport.abortConnection()
+        else:
+            if now - self.last_sent_command >= PING_TIME:
+                self.send_command(PingCommand(now))
+
+            if self.received_ping and now - self.last_received_command > PING_TIMEOUT_MS:
+                logger.info(
+                    "[%s] Connection hasn't received command in %r ms. Closing.",
+                    self.id(), now - self.last_received_command
+                )
+                self.send_error("ping timeout")
+
+    def lineReceived(self, line):
+        """Called when we've received a line
+        """
+        if line.strip() == "":
+            # Ignore blank lines
+            return
+
+        line = line.decode("utf-8")
+        cmd_name, rest_of_line = line.split(" ", 1)
+
+        if cmd_name not in self.VALID_INBOUND_COMMANDS:
+            logger.error("[%s] invalid command %s", self.id(), cmd_name)
+            self.send_error("invalid command: %s", cmd_name)
+            return
+
+        self.last_received_command = self.clock.time_msec()
+
+        self.inbound_commands_counter.inc(cmd_name)
+
+        cmd_cls = COMMAND_MAP[cmd_name]
+        try:
+            cmd = cmd_cls.from_line(rest_of_line)
+        except Exception as e:
+            logger.exception(
+                "[%s] failed to parse line %r: %r", self.id(), cmd_name, rest_of_line
+            )
+            self.send_error(
+                "failed to parse line for  %r: %r (%r):" % (cmd_name, e, rest_of_line)
+            )
+            return
+
+        # Now lets try and call on_<CMD_NAME> function
+        try:
+            getattr(self, "on_%s" % (cmd_name,))(cmd)
+        except Exception:
+            logger.exception("[%s] Failed to handle line: %r", self.id(), line)
+
+    def close(self):
+        logger.warn("[%s] Closing connection", self.id())
+        self.time_we_closed = self.clock.time_msec()
+        self.transport.loseConnection()
+        self.on_connection_closed()
+
+    def send_error(self, error_string, *args):
+        """Send an error to remote and close the connection.
+        """
+        self.send_command(ErrorCommand(error_string % args))
+        self.close()
+
+    def send_command(self, cmd, do_buffer=True):
+        """Send a command if connection has been established.
+
+        Args:
+            cmd (Command)
+            do_buffer (bool): Whether to buffer the message or always attempt
+                to send the command. This is mostly used to send an error
+                message if we're about to close the connection due our buffers
+                becoming full.
+        """
+        if self.state == ConnectionStates.CLOSED:
+            logger.info("[%s] Not sending, connection closed", self.id())
+            return
+
+        if do_buffer and self.state != ConnectionStates.ESTABLISHED:
+            self._queue_command(cmd)
+            return
+
+        self.outbound_commands_counter.inc(cmd.NAME)
+
+        string = "%s %s" % (cmd.NAME, cmd.to_line(),)
+        if "\n" in string:
+            raise Exception("Unexpected newline in command: %r", string)
+
+        self.sendLine(string.encode("utf-8"))
+
+        self.last_sent_command = self.clock.time_msec()
+
+    def _queue_command(self, cmd):
+        """Queue the command until the connection is ready to write to again.
+        """
+        logger.info("[%s] Queing as conn %r, cmd: %r", self.id(), self.state, cmd)
+        self.pending_commands.append(cmd)
+
+        if len(self.pending_commands) > self.max_line_buffer:
+            # The other side is failing to keep up and out buffers are becoming
+            # full, so lets close the connection.
+            # XXX: should we squawk more loudly?
+            logger.error("[%s] Remote failed to keep up", self.id())
+            self.send_command(ErrorCommand("Failed to keep up"), do_buffer=False)
+            self.close()
+
+    def _send_pending_commands(self):
+        """Send any queued commandes
+        """
+        pending = self.pending_commands
+        self.pending_commands = []
+        for cmd in pending:
+            self.send_command(cmd)
+
+    def on_PING(self, line):
+        self.received_ping = True
+
+    def on_ERROR(self, cmd):
+        logger.error("[%s] Remote reported error: %r", self.id(), cmd.data)
+
+    def pauseProducing(self):
+        """This is called when both the kernel send buffer and the twisted
+        tcp connection send buffers have become full.
+
+        We don't actually have any control over those sizes, so we buffer some
+        commands ourselves before knifing the connection due to the remote
+        failing to keep up.
+        """
+        logger.info("[%s] Pause producing", self.id())
+        self.state = ConnectionStates.PAUSED
+
+    def resumeProducing(self):
+        """The remote has caught up after we started buffering!
+        """
+        logger.info("[%s] Resume producing", self.id())
+        self.state = ConnectionStates.ESTABLISHED
+        self._send_pending_commands()
+
+    def stopProducing(self):
+        """We're never going to send any more data (normally because either
+        we or the remote has closed the connection)
+        """
+        logger.info("[%s] Stop producing", self.id())
+        self.on_connection_closed()
+
+    def connectionLost(self, reason):
+        logger.info("[%s] Replication connection closed: %r", self.id(), reason)
+        if isinstance(reason, Failure):
+            connection_close_counter.inc(reason.type.__name__)
+        else:
+            connection_close_counter.inc(reason.__class__.__name__)
+
+        try:
+            # Remove us from list of connections to be monitored
+            connected_connections.remove(self)
+        except ValueError:
+            pass
+
+        # Stop the looping call sending pings.
+        if self._send_ping_loop and self._send_ping_loop.running:
+            self._send_ping_loop.stop()
+
+        self.on_connection_closed()
+
+    def on_connection_closed(self):
+        logger.info("[%s] Connection was closed", self.id())
+
+        self.state = ConnectionStates.CLOSED
+        self.pending_commands = []
+
+        if self.transport:
+            self.transport.unregisterProducer()
+
+    def __str__(self):
+        return "ReplicationConnection<name=%s,conn_id=%s,addr=%s>" % (
+            self.name, self.conn_id, self.addr,
+        )
+
+    def id(self):
+        return "%s-%s" % (self.name, self.conn_id)
+
+
+class ServerReplicationStreamProtocol(BaseReplicationStreamProtocol):
+    VALID_INBOUND_COMMANDS = VALID_CLIENT_COMMANDS
+    VALID_OUTBOUND_COMMANDS = VALID_SERVER_COMMANDS
+
+    def __init__(self, server_name, clock, streamer, addr):
+        BaseReplicationStreamProtocol.__init__(self, clock)  # Old style class
+
+        self.server_name = server_name
+        self.streamer = streamer
+        self.addr = addr
+
+        # The streams the client has subscribed to and is up to date with
+        self.replication_streams = set()
+
+        # The streams the client is currently subscribing to.
+        self.connecting_streams = set()
+
+        # Map from stream name to list of updates to send once we've finished
+        # subscribing the client to the stream.
+        self.pending_rdata = {}
+
+    def connectionMade(self):
+        self.send_command(ServerCommand(self.server_name))
+        BaseReplicationStreamProtocol.connectionMade(self)
+        self.streamer.new_connection(self)
+
+    def on_NAME(self, cmd):
+        logger.info("[%s] Renamed to %r", self.id(), cmd.data)
+        self.name = cmd.data
+
+    def on_USER_SYNC(self, cmd):
+        self.streamer.on_user_sync(
+            self.conn_id, cmd.user_id, cmd.is_syncing, cmd.last_sync_ms,
+        )
+
+    def on_REPLICATE(self, cmd):
+        stream_name = cmd.stream_name
+        token = cmd.token
+
+        if stream_name == "ALL":
+            # Subscribe to all streams we're publishing to.
+            for stream in self.streamer.streams_by_name.iterkeys():
+                self.subscribe_to_stream(stream, token)
+        else:
+            self.subscribe_to_stream(stream_name, token)
+
+    def on_FEDERATION_ACK(self, cmd):
+        self.streamer.federation_ack(cmd.token)
