summary refs log tree commit diff
path: root/synapse
diff options
context:
space:
mode:
Diffstat (limited to 'synapse')
-rw-r--r--synapse/federation/federation_client.py2
-rw-r--r--synapse/handlers/federation.py19
-rw-r--r--synapse/handlers/message.py32
-rw-r--r--synapse/handlers/sync.py21
-rw-r--r--synapse/metrics/__init__.py29
-rw-r--r--synapse/state.py12
-rw-r--r--synapse/storage/_base.py213
-rw-r--r--synapse/storage/directory.py3
-rw-r--r--synapse/storage/event_federation.py3
-rw-r--r--synapse/storage/keys.py3
-rw-r--r--synapse/storage/presence.py3
-rw-r--r--synapse/storage/push_rule.py3
-rw-r--r--synapse/storage/receipts.py3
-rw-r--r--synapse/storage/registration.py3
-rw-r--r--synapse/storage/room.py3
-rw-r--r--synapse/storage/roommember.py3
-rw-r--r--synapse/storage/state.py355
-rw-r--r--synapse/storage/stream.py6
-rw-r--r--synapse/storage/transactions.py3
-rw-r--r--synapse/util/caches/__init__.py27
-rw-r--r--synapse/util/caches/descriptors.py377
-rw-r--r--synapse/util/caches/dictionary_cache.py103
-rw-r--r--synapse/util/caches/expiringcache.py (renamed from synapse/util/expiringcache.py)0
-rw-r--r--synapse/util/caches/lrucache.py (renamed from synapse/util/lrucache.py)0
24 files changed, 882 insertions, 344 deletions
diff --git a/synapse/federation/federation_client.py b/synapse/federation/federation_client.py
index 7736d14fb5..58a6d6a0ed 100644
--- a/synapse/federation/federation_client.py
+++ b/synapse/federation/federation_client.py
@@ -23,7 +23,7 @@ from synapse.api.errors import (
     CodeMessageException, HttpResponseException, SynapseError,
 )
 from synapse.util import unwrapFirstError
-from synapse.util.expiringcache import ExpiringCache
+from synapse.util.caches.expiringcache import ExpiringCache
 from synapse.util.logutils import log_function
 from synapse.events import FrozenEvent
 import synapse.metrics
diff --git a/synapse/handlers/federation.py b/synapse/handlers/federation.py
index f7155fd8d3..2bfd0a40e0 100644
--- a/synapse/handlers/federation.py
+++ b/synapse/handlers/federation.py
@@ -229,12 +229,14 @@ class FederationHandler(BaseHandler):
 
     @defer.inlineCallbacks
     def _filter_events_for_server(self, server_name, room_id, events):
-        states = yield self.store.get_state_for_events(
-            room_id, [e.event_id for e in events],
+        event_to_state = yield self.store.get_state_for_events(
+            room_id, frozenset(e.event_id for e in events),
+            types=(
+                (EventTypes.RoomHistoryVisibility, ""),
+                (EventTypes.Member, None),
+            )
         )
 
-        events_and_states = zip(events, states)
-
         def redact_disallowed(event_and_state):
             event, state = event_and_state
 
@@ -271,9 +273,10 @@ class FederationHandler(BaseHandler):
 
             return event
 
-        res = map(redact_disallowed, events_and_states)
-
-        logger.info("_filter_events_for_server %r", res)
+        res = map(redact_disallowed, [
+            (e, event_to_state[e.event_id])
+            for e in events
+        ])
 
         defer.returnValue(res)
 
@@ -503,7 +506,7 @@ class FederationHandler(BaseHandler):
         event_ids = list(extremities.keys())
 
         states = yield defer.gatherResults([
-            self.state_handler.resolve_state_groups([e])
+            self.state_handler.resolve_state_groups(room_id, [e])
             for e in event_ids
         ])
         states = dict(zip(event_ids, [s[1] for s in states]))
diff --git a/synapse/handlers/message.py b/synapse/handlers/message.py
index 9d6d4f0978..29e81085d1 100644
--- a/synapse/handlers/message.py
+++ b/synapse/handlers/message.py
@@ -137,15 +137,15 @@ class MessageHandler(BaseHandler):
 
     @defer.inlineCallbacks
     def _filter_events_for_client(self, user_id, room_id, events):
-        states = yield self.store.get_state_for_events(
-            room_id, [e.event_id for e in events],
+        event_id_to_state = yield self.store.get_state_for_events(
+            room_id, frozenset(e.event_id for e in events),
+            types=(
+                (EventTypes.RoomHistoryVisibility, ""),
+                (EventTypes.Member, user_id),
+            )
         )
 
-        events_and_states = zip(events, states)
-
-        def allowed(event_and_state):
-            event, state = event_and_state
-
+        def allowed(event, state):
             if event.type == EventTypes.RoomHistoryVisibility:
                 return True
 
@@ -175,10 +175,10 @@ class MessageHandler(BaseHandler):
 
             return True
 
-        events_and_states = filter(allowed, events_and_states)
         defer.returnValue([
-            ev
-            for ev, _ in events_and_states
+            event
+            for event in events
+            if allowed(event, event_id_to_state[event.event_id])
         ])
 
     @defer.inlineCallbacks
@@ -401,10 +401,14 @@ class MessageHandler(BaseHandler):
             except:
                 logger.exception("Failed to get snapshot")
 
-        yield defer.gatherResults(
-            [handle_room(e) for e in room_list],
-            consumeErrors=True
-        ).addErrback(unwrapFirstError)
+        # Only do N rooms at once
+        n = 5
+        d_list = [handle_room(e) for e in room_list]
+        for i in range(0, len(d_list), n):
+            yield defer.gatherResults(
+                d_list[i:i + n],
+                consumeErrors=True
+            ).addErrback(unwrapFirstError)
 
         ret = {
             "rooms": rooms_ret,
diff --git a/synapse/handlers/sync.py b/synapse/handlers/sync.py
index 6cff6230c1..7206ae23d7 100644
--- a/synapse/handlers/sync.py
+++ b/synapse/handlers/sync.py
@@ -294,15 +294,15 @@ class SyncHandler(BaseHandler):
 
     @defer.inlineCallbacks
     def _filter_events_for_client(self, user_id, room_id, events):
-        states = yield self.store.get_state_for_events(
-            room_id, [e.event_id for e in events],
+        event_id_to_state = yield self.store.get_state_for_events(
+            room_id, frozenset(e.event_id for e in events),
+            types=(
+                (EventTypes.RoomHistoryVisibility, ""),
+                (EventTypes.Member, user_id),
+            )
         )
 
