Merge pull request #3673 from matrix-org/erikj/refactor_state_handler

Refactor state module to support multiple room versions

2 files changed, 971 insertions, 0 deletions

diff --git a/synapse/state/__init__.py b/synapse/state/__init__.py
new file mode 100644
index 0000000000..b34970e4d1
--- /dev/null
+++ b/synapse/state/__init__.py
@@ -0,0 +1,650 @@
+# -*- coding: utf-8 -*-
+# Copyright 2014-2016 OpenMarket Ltd
+# Copyright 2018 New Vector 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 collections import namedtuple
+
+from six import iteritems, itervalues
+
+from frozendict import frozendict
+
+from twisted.internet import defer
+
+from synapse.api.constants import EventTypes, RoomVersions
+from synapse.events.snapshot import EventContext
+from synapse.state import v1
+from synapse.util.async_helpers import Linearizer
+from synapse.util.caches import get_cache_factor_for
+from synapse.util.caches.expiringcache import ExpiringCache
+from synapse.util.logutils import log_function
+from synapse.util.metrics import Measure
+
+logger = logging.getLogger(__name__)
+
+
+KeyStateTuple = namedtuple("KeyStateTuple", ("context", "type", "state_key"))
+
+
+SIZE_OF_CACHE = 100000 * get_cache_factor_for("state_cache")
+EVICTION_TIMEOUT_SECONDS = 60 * 60
+
+
+_NEXT_STATE_ID = 1
+
+POWER_KEY = (EventTypes.PowerLevels, "")
+
+
+def _gen_state_id():
+    global _NEXT_STATE_ID
+    s = "X%d" % (_NEXT_STATE_ID,)
+    _NEXT_STATE_ID += 1
+    return s
+
+
+class _StateCacheEntry(object):
+    __slots__ = ["state", "state_group", "state_id", "prev_group", "delta_ids"]
+
+    def __init__(self, state, state_group, prev_group=None, delta_ids=None):
+        # dict[(str, str), str] map  from (type, state_key) to event_id
+        self.state = frozendict(state)
+
+        # the ID of a state group if one and only one is involved.
+        # otherwise, None otherwise?
+        self.state_group = state_group
+
+        self.prev_group = prev_group
+        self.delta_ids = frozendict(delta_ids) if delta_ids is not None else None
+
+        # The `state_id` is a unique ID we generate that can be used as ID for
+        # this collection of state. Usually this would be the same as the
+        # state group, but on worker instances we can't generate a new state
+        # group each time we resolve state, so we generate a separate one that
+        # isn't persisted and is used solely for caches.
+        # `state_id` is either a state_group (and so an int) or a string. This
+        # ensures we don't accidentally persist a state_id as a stateg_group
+        if state_group:
+            self.state_id = state_group
+        else:
+            self.state_id = _gen_state_id()
+
+    def __len__(self):
+        return len(self.state)
+
+
+class StateHandler(object):
+    """Fetches bits of state from the stores, and does state resolution
+    where necessary
+    """
+
+    def __init__(self, hs):
+        self.clock = hs.get_clock()
+        self.store = hs.get_datastore()
+        self.hs = hs
+        self._state_resolution_handler = hs.get_state_resolution_handler()
+
+    def start_caching(self):
+        # TODO: remove this shim
+        self._state_resolution_handler.start_caching()
+
+    @defer.inlineCallbacks
+    def get_current_state(self, room_id, event_type=None, state_key="",
+                          latest_event_ids=None):
+        """ Retrieves the current state for the room. This is done by
+        calling `get_latest_events_in_room` to get the leading edges of the
+        event graph and then resolving any of the state conflicts.
+
+        This is equivalent to getting the state of an event that were to send
+        next before receiving any new events.
+
+        If `event_type` is specified, then the method returns only the one
+        event (or None) with that `event_type` and `state_key`.
