Merge branch 'release-v0.16.0' of github.com:matrix-org/synapse v0.16.0

1 files changed, 463 insertions, 287 deletions

diff --git a/synapse/storage/events.py b/synapse/storage/events.py
index 5233430028..6d978ffcd5 100644
--- a/synapse/storage/events.py
+++ b/synapse/storage/events.py
@@ -19,12 +19,17 @@ from twisted.internet import defer, reactor
 from synapse.events import FrozenEvent, USE_FROZEN_DICTS
 from synapse.events.utils import prune_event
 
+from synapse.util.async import ObservableDeferred
 from synapse.util.logcontext import preserve_fn, PreserveLoggingContext
 from synapse.util.logutils import log_function
 from synapse.api.constants import EventTypes
 
 from canonicaljson import encode_canonical_json
-from contextlib import contextmanager
+from collections import deque, namedtuple
+
+import synapse
+import synapse.metrics
+
 
 import logging
 import math
@@ -33,6 +38,10 @@ import ujson as json
 logger = logging.getLogger(__name__)
 
 
+metrics = synapse.metrics.get_metrics_for(__name__)
+persist_event_counter = metrics.register_counter("persisted_events")
+
+
 def encode_json(json_object):
     if USE_FROZEN_DICTS:
         # ujson doesn't like frozen_dicts
@@ -50,76 +59,237 @@ EVENT_QUEUE_ITERATIONS = 3  # No. times we block waiting for requests for events
 EVENT_QUEUE_TIMEOUT_S = 0.1  # Timeout when waiting for requests for events
 
 
+class _EventPeristenceQueue(object):
+    """Queues up events so that they can be persisted in bulk with only one
+    concurrent transaction per room.
+    """
+
+    _EventPersistQueueItem = namedtuple("_EventPersistQueueItem", (
+        "events_and_contexts", "current_state", "backfilled", "deferred",
+    ))
+
+    def __init__(self):
+        self._event_persist_queues = {}
+        self._currently_persisting_rooms = set()
+
+    def add_to_queue(self, room_id, events_and_contexts, backfilled, current_state):
+        """Add events to the queue, with the given persist_event options.
+        """
+        queue = self._event_persist_queues.setdefault(room_id, deque())
+        if queue:
+            end_item = queue[-1]
+            if end_item.current_state or current_state:
+                # We perist events with current_state set to True one at a time
+                pass
+            if end_item.backfilled == backfilled:
+                end_item.events_and_contexts.extend(events_and_contexts)
+                return end_item.deferred.observe()
+
+        deferred = ObservableDeferred(defer.Deferred())
+
+        queue.append(self._EventPersistQueueItem(
+            events_and_contexts=events_and_contexts,
+            backfilled=backfilled,
+            current_state=current_state,
+            deferred=deferred,
+        ))
+
+        return deferred.observe()
+
+    def handle_queue(self, room_id, per_item_callback):
+        """Attempts to handle the queue for a room if not already being handled.
+
+        The given callback will be invoked with for each item in the queue,1
+        of type _EventPersistQueueItem. The per_item_callback will continuously
+        be called with new items, unless the queue becomnes empty. The return
+        value of the function will be given to the deferreds waiting on the item,
+        exceptions will be passed to the deferres as well.
+
+        This function should therefore be called whenever anything is added
+        to the queue.
+
+        If another callback is currently handling the queue then it will not be
+        invoked.
+        """
+
+        if room_id in self._currently_persisting_rooms:
+            return
+
+        self._currently_persisting_rooms.add(room_id)
+
+        @defer.inlineCallbacks
+        def handle_queue_loop():
+            try:
+                queue = self._get_drainining_queue(room_id)
+                for item in queue:
+                    try:
+                        ret = yield per_item_callback(item)
+                        item.deferred.callback(ret)
+                    except Exception as e:
+                        item.deferred.errback(e)
+            finally:
+                queue = self._event_persist_queues.pop(room_id, None)
+                if queue:
+                    self._event_persist_queues[room_id] = queue
+                self._currently_persisting_rooms.discard(room_id)
+
+        preserve_fn(handle_queue_loop)()
+
+    def _get_drainining_queue(self, room_id):
+        queue = self._event_persist_queues.setdefault(room_id, deque())
+
+        try:
+            while True:
+                yield queue.popleft()
+        except IndexError:
+            # Queue has been drained.
+            pass
+
+
+_EventCacheEntry = namedtuple("_EventCacheEntry", ("event", "redacted_event"))
+
+
 class EventsStore(SQLBaseStore):
     EVENT_ORIGIN_SERVER_TS_NAME = "event_origin_server_ts"
 
     def __init__(self, hs):
         super(EventsStore, self).__init__(hs)
+        self._clock = hs.get_clock()
         self.register_background_update_handler(
             self.EVENT_ORIGIN_SERVER_TS_NAME, self._background_reindex_origin_server_ts
         )
 