-        events_and_states = zip(events, states)
-
-        def allowed(event_and_state):
-            event, state = event_and_state
-
+        def allowed(event, state):
             if event.type == EventTypes.RoomHistoryVisibility:
                 return True
 
@@ -331,10 +331,11 @@ class SyncHandler(BaseHandler):
                 return membership == Membership.INVITE
 
             return True
-        events_and_states = filter(allowed, events_and_states)
+
         defer.returnValue([
-            ev
-            for ev, _ in events_and_states
+            event
+            for event in events
+            if allowed(event, event_id_to_state[event.event_id])
         ])
 
     @defer.inlineCallbacks
diff --git a/synapse/metrics/__init__.py b/synapse/metrics/__init__.py
index 9233ea3da9..0be9772991 100644
--- a/synapse/metrics/__init__.py
+++ b/synapse/metrics/__init__.py
@@ -18,8 +18,12 @@ from __future__ import absolute_import
 
 import logging
 from resource import getrusage, getpagesize, RUSAGE_SELF
+import functools
 import os
 import stat
+import time
+
+from twisted.internet import reactor
 
 from .metric import (
     CounterMetric, CallbackMetric, DistributionMetric, CacheMetric
@@ -144,3 +148,28 @@ def _process_fds():
     return counts
 
 get_metrics_for("process").register_callback("fds", _process_fds, labels=["type"])
+
+reactor_metrics = get_metrics_for("reactor")
+tick_time = reactor_metrics.register_distribution("tick_time")
+pending_calls_metric = reactor_metrics.register_distribution("pending_calls")
+
+
+def runUntilCurrentTimer(func):
+
+    @functools.wraps(func)
+    def f(*args, **kwargs):
+        pending_calls = len(reactor.getDelayedCalls())
+        start = time.time() * 1000
+        ret = func(*args, **kwargs)
+        end = time.time() * 1000
+        tick_time.inc_by(end - start)
+        pending_calls_metric.inc_by(pending_calls)
+        return ret
+
+    return f
+
+
+if hasattr(reactor, "runUntilCurrent"):
+    # runUntilCurrent is called when we have pending calls. It is called once
+    # per iteratation after fd polling.
+    reactor.runUntilCurrent = runUntilCurrentTimer(reactor.runUntilCurrent)
diff --git a/synapse/state.py b/synapse/state.py
index 80da90a72c..1fe4d066bd 100644
--- a/synapse/state.py
+++ b/synapse/state.py
@@ -18,7 +18,7 @@ from twisted.internet import defer
 
 from synapse.util.logutils import log_function
 from synapse.util.async import run_on_reactor
-from synapse.util.expiringcache import ExpiringCache
+from synapse.util.caches.expiringcache import ExpiringCache
 from synapse.api.constants import EventTypes
 from synapse.api.errors import AuthError
 from synapse.api.auth import AuthEventTypes
@@ -96,7 +96,7 @@ class StateHandler(object):
             cache.ts = self.clock.time_msec()
             state = cache.state
         else:
-            res = yield self.resolve_state_groups(event_ids)
+            res = yield self.resolve_state_groups(room_id, event_ids)
             state = res[1]
 
         if event_type:
@@ -155,13 +155,13 @@ class StateHandler(object):
 
         if event.is_state():
             ret = yield self.resolve_state_groups(
-                [e for e, _ in event.prev_events],
+                event.room_id, [e for e, _ in event.prev_events],
                 event_type=event.type,
                 state_key=event.state_key,
             )
         else:
             ret = yield self.resolve_state_groups(
-                [e for e, _ in event.prev_events],
+                event.room_id, [e for e, _ in event.prev_events],
             )
 
         group, curr_state, prev_state = ret
@@ -180,7 +180,7 @@ class StateHandler(object):
 
     @defer.inlineCallbacks
     @log_function
-    def resolve_state_groups(self, event_ids, event_type=None, state_key=""):
+    def resolve_state_groups(self, room_id, event_ids, event_type=None, state_key=""):
         """ Given a list of event_ids this method fetches the state at each
         event, resolves conflicts between them and returns them.
 
@@ -205,7 +205,7 @@ class StateHandler(object):
                 )
 
         state_groups = yield self.store.get_state_groups(
-            event_ids
+            room_id, event_ids
         )
 
         logger.debug(
diff --git a/synapse/storage/_base.py b/synapse/storage/_base.py
index 73eea157a4..1444767a52 100644
--- a/synapse/storage/_base.py
+++ b/synapse/storage/_base.py
@@ -15,25 +15,22 @@
 import logging
 
 from synapse.api.errors import StoreError
-from synapse.util.async import ObservableDeferred
 from synapse.util.logutils import log_function
 from synapse.util.logcontext import preserve_context_over_fn, LoggingContext
-from synapse.util.lrucache import LruCache
+from synapse.util.caches.dictionary_cache import DictionaryCache
+from synapse.util.caches.descriptors import Cache
 import synapse.metrics
 
 from util.id_generators import IdGenerator, StreamIdGenerator
 
 from twisted.internet import defer
 
-from collections import namedtuple, OrderedDict
+from collections import namedtuple
 
-import functools
-import inspect
 import sys
 import time
 import threading
 
-DEBUG_CACHES = False
 
 logger = logging.getLogger(__name__)
 
@@ -49,208 +46,6 @@ sql_scheduling_timer = metrics.register_distribution("schedule_time")
 sql_query_timer = metrics.register_distribution("query_time", labels=["verb"])
 sql_txn_timer = metrics.register_distribution("transaction_time", labels=["desc"])
 