+
+    def on_REMOVE_PUSHER(self, cmd):
+        self.streamer.on_remove_pusher(cmd.app_id, cmd.push_key, cmd.user_id)
+
+    def on_INVALIDATE_CACHE(self, cmd):
+        self.streamer.on_invalidate_cache(cmd.cache_func, cmd.keys)
+
+    @defer.inlineCallbacks
+    def subscribe_to_stream(self, stream_name, token):
+        """Subscribe the remote to a streams.
+
+        This invloves checking if they've missed anything and sending those
+        updates down if they have. During that time new updates for the stream
+        are queued and sent once we've sent down any missed updates.
+        """
+        self.replication_streams.discard(stream_name)
+        self.connecting_streams.add(stream_name)
+
+        try:
+            # Get missing updates
+            updates, current_token = yield self.streamer.get_stream_updates(
+                stream_name, token,
+            )
+
+            # Send all the missing updates
+            for update in updates:
+                token, row = update[0], update[1]
+                self.send_command(RdataCommand(stream_name, token, row))
+
+            # We send a POSITION command to ensure that they have an up to
+            # date token (especially useful if we didn't send any updates
+            # above)
+            self.send_command(PositionCommand(stream_name, current_token))
+
+            # Now we can send any updates that came in while we were subscribing
+            pending_rdata = self.pending_rdata.pop(stream_name, [])
+            for token, update in pending_rdata:
+                # Only send updates newer than the current token
+                if token > current_token:
+                    self.send_command(RdataCommand(stream_name, token, update))
+
+            # They're now fully subscribed
+            self.replication_streams.add(stream_name)
+        except Exception as e:
+            logger.exception("[%s] Failed to handle REPLICATE command", self.id())
+            self.send_error("failed to handle replicate: %r", e)
+        finally:
+            self.connecting_streams.discard(stream_name)
+
+    def stream_update(self, stream_name, token, data):
+        """Called when a new update is available to stream to clients.
+
+        We need to check if the client is interested in the stream or not
+        """
+        if stream_name in self.replication_streams:
+            # The client is subscribed to the stream
+            self.send_command(RdataCommand(stream_name, token, data))
+        elif stream_name in self.connecting_streams:
+            # The client is being subscribed to the stream
+            logger.debug("[%s] Queuing RDATA %r %r", self.id(), stream_name, token)
+            self.pending_rdata.setdefault(stream_name, []).append((token, data))
+        else:
+            # The client isn't subscribed
+            logger.debug("[%s] Dropping RDATA %r %r", self.id(), stream_name, token)
+
+    def send_sync(self, data):
+        self.send_command(SyncCommand(data))
+
+    def on_connection_closed(self):
+        BaseReplicationStreamProtocol.on_connection_closed(self)
+        self.streamer.lost_connection(self)
+
+
+class ClientReplicationStreamProtocol(BaseReplicationStreamProtocol):
+    VALID_INBOUND_COMMANDS = VALID_SERVER_COMMANDS
+    VALID_OUTBOUND_COMMANDS = VALID_CLIENT_COMMANDS
+
+    def __init__(self, client_name, server_name, clock, handler):
+        BaseReplicationStreamProtocol.__init__(self, clock)
+
+        self.client_name = client_name
+        self.server_name = server_name
+        self.handler = handler
+
+        # Map of stream to batched updates. See RdataCommand for info on how
+        # batching works.
+        self.pending_batches = {}
+
+    def connectionMade(self):
+        self.send_command(NameCommand(self.client_name))
+        BaseReplicationStreamProtocol.connectionMade(self)
+
+        # Once we've connected subscribe to the necessary streams
+        for stream_name, token in self.handler.get_streams_to_replicate().iteritems():
+            self.replicate(stream_name, token)
+
+        # Tell the server if we have any users currently syncing (should only
+        # happen on synchrotrons)
+        currently_syncing = self.handler.get_currently_syncing_users()
+        now = self.clock.time_msec()
+        for user_id in currently_syncing:
+            self.send_command(UserSyncCommand(user_id, True, now))
+
+        # We've now finished connecting to so inform the client handler
+        self.handler.update_connection(self)
+
+    def on_SERVER(self, cmd):
+        if cmd.data != self.server_name:
+            logger.error("[%s] Connected to wrong remote: %r", self.id(), cmd.data)
+            self.send_error("Wrong remote")
+
+    def on_RDATA(self, cmd):
+        try:
+            row = STREAMS_MAP[cmd.stream_name].ROW_TYPE(*cmd.row)
+        except Exception:
+            logger.exception(
+                "[%s] Failed to parse RDATA: %r %r",
+                self.id(), cmd.stream_name, cmd.row
+            )
+            raise
+
+        if cmd.token is None:
+            # I.e. this is part of a batch of updates for this stream. Batch
+            # until we get an update for the stream with a non None token
+            self.pending_batches.setdefault(cmd.stream_name, []).append(row)
+        else:
+            # Check if this is the last of a batch of updates
+            rows = self.pending_batches.pop(cmd.stream_name, [])
+            rows.append(row)
+
+            self.handler.on_rdata(cmd.stream_name, cmd.token, rows)
+
+    def on_POSITION(self, cmd):
+        self.handler.on_position(cmd.stream_name, cmd.token)
+
+    def on_SYNC(self, cmd):
+        self.handler.on_sync(cmd.data)
+
+    def replicate(self, stream_name, token):
+        """Send the subscription request to the server
+        """
+        if stream_name not in STREAMS_MAP:
+            raise Exception("Invalid stream name %r" % (stream_name,))
+
+        logger.info(
+            "[%s] Subscribing to replication stream: %r from %r",
+            self.id(), stream_name, token
+        )
+
+        self.send_command(ReplicateCommand(stream_name, token))
+
+    def on_connection_closed(self):
+        BaseReplicationStreamProtocol.on_connection_closed(self)
+        self.handler.update_connection(None)
+
+
+# The following simply registers metrics for the replication connections
+
+metrics.register_callback(
+    "pending_commands",
+    lambda: {
+        (p.name, p.conn_id): len(p.pending_commands)
+        for p in connected_connections
+    },
+    labels=["name", "conn_id"],
+)
+
+
+def transport_buffer_size(protocol):
+    if protocol.transport:
+        size = len(protocol.transport.dataBuffer) + protocol.transport._tempDataLen
+        return size
+    return 0
+
+
+metrics.register_callback(
+    "transport_send_buffer",
+    lambda: {
+        (p.name, p.conn_id): transport_buffer_size(p)
+        for p in connected_connections
+    },
+    labels=["name", "conn_id"],
+)
+
+
+def transport_kernel_read_buffer_size(protocol, read=True):
+    SIOCINQ = 0x541B
+    SIOCOUTQ = 0x5411
+
+    if protocol.transport:
+        fileno = protocol.transport.getHandle().fileno()
+        if read:
+            op = SIOCINQ
+        else:
+            op = SIOCOUTQ
+        size = struct.unpack("I", fcntl.ioctl(fileno, op, '\0\0\0\0'))[0]
+        return size
+    return 0
+
+
+metrics.register_callback(
+    "transport_kernel_send_buffer",
+    lambda: {
+        (p.name, p.conn_id): transport_kernel_read_buffer_size(p, False)
+        for p in connected_connections
+    },
+    labels=["name", "conn_id"],
+)
+
+
+metrics.register_callback(
+    "transport_kernel_read_buffer",
+    lambda: {
+        (p.name, p.conn_id): transport_kernel_read_buffer_size(p, True)
+        for p in connected_connections
+    },
+    labels=["name", "conn_id"],
+)
+
+
+metrics.register_callback(
+    "inbound_commands",
+    lambda: {
+        (k[0], p.name, p.conn_id): count
+        for p in connected_connections
+        for k, count in p.inbound_commands_counter.counts.iteritems()
+    },
+    labels=["command", "name", "conn_id"],
+)
+
+metrics.register_callback(
+    "outbound_commands",
+    lambda: {