+        self._event_persist_queue = _EventPeristenceQueue()
+
+    def persist_events(self, events_and_contexts, backfilled=False):
+        """
+        Write events to the database
+        Args:
+            events_and_contexts: list of tuples of (event, context)
+            backfilled: ?
+        """
+        partitioned = {}
+        for event, ctx in events_and_contexts:
+            partitioned.setdefault(event.room_id, []).append((event, ctx))
+
+        deferreds = []
+        for room_id, evs_ctxs in partitioned.items():
+            d = self._event_persist_queue.add_to_queue(
+                room_id, evs_ctxs,
+                backfilled=backfilled,
+                current_state=None,
+            )
+            deferreds.append(d)
+
+        for room_id in partitioned.keys():
+            self._maybe_start_persisting(room_id)
+
+        return defer.gatherResults(deferreds, consumeErrors=True)
+
     @defer.inlineCallbacks
-    def persist_events(self, events_and_contexts, backfilled=False,
-                       is_new_state=True):
+    @log_function
+    def persist_event(self, event, context, current_state=None, backfilled=False):
+        deferred = self._event_persist_queue.add_to_queue(
+            event.room_id, [(event, context)],
+            backfilled=backfilled,
+            current_state=current_state,
+        )
+
+        self._maybe_start_persisting(event.room_id)
+
+        yield deferred
+
+        max_persisted_id = yield self._stream_id_gen.get_current_token()
+        defer.returnValue((event.internal_metadata.stream_ordering, max_persisted_id))
+
+    def _maybe_start_persisting(self, room_id):
+        @defer.inlineCallbacks
+        def persisting_queue(item):
+            if item.current_state:
+                for event, context in item.events_and_contexts:
+                    # There should only ever be one item in
+                    # events_and_contexts when current_state is
+                    # not None
+                    yield self._persist_event(
+                        event, context,
+                        current_state=item.current_state,
+                        backfilled=item.backfilled,
+                    )
+            else:
+                yield self._persist_events(
+                    item.events_and_contexts,
+                    backfilled=item.backfilled,
+                )
+
+        self._event_persist_queue.handle_queue(room_id, persisting_queue)
+
+    @defer.inlineCallbacks
+    def _persist_events(self, events_and_contexts, backfilled=False):
         if not events_and_contexts:
             return
 
         if backfilled:
-            start = self.min_stream_token - 1
-            self.min_stream_token -= len(events_and_contexts) + 1
-            stream_orderings = range(start, self.min_stream_token, -1)
-
-            @contextmanager
-            def stream_ordering_manager():
-                yield stream_orderings
-            stream_ordering_manager = stream_ordering_manager()
+            stream_ordering_manager = self._backfill_id_gen.get_next_mult(
+                len(events_and_contexts)
+            )
         else:
             stream_ordering_manager = self._stream_id_gen.get_next_mult(
                 len(events_and_contexts)
             )
 
+        state_group_id_manager = self._state_groups_id_gen.get_next_mult(
+            len(events_and_contexts)
+        )
         with stream_ordering_manager as stream_orderings:
-            for (event, _), stream in zip(events_and_contexts, stream_orderings):
-                event.internal_metadata.stream_ordering = stream
-
-            chunks = [
-                events_and_contexts[x:x + 100]
-                for x in xrange(0, len(events_and_contexts), 100)
-            ]
-
-            for chunk in chunks:
-                # We can't easily parallelize these since different chunks
-                # might contain the same event. :(
-                yield self.runInteraction(
-                    "persist_events",
-                    self._persist_events_txn,
-                    events_and_contexts=chunk,
-                    backfilled=backfilled,
-                    is_new_state=is_new_state,
-                )
+            with state_group_id_manager as state_group_ids:
+                for (event, context), stream, state_group_id in zip(
+                    events_and_contexts, stream_orderings, state_group_ids
+                ):
+                    event.internal_metadata.stream_ordering = stream
+                    # Assign a state group_id in case a new id is needed for
+                    # this context. In theory we only need to assign this
+                    # for contexts that have current_state and aren't outliers
+                    # but that make the code more complicated. Assigning an ID
+                    # per event only causes the state_group_ids to grow as fast
+                    # as the stream_ordering so in practise shouldn't be a problem.
+                    context.new_state_group_id = state_group_id
+
+                chunks = [
+                    events_and_contexts[x:x + 100]
+                    for x in xrange(0, len(events_and_contexts), 100)
+                ]
+
+                for chunk in chunks:
+                    # We can't easily parallelize these since different chunks
+                    # might contain the same event. :(
+                    yield self.runInteraction(
+                        "persist_events",
+                        self._persist_events_txn,
+                        events_and_contexts=chunk,
+                        backfilled=backfilled,
+                    )
+                    persist_event_counter.inc_by(len(chunk))
 