-caches_by_name = {}
-cache_counter = metrics.register_cache(
-    "cache",
-    lambda: {(name,): len(caches_by_name[name]) for name in caches_by_name.keys()},
-    labels=["name"],
-)
-
-
-_CacheSentinel = object()
-
-
-class Cache(object):
-
-    def __init__(self, name, max_entries=1000, keylen=1, lru=True):
-        if lru:
-            self.cache = LruCache(max_size=max_entries)
-            self.max_entries = None
-        else:
-            self.cache = OrderedDict()
-            self.max_entries = max_entries
-
-        self.name = name
-        self.keylen = keylen
-        self.sequence = 0
-        self.thread = None
-        caches_by_name[name] = self.cache
-
-    def check_thread(self):
-        expected_thread = self.thread
-        if expected_thread is None:
-            self.thread = threading.current_thread()
-        else:
-            if expected_thread is not threading.current_thread():
-                raise ValueError(
-                    "Cache objects can only be accessed from the main thread"
-                )
-
-    def get(self, key, default=_CacheSentinel):
-        val = self.cache.get(key, _CacheSentinel)
-        if val is not _CacheSentinel:
-            cache_counter.inc_hits(self.name)
-            return val
-
-        cache_counter.inc_misses(self.name)
-
-        if default is _CacheSentinel:
-            raise KeyError()
-        else:
-            return default
-
-    def update(self, sequence, key, value):
-        self.check_thread()
-        if self.sequence == sequence:
-            # Only update the cache if the caches sequence number matches the
-            # number that the cache had before the SELECT was started (SYN-369)
-            self.prefill(key, value)
-
-    def prefill(self, key, value):
-        if self.max_entries is not None:
-            while len(self.cache) >= self.max_entries:
-                self.cache.popitem(last=False)
-
-        self.cache[key] = value
-
-    def invalidate(self, key):
-        self.check_thread()
-        if not isinstance(key, tuple):
-            raise TypeError(
-                "The cache key must be a tuple not %r" % (type(key),)
-            )
-
-        # Increment the sequence number so that any SELECT statements that
-        # raced with the INSERT don't update the cache (SYN-369)
-        self.sequence += 1
-        self.cache.pop(key, None)
-
-    def invalidate_all(self):
-        self.check_thread()
-        self.sequence += 1
-        self.cache.clear()
-
-
-class CacheDescriptor(object):
-    """ A method decorator that applies a memoizing cache around the function.
-
-    This caches deferreds, rather than the results themselves. Deferreds that
-    fail are removed from the cache.
-
-    The function is presumed to take zero or more arguments, which are used in
-    a tuple as the key for the cache. Hits are served directly from the cache;
-    misses use the function body to generate the value.
-
-    The wrapped function has an additional member, a callable called
-    "invalidate". This can be used to remove individual entries from the cache.
-
-    The wrapped function has another additional callable, called "prefill",
-    which can be used to insert values into the cache specifically, without
-    calling the calculation function.
-    """
-    def __init__(self, orig, max_entries=1000, num_args=1, lru=True,
-                 inlineCallbacks=False):
-        self.orig = orig
-
-        if inlineCallbacks:
-            self.function_to_call = defer.inlineCallbacks(orig)
-        else:
-            self.function_to_call = orig
-
-        self.max_entries = max_entries
-        self.num_args = num_args
-        self.lru = lru
-
-        self.arg_names = inspect.getargspec(orig).args[1:num_args+1]
-
-        if len(self.arg_names) < self.num_args:
-            raise Exception(
-                "Not enough explicit positional arguments to key off of for %r."
-                " (@cached cannot key off of *args or **kwars)"
-                % (orig.__name__,)
-            )
-
-        self.cache = Cache(
-            name=self.orig.__name__,
-            max_entries=self.max_entries,
-            keylen=self.num_args,
-            lru=self.lru,
-        )
-
-    def __get__(self, obj, objtype=None):
-
-        @functools.wraps(self.orig)
-        def wrapped(*args, **kwargs):
-            arg_dict = inspect.getcallargs(self.orig, obj, *args, **kwargs)
-            cache_key = tuple(arg_dict[arg_nm] for arg_nm in self.arg_names)
-            try:
-                cached_result_d = self.cache.get(cache_key)
-
-                observer = cached_result_d.observe()
-                if DEBUG_CACHES:
-                    @defer.inlineCallbacks
-                    def check_result(cached_result):
-                        actual_result = yield self.function_to_call(obj, *args, **kwargs)
-                        if actual_result != cached_result:
-                            logger.error(
-                                "Stale cache entry %s%r: cached: %r, actual %r",
-                                self.orig.__name__, cache_key,
-                                cached_result, actual_result,
-                            )
-                            raise ValueError("Stale cache entry")
-                        defer.returnValue(cached_result)
-                    observer.addCallback(check_result)
-
-                return observer
-            except KeyError:
-                # Get the sequence number of the cache before reading from the
-                # database so that we can tell if the cache is invalidated
-                # while the SELECT is executing (SYN-369)
-                sequence = self.cache.sequence
-
-                ret = defer.maybeDeferred(
-                    self.function_to_call,
-                    obj, *args, **kwargs
-                )
-
-                def onErr(f):
-                    self.cache.invalidate(cache_key)
-                    return f
-
-                ret.addErrback(onErr)
-
-                ret = ObservableDeferred(ret, consumeErrors=True)
-                self.cache.update(sequence, cache_key, ret)
-
-                return ret.observe()
-
-        wrapped.invalidate = self.cache.invalidate
-        wrapped.invalidate_all = self.cache.invalidate_all
-        wrapped.prefill = self.cache.prefill
-
-        obj.__dict__[self.orig.__name__] = wrapped
-
-        return wrapped
-
-
-def cached(max_entries=1000, num_args=1, lru=True):
-    return lambda orig: CacheDescriptor(
-        orig,
-        max_entries=max_entries,
-        num_args=num_args,
-        lru=lru
-    )
-
-
-def cachedInlineCallbacks(max_entries=1000, num_args=1, lru=False):
-    return lambda orig: CacheDescriptor(
-        orig,
-        max_entries=max_entries,
-        num_args=num_args,
-        lru=lru,
-        inlineCallbacks=True,
-    )
-
 
 class LoggingTransaction(object):
     """An object that almost-transparently proxies for the 'txn' object
@@ -372,6 +167,8 @@ class SQLBaseStore(object):
         self._get_event_cache = Cache("*getEvent*", keylen=3, lru=True,
                                       max_entries=hs.config.event_cache_size)
 