+        (k[0], p.name, p.conn_id): count
+        for p in connected_connections
+        for k, count in p.outbound_commands_counter.counts.iteritems()
+    },
+    labels=["command", "name", "conn_id"],
+)
diff --git a/synapse/replication/tcp/resource.py b/synapse/replication/tcp/resource.py
new file mode 100644
index 0000000000..8b2c4c3043
--- /dev/null
+++ b/synapse/replication/tcp/resource.py
@@ -0,0 +1,290 @@
+# -*- coding: utf-8 -*-
+# Copyright 2017 Vector Creations Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""The server side of the replication stream.
+"""
+
+from twisted.internet import defer, reactor
+from twisted.internet.protocol import Factory
+
+from streams import STREAMS_MAP, FederationStream
+from protocol import ServerReplicationStreamProtocol
+
+from synapse.util.metrics import Measure, measure_func
+
+import logging
+import synapse.metrics
+
+
+metrics = synapse.metrics.get_metrics_for(__name__)
+stream_updates_counter = metrics.register_counter(
+    "stream_updates", labels=["stream_name"]
+)
+user_sync_counter = metrics.register_counter("user_sync")
+federation_ack_counter = metrics.register_counter("federation_ack")
+remove_pusher_counter = metrics.register_counter("remove_pusher")
+invalidate_cache_counter = metrics.register_counter("invalidate_cache")
+
+logger = logging.getLogger(__name__)
+
+
+class ReplicationStreamProtocolFactory(Factory):
+    """Factory for new replication connections.
+    """
+    def __init__(self, hs):
+        self.streamer = ReplicationStreamer(hs)
+        self.clock = hs.get_clock()
+        self.server_name = hs.config.server_name
+
+    def buildProtocol(self, addr):
+        return ServerReplicationStreamProtocol(
+            self.server_name,
+            self.clock,
+            self.streamer,
+            addr
+        )
+
+
+class ReplicationStreamer(object):
+    """Handles replication connections.
+
+    This needs to be poked when new replication data may be available. When new
+    data is available it will propagate to all connected clients.
+    """
+
+    def __init__(self, hs):
+        self.store = hs.get_datastore()
+        self.presence_handler = hs.get_presence_handler()
+        self.clock = hs.get_clock()
+
+        # Current connections.
+        self.connections = []
+
+        metrics.register_callback("total_connections", lambda: len(self.connections))
+
+        # List of streams that clients can subscribe to.
+        # We only support federation stream if federation sending hase been
+        # disabled on the master.
+        self.streams = [
+            stream(hs) for stream in STREAMS_MAP.itervalues()
+            if stream != FederationStream or not hs.config.send_federation
+        ]
+
+        self.streams_by_name = {stream.NAME: stream for stream in self.streams}
+
+        metrics.register_callback(
+            "connections_per_stream",
+            lambda: {
+                (stream_name,): len([
+                    conn for conn in self.connections
+                    if stream_name in conn.replication_streams
+                ])
+                for stream_name in self.streams_by_name
+            },
+            labels=["stream_name"],
+        )
+
+        self.federation_sender = None
+        if not hs.config.send_federation:
+            self.federation_sender = hs.get_federation_sender()
+
+        hs.get_notifier().add_replication_callback(self.on_notifier_poke)
+
+        # Keeps track of whether we are currently checking for updates
+        self.is_looping = False
+        self.pending_updates = False
+
+        reactor.addSystemEventTrigger("before", "shutdown", self.on_shutdown)
+
+    def on_shutdown(self):
+        # close all connections on shutdown
+        for conn in self.connections:
+            conn.send_error("server shutting down")
+
+    @defer.inlineCallbacks
+    def on_notifier_poke(self):
+        """Checks if there is actually any new data and sends it to the
+        connections if there are.
+
+        This should get called each time new data is available, even if it
+        is currently being executed, so that nothing gets missed
+        """
+        if not self.connections:
+            # Don't bother if nothing is listening. We still need to advance
+            # the stream tokens otherwise they'll fall beihind forever
+            for stream in self.streams:
+                stream.discard_updates_and_advance()
+            return
+
+        # If we're in the process of checking for new updates, mark that fact
+        # and return
+        if self.is_looping:
+            logger.debug("Noitifier poke loop already running")
+            self.pending_updates = True
+            return
+
+        self.pending_updates = True
+        self.is_looping = True
+
+        try:
+            # Keep looping while there have been pokes about potential updates.
+            # This protects against the race where a stream we already checked
+            # gets an update while we're handling other streams.
+            while self.pending_updates:
+                self.pending_updates = False
+
+                with Measure(self.clock, "repl.stream.get_updates"):
+                    # First we tell the streams that they should update their
+                    # current tokens.
+                    for stream in self.streams:
+                        stream.advance_current_token()
+
+                    for stream in self.streams:
+                        if stream.last_token == stream.upto_token:
+                            continue
+
+                        logger.debug(
+                            "Getting stream: %s: %s -> %s",
+                            stream.NAME, stream.last_token, stream.upto_token
+                        )
+                        updates, current_token = yield stream.get_updates()
+
+                        logger.debug(
+                            "Sending %d updates to %d connections",
+                            len(updates), len(self.connections),
+                        )
+
+                        if updates:
+                            logger.info(
+                                "Streaming: %s -> %s", stream.NAME, updates[-1][0]
+                            )
+                            stream_updates_counter.inc_by(len(updates), stream.NAME)
+
+                        # Some streams return multiple rows with the same stream IDs,
+                        # we need to make sure they get sent out in batches. We do
+                        # this by setting the current token to all but the last of
+                        # a series of updates with the same token to have a None
+                        # token. See RdataCommand for more details.
+                        batched_updates = _batch_updates(updates)
+
+                        for conn in self.connections:
+                            for token, row in batched_updates:
+                                try:
+                                    conn.stream_update(stream.NAME, token, row)
+                                except Exception:
+                                    logger.exception("Failed to replicate")
+
+            logger.debug("No more pending updates, breaking poke loop")
+        finally:
+            self.pending_updates = False
+            self.is_looping = False
+
+    @measure_func("repl.get_stream_updates")
+    def get_stream_updates(self, stream_name, token):
+        """For a given stream get all updates since token. This is called when
+        a client first subscribes to a stream.
+        """
+        stream = self.streams_by_name.get(stream_name, None)
+        if not stream:
+            raise Exception("unknown stream %s", stream_name)
+
+        return stream.get_updates_since(token)
+
+    @measure_func("repl.federation_ack")
+    def federation_ack(self, token):
+        """We've received an ack for federation stream from a client.
+        """
+        federation_ack_counter.inc()
+        if self.federation_sender:
+            self.federation_sender.federation_ack(token)
+
+    @measure_func("repl.on_user_sync")
+    def on_user_sync(self, conn_id, user_id, is_syncing, last_sync_ms):
+        """A client has started/stopped syncing on a worker.