+
+        Returns:
+            map from (type, state_key) to event
+        """
+        if not latest_event_ids:
+            latest_event_ids = yield self.store.get_latest_event_ids_in_room(room_id)
+
+        logger.debug("calling resolve_state_groups from get_current_state")
+        ret = yield self.resolve_state_groups_for_events(room_id, latest_event_ids)
+        state = ret.state
+
+        if event_type:
+            event_id = state.get((event_type, state_key))
+            event = None
+            if event_id:
+                event = yield self.store.get_event(event_id, allow_none=True)
+            defer.returnValue(event)
+            return
+
+        state_map = yield self.store.get_events(list(state.values()),
+                                                get_prev_content=False)
+        state = {
+            key: state_map[e_id] for key, e_id in iteritems(state) if e_id in state_map
+        }
+
+        defer.returnValue(state)
+
+    @defer.inlineCallbacks
+    def get_current_state_ids(self, room_id, latest_event_ids=None):
+        """Get the current state, or the state at a set of events, for a room
+
+        Args:
+            room_id (str):
+
+            latest_event_ids (iterable[str]|None): if given, the forward
+                extremities to resolve. If None, we look them up from the
+                database (via a cache)
+
+        Returns:
+            Deferred[dict[(str, str), str)]]: the state dict, mapping from
+                (event_type, state_key) -> event_id
+        """
+        if not latest_event_ids:
+            latest_event_ids = yield self.store.get_latest_event_ids_in_room(room_id)
+
+        logger.debug("calling resolve_state_groups from get_current_state_ids")
+        ret = yield self.resolve_state_groups_for_events(room_id, latest_event_ids)
+        state = ret.state
+
+        defer.returnValue(state)
+
+    @defer.inlineCallbacks
+    def get_current_user_in_room(self, room_id, latest_event_ids=None):
+        if not latest_event_ids:
+            latest_event_ids = yield self.store.get_latest_event_ids_in_room(room_id)
+        logger.debug("calling resolve_state_groups from get_current_user_in_room")
+        entry = yield self.resolve_state_groups_for_events(room_id, latest_event_ids)
+        joined_users = yield self.store.get_joined_users_from_state(room_id, entry)
+        defer.returnValue(joined_users)
+
+    @defer.inlineCallbacks
+    def get_current_hosts_in_room(self, room_id, latest_event_ids=None):
+        if not latest_event_ids:
+            latest_event_ids = yield self.store.get_latest_event_ids_in_room(room_id)
+        logger.debug("calling resolve_state_groups from get_current_hosts_in_room")
+        entry = yield self.resolve_state_groups_for_events(room_id, latest_event_ids)
+        joined_hosts = yield self.store.get_joined_hosts(room_id, entry)
+        defer.returnValue(joined_hosts)
+
+    @defer.inlineCallbacks
+    def compute_event_context(self, event, old_state=None):
+        """Build an EventContext structure for the event.
+
+        This works out what the current state should be for the event, and
+        generates a new state group if necessary.
+
+        Args:
+            event (synapse.events.EventBase):
+            old_state (dict|None): The state at the event if it can't be
+                calculated from existing events. This is normally only specified
+                when receiving an event from federation where we don't have the
+                prev events for, e.g. when backfilling.
+        Returns:
+            synapse.events.snapshot.EventContext:
+        """
+
+        if event.internal_metadata.is_outlier():
+            # If this is an outlier, then we know it shouldn't have any current
+            # state. Certainly store.get_current_state won't return any, and
+            # persisting the event won't store the state group.
+            if old_state:
+                prev_state_ids = {
+                    (s.type, s.state_key): s.event_id for s in old_state
+                }
+                if event.is_state():
+                    current_state_ids = dict(prev_state_ids)
+                    key = (event.type, event.state_key)
+                    current_state_ids[key] = event.event_id
+                else:
+                    current_state_ids = prev_state_ids
+            else:
+                current_state_ids = {}
+                prev_state_ids = {}
+
+            # We don't store state for outliers, so we don't generate a state
+            # group for it.