+        self._state_group_cache = DictionaryCache("*stateGroupCache*", 100000)
+
         self._event_fetch_lock = threading.Condition()
         self._event_fetch_list = []
         self._event_fetch_ongoing = 0
diff --git a/synapse/storage/directory.py b/synapse/storage/directory.py
index f3947bbe89..d92028ea43 100644
--- a/synapse/storage/directory.py
+++ b/synapse/storage/directory.py
@@ -13,7 +13,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ._base import SQLBaseStore, cached
+from ._base import SQLBaseStore
+from synapse.util.caches.descriptors import cached
 
 from synapse.api.errors import SynapseError
 
diff --git a/synapse/storage/event_federation.py b/synapse/storage/event_federation.py
index 910b6598a7..25cc84eb95 100644
--- a/synapse/storage/event_federation.py
+++ b/synapse/storage/event_federation.py
@@ -15,7 +15,8 @@
 
 from twisted.internet import defer
 
-from ._base import SQLBaseStore, cached
+from ._base import SQLBaseStore
+from synapse.util.caches.descriptors import cached
 from syutil.base64util import encode_base64
 
 import logging
diff --git a/synapse/storage/keys.py b/synapse/storage/keys.py
index 49b8e37cfd..ffd6daa880 100644
--- a/synapse/storage/keys.py
+++ b/synapse/storage/keys.py
@@ -13,7 +13,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from _base import SQLBaseStore, cachedInlineCallbacks
+from _base import SQLBaseStore
+from synapse.util.caches.descriptors import cachedInlineCallbacks
 
 from twisted.internet import defer
 
diff --git a/synapse/storage/presence.py b/synapse/storage/presence.py
index 576cf670cc..4f91a2b87c 100644
--- a/synapse/storage/presence.py
+++ b/synapse/storage/presence.py
@@ -13,7 +13,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ._base import SQLBaseStore, cached
+from ._base import SQLBaseStore
+from synapse.util.caches.descriptors import cached
 
 from twisted.internet import defer
 
diff --git a/synapse/storage/push_rule.py b/synapse/storage/push_rule.py
index 9b88ca7b39..5305b7e122 100644
--- a/synapse/storage/push_rule.py
+++ b/synapse/storage/push_rule.py
@@ -13,7 +13,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ._base import SQLBaseStore, cachedInlineCallbacks
+from ._base import SQLBaseStore
+from synapse.util.caches.descriptors import cachedInlineCallbacks
 from twisted.internet import defer
 
 import logging
diff --git a/synapse/storage/receipts.py b/synapse/storage/receipts.py
index b79d6683ca..cac1a5657e 100644
--- a/synapse/storage/receipts.py
+++ b/synapse/storage/receipts.py
@@ -13,7 +13,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ._base import SQLBaseStore, cachedInlineCallbacks
+from ._base import SQLBaseStore
+from synapse.util.caches.descriptors import cachedInlineCallbacks
 
 from twisted.internet import defer
 
diff --git a/synapse/storage/registration.py b/synapse/storage/registration.py
index 4eaa088b36..aa446f94c6 100644
--- a/synapse/storage/registration.py
+++ b/synapse/storage/registration.py
@@ -17,7 +17,8 @@ from twisted.internet import defer
 
 from synapse.api.errors import StoreError, Codes
 
-from ._base import SQLBaseStore, cached
+from ._base import SQLBaseStore
+from synapse.util.caches.descriptors import cached
 
 
 class RegistrationStore(SQLBaseStore):
diff --git a/synapse/storage/room.py b/synapse/storage/room.py
index dd5bc2c8fb..5e07b7e0e5 100644
--- a/synapse/storage/room.py
+++ b/synapse/storage/room.py
@@ -17,7 +17,8 @@ from twisted.internet import defer
 
 from synapse.api.errors import StoreError
 
-from ._base import SQLBaseStore, cachedInlineCallbacks
+from ._base import SQLBaseStore
+from synapse.util.caches.descriptors import cachedInlineCallbacks
 
 import collections
 import logging
diff --git a/synapse/storage/roommember.py b/synapse/storage/roommember.py
index 9f14f38f24..8eee2dfbcc 100644
--- a/synapse/storage/roommember.py
+++ b/synapse/storage/roommember.py
@@ -17,7 +17,8 @@ from twisted.internet import defer
 
 from collections import namedtuple
 
-from ._base import SQLBaseStore, cached
+from ._base import SQLBaseStore
+from synapse.util.caches.descriptors import cached
 
 from synapse.api.constants import Membership
 from synapse.types import UserID
diff --git a/synapse/storage/state.py b/synapse/storage/state.py
index 7ce51b9bdc..abc5b66431 100644
--- a/synapse/storage/state.py
+++ b/synapse/storage/state.py
@@ -13,7 +13,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ._base import SQLBaseStore, cachedInlineCallbacks
+from ._base import SQLBaseStore
+from synapse.util.caches.descriptors import (
+    cached, cachedInlineCallbacks, cachedList
+)
 
 from twisted.internet import defer
 
@@ -44,59 +47,26 @@ class StateStore(SQLBaseStore):
     """
 
     @defer.inlineCallbacks
-    def get_state_groups(self, event_ids):
+    def get_state_groups(self, room_id, event_ids):
         """ Get the state groups for the given list of event_ids
 
         The return value is a dict mapping group names to lists of events.
         """
+        if not event_ids:
+            defer.returnValue({})
 
-        def f(txn):
-            groups = set()
-            for event_id in event_ids:
-                group = self._simple_select_one_onecol_txn(
-                    txn,
-                    table="event_to_state_groups",
-                    keyvalues={"event_id": event_id},
-                    retcol="state_group",
-                    allow_none=True,
-                )
-                if group:
-                    groups.add(group)
-
-            res = {}
-            for group in groups:
-                state_ids = self._simple_select_onecol_txn(
-                    txn,
-                    table="state_groups_state",
-                    keyvalues={"state_group": group},
-                    retcol="event_id",
-                )
-
-                res[group] = state_ids
-
-            return res
-
-        states = yield self.runInteraction(
-            "get_state_groups",
-            f,
+        event_to_groups = yield self._get_state_group_for_events(
+            room_id, event_ids,
         )
 
-        state_list = yield defer.gatherResults(
-            [
-                self._fetch_events_for_group(group, vals)
-                for group, vals in states.items()
-            ],
-            consumeErrors=True,
-        )
+        groups = set(event_to_groups.values())
 
-        defer.returnValue(dict(state_list))
+        group_to_state = yield self._get_state_for_groups(groups)
 
-    def _fetch_events_for_group(self, key, events):
-        return self._get_events(
-            events, get_prev_content=False
-        ).addCallback(
-            lambda evs: (key, evs)
-        )
+        defer.returnValue({
+            group: state_map.values()
+            for group, state_map in group_to_state.items()
+        })
 
     def _store_state_groups_txn(self, txn, event, context):
         return self._store_mult_state_groups_txn(txn, [(event, context)])
@@ -204,64 +174,281 @@ class StateStore(SQLBaseStore):
         events = yield self._get_events(event_ids, get_prev_content=False)
         defer.returnValue(events)
 