+        """
+        user_sync_counter.inc()
+        self.presence_handler.update_external_syncs_row(
+            conn_id, user_id, is_syncing, last_sync_ms,
+        )
+
+    @measure_func("repl.on_remove_pusher")
+    @defer.inlineCallbacks
+    def on_remove_pusher(self, app_id, push_key, user_id):
+        """A client has asked us to remove a pusher
+        """
+        remove_pusher_counter.inc()
+        yield self.store.delete_pusher_by_app_id_pushkey_user_id(
+            app_id=app_id, pushkey=push_key, user_id=user_id
+        )
+
+        self.notifier.on_new_replication_data()
+
+    @measure_func("repl.on_invalidate_cache")
+    def on_invalidate_cache(self, cache_func, keys):
+        """The client has asked us to invalidate a cache
+        """
+        invalidate_cache_counter.inc()
+        getattr(self.store, cache_func).invalidate(tuple(keys))
+
+    def send_sync_to_all_connections(self, data):
+        """Sends a SYNC command to all clients.
+
+        Used in tests.
+        """
+        for conn in self.connections:
+            conn.send_sync(data)
+
+    def new_connection(self, connection):
+        """A new client connection has been established
+        """
+        self.connections.append(connection)
+
+    def lost_connection(self, connection):
+        """A client connection has been lost
+        """
+        try:
+            self.connections.remove(connection)
+        except ValueError:
+            pass
+
+        # We need to tell the presence handler that the connection has been
+        # lost so that it can handle any ongoing syncs on that connection.
+        self.presence_handler.update_external_syncs_clear(connection.conn_id)
+
+
+def _batch_updates(updates):
+    """Takes a list of updates of form [(token, row)] and sets the token to
+    None for all rows where the next row has the same token. This is used to
+    implement batching.
+
+    For example:
+
+        [(1, _), (1, _), (2, _), (3, _), (3, _)]
+
+    becomes:
+
+        [(None, _), (1, _), (2, _), (None, _), (3, _)]
+    """
+    if not updates:
+        return []
+
+    new_updates = []
+    for i, update in enumerate(updates[:-1]):
+        if update[0] == updates[i + 1][0]:
+            new_updates.append((None, update[1]))
+        else:
+            new_updates.append(update)
+
+    new_updates.append(updates[-1])
+    return new_updates
diff --git a/synapse/replication/tcp/streams.py b/synapse/replication/tcp/streams.py
new file mode 100644
index 0000000000..369d5f2428
--- /dev/null
+++ b/synapse/replication/tcp/streams.py
@@ -0,0 +1,464 @@
+# -*- coding: utf-8 -*-
+# Copyright 2017 Vector Creations Ltd
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Defines all the valid streams that clients can subscribe to, and the format
+of the rows returned by each stream.
+
+Each stream is defined by the following information:
+
+    stream name:        The name of the stream
+    row type:           The type that is used to serialise/deserialse the row
+    current_token:      The function that returns the current token for the stream
+    update_function:    The function that returns a list of updates between two tokens
+"""
+
+from twisted.internet import defer
+from collections import namedtuple
+
+import logging
+
+
+logger = logging.getLogger(__name__)
+
+
+MAX_EVENTS_BEHIND = 10000
+
+
+EventStreamRow = namedtuple("EventStreamRow", (
+    "event_id",  # str
+    "room_id",  # str
+    "type",  # str
+    "state_key",  # str, optional
+    "redacts",  # str, optional
+))
+BackfillStreamRow = namedtuple("BackfillStreamRow", (
+    "event_id",  # str
+    "room_id",  # str
+    "type",  # str
+    "state_key",  # str, optional
+    "redacts",  # str, optional
+))
+PresenceStreamRow = namedtuple("PresenceStreamRow", (
+    "user_id",  # str
+    "state",  # str
+    "last_active_ts",  # int
+    "last_federation_update_ts",  # int
+    "last_user_sync_ts",  # int
+    "status_msg",   # str
+    "currently_active",  # bool
+))
+TypingStreamRow = namedtuple("TypingStreamRow", (
+    "room_id",  # str
+    "user_ids",  # list(str)
+))
+ReceiptsStreamRow = namedtuple("ReceiptsStreamRow", (
+    "room_id",  # str
+    "receipt_type",  # str
+    "user_id",  # str
+    "event_id",  # str
+    "data",  # dict
+))
+PushRulesStreamRow = namedtuple("PushRulesStreamRow", (
+    "user_id",  # str
+))
+PushersStreamRow = namedtuple("PushersStreamRow", (
+    "user_id",  # str
+    "app_id",  # str
+    "pushkey",  # str
+    "deleted",  # bool
+))
+CachesStreamRow = namedtuple("CachesStreamRow", (
+    "cache_func",  # str
+    "keys",  # list(str)
+    "invalidation_ts",  # int
+))
+PublicRoomsStreamRow = namedtuple("PublicRoomsStreamRow", (
+    "room_id",  # str
+    "visibility",  # str
+    "appservice_id",  # str, optional
+    "network_id",  # str, optional
+))
+DeviceListsStreamRow = namedtuple("DeviceListsStreamRow", (
+    "user_id",  # str
+    "destination",  # str
+))
+ToDeviceStreamRow = namedtuple("ToDeviceStreamRow", (
+    "entity",  # str
+))
+FederationStreamRow = namedtuple("FederationStreamRow", (
+    "type",  # str, the type of data as defined in the BaseFederationRows
+    "data",  # dict, serialization of a federation.send_queue.BaseFederationRow
+))
+TagAccountDataStreamRow = namedtuple("TagAccountDataStreamRow", (
+    "user_id",  # str
+    "room_id",  # str
+    "data",  # dict
+))
+AccountDataStreamRow = namedtuple("AccountDataStream", (
+    "user_id",  # str
+    "room_id",  # str
+    "data_type",  # str
+    "data",  # dict
+))
+
+
+class Stream(object):
+    """Base class for the streams.
+
+    Provides a `get_updates()` function that returns new updates since the last
+    time it was called up until the point `advance_current_token` was called.
+    """
+    NAME = None  # The name of the stream
+    ROW_TYPE = None  # The type of the row
+    _LIMITED = True  # Whether the update function takes a limit
+
+    def __init__(self, hs):
+        # The token from which we last asked for updates
+        self.last_token = self.current_token()
+
+        # The token that we will get updates up to
+        self.upto_token = self.current_token()
+
+    def advance_current_token(self):
+        """Updates `upto_token` to "now", which updates up until which point
+        get_updates[_since] will fetch rows till.
+        """
+        self.upto_token = self.current_token()
+
+    def discard_updates_and_advance(self):
+        """Called when the stream should advance but the updates would be discarded,
+        e.g. when there are no currently connected workers.
+        """
+        self.upto_token = self.current_token()
+        self.last_token = self.upto_token
+
+    @defer.inlineCallbacks
+    def get_updates(self):
+        """Gets all updates since the last time this function was called (or
+        since the stream was constructed if it hadn't been called before),
+        until the `upto_token`
+
+        Returns:
+            (list(ROW_TYPE), int): list of updates plus the token used as an
+                upper bound of the updates (i.e. the "current token")
+        """
+        updates, current_token = yield self.get_updates_since(self.last_token)
+        self.last_token = current_token
+
+        defer.returnValue((updates, current_token))
+
+    @defer.inlineCallbacks
+    def get_updates_since(self, from_token):
+        """Like get_updates except allows specifying from when we should
+        stream updates
+
+        Returns:
+            (list(ROW_TYPE), int): list of updates plus the token used as an
+                upper bound of the updates (i.e. the "current token")
+        """
+        if from_token in ("NOW", "now"):
+            defer.returnValue(([], self.upto_token))
+
+        current_token = self.upto_token
+
+        from_token = int(from_token)
+
+        if from_token == current_token:
+            defer.returnValue(([], current_token))
+
+        if self._LIMITED:
+            rows = yield self.update_function(
+                from_token, current_token,
+                limit=MAX_EVENTS_BEHIND + 1,
+            )
+
+            if len(rows) >= MAX_EVENTS_BEHIND:
+                raise Exception("stream %s has fallen behined" % (self.NAME))
+        else:
+            rows = yield self.update_function(
+                from_token, current_token,
+            )
+
+        updates = [(row[0], self.ROW_TYPE(*row[1:])) for row in rows]
+
+        defer.returnValue((updates, current_token))
+
+    def current_token(self):
+        """Gets the current token of the underlying streams. Should be provided
+        by the sub classes
+
+        Returns:
+            int
+        """
+        raise NotImplementedError()
+
+    def update_function(self, from_token, current_token, limit=None):
+        """Get updates between from_token and to_token. If Stream._LIMITED is
+        True then limit is provided, otherwise it's not.
+
+        Returns:
+            Deferred(list(tuple)): the first entry in the tuple is the token for
+                that update, and the rest of the tuple gets used to construct
+                a ``ROW_TYPE`` instance
+        """
+        raise NotImplementedError()
+
+
+class EventsStream(Stream):
+    """We received a new event, or an event went from being an outlier to not
+    """
+    NAME = "events"
+    ROW_TYPE = EventStreamRow
+
+    def __init__(self, hs):
+        store = hs.get_datastore()
+        self.current_token = store.get_current_events_token
+        self.update_function = store.get_all_new_forward_event_rows
+
+        super(EventsStream, self).__init__(hs)
+
+
+class BackfillStream(Stream):
+    """We fetched some old events and either we had never seen that event before
+    or it went from being an outlier to not.
+    """
+    NAME = "backfill"
+    ROW_TYPE = BackfillStreamRow
+
+    def __init__(self, hs):
+        store = hs.get_datastore()
+        self.current_token = store.get_current_backfill_token
+        self.update_function = store.get_all_new_backfill_event_rows
+
+        super(BackfillStream, self).__init__(hs)
+
+
+class PresenceStream(Stream):
+    NAME = "presence"
+    _LIMITED = False
+    ROW_TYPE = PresenceStreamRow
+
+    def __init__(self, hs):
+        store = hs.get_datastore()
+        presence_handler = hs.get_presence_handler()
+
+        self.current_token = store.get_current_presence_token
+        self.update_function = presence_handler.get_all_presence_updates
+
+        super(PresenceStream, self).__init__(hs)
+
+
+class TypingStream(Stream):
+    NAME = "typing"
+    _LIMITED = False
+    ROW_TYPE = TypingStreamRow
+
+    def __init__(self, hs):
+        typing_handler = hs.get_typing_handler()
+
+        self.current_token = typing_handler.get_current_token
+        self.update_function = typing_handler.get_all_typing_updates
+
+        super(TypingStream, self).__init__(hs)
+
+
+class ReceiptsStream(Stream):
+    NAME = "receipts"
+    ROW_TYPE = ReceiptsStreamRow
+
+    def __init__(self, hs):
+        store = hs.get_datastore()
+
+        self.current_token = store.get_max_receipt_stream_id
+        self.update_function = store.get_all_updated_receipts
+
+        super(ReceiptsStream, self).__init__(hs)
+
+
+class PushRulesStream(Stream):
+    """A user has changed their push rules
+    """
+    NAME = "push_rules"
+    ROW_TYPE = PushRulesStreamRow
+
+    def __init__(self, hs):
+        self.store = hs.get_datastore()
+        super(PushRulesStream, self).__init__(hs)
+
+    def current_token(self):
+        push_rules_token, _ = self.store.get_push_rules_stream_token()
+        return push_rules_token
+
+    @defer.inlineCallbacks
+    def update_function(self, from_token, to_token, limit):
+        rows = yield self.store.get_all_push_rule_updates(from_token, to_token, limit)
+        defer.returnValue([(row[0], row[2]) for row in rows])
+
+
+class PushersStream(Stream):
+    """A user has added/changed/removed a pusher
+    """
+    NAME = "pushers"
+    ROW_TYPE = PushersStreamRow
+
+    def __init__(self, hs):
+        store = hs.get_datastore()
+
+        self.current_token = store.get_pushers_stream_token
+        self.update_function = store.get_all_updated_pushers_rows
+
+        super(PushersStream, self).__init__(hs)
+
+
+class CachesStream(Stream):
+    """A cache was invalidated on the master and no other stream would invalidate
+    the cache on the workers
+    """
+    NAME = "caches"
+    ROW_TYPE = CachesStreamRow
+
+    def __init__(self, hs):
+        store = hs.get_datastore()
+
+        self.current_token = store.get_cache_stream_token
+        self.update_function = store.get_all_updated_caches
+
+        super(CachesStream, self).__init__(hs)
+
+
+class PublicRoomsStream(Stream):
+    """The public rooms list changed
+    """
+    NAME = "public_rooms"
+    ROW_TYPE = PublicRoomsStreamRow
+
+    def __init__(self, hs):
+        store = hs.get_datastore()
+
+        self.current_token = store.get_current_public_room_stream_id
+        self.update_function = store.get_all_new_public_rooms
+
+        super(PublicRoomsStream, self).__init__(hs)
+
+
+class DeviceListsStream(Stream):
+    """Someone added/changed/removed a device
+    """
+    NAME = "device_lists"
+    _LIMITED = False
+    ROW_TYPE = DeviceListsStreamRow
+
+    def __init__(self, hs):
+        store = hs.get_datastore()
+
+        self.current_token = store.get_device_stream_token
+        self.update_function = store.get_all_device_list_changes_for_remotes
+
+        super(DeviceListsStream, self).__init__(hs)
+
+
+class ToDeviceStream(Stream):
+    """New to_device messages for a client
+    """
+    NAME = "to_device"
+    ROW_TYPE = ToDeviceStreamRow
+
+    def __init__(self, hs):
+        store = hs.get_datastore()
+
+        self.current_token = store.get_to_device_stream_token
+        self.update_function = store.get_all_new_device_messages
+
+        super(ToDeviceStream, self).__init__(hs)
+
+
+class FederationStream(Stream):
+    """Data to be sent over federation. Only available when master has federation
+    sending disabled.
+    """
+    NAME = "federation"
+    ROW_TYPE = FederationStreamRow
+
+    def __init__(self, hs):
+        federation_sender = hs.get_federation_sender()
+
+        self.current_token = federation_sender.get_current_token
+        self.update_function = federation_sender.get_replication_rows
+
+        super(FederationStream, self).__init__(hs)
+
+
+class TagAccountDataStream(Stream):
+    """Someone added/removed a tag for a room
+    """
+    NAME = "tag_account_data"
+    ROW_TYPE = TagAccountDataStreamRow
+
+    def __init__(self, hs):
+        store = hs.get_datastore()
+
+        self.current_token = store.get_max_account_data_stream_id
+        self.update_function = store.get_all_updated_tags
+
+        super(TagAccountDataStream, self).__init__(hs)
+
+
+class AccountDataStream(Stream):
+    """Global or per room account data was changed
+    """
+    NAME = "account_data"
+    ROW_TYPE = AccountDataStreamRow
+
+    def __init__(self, hs):
+        self.store = hs.get_datastore()
+
+        self.current_token = self.store.get_max_account_data_stream_id
+
+        super(AccountDataStream, self).__init__(hs)
+
+    @defer.inlineCallbacks
+    def update_function(self, from_token, to_token, limit):
+        global_results, room_results = yield self.store.get_all_updated_account_data(
+            from_token, from_token, to_token, limit
+        )
+
+        results = list(room_results)
+        results.extend(
+            (stream_id, user_id, None, account_data_type, content,)
+            for stream_id, user_id, account_data_type, content in global_results
+        )
+
+        defer.returnValue(results)
+
+
+STREAMS_MAP = {
+    stream.NAME: stream
+    for stream in (
+        EventsStream,
+        BackfillStream,
+        PresenceStream,
+        TypingStream,
+        ReceiptsStream,
+        PushRulesStream,
+        PushersStream,
+        CachesStream,
+        PublicRoomsStream,
+        DeviceListsStream,
+        ToDeviceStream,
+        FederationStream,
+        TagAccountDataStream,
+        AccountDataStream,
+    )
+}
diff --git a/synapse/rest/client/v1/voip.py b/synapse/rest/client/v1/voip.py
index 03141c623c..c43b30b73a 100644
--- a/synapse/rest/client/v1/voip.py
+++ b/synapse/rest/client/v1/voip.py
@@ -28,7 +28,10 @@ class VoipRestServlet(ClientV1RestServlet):
 