     @defer.inlineCallbacks
     @log_function
-    def persist_event(self, event, context,
-                      is_new_state=True, current_state=None):
+    def _persist_event(self, event, context, current_state=None, backfilled=False):
         try:
             with self._stream_id_gen.get_next() as stream_ordering:
-                event.internal_metadata.stream_ordering = stream_ordering
-                yield self.runInteraction(
-                    "persist_event",
-                    self._persist_event_txn,
-                    event=event,
-                    context=context,
-                    is_new_state=is_new_state,
-                    current_state=current_state,
-                )
+                with self._state_groups_id_gen.get_next() as state_group_id:
+                    event.internal_metadata.stream_ordering = stream_ordering
+                    context.new_state_group_id = state_group_id
+                    yield self.runInteraction(
+                        "persist_event",
+                        self._persist_event_txn,
+                        event=event,
+                        context=context,
+                        current_state=current_state,
+                        backfilled=backfilled,
+                    )
+                    persist_event_counter.inc()
         except _RollbackButIsFineException:
             pass
 
-        max_persisted_id = yield self._stream_id_gen.get_max_token()
-        defer.returnValue((stream_ordering, max_persisted_id))
-
     @defer.inlineCallbacks
     def get_event(self, event_id, check_redacted=True,
                   get_prev_content=False, allow_rejected=False,
@@ -177,8 +347,7 @@ class EventsStore(SQLBaseStore):
         defer.returnValue({e.event_id: e for e in events})
 
     @log_function
-    def _persist_event_txn(self, txn, event, context,
-                           is_new_state=True, current_state=None):
+    def _persist_event_txn(self, txn, event, context, current_state, backfilled=False):
         # We purposefully do this first since if we include a `current_state`
         # key, we *want* to update the `current_state_events` table
         if current_state:
@@ -186,7 +355,16 @@ class EventsStore(SQLBaseStore):
             txn.call_after(self.get_rooms_for_user.invalidate_all)
             txn.call_after(self.get_users_in_room.invalidate, (event.room_id,))
             txn.call_after(self.get_joined_hosts_for_room.invalidate, (event.room_id,))
-            txn.call_after(self.get_room_name_and_aliases, event.room_id)
+            txn.call_after(self.get_room_name_and_aliases.invalidate, (event.room_id,))
+
+            # Add an entry to the current_state_resets table to record the point
+            # where we clobbered the current state
+            stream_order = event.internal_metadata.stream_ordering
+            self._simple_insert_txn(
+                txn,
+                table="current_state_resets",
+                values={"event_stream_ordering": stream_order}
+            )
 
             self._simple_delete_txn(
                 txn,
@@ -209,13 +387,11 @@ class EventsStore(SQLBaseStore):
         return self._persist_events_txn(
             txn,
             [(event, context)],
-            backfilled=False,
-            is_new_state=is_new_state,
+            backfilled=backfilled,
         )
 
     @log_function
-    def _persist_events_txn(self, txn, events_and_contexts, backfilled,
-                            is_new_state=True):
+    def _persist_events_txn(self, txn, events_and_contexts, backfilled):
         depth_updates = {}
         for event, context in events_and_contexts:
             # Remove the any existing cache entries for the event_ids
@@ -282,9 +458,7 @@ class EventsStore(SQLBaseStore):
 
             outlier_persisted = have_persisted[event.event_id]
             if not event.internal_metadata.is_outlier() and outlier_persisted:
-                self._store_state_groups_txn(
-                    txn, event, context,
-                )
+                self._store_mult_state_groups_txn(txn, ((event, context),))
 
                 metadata_json = encode_json(
                     event.internal_metadata.get_dict()
@@ -299,6 +473,18 @@ class EventsStore(SQLBaseStore):
                     (metadata_json, event.event_id,)
                 )
 