+    def _get_state_groups_from_groups(self, groups_and_types):
+        """Returns dictionary state_group -> state event ids
+
+        Args:
+            groups_and_types (list): list of 2-tuple (`group`, `types`)
+        """
+        def f(txn):
+            results = {}
+            for group, types in groups_and_types:
+                if types is not None:
+                    where_clause = "AND (%s)" % (
+                        " OR ".join(["(type = ? AND state_key = ?)"] * len(types)),
+                    )
+                else:
+                    where_clause = ""
+
+                sql = (
+                    "SELECT event_id FROM state_groups_state WHERE"
+                    " state_group = ? %s"
+                ) % (where_clause,)
+
+                args = [group]
+                if types is not None:
+                    args.extend([i for typ in types for i in typ])
+
+                txn.execute(sql, args)
+
+                results[group] = [
+                    r[0]
+                    for r in txn.fetchall()
+                ]
+
+            return results
+
+        return self.runInteraction(
+            "_get_state_groups_from_groups",
+            f,
+        )
+
     @defer.inlineCallbacks
-    def get_state_for_events(self, room_id, event_ids):
+    def get_state_for_events(self, room_id, event_ids, types):
+        """Given a list of event_ids and type tuples, return a list of state
+        dicts for each event. The state dicts will only have the type/state_keys
+        that are in the `types` list.
+
+        Args:
+            room_id (str)
+            event_ids (list)
+            types (list): List of (type, state_key) tuples which are used to
+                filter the state fetched. `state_key` may be None, which matches
+                any `state_key`
+
+        Returns:
+            deferred: A list of dicts corresponding to the event_ids given.
+            The dicts are mappings from (type, state_key) -> state_events
+        """
+        event_to_groups = yield self._get_state_group_for_events(
+            room_id, event_ids,
+        )
+
+        groups = set(event_to_groups.values())
+
+        group_to_state = yield self._get_state_for_groups(
+            groups, types
+        )
+
+        event_to_state = {
+            event_id: group_to_state[group]
+            for event_id, group in event_to_groups.items()
+        }
+
+        defer.returnValue({
+            event: event_to_state[event]
+            for event in event_ids
+        })
+
+    @cached(num_args=2, lru=True, max_entries=100000)
+    def _get_state_group_for_event(self, room_id, event_id):
+        return self._simple_select_one_onecol(
+            table="event_to_state_groups",
+            keyvalues={
+                "event_id": event_id,
+            },
+            retcol="state_group",
+            allow_none=True,
+            desc="_get_state_group_for_event",
+        )
+
+    @cachedList(cache=_get_state_group_for_event.cache, list_name="event_ids",
+                num_args=2)
+    def _get_state_group_for_events(self, room_id, event_ids):
+        """Returns mapping event_id -> state_group
+        """
         def f(txn):
-            groups = set()
-            event_to_group = {}
+            results = {}
             for event_id in event_ids:
-                # TODO: Remove this loop.
-                group = self._simple_select_one_onecol_txn(
+                results[event_id] = self._simple_select_one_onecol_txn(
                     txn,
                     table="event_to_state_groups",
-                    keyvalues={"event_id": event_id},
+                    keyvalues={
+                        "event_id": event_id,
+                    },
                     retcol="state_group",
                     allow_none=True,
                 )
-                if group:
-                    event_to_group[event_id] = group
-                    groups.add(group)
 
-            group_to_state_ids = {}
-            for group in groups:
-                state_ids = self._simple_select_onecol_txn(
-                    txn,
-                    table="state_groups_state",
-                    keyvalues={"state_group": group},
-                    retcol="event_id",
+            return results
+
+        return self.runInteraction(
+            "_get_state_group_for_events",
+            f,
+        )
+
+    def _get_some_state_from_cache(self, group, types):
+        """Checks if group is in cache. See `_get_state_for_groups`
+
+        Returns 3-tuple (`state_dict`, `missing_types`, `got_all`).
+        `missing_types` is the list of types that aren't in the cache for that
+        group. `got_all` is a bool indicating if we successfully retrieved all
+        requests state from the cache, if False we need to query the DB for the
+        missing state.
+
+        Args:
+            group: The state group to lookup
+            types (list): List of 2-tuples of the form (`type`, `state_key`),
+                where a `state_key` of `None` matches all state_keys for the
+                `type`.
+        """
+        is_all, state_dict = self._state_group_cache.get(group)
+
+        type_to_key = {}
+        missing_types = set()
+        for typ, state_key in types:
+            if state_key is None:
+                type_to_key[typ] = None
+                missing_types.add((typ, state_key))
+            else:
+                if type_to_key.get(typ, object()) is not None:
+                    type_to_key.setdefault(typ, set()).add(state_key)
+
+                if (typ, state_key) not in state_dict:
+                    missing_types.add((typ, state_key))
+
+        sentinel = object()
+
+        def include(typ, state_key):
+            valid_state_keys = type_to_key.get(typ, sentinel)
+            if valid_state_keys is sentinel:
+                return False
+            if valid_state_keys is None:
+                return True
+            if state_key in valid_state_keys:
+                return True
+            return False
+
+        got_all = not (missing_types or types is None)
+
+        return {
+            k: v
+            for k, v in state_dict.items()
+            if include(k[0], k[1])
+        }, missing_types, got_all
+
+    def _get_all_state_from_cache(self, group):
+        """Checks if group is in cache. See `_get_state_for_groups`
+
+        Returns 2-tuple (`state_dict`, `got_all`). `got_all` is a bool
+        indicating if we successfully retrieved all requests state from the
+        cache, if False we need to query the DB for the missing state.
+
+        Args:
+            group: The state group to lookup
+        """
+        is_all, state_dict = self._state_group_cache.get(group)
+        return state_dict, is_all
+
+    @defer.inlineCallbacks
+    def _get_state_for_groups(self, groups, types=None):
+        """Given list of groups returns dict of group -> list of state events
+        with matching types. `types` is a list of `(type, state_key)`, where
+        a `state_key` of None matches all state_keys. If `types` is None then
+        all events are returned.
+        """
+        results = {}
+        missing_groups_and_types = []
+        if types is not None:
+            for group in set(groups):
+                state_dict, missing_types, got_all = self._get_some_state_from_cache(
+                    group, types
                 )
 
-                group_to_state_ids[group] = state_ids
+                results[group] = state_dict
+
+                if not got_all:
+                    missing_groups_and_types.append((
+                        group,
+                        missing_types
+                    ))
+        else:
+            for group in set(groups):
+                state_dict, got_all = self._get_all_state_from_cache(
+                    group
+                )
 
-            return event_to_group, group_to_state_ids
+                results[group] = state_dict
 
-        res = yield self.runInteraction(
-            "annotate_events_with_state_groups",
-            f,
-        )
+                if not got_all:
+                    missing_groups_and_types.append((group, None))
 