     @defer.inlineCallbacks
     def on_GET(self, request):
-        requester = yield self.auth.get_user_by_req(request)
+        requester = yield self.auth.get_user_by_req(
+            request,
+            self.hs.config.turn_allow_guests
+        )
 
         turnUris = self.hs.config.turn_uris
         turnSecret = self.hs.config.turn_shared_secret
diff --git a/synapse/rest/client/v2_alpha/thirdparty.py b/synapse/rest/client/v2_alpha/thirdparty.py
index 31f94bc6e9..6fceb23e26 100644
--- a/synapse/rest/client/v2_alpha/thirdparty.py
+++ b/synapse/rest/client/v2_alpha/thirdparty.py
@@ -36,7 +36,7 @@ class ThirdPartyProtocolsServlet(RestServlet):
 
     @defer.inlineCallbacks
     def on_GET(self, request):
-        yield self.auth.get_user_by_req(request)
+        yield self.auth.get_user_by_req(request, allow_guest=True)
 
         protocols = yield self.appservice_handler.get_3pe_protocols()
         defer.returnValue((200, protocols))
@@ -54,7 +54,7 @@ class ThirdPartyProtocolServlet(RestServlet):
 
     @defer.inlineCallbacks
     def on_GET(self, request, protocol):
-        yield self.auth.get_user_by_req(request)
+        yield self.auth.get_user_by_req(request, allow_guest=True)
 
         protocols = yield self.appservice_handler.get_3pe_protocols(
             only_protocol=protocol,
@@ -77,7 +77,7 @@ class ThirdPartyUserServlet(RestServlet):
 
     @defer.inlineCallbacks
     def on_GET(self, request, protocol):
-        yield self.auth.get_user_by_req(request)
+        yield self.auth.get_user_by_req(request, allow_guest=True)
 
         fields = request.args
         fields.pop("access_token", None)
@@ -101,7 +101,7 @@ class ThirdPartyLocationServlet(RestServlet):
 
     @defer.inlineCallbacks
     def on_GET(self, request, protocol):
-        yield self.auth.get_user_by_req(request)
+        yield self.auth.get_user_by_req(request, allow_guest=True)
 
         fields = request.args
         fields.pop("access_token", None)
diff --git a/synapse/server.py b/synapse/server.py
index c577032041..6310152560 100644
--- a/synapse/server.py
+++ b/synapse/server.py
@@ -132,6 +132,7 @@ class HomeServer(object):
         'federation_sender',
         'receipts_handler',
         'macaroon_generator',
+        'tcp_replication',
     ]
 
     def __init__(self, hostname, **kwargs):
@@ -290,6 +291,9 @@ class HomeServer(object):
     def build_receipts_handler(self):
         return ReceiptsHandler(self)
 
+    def build_tcp_replication(self):
+        raise NotImplementedError()
+
     def remove_pusher(self, app_id, push_key, user_id):
         return self.get_pusherpool().remove_pusher(app_id, push_key, user_id)
 
diff --git a/synapse/storage/deviceinbox.py b/synapse/storage/deviceinbox.py
index 2714519d21..0b62b493d5 100644
--- a/synapse/storage/deviceinbox.py
+++ b/synapse/storage/deviceinbox.py
@@ -325,23 +325,26 @@ class DeviceInboxStore(BackgroundUpdateStore):
             # we return.
             upper_pos = min(current_pos, last_pos + limit)
             sql = (
-                "SELECT stream_id, user_id"
+                "SELECT max(stream_id), user_id"
                 " FROM device_inbox"
                 " WHERE ? < stream_id AND stream_id <= ?"
-                " ORDER BY stream_id ASC"
+                " GROUP BY user_id"
             )
             txn.execute(sql, (last_pos, upper_pos))
             rows = txn.fetchall()
 
             sql = (
-                "SELECT stream_id, destination"
+                "SELECT max(stream_id), destination"
                 " FROM device_federation_outbox"
                 " WHERE ? < stream_id AND stream_id <= ?"
-                " ORDER BY stream_id ASC"
+                " GROUP BY destination"
             )
             txn.execute(sql, (last_pos, upper_pos))
             rows.extend(txn)
 
+            # Order by ascending stream ordering
+            rows.sort()
+
             return rows
 
         return self.runInteraction(
diff --git a/synapse/storage/devices.py b/synapse/storage/devices.py
index 53e36791d5..c8d5f5ba8b 100644
--- a/synapse/storage/devices.py
+++ b/synapse/storage/devices.py
@@ -533,7 +533,7 @@ class DeviceStore(SQLBaseStore):
         rows = yield self._execute("get_user_whose_devices_changed", None, sql, from_key)
         defer.returnValue(set(row[0] for row in rows))
 
-    def get_all_device_list_changes_for_remotes(self, from_key):
+    def get_all_device_list_changes_for_remotes(self, from_key, to_key):
         """Return a list of `(stream_id, user_id, destination)` which is the
         combined list of changes to devices, and which destinations need to be
         poked. `destination` may be None if no destinations need to be poked.
@@ -541,11 +541,11 @@ class DeviceStore(SQLBaseStore):
         sql = """
             SELECT stream_id, user_id, destination FROM device_lists_stream
             LEFT JOIN device_lists_outbound_pokes USING (stream_id, user_id, device_id)
-            WHERE stream_id > ?
+            WHERE ? < stream_id AND stream_id <= ?
         """
         return self._execute(
             "get_all_device_list_changes_for_remotes", None,
-            sql, from_key,
+            sql, from_key, to_key
         )
 
     @defer.inlineCallbacks
diff --git a/synapse/storage/events.py b/synapse/storage/events.py
index 3f6833fad2..64fe937bdc 100644
--- a/synapse/storage/events.py
+++ b/synapse/storage/events.py
@@ -1771,6 +1771,94 @@ class EventsStore(SQLBaseStore):
         """The current minimum token that backfilled events have reached"""
         return -self._backfill_id_gen.get_current_token()
 
+    def get_current_events_token(self):
+        """The current maximum token that events have reached"""
+        return self._stream_id_gen.get_current_token()
+
+    def get_all_new_forward_event_rows(self, last_id, current_id, limit):
+        if last_id == current_id:
+            return defer.succeed([])
+
+        def get_all_new_forward_event_rows(txn):
+            sql = (
+                "SELECT e.stream_ordering, e.event_id, e.room_id, e.type,"
+                " state_key, redacts"
+                " FROM events AS e"
+                " LEFT JOIN redactions USING (event_id)"
+                " LEFT JOIN state_events USING (event_id)"
+                " WHERE ? < stream_ordering AND stream_ordering <= ?"
+                " ORDER BY stream_ordering ASC"
+                " LIMIT ?"
+            )
+            txn.execute(sql, (last_id, current_id, limit))
+            new_event_updates = txn.fetchall()
+
+            if len(new_event_updates) == limit:
+                upper_bound = new_event_updates[-1][0]
+            else:
+                upper_bound = current_id
+
+            sql = (
+                "SELECT event_stream_ordering, e.event_id, e.room_id, e.type,"
+                " state_key, redacts"
+                " FROM events AS e"
+                " INNER JOIN ex_outlier_stream USING (event_id)"
+                " LEFT JOIN redactions USING (event_id)"
+                " LEFT JOIN state_events USING (event_id)"
+                " WHERE ? < event_stream_ordering"
+                " AND event_stream_ordering <= ?"
+                " ORDER BY event_stream_ordering DESC"
+            )
+            txn.execute(sql, (last_id, upper_bound))
+            new_event_updates.extend(txn)
+
+            return new_event_updates
+        return self.runInteraction(
+            "get_all_new_forward_event_rows", get_all_new_forward_event_rows
+        )
+
+    def get_all_new_backfill_event_rows(self, last_id, current_id, limit):
+        if last_id == current_id:
+            return defer.succeed([])
+
+        def get_all_new_backfill_event_rows(txn):
+            sql = (
+                "SELECT -e.stream_ordering, e.event_id, e.room_id, e.type,"
+                " state_key, redacts"
+                " FROM events AS e"
+                " LEFT JOIN redactions USING (event_id)"
+                " LEFT JOIN state_events USING (event_id)"
+                " WHERE ? > stream_ordering AND stream_ordering >= ?"
+                " ORDER BY stream_ordering ASC"
+                " LIMIT ?"
+            )
+            txn.execute(sql, (-last_id, -current_id, limit))
+            new_event_updates = txn.fetchall()
+
+            if len(new_event_updates) == limit:
+                upper_bound = new_event_updates[-1][0]
+            else:
+                upper_bound = current_id
+
+            sql = (
+                "SELECT -event_stream_ordering, e.event_id, e.room_id, e.type,"
+                " state_key, redacts"
+                " FROM events AS e"
+                " INNER JOIN ex_outlier_stream USING (event_id)"
+                " LEFT JOIN redactions USING (event_id)"
+                " LEFT JOIN state_events USING (event_id)"
+                " WHERE ? > event_stream_ordering"
+                " AND event_stream_ordering >= ?"
+                " ORDER BY event_stream_ordering DESC"
+            )
+            txn.execute(sql, (-last_id, -upper_bound))
+            new_event_updates.extend(txn.fetchall())
+
+            return new_event_updates
+        return self.runInteraction(
+            "get_all_new_backfill_event_rows", get_all_new_backfill_event_rows
+        )
+
     @cached(num_args=5, max_entries=10)
     def get_all_new_events(self, last_backfill_id, last_forward_id,
                            current_backfill_id, current_forward_id, limit):
diff --git a/synapse/storage/pusher.py b/synapse/storage/pusher.py
index 8cc9f0353b..34d2f82b7f 100644
--- a/synapse/storage/pusher.py
+++ b/synapse/storage/pusher.py
@@ -135,6 +135,48 @@ class PusherStore(SQLBaseStore):
             "get_all_updated_pushers", get_all_updated_pushers_txn
         )
 