+                stream_order = event.internal_metadata.stream_ordering
+                state_group_id = context.state_group or context.new_state_group_id
+                self._simple_insert_txn(
+                    txn,
+                    table="ex_outlier_stream",
+                    values={
+                        "event_stream_ordering": stream_order,
+                        "event_id": event.event_id,
+                        "state_group": state_group_id,
+                    }
+                )
+
                 sql = (
                     "UPDATE events SET outlier = ?"
                     " WHERE event_id = ?"
@@ -310,19 +496,14 @@ class EventsStore(SQLBaseStore):
 
                 self._update_extremeties(txn, [event])
 
-        events_and_contexts = filter(
-            lambda ec: ec[0] not in to_remove,
-            events_and_contexts
-        )
+        events_and_contexts = [
+            ec for ec in events_and_contexts if ec[0] not in to_remove
+        ]
 
         if not events_and_contexts:
             return
 
-        self._store_mult_state_groups_txn(txn, [
-            (event, context)
-            for event, context in events_and_contexts
-            if not event.internal_metadata.is_outlier()
-        ])
+        self._store_mult_state_groups_txn(txn, events_and_contexts)
 
         self._handle_mult_prev_events(
             txn,
@@ -349,7 +530,8 @@ class EventsStore(SQLBaseStore):
                 event
                 for event, _ in events_and_contexts
                 if event.type == EventTypes.Member
-            ]
+            ],
+            backfilled=backfilled,
         )
 
         def event_dict(event):
@@ -393,6 +575,7 @@ class EventsStore(SQLBaseStore):
                     "outlier": event.internal_metadata.is_outlier(),
                     "content": encode_json(event.content).decode("UTF-8"),
                     "origin_server_ts": int(event.origin_server_ts),
+                    "received_ts": self._clock.time_msec(),
                 }
                 for event, _ in events_and_contexts
             ],
@@ -421,10 +604,9 @@ class EventsStore(SQLBaseStore):
             txn, [event for event, _ in events_and_contexts]
         )
 
-        state_events_and_contexts = filter(
-            lambda i: i[0].is_state(),
-            events_and_contexts,
-        )
+        state_events_and_contexts = [
+            ec for ec in events_and_contexts if ec[0].is_state()
+        ]
 
         state_values = []
         for event, context in state_events_and_contexts:
@@ -462,35 +644,88 @@ class EventsStore(SQLBaseStore):
             ],
         )
 
-        if is_new_state:
-            for event, _ in state_events_and_contexts:
-                if not context.rejected:
-                    txn.call_after(
-                        self._get_current_state_for_key.invalidate,
-                        (event.room_id, event.type, event.state_key,)
-                    )
+        self._add_to_cache(txn, events_and_contexts)
 
-                    if event.type in [EventTypes.Name, EventTypes.Aliases]:
-                        txn.call_after(
-                            self.get_room_name_and_aliases.invalidate,
-                            (event.room_id,)
-                        )
+        if backfilled:
+            # Backfilled events come before the current state so we don't need
+            # to update the current state table
+            return
 
-                    self._simple_upsert_txn(
-                        txn,
-                        "current_state_events",
-                        keyvalues={
-                            "room_id": event.room_id,
-                            "type": event.type,
-                            "state_key": event.state_key,
-                        },
-                        values={
-                            "event_id": event.event_id,
-                        }
-                    )
+        for event, _ in state_events_and_contexts:
+            if event.internal_metadata.is_outlier():
+                # Outlier events shouldn't clobber the current state.
+                continue
+
+            if context.rejected:
+                # If the event failed it's auth checks then it shouldn't
+                # clobbler the current state.
+                continue
+
+            txn.call_after(
+                self._get_current_state_for_key.invalidate,
+                (event.room_id, event.type, event.state_key,)
+            )
+
+            if event.type in [EventTypes.Name, EventTypes.Aliases]:
+                txn.call_after(
+                    self.get_room_name_and_aliases.invalidate,
+                    (event.room_id,)
+                )
+
+            self._simple_upsert_txn(
+                txn,
+                "current_state_events",
+                keyvalues={
+                    "room_id": event.room_id,
+                    "type": event.type,
+                    "state_key": event.state_key,
+                },
+                values={
+                    "event_id": event.event_id,
+                }
+            )
 