-        event_to_group, group_to_state_ids = res
+        if not missing_groups_and_types:
+            defer.returnValue({
+                group: {
+                    type_tuple: event
+                    for type_tuple, event in state.items()
+                    if event
+                }
+                for group, state in results.items()
+            })
 
-        state_list = yield defer.gatherResults(
-            [
-                self._fetch_events_for_group(group, vals)
-                for group, vals in group_to_state_ids.items()
-            ],
-            consumeErrors=True,
+        # Okay, so we have some missing_types, lets fetch them.
+        cache_seq_num = self._state_group_cache.sequence
+
+        group_state_dict = yield self._get_state_groups_from_groups(
+            missing_groups_and_types
+        )
+
+        state_events = yield self._get_events(
+            [e_id for l in group_state_dict.values() for e_id in l],
+            get_prev_content=False
         )
 
-        state_dict = {
-            group: {
-                (ev.type, ev.state_key): ev
-                for ev in state
-            }
-            for group, state in state_list
+        state_events = {
+            e.event_id: e
+            for e in state_events
         }
 
-        defer.returnValue([
-            state_dict.get(event_to_group.get(event, None), None)
-            for event in event_ids
-        ])
+        # Now we want to update the cache with all the things we fetched
+        # from the database.
+        for group, state_ids in group_state_dict.items():
+            if types:
+                # We delibrately put key -> None mappings into the cache to
+                # cache absence of the key, on the assumption that if we've
+                # explicitly asked for some types then we will probably ask
+                # for them again.
+                state_dict = {
+                    key: None
+                    for key in types
+                }
+                state_dict.update(results[group])
+            else:
+                state_dict = results[group]
+
+            evs = [
+                state_events[e_id] for e_id in state_ids
+                if e_id in state_events  # This can happen if event is rejected.
+            ]
+            state_dict.update({
+                (e.type, e.state_key): e
+                for e in evs
+            })
+
+            self._state_group_cache.update(
+                cache_seq_num,
+                key=group,
+                value=state_dict,
+                full=(types is None),
+            )
+
+            results[group].update({
+                key: value
+                for key, value in state_dict.items()
+                if value
+            })
+
+        defer.returnValue(results)
 
 
 def _make_group_id(clock):
diff --git a/synapse/storage/stream.py b/synapse/storage/stream.py
index af45fc5619..d7fe423f5a 100644
--- a/synapse/storage/stream.py
+++ b/synapse/storage/stream.py
@@ -36,6 +36,7 @@ what sort order was used:
 from twisted.internet import defer
 
 from ._base import SQLBaseStore
+from synapse.util.caches.descriptors import cachedInlineCallbacks
 from synapse.api.constants import EventTypes
 from synapse.types import RoomStreamToken
 from synapse.util.logutils import log_function
@@ -299,9 +300,8 @@ class StreamStore(SQLBaseStore):
 
         defer.returnValue((events, token))
 
-    @defer.inlineCallbacks
-    def get_recent_events_for_room(self, room_id, limit, end_token,
-                                   with_feedback=False, from_token=None):
+    @cachedInlineCallbacks(num_args=4)
+    def get_recent_events_for_room(self, room_id, limit, end_token, from_token=None):
         # TODO (erikj): Handle compressed feedback
 
         end_token = RoomStreamToken.parse_stream_token(end_token)
diff --git a/synapse/storage/transactions.py b/synapse/storage/transactions.py
index 624da4a9dc..c8c7e6591a 100644
--- a/synapse/storage/transactions.py
+++ b/synapse/storage/transactions.py
@@ -13,7 +13,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ._base import SQLBaseStore, cached
+from ._base import SQLBaseStore
+from synapse.util.caches.descriptors import cached
 