+    def get_all_updated_pushers_rows(self, last_id, current_id, limit):
+        """Get all the pushers that have changed between the given tokens.
+
+        Returns:
+            Deferred(list(tuple)): each tuple consists of:
+                stream_id (str)
+                user_id (str)
+                app_id (str)
+                pushkey (str)
+                was_deleted (bool): whether the pusher was added/updated (False)
+                    or deleted (True)
+        """
+
+        if last_id == current_id:
+            return defer.succeed([])
+
+        def get_all_updated_pushers_rows_txn(txn):
+            sql = (
+                "SELECT id, user_name, app_id, pushkey"
+                " FROM pushers"
+                " WHERE ? < id AND id <= ?"
+                " ORDER BY id ASC LIMIT ?"
+            )
+            txn.execute(sql, (last_id, current_id, limit))
+            results = [list(row) + [False] for row in txn]
+
+            sql = (
+                "SELECT stream_id, user_id, app_id, pushkey"
+                " FROM deleted_pushers"
+                " WHERE ? < stream_id AND stream_id <= ?"
+                " ORDER BY stream_id ASC LIMIT ?"
+            )
+            txn.execute(sql, (last_id, current_id, limit))
+
+            results.extend(list(row) + [True] for row in txn)
+            results.sort()  # Sort so that they're ordered by stream id
+
+            return results
+        return self.runInteraction(
+            "get_all_updated_pushers_rows", get_all_updated_pushers_rows_txn
+        )
+
     @cachedInlineCallbacks(num_args=1, max_entries=15000)
     def get_if_user_has_pusher(self, user_id):
         # This only exists for the cachedList decorator
diff --git a/synapse/storage/state.py b/synapse/storage/state.py
index fb23f6f462..acd69944c4 100644
--- a/synapse/storage/state.py
+++ b/synapse/storage/state.py
@@ -14,7 +14,7 @@
 # limitations under the License.
 
 from ._base import SQLBaseStore
-from synapse.util.caches.descriptors import cached, cachedList, cachedInlineCallbacks
+from synapse.util.caches.descriptors import cached, cachedList
 from synapse.util.caches import intern_string
 from synapse.storage.engines import PostgresEngine
 
@@ -69,17 +69,33 @@ class StateStore(SQLBaseStore):
             where_clause="type='m.room.member'",
         )
 
-    @cachedInlineCallbacks(max_entries=100000, iterable=True)
+    @cached(max_entries=100000, iterable=True)
     def get_current_state_ids(self, room_id):
-        rows = yield self._simple_select_list(
-            table="current_state_events",
-            keyvalues={"room_id": room_id},
-            retcols=["event_id", "type", "state_key"],
-            desc="_calculate_state_delta",
+        """Get the current state event ids for a room based on the
+        current_state_events table.
+
+        Args:
+            room_id (str)
+
+        Returns:
+            deferred: dict of (type, state_key) -> event_id
+        """
+        def _get_current_state_ids_txn(txn):
+            txn.execute(
+                """SELECT type, state_key, event_id FROM current_state_events
+                WHERE room_id = ?
+                """,
+                (room_id,)
+            )
+
+            return {
+                (r[0], r[1]): r[2] for r in txn
+            }
+
+        return self.runInteraction(
+            "get_current_state_ids",
+            _get_current_state_ids_txn,
         )
-        defer.returnValue({
-            (r["type"], r["state_key"]): r["event_id"] for r in rows
-        })
 
     @defer.inlineCallbacks
     def get_state_groups_ids(self, room_id, event_ids):
diff --git a/synapse/types.py b/synapse/types.py
index 9666f9d73f..c87ed813b9 100644
--- a/synapse/types.py
+++ b/synapse/types.py
@@ -216,9 +216,7 @@ class StreamToken(
             return self
 
     def copy_and_replace(self, key, new_value):
-        d = self._asdict()
-        d[key] = new_value
-        return StreamToken(**d)
+        return self._replace(**{key: new_value})
 
 
 StreamToken.START = StreamToken(
diff --git a/synapse/util/async.py b/synapse/util/async.py
index 35380bf8ed..1453faf0ef 100644
--- a/synapse/util/async.py
+++ b/synapse/util/async.py
@@ -89,6 +89,11 @@ class ObservableDeferred(object):
         deferred.addCallbacks(callback, errback)
 
     def observe(self):
+        """Observe the underlying deferred.
+
+        Can return either a deferred if the underlying deferred is still pending
+        (or has failed), or the actual value. Callers may need to use maybeDeferred.
+        """
         if not self._result:
             d = defer.Deferred()
 
@@ -101,7 +106,7 @@ class ObservableDeferred(object):
             return d
         else:
             success, res = self._result
-            return defer.succeed(res) if success else defer.fail(res)
+            return res if success else defer.fail(res)
 
     def observers(self):
         return self._observers
diff --git a/synapse/util/caches/descriptors.py b/synapse/util/caches/descriptors.py
index 5c30ed235d..9d0d0be1f9 100644
--- a/synapse/util/caches/descriptors.py
+++ b/synapse/util/caches/descriptors.py
@@ -224,8 +224,20 @@ class _CacheDescriptorBase(object):
             )
 
         self.num_args = num_args
+
+        # list of the names of the args used as the cache key
         self.arg_names = all_args[1:num_args + 1]
 
+        # self.arg_defaults is a map of arg name to its default value for each
+        # argument that has a default value
+        if arg_spec.defaults:
+            self.arg_defaults = dict(zip(
+                all_args[-len(arg_spec.defaults):],
+                arg_spec.defaults
+            ))
+        else:
+            self.arg_defaults = {}
+
         if "cache_context" in self.arg_names:
             raise Exception(
                 "cache_context arg cannot be included among the cache keys"
@@ -289,18 +301,31 @@ class CacheDescriptor(_CacheDescriptorBase):
             iterable=self.iterable,
         )
 
+        def get_cache_key(args, kwargs):
+            """Given some args/kwargs return a generator that resolves into
+            the cache_key.
+
+            We loop through each arg name, looking up if its in the `kwargs`,
+            otherwise using the next argument in `args`. If there are no more
+            args then we try looking the arg name up in the defaults
+            """
+            pos = 0
+            for nm in self.arg_names:
+                if nm in kwargs:
+                    yield kwargs[nm]
+                elif pos < len(args):
+                    yield args[pos]
+                    pos += 1
+                else:
+                    yield self.arg_defaults[nm]
+
         @functools.wraps(self.orig)
         def wrapped(*args, **kwargs):
             # If we're passed a cache_context then we'll want to call its invalidate()
             # whenever we are invalidated
             invalidate_callback = kwargs.pop("on_invalidate", None)
 
-            # Add temp cache_context so inspect.getcallargs doesn't explode
-            if self.add_cache_context:
-                kwargs["cache_context"] = None
-
-            arg_dict = inspect.getcallargs(self.orig, obj, *args, **kwargs)
-            cache_key = tuple(arg_dict[arg_nm] for arg_nm in self.arg_names)
+            cache_key = tuple(get_cache_key(args, kwargs))
 