         return
 
+    def _add_to_cache(self, txn, events_and_contexts):
+        to_prefill = []
+
+        rows = []
+        N = 200
+        for i in range(0, len(events_and_contexts), N):
+            ev_map = {
+                e[0].event_id: e[0]
+                for e in events_and_contexts[i:i + N]
+            }
+            if not ev_map:
+                break
+
+            sql = (
+                "SELECT "
+                " e.event_id as event_id, "
+                " r.redacts as redacts,"
+                " rej.event_id as rejects "
+                " FROM events as e"
+                " LEFT JOIN rejections as rej USING (event_id)"
+                " LEFT JOIN redactions as r ON e.event_id = r.redacts"
+                " WHERE e.event_id IN (%s)"
+            ) % (",".join(["?"] * len(ev_map)),)
+
+            txn.execute(sql, ev_map.keys())
+            rows = self.cursor_to_dict(txn)
+            for row in rows:
+                event = ev_map[row["event_id"]]
+                if not row["rejects"] and not row["redacts"]:
+                    to_prefill.append(_EventCacheEntry(
+                        event=event,
+                        redacted_event=None,
+                    ))
+
+        def prefill():
+            for cache_entry in to_prefill:
+                self._get_event_cache.prefill((cache_entry[0].event_id,), cache_entry)
+        txn.call_after(prefill)
+
     def _store_redaction(self, txn, event):
         # invalidate the cache for the redacted event
         txn.call_after(self._invalidate_get_event_cache, event.redacts)
@@ -499,6 +734,22 @@ class EventsStore(SQLBaseStore):
             (event.event_id, event.redacts)
         )
 
+    @defer.inlineCallbacks
+    def have_events_in_timeline(self, event_ids):
+        """Given a list of event ids, check if we have already processed and
+        stored them as non outliers.
+        """
+        rows = yield self._simple_select_many_batch(
+            table="events",
+            retcols=("event_id",),
+            column="event_id",
+            iterable=list(event_ids),
+            keyvalues={"outlier": False},
+            desc="have_events_in_timeline",
+        )
+
+        defer.returnValue(set(r["event_id"] for r in rows))
+
     def have_events(self, event_ids):
         """Given a list of event ids, check if we have already processed them.
 
@@ -540,100 +791,65 @@ class EventsStore(SQLBaseStore):
         event_id_list = event_ids
         event_ids = set(event_ids)
 
-        event_map = self._get_events_from_cache(
+        event_entry_map = self._get_events_from_cache(
             event_ids,
-            check_redacted=check_redacted,
-            get_prev_content=get_prev_content,
             allow_rejected=allow_rejected,
         )
 
-        missing_events_ids = [e for e in event_ids if e not in event_map]
+        missing_events_ids = [e for e in event_ids if e not in event_entry_map]
 
         if missing_events_ids:
             missing_events = yield self._enqueue_events(
                 missing_events_ids,
                 check_redacted=check_redacted,
-                get_prev_content=get_prev_content,
                 allow_rejected=allow_rejected,
             )
 
-            event_map.update(missing_events)
-
-        defer.returnValue([
-            event_map[e_id] for e_id in event_id_list
-            if e_id in event_map and event_map[e_id]
-        ])
-
-    def _get_events_txn(self, txn, event_ids, check_redacted=True,
-                        get_prev_content=False, allow_rejected=False):
-        if not event_ids:
-            return []
-
-        event_map = self._get_events_from_cache(
-            event_ids,
-            check_redacted=check_redacted,
-            get_prev_content=get_prev_content,
-            allow_rejected=allow_rejected,
-        )
+            event_entry_map.update(missing_events)
 
-        missing_events_ids = [e for e in event_ids if e not in event_map]
+        events = []
+        for event_id in event_id_list:
+            entry = event_entry_map.get(event_id, None)
+            if not entry:
+                continue
 
-        if not missing_events_ids:
-            return [
-                event_map[e_id] for e_id in event_ids
-                if e_id in event_map and event_map[e_id]
-            ]
+            if allow_rejected or not entry.event.rejected_reason:
+                if check_redacted and entry.redacted_event:
+                    event = entry.redacted_event
+                else:
+                    event = entry.event
 
-        missing_events = self._fetch_events_txn(
-            txn,
-            missing_events_ids,
-            check_redacted=check_redacted,
-            get_prev_content=get_prev_content,
-            allow_rejected=allow_rejected,
-        )
+                events.append(event)
 