 from collections import namedtuple
 
diff --git a/synapse/util/caches/__init__.py b/synapse/util/caches/__init__.py
new file mode 100644
index 0000000000..da0e06a468
--- /dev/null
+++ b/synapse/util/caches/__init__.py
@@ -0,0 +1,27 @@
+# -*- coding: utf-8 -*-
+# Copyright 2015 OpenMarket 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.
+
+import synapse.metrics
+
+DEBUG_CACHES = False
+
+metrics = synapse.metrics.get_metrics_for("synapse.util.caches")
+
+caches_by_name = {}
+cache_counter = metrics.register_cache(
+    "cache",
+    lambda: {(name,): len(caches_by_name[name]) for name in caches_by_name.keys()},
+    labels=["name"],
+)
diff --git a/synapse/util/caches/descriptors.py b/synapse/util/caches/descriptors.py
new file mode 100644
index 0000000000..83bfec2f02
--- /dev/null
+++ b/synapse/util/caches/descriptors.py
@@ -0,0 +1,377 @@
+# -*- coding: utf-8 -*-
+# Copyright 2015 OpenMarket 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.
+import logging
+
+from synapse.util.async import ObservableDeferred
+from synapse.util import unwrapFirstError
+from synapse.util.caches.lrucache import LruCache
+
+from . import caches_by_name, DEBUG_CACHES, cache_counter
+
+from twisted.internet import defer
+
+from collections import OrderedDict
+
+import functools
+import inspect
+import threading
+
+logger = logging.getLogger(__name__)
+
+
+_CacheSentinel = object()
+
+
+class Cache(object):
+
+    def __init__(self, name, max_entries=1000, keylen=1, lru=True):
+        if lru:
+            self.cache = LruCache(max_size=max_entries)
+            self.max_entries = None
+        else:
+            self.cache = OrderedDict()
+            self.max_entries = max_entries
+
+        self.name = name
+        self.keylen = keylen
+        self.sequence = 0
+        self.thread = None
+        caches_by_name[name] = self.cache
+
+    def check_thread(self):
+        expected_thread = self.thread
+        if expected_thread is None:
+            self.thread = threading.current_thread()
+        else:
+            if expected_thread is not threading.current_thread():
+                raise ValueError(
+                    "Cache objects can only be accessed from the main thread"
+                )
+
+    def get(self, key, default=_CacheSentinel):
+        val = self.cache.get(key, _CacheSentinel)
+        if val is not _CacheSentinel:
+            cache_counter.inc_hits(self.name)
+            return val
+
+        cache_counter.inc_misses(self.name)
+
+        if default is _CacheSentinel:
+            raise KeyError()
+        else:
+            return default
+
+    def update(self, sequence, key, value):
+        self.check_thread()
+        if self.sequence == sequence:
+            # Only update the cache if the caches sequence number matches the
+            # number that the cache had before the SELECT was started (SYN-369)
+            self.prefill(key, value)
+
+    def prefill(self, key, value):
+        if self.max_entries is not None:
+            while len(self.cache) >= self.max_entries:
+                self.cache.popitem(last=False)
+
+        self.cache[key] = value
+
+    def invalidate(self, key):
+        self.check_thread()
+        if not isinstance(key, tuple):
+            raise TypeError(
+                "The cache key must be a tuple not %r" % (type(key),)
+            )
+
+        # Increment the sequence number so that any SELECT statements that
+        # raced with the INSERT don't update the cache (SYN-369)
+        self.sequence += 1
+        self.cache.pop(key, None)
+
+    def invalidate_all(self):
+        self.check_thread()
+        self.sequence += 1
+        self.cache.clear()
+
+
+class CacheDescriptor(object):
+    """ A method decorator that applies a memoizing cache around the function.
+
+    This caches deferreds, rather than the results themselves. Deferreds that
+    fail are removed from the cache.
+
+    The function is presumed to take zero or more arguments, which are used in
+    a tuple as the key for the cache. Hits are served directly from the cache;
+    misses use the function body to generate the value.
+
+    The wrapped function has an additional member, a callable called
+    "invalidate". This can be used to remove individual entries from the cache.
+
+    The wrapped function has another additional callable, called "prefill",
+    which can be used to insert values into the cache specifically, without
+    calling the calculation function.
+    """
+    def __init__(self, orig, max_entries=1000, num_args=1, lru=True,
+                 inlineCallbacks=False):
+        self.orig = orig
+
+        if inlineCallbacks:
+            self.function_to_call = defer.inlineCallbacks(orig)
+        else:
+            self.function_to_call = orig
+
+        self.max_entries = max_entries
+        self.num_args = num_args
+        self.lru = lru
+
+        self.arg_names = inspect.getargspec(orig).args[1:num_args+1]
+
+        if len(self.arg_names) < self.num_args:
+            raise Exception(
+                "Not enough explicit positional arguments to key off of for %r."
+                " (@cached cannot key off of *args or **kwars)"
+                % (orig.__name__,)
+            )
+
+        self.cache = Cache(
+            name=self.orig.__name__,
+            max_entries=self.max_entries,
+            keylen=self.num_args,
+            lru=self.lru,
+        )
+
+    def __get__(self, obj, objtype=None):
+
+        @functools.wraps(self.orig)
+        def wrapped(*args, **kwargs):
+            arg_dict = inspect.getcallargs(self.orig, obj, *args, **kwargs)
+            cache_key = tuple(arg_dict[arg_nm] for arg_nm in self.arg_names)
+            try:
+                cached_result_d = self.cache.get(cache_key)
+
+                observer = cached_result_d.observe()
+                if DEBUG_CACHES:
+                    @defer.inlineCallbacks
+                    def check_result(cached_result):
+                        actual_result = yield self.function_to_call(obj, *args, **kwargs)
+                        if actual_result != cached_result:
+                            logger.error(
+                                "Stale cache entry %s%r: cached: %r, actual %r",
+                                self.orig.__name__, cache_key,
+                                cached_result, actual_result,
+                            )
+                            raise ValueError("Stale cache entry")
+                        defer.returnValue(cached_result)
+                    observer.addCallback(check_result)
+
+                return observer
+            except KeyError:
+                # Get the sequence number of the cache before reading from the
+                # database so that we can tell if the cache is invalidated
+                # while the SELECT is executing (SYN-369)
+                sequence = self.cache.sequence
+
+                ret = defer.maybeDeferred(
+                    self.function_to_call,
+                    obj, *args, **kwargs
+                )
+
+                def onErr(f):
+                    self.cache.invalidate(cache_key)
+                    return f
+
+                ret.addErrback(onErr)
+
+                ret = ObservableDeferred(ret, consumeErrors=True)
+                self.cache.update(sequence, cache_key, ret)
+
+                return ret.observe()
+
+        wrapped.invalidate = self.cache.invalidate
+        wrapped.invalidate_all = self.cache.invalidate_all
+        wrapped.prefill = self.cache.prefill
+
+        obj.__dict__[self.orig.__name__] = wrapped
+
+        return wrapped
+
+
+class CacheListDescriptor(object):
+    """Wraps an existing cache to support bulk fetching of keys.
+
+    Given a list of keys it looks in the cache to find any hits, then passes
+    the list of missing keys to the wrapped fucntion.
+    """
+
+    def __init__(self, orig, cache, list_name, num_args=1, inlineCallbacks=False):
+        """
+        Args:
+            orig (function)
+            cache (Cache)
+            list_name (str): Name of the argument which is the bulk lookup list
+            num_args (int)
+            inlineCallbacks (bool): Whether orig is a generator that should
+                be wrapped by defer.inlineCallbacks
+        """
+        self.orig = orig
+
+        if inlineCallbacks:
+            self.function_to_call = defer.inlineCallbacks(orig)
+        else:
+            self.function_to_call = orig
+
+        self.num_args = num_args