             # Add our own `cache_context` to argument list if the wrapped function
             # has asked for one
@@ -341,7 +366,10 @@ class CacheDescriptor(_CacheDescriptorBase):
                 cache.set(cache_key, result_d, callback=invalidate_callback)
                 observer = result_d.observe()
 
-            return logcontext.make_deferred_yieldable(observer)
+            if isinstance(observer, defer.Deferred):
+                return logcontext.make_deferred_yieldable(observer)
+            else:
+                return observer
 
         wrapped.invalidate = cache.invalidate
         wrapped.invalidate_all = cache.invalidate_all
diff --git a/synapse/util/logcontext.py b/synapse/util/logcontext.py
index 857afee7cb..990216145e 100644
--- a/synapse/util/logcontext.py
+++ b/synapse/util/logcontext.py
@@ -334,12 +334,8 @@ def preserve_fn(f):
         LoggingContext.set_current_context(LoggingContext.sentinel)
         return result
 
-    # XXX: why is this here rather than inside g? surely we want to preserve
-    # the context from the time the function was called, not when it was
-    # wrapped?
-    current = LoggingContext.current_context()
-
     def g(*args, **kwargs):
+        current = LoggingContext.current_context()
         res = f(*args, **kwargs)
         if isinstance(res, defer.Deferred) and not res.called:
             # The function will have reset the context before returning, so
diff --git a/synapse/visibility.py b/synapse/visibility.py
index 31659156ae..c4dd9ae2c7 100644
--- a/synapse/visibility.py
+++ b/synapse/visibility.py
@@ -56,7 +56,8 @@ def filter_events_for_clients(store, user_tuples, events, event_id_to_state):
         events ([synapse.events.EventBase]): list of events to filter
     """
     forgotten = yield preserve_context_over_deferred(defer.gatherResults([
-        preserve_fn(store.who_forgot_in_room)(
+        defer.maybeDeferred(
+            preserve_fn(store.who_forgot_in_room),
             room_id,
         )
         for room_id in frozenset(e.room_id for e in events)
diff --git a/tests/appservice/test_appservice.py b/tests/appservice/test_appservice.py
index aa8cc50550..7586ea9053 100644
--- a/tests/appservice/test_appservice.py
+++ b/tests/appservice/test_appservice.py
@@ -19,10 +19,12 @@ from twisted.internet import defer
 from mock import Mock
 from tests import unittest
 
+import re
+
 
 def _regex(regex, exclusive=True):
     return {
-        "regex": regex,
+        "regex": re.compile(regex),
         "exclusive": exclusive
     }
 
diff --git a/tests/replication/slave/storage/_base.py b/tests/replication/slave/storage/_base.py
index b82868054d..81063f19a1 100644
--- a/tests/replication/slave/storage/_base.py
+++ b/tests/replication/slave/storage/_base.py
@@ -12,12 +12,15 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from twisted.internet import defer
+from twisted.internet import defer, reactor
 from tests import unittest
 
 from mock import Mock, NonCallableMock
 from tests.utils import setup_test_homeserver
-from synapse.replication.resource import ReplicationResource
+from synapse.replication.tcp.resource import ReplicationStreamProtocolFactory
+from synapse.replication.tcp.client import (
+    ReplicationClientHandler, ReplicationClientFactory,
+)
 
 
 class BaseSlavedStoreTestCase(unittest.TestCase):
@@ -33,18 +36,29 @@ class BaseSlavedStoreTestCase(unittest.TestCase):
         )
         self.hs.get_ratelimiter().send_message.return_value = (True, 0)
 
-        self.replication = ReplicationResource(self.hs)
-
         self.master_store = self.hs.get_datastore()
         self.slaved_store = self.STORE_TYPE(self.hs.get_db_conn(), self.hs)
         self.event_id = 0
 
+        server_factory = ReplicationStreamProtocolFactory(self.hs)
+        listener = reactor.listenUNIX("\0xxx", server_factory)
+        self.addCleanup(listener.stopListening)
+        self.streamer = server_factory.streamer
+
+        self.replication_handler = ReplicationClientHandler(self.slaved_store)
+        client_factory = ReplicationClientFactory(
+            self.hs, "client_name", self.replication_handler
+        )
+        client_connector = reactor.connectUNIX("\0xxx", client_factory)
+        self.addCleanup(client_factory.stopTrying)
+        self.addCleanup(client_connector.disconnect)
+
     @defer.inlineCallbacks
     def replicate(self):
-        streams = self.slaved_store.stream_positions()
-        writer = yield self.replication.replicate(streams, 100)
-        result = writer.finish()
-        yield self.slaved_store.process_replication(result)
+        yield self.streamer.on_notifier_poke()
+        d = self.replication_handler.await_sync("replication_test")
+        self.streamer.send_sync_to_all_connections("replication_test")
+        yield d
 
     @defer.inlineCallbacks
     def check(self, method, args, expected_result=None):
diff --git a/tests/storage/test__base.py b/tests/storage/test__base.py
index 8361dd8cee..281eb16254 100644
--- a/tests/storage/test__base.py
+++ b/tests/storage/test__base.py
@@ -199,7 +199,7 @@ class CacheDecoratorTestCase(unittest.TestCase):
 
         a.func.prefill(("foo",), ObservableDeferred(d))
 
-        self.assertEquals(a.func("foo").result, d.result)
+        self.assertEquals(a.func("foo"), d.result)
         self.assertEquals(callcount[0], 0)
 
     @defer.inlineCallbacks
diff --git a/tests/util/caches/test_descriptors.py b/tests/util/caches/test_descriptors.py
index 4414e86771..3f14ab503f 100644
--- a/tests/util/caches/test_descriptors.py
+++ b/tests/util/caches/test_descriptors.py
@@ -175,3 +175,41 @@ class DescriptorTestCase(unittest.TestCase):
                          logcontext.LoggingContext.sentinel)
 
         return d1
+
+    @defer.inlineCallbacks
+    def test_cache_default_args(self):
+        class Cls(object):
+            def __init__(self):
+                self.mock = mock.Mock()
+
+            @descriptors.cached()
+            def fn(self, arg1, arg2=2, arg3=3):
+                return self.mock(arg1, arg2, arg3)
+
+        obj = Cls()
+
+        obj.mock.return_value = 'fish'
+        r = yield obj.fn(1, 2, 3)
+        self.assertEqual(r, 'fish')
+        obj.mock.assert_called_once_with(1, 2, 3)
+        obj.mock.reset_mock()
+
+        # a call with same params shouldn't call the mock again
+        r = yield obj.fn(1, 2)
+        self.assertEqual(r, 'fish')
+        obj.mock.assert_not_called()
+        obj.mock.reset_mock()
+
+        # a call with different params should call the mock again
+        obj.mock.return_value = 'chips'
+        r = yield obj.fn(2, 3)
+        self.assertEqual(r, 'chips')
+        obj.mock.assert_called_once_with(2, 3, 3)
+        obj.mock.reset_mock()
+
+        # the two values should now be cached
+        r = yield obj.fn(1, 2)
+        self.assertEqual(r, 'fish')
+        r = yield obj.fn(2, 3)
+        self.assertEqual(r, 'chips')
+        obj.mock.assert_not_called()
diff --git a/tests/util/test_snapshot_cache.py b/tests/util/test_snapshot_cache.py
index 7e289715ba..d3a8630c2f 100644
--- a/tests/util/test_snapshot_cache.py
+++ b/tests/util/test_snapshot_cache.py
@@ -53,7 +53,9 @@ class SnapshotCacheTestCase(unittest.TestCase):
         # before the cache expires returns a resolved deferred.
         get_result_at_11 = self.cache.get(11, "key")
         self.assertIsNotNone(get_result_at_11)
-        self.assertTrue(get_result_at_11.called)
+        if isinstance(get_result_at_11, Deferred):
+            # The cache may return the actual result rather than a deferred
+            self.assertTrue(get_result_at_11.called)
 
         # Check that getting the key after the deferred has resolved
         # after the cache expires returns None