-        event_map.update(missing_events)
+                if get_prev_content:
+                    if "replaces_state" in event.unsigned:
+                        prev = yield self.get_event(
+                            event.unsigned["replaces_state"],
+                            get_prev_content=False,
+                            allow_none=True,
+                        )
+                        if prev:
+                            event.unsigned = dict(event.unsigned)
+                            event.unsigned["prev_content"] = prev.content
+                            event.unsigned["prev_sender"] = prev.sender
 
-        return [
-            event_map[e_id] for e_id in event_ids
-            if e_id in event_map and event_map[e_id]
-        ]
+        defer.returnValue(events)
 
     def _invalidate_get_event_cache(self, event_id):
-        for check_redacted in (False, True):
-            for get_prev_content in (False, True):
-                self._get_event_cache.invalidate(
-                    (event_id, check_redacted, get_prev_content)
-                )
+            self._get_event_cache.invalidate((event_id,))
 
-    def _get_event_txn(self, txn, event_id, check_redacted=True,
-                       get_prev_content=False, allow_rejected=False):
-
-        events = self._get_events_txn(
-            txn, [event_id],
-            check_redacted=check_redacted,
-            get_prev_content=get_prev_content,
-            allow_rejected=allow_rejected,
-        )
-
-        return events[0] if events else None
-
-    def _get_events_from_cache(self, events, check_redacted, get_prev_content,
-                               allow_rejected):
+    def _get_events_from_cache(self, events, allow_rejected):
         event_map = {}
 
         for event_id in events:
-            try:
-                ret = self._get_event_cache.get(
-                    (event_id, check_redacted, get_prev_content,)
-                )
+            ret = self._get_event_cache.get((event_id,), None)
+            if not ret:
+                continue
 
-                if allow_rejected or not ret.rejected_reason:
-                    event_map[event_id] = ret
-                else:
-                    event_map[event_id] = None
-            except KeyError:
-                pass
+            if allow_rejected or not ret.event.rejected_reason:
+                event_map[event_id] = ret
+            else:
+                event_map[event_id] = None
 
         return event_map
 
@@ -704,8 +920,7 @@ class EventsStore(SQLBaseStore):
                         reactor.callFromThread(fire, event_list)
 
     @defer.inlineCallbacks
-    def _enqueue_events(self, events, check_redacted=True,
-                        get_prev_content=False, allow_rejected=False):
+    def _enqueue_events(self, events, check_redacted=True, allow_rejected=False):
         """Fetches events from the database using the _event_fetch_list. This
         allows batch and bulk fetching of events - it allows us to fetch events
         without having to create a new transaction for each request for events.
@@ -743,8 +958,6 @@ class EventsStore(SQLBaseStore):
             [
                 preserve_fn(self._get_event_from_row)(
                     row["internal_metadata"], row["json"], row["redacts"],
-                    check_redacted=check_redacted,
-                    get_prev_content=get_prev_content,
                     rejected_reason=row["rejects"],
                 )
                 for row in rows
@@ -753,7 +966,7 @@ class EventsStore(SQLBaseStore):
         )
 
         defer.returnValue({
-            e.event_id: e
+            e.event.event_id: e
             for e in res if e
         })
 
@@ -783,37 +996,8 @@ class EventsStore(SQLBaseStore):
 
         return rows
 
-    def _fetch_events_txn(self, txn, events, check_redacted=True,
-                          get_prev_content=False, allow_rejected=False):
-        if not events:
-            return {}
-
-        rows = self._fetch_event_rows(
-            txn, events,
-        )
-
-        if not allow_rejected:
-            rows[:] = [r for r in rows if not r["rejects"]]
-
-        res = [
-            self._get_event_from_row_txn(
-                txn,
-                row["internal_metadata"], row["json"], row["redacts"],
-                check_redacted=check_redacted,
-                get_prev_content=get_prev_content,
-                rejected_reason=row["rejects"],
-            )
-            for row in rows
-        ]
-
-        return {
-            r.event_id: r
-            for r in res
-        }
-
     @defer.inlineCallbacks
     def _get_event_from_row(self, internal_metadata, js, redacted,
-                            check_redacted=True, get_prev_content=False,
                             rejected_reason=None):
         d = json.loads(js)
         internal_metadata = json.loads(internal_metadata)
@@ -823,26 +1007,27 @@ class EventsStore(SQLBaseStore):
                 table="rejections",
                 keyvalues={"event_id": rejected_reason},
                 retcol="reason",
-                desc="_get_event_from_row",
+                desc="_get_event_from_row_rejected_reason",
             )
 