+        self.list_name = list_name
+
+        self.arg_names = inspect.getargspec(orig).args[1:num_args+1]
+        self.list_pos = self.arg_names.index(self.list_name)
+
+        self.cache = cache
+
+        self.sentinel = object()
+
+        if len(self.arg_names) < self.num_args:
+            raise Exception(
+                "Not enough explicit positional arguments to key off of for %r."
+                " (@cached cannot key off of *args or **kwars)"
+                % (orig.__name__,)
+            )
+
+        if self.list_name not in self.arg_names:
+            raise Exception(
+                "Couldn't see arguments %r for %r."
+                % (self.list_name, cache.name,)
+            )
+
+    def __get__(self, obj, objtype=None):
+
+        @functools.wraps(self.orig)
+        def wrapped(*args, **kwargs):
+            arg_dict = inspect.getcallargs(self.orig, obj, *args, **kwargs)
+            keyargs = [arg_dict[arg_nm] for arg_nm in self.arg_names]
+            list_args = arg_dict[self.list_name]
+
+            # cached is a dict arg -> deferred, where deferred results in a
+            # 2-tuple (`arg`, `result`)
+            cached = {}
+            missing = []
+            for arg in list_args:
+                key = list(keyargs)
+                key[self.list_pos] = arg
+
+                try:
+                    res = self.cache.get(tuple(key)).observe()
+                    res.addCallback(lambda r, arg: (arg, r), arg)
+                    cached[arg] = res
+                except KeyError:
+                    missing.append(arg)
+
+            if missing:
+                sequence = self.cache.sequence
+                args_to_call = dict(arg_dict)
+                args_to_call[self.list_name] = missing
+
+                ret_d = defer.maybeDeferred(
+                    self.function_to_call,
+                    **args_to_call
+                )
+
+                ret_d = ObservableDeferred(ret_d)
+
+                # We need to create deferreds for each arg in the list so that
+                # we can insert the new deferred into the cache.
+                for arg in missing:
+                    observer = ret_d.observe()
+                    observer.addCallback(lambda r, arg: r[arg], arg)
+
+                    observer = ObservableDeferred(observer)
+
+                    key = list(keyargs)
+                    key[self.list_pos] = arg
+                    self.cache.update(sequence, tuple(key), observer)
+
+                    def invalidate(f, key):
+                        self.cache.invalidate(key)
+                        return f
+                    observer.addErrback(invalidate, tuple(key))
+
+                    res = observer.observe()
+                    res.addCallback(lambda r, arg: (arg, r), arg)
+
+                    cached[arg] = res
+
+            return defer.gatherResults(
+                cached.values(),
+                consumeErrors=True,
+            ).addErrback(unwrapFirstError).addCallback(lambda res: dict(res))
+
+        obj.__dict__[self.orig.__name__] = wrapped
+
+        return wrapped
+
+
+def cached(max_entries=1000, num_args=1, lru=True):
+    return lambda orig: CacheDescriptor(
+        orig,
+        max_entries=max_entries,
+        num_args=num_args,
+        lru=lru
+    )
+
+
+def cachedInlineCallbacks(max_entries=1000, num_args=1, lru=False):
+    return lambda orig: CacheDescriptor(
+        orig,
+        max_entries=max_entries,
+        num_args=num_args,
+        lru=lru,
+        inlineCallbacks=True,
+    )
+
+
+def cachedList(cache, list_name, num_args=1, inlineCallbacks=False):
+    """Creates a descriptor that wraps a function in a `CacheListDescriptor`.
+
+    Used to do batch lookups for an already created cache. A single argument
+    is specified as a list that is iterated through to lookup keys in the
+    original cache. A new list consisting of the keys that weren't in the cache
+    get passed to the original function, the result of which is stored in the
+    cache.
+
+    Args:
+        cache (Cache): The underlying cache to use.
+        list_name (str): The name of the argument that is the list to use to
+            do batch lookups in the cache.
+        num_args (int): Number of arguments to use as the key in the cache.
+        inlineCallbacks (bool): Should the function be wrapped in an
+            `defer.inlineCallbacks`?
+
+    Example:
+
+        class Example(object):
+            @cached(num_args=2)
+            def do_something(self, first_arg):
+                ...
+
+            @cachedList(do_something.cache, list_name="second_args", num_args=2)
+            def batch_do_something(self, first_arg, second_args):
+                ...
+    """
+    return lambda orig: CacheListDescriptor(
+        orig,
+        cache=cache,
+        list_name=list_name,
+        num_args=num_args,
+        inlineCallbacks=inlineCallbacks,
+    )
diff --git a/synapse/util/caches/dictionary_cache.py b/synapse/util/caches/dictionary_cache.py
new file mode 100644
index 0000000000..e69adf62fe
--- /dev/null
+++ b/synapse/util/caches/dictionary_cache.py
@@ -0,0 +1,103 @@
+# -*- coding: utf-8 -*-
+# Copyright 2015 OpenMarket 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.
+
+from synapse.util.caches.lrucache import LruCache
+from collections import namedtuple
+from . import caches_by_name, cache_counter
+import threading
+import logging
+
+
+logger = logging.getLogger(__name__)
+
+
+DictionaryEntry = namedtuple("DictionaryEntry", ("full", "value"))
+
+
+class DictionaryCache(object):
+    """Caches key -> dictionary lookups, supporting caching partial dicts, i.e.
+    fetching a subset of dictionary keys for a particular key.
+    """
+
+    def __init__(self, name, max_entries=1000):
+        self.cache = LruCache(max_size=max_entries)
+
+        self.name = name
+        self.sequence = 0
+        self.thread = None
+        # caches_by_name[name] = self.cache
+
+        class Sentinel(object):
+            __slots__ = []
+
+        self.sentinel = Sentinel()
+        caches_by_name[name] = self.cache
+
+    def check_thread(self):
+        expected_thread = self.thread
+        if expected_thread is None:
+            self.thread = threading.current_thread()
+        else:
+            if expected_thread is not threading.current_thread():
+                raise ValueError(
+                    "Cache objects can only be accessed from the main thread"
+                )
+
+    def get(self, key, dict_keys=None):
+        entry = self.cache.get(key, self.sentinel)
+        if entry is not self.sentinel:
+            cache_counter.inc_hits(self.name)
+
+            if dict_keys is None:
+                return DictionaryEntry(entry.full, dict(entry.value))
+            else:
+                return DictionaryEntry(entry.full, {
+                    k: entry.value[k]
+                    for k in dict_keys
+                    if k in entry.value
+                })
+
+        cache_counter.inc_misses(self.name)
+        return DictionaryEntry(False, {})
+
+    def invalidate(self, key):
+        self.check_thread()
+
+        # Increment the sequence number so that any SELECT statements that
+        # raced with the INSERT don't update the cache (SYN-369)
+        self.sequence += 1
+        self.cache.pop(key, None)
+
+    def invalidate_all(self):
+        self.check_thread()
+        self.sequence += 1
+        self.cache.clear()
+
+    def update(self, sequence, key, value, full=False):
+        self.check_thread()
+        if self.sequence == sequence:
+            # Only update the cache if the caches sequence number matches the
+            # number that the cache had before the SELECT was started (SYN-369)
+            if full:
+                self._insert(key, value)
+            else:
+                self._update_or_insert(key, value)
+
+    def _update_or_insert(self, key, value):
+        entry = self.cache.setdefault(key, DictionaryEntry(False, {}))
+        entry.value.update(value)
+
+    def _insert(self, key, value):
+        self.cache[key] = DictionaryEntry(True, value)
diff --git a/synapse/util/expiringcache.py b/synapse/util/caches/expiringcache.py
index 06d1eea01b..06d1eea01b 100644
--- a/synapse/util/expiringcache.py
+++ b/synapse/util/caches/expiringcache.py
diff --git a/synapse/util/lrucache.py b/synapse/util/caches/lrucache.py
index cacd7e45fa..cacd7e45fa 100644
--- a/synapse/util/lrucache.py
+++ b/synapse/util/caches/lrucache.py