-        ev = FrozenEvent(
+        original_ev = FrozenEvent(
             d,
             internal_metadata_dict=internal_metadata,
             rejected_reason=rejected_reason,
         )
 
-        if check_redacted and redacted:
-            ev = prune_event(ev)
+        redacted_event = None
+        if redacted:
+            redacted_event = prune_event(original_ev)
 
             redaction_id = yield self._simple_select_one_onecol(
                 table="redactions",
-                keyvalues={"redacts": ev.event_id},
+                keyvalues={"redacts": redacted_event.event_id},
                 retcol="event_id",
-                desc="_get_event_from_row",
+                desc="_get_event_from_row_redactions",
             )
 
-            ev.unsigned["redacted_by"] = redaction_id
+            redacted_event.unsigned["redacted_by"] = redaction_id
             # Get the redaction event.
 
             because = yield self.get_event(
@@ -854,86 +1039,16 @@ class EventsStore(SQLBaseStore):
             if because:
                 # It's fine to do add the event directly, since get_pdu_json
                 # will serialise this field correctly
-                ev.unsigned["redacted_because"] = because
-
-        if get_prev_content and "replaces_state" in ev.unsigned:
-            prev = yield self.get_event(
-                ev.unsigned["replaces_state"],
-                get_prev_content=False,
-                allow_none=True,
-            )
-            if prev:
-                ev.unsigned["prev_content"] = prev.content
-                ev.unsigned["prev_sender"] = prev.sender
+                redacted_event.unsigned["redacted_because"] = because
 
-        self._get_event_cache.prefill(
-            (ev.event_id, check_redacted, get_prev_content), ev
+        cache_entry = _EventCacheEntry(
+            event=original_ev,
+            redacted_event=redacted_event,
         )
 
-        defer.returnValue(ev)
+        self._get_event_cache.prefill((original_ev.event_id,), cache_entry)
 
-    def _get_event_from_row_txn(self, txn, internal_metadata, js, redacted,
-                                check_redacted=True, get_prev_content=False,
-                                rejected_reason=None):
-        d = json.loads(js)
-        internal_metadata = json.loads(internal_metadata)
-
-        if rejected_reason:
-            rejected_reason = self._simple_select_one_onecol_txn(
-                txn,
-                table="rejections",
-                keyvalues={"event_id": rejected_reason},
-                retcol="reason",
-            )
-
-        ev = FrozenEvent(
-            d,
-            internal_metadata_dict=internal_metadata,
-            rejected_reason=rejected_reason,
-        )
-
-        if check_redacted and redacted:
-            ev = prune_event(ev)
-
-            redaction_id = self._simple_select_one_onecol_txn(
-                txn,
-                table="redactions",
-                keyvalues={"redacts": ev.event_id},
-                retcol="event_id",
-            )
-
-            ev.unsigned["redacted_by"] = redaction_id
-            # Get the redaction event.
-
-            because = self._get_event_txn(
-                txn,
-                redaction_id,
-                check_redacted=False
-            )
-
-            if because:
-                ev.unsigned["redacted_because"] = because
-
-        if get_prev_content and "replaces_state" in ev.unsigned:
-            prev = self._get_event_txn(
-                txn,
-                ev.unsigned["replaces_state"],
-                get_prev_content=False,
-            )
-            if prev:
-                ev.unsigned["prev_content"] = prev.content
-                ev.unsigned["prev_sender"] = prev.sender
-
-        self._get_event_cache.prefill(
-            (ev.event_id, check_redacted, get_prev_content), ev
-        )
-
-        return ev
-
-    def _parse_events_txn(self, txn, rows):
-        event_ids = [r["event_id"] for r in rows]
-
-        return self._get_events_txn(txn, event_ids)
+        defer.returnValue(cache_entry)
 
     @defer.inlineCallbacks
     def count_daily_messages(self):
@@ -1076,45 +1191,106 @@ class EventsStore(SQLBaseStore):
 
     def get_current_backfill_token(self):
         """The current minimum token that backfilled events have reached"""
-
-        # TODO: Fix race with the persit_event txn by using one of the
-        # stream id managers
-        return -self.min_stream_token
+        return -self._backfill_id_gen.get_current_token()
 
     def get_all_new_events(self, last_backfill_id, last_forward_id,
                            current_backfill_id, current_forward_id, limit):
         """Get all the new events that have arrived at the server either as
         new events or as backfilled events"""
+        have_backfill_events = last_backfill_id != current_backfill_id
+        have_forward_events = last_forward_id != current_forward_id
+
+        if not have_backfill_events and not have_forward_events:
+            return defer.succeed(AllNewEventsResult([], [], [], [], []))
+
         def get_all_new_events_txn(txn):
             sql = (
-                "SELECT e.stream_ordering, ej.internal_metadata, ej.json"
+                "SELECT e.stream_ordering, ej.internal_metadata, ej.json, eg.state_group"
                 " FROM events as e"
                 " JOIN event_json as ej"
                 " ON e.event_id = ej.event_id AND e.room_id = ej.room_id"
+                " LEFT JOIN event_to_state_groups as eg"
+                " ON e.event_id = eg.event_id"
                 " WHERE ? < e.stream_ordering AND e.stream_ordering <= ?"
                 " ORDER BY e.stream_ordering ASC"
                 " LIMIT ?"
             )
-            if last_forward_id != current_forward_id:
+            if have_forward_events:
                 txn.execute(sql, (last_forward_id, current_forward_id, limit))
                 new_forward_events = txn.fetchall()
+
+                if len(new_forward_events) == limit:
+                    upper_bound = new_forward_events[-1][0]
+                else:
+                    upper_bound = current_forward_id
+
+                sql = (
+                    "SELECT event_stream_ordering FROM current_state_resets"
+                    " WHERE ? < event_stream_ordering"
+                    " AND event_stream_ordering <= ?"
+                    " ORDER BY event_stream_ordering ASC"
+                )
+                txn.execute(sql, (last_forward_id, upper_bound))
+                state_resets = txn.fetchall()
+
+                sql = (
+                    "SELECT event_stream_ordering, event_id, state_group"
+                    " FROM ex_outlier_stream"
+                    " WHERE ? > event_stream_ordering"
+                    " AND event_stream_ordering >= ?"
+                    " ORDER BY event_stream_ordering DESC"
+                )
+                txn.execute(sql, (last_forward_id, upper_bound))
+                forward_ex_outliers = txn.fetchall()
             else:
                 new_forward_events = []
+                state_resets = []
+                forward_ex_outliers = []
 
             sql = (
-                "SELECT -e.stream_ordering, ej.internal_metadata, ej.json"
+                "SELECT -e.stream_ordering, ej.internal_metadata, ej.json,"
+                " eg.state_group"
                 " FROM events as e"
                 " JOIN event_json as ej"
                 " ON e.event_id = ej.event_id AND e.room_id = ej.room_id"
+                " LEFT JOIN event_to_state_groups as eg"
+                " ON e.event_id = eg.event_id"
                 " WHERE ? > e.stream_ordering AND e.stream_ordering >= ?"
                 " ORDER BY e.stream_ordering DESC"
                 " LIMIT ?"
             )
-            if last_backfill_id != current_backfill_id:
+            if have_backfill_events:
                 txn.execute(sql, (-last_backfill_id, -current_backfill_id, limit))
                 new_backfill_events = txn.fetchall()
+
+                if len(new_backfill_events) == limit:
+                    upper_bound = new_backfill_events[-1][0]
+                else:
+                    upper_bound = current_backfill_id
+
+                sql = (
+                    "SELECT -event_stream_ordering, event_id, state_group"
+                    " FROM ex_outlier_stream"
+                    " WHERE ? > event_stream_ordering"
+                    " AND event_stream_ordering >= ?"
+                    " ORDER BY event_stream_ordering DESC"
+                )
+                txn.execute(sql, (-last_backfill_id, -upper_bound))
+                backward_ex_outliers = txn.fetchall()
             else:
                 new_backfill_events = []
+                backward_ex_outliers = []
 
-            return (new_forward_events, new_backfill_events)
+            return AllNewEventsResult(
+                new_forward_events, new_backfill_events,
+                forward_ex_outliers, backward_ex_outliers,
+                state_resets,
+            )
         return self.runInteraction("get_all_new_events", get_all_new_events_txn)
+
+
+AllNewEventsResult = namedtuple("AllNewEventsResult", [
+    "new_forward_events", "new_backfill_events",
+    "forward_ex_outliers", "backward_ex_outliers",
+    "state_resets"
+])