+            context = EventContext.with_state(
+                state_group=None,
+                current_state_ids=current_state_ids,
+                prev_state_ids=prev_state_ids,
+            )
+
+            defer.returnValue(context)
+
+        if old_state:
+            # We already have the state, so we don't need to calculate it.
+            # Let's just correctly fill out the context and create a
+            # new state group for it.
+
+            prev_state_ids = {
+                (s.type, s.state_key): s.event_id for s in old_state
+            }
+
+            if event.is_state():
+                key = (event.type, event.state_key)
+                if key in prev_state_ids:
+                    replaces = prev_state_ids[key]
+                    if replaces != event.event_id:  # Paranoia check
+                        event.unsigned["replaces_state"] = replaces
+                current_state_ids = dict(prev_state_ids)
+                current_state_ids[key] = event.event_id
+            else:
+                current_state_ids = prev_state_ids
+
+            state_group = yield self.store.store_state_group(
+                event.event_id,
+                event.room_id,
+                prev_group=None,
+                delta_ids=None,
+                current_state_ids=current_state_ids,
+            )
+
+            context = EventContext.with_state(
+                state_group=state_group,
+                current_state_ids=current_state_ids,
+                prev_state_ids=prev_state_ids,
+            )
+
+            defer.returnValue(context)
+
+        logger.debug("calling resolve_state_groups from compute_event_context")
+
+        entry = yield self.resolve_state_groups_for_events(
+            event.room_id, [e for e, _ in event.prev_events],
+        )
+
+        prev_state_ids = entry.state
+        prev_group = None
+        delta_ids = None
+
+        if event.is_state():
+            # If this is a state event then we need to create a new state
+            # group for the state after this event.
+
+            key = (event.type, event.state_key)
+            if key in prev_state_ids:
+                replaces = prev_state_ids[key]
+                event.unsigned["replaces_state"] = replaces
+
+            current_state_ids = dict(prev_state_ids)
+            current_state_ids[key] = event.event_id
+
+            if entry.state_group:
+                # If the state at the event has a state group assigned then
+                # we can use that as the prev group
+                prev_group = entry.state_group
+                delta_ids = {
+                    key: event.event_id
+                }
+            elif entry.prev_group:
+                # If the state at the event only has a prev group, then we can
+                # use that as a prev group too.
+                prev_group = entry.prev_group
+                delta_ids = dict(entry.delta_ids)
+                delta_ids[key] = event.event_id
+
+            state_group = yield self.store.store_state_group(
+                event.event_id,
+                event.room_id,
+                prev_group=prev_group,
+                delta_ids=delta_ids,
+                current_state_ids=current_state_ids,
+            )
+        else:
+            current_state_ids = prev_state_ids
+            prev_group = entry.prev_group
+            delta_ids = entry.delta_ids
+
+            if entry.state_group is None:
+                entry.state_group = yield self.store.store_state_group(
+                    event.event_id,
+                    event.room_id,
+                    prev_group=entry.prev_group,
+                    delta_ids=entry.delta_ids,
+                    current_state_ids=current_state_ids,
+                )
+                entry.state_id = entry.state_group
+
+            state_group = entry.state_group
+
+        context = EventContext.with_state(
+            state_group=state_group,
+            current_state_ids=current_state_ids,
+            prev_state_ids=prev_state_ids,
+            prev_group=prev_group,
+            delta_ids=delta_ids,
+        )
+
+        defer.returnValue(context)
+
+    @defer.inlineCallbacks
+    def resolve_state_groups_for_events(self, room_id, event_ids):
+        """ Given a list of event_ids this method fetches the state at each
+        event, resolves conflicts between them and returns them.
+
+        Args:
+            room_id (str)
+            event_ids (list[str])
+            explicit_room_version (str|None): If set uses the the given room
+                version to choose the resolution algorithm. If None, then
+                checks the database for room version.
+
+        Returns:
+            Deferred[_StateCacheEntry]: resolved state
+        """
+        logger.debug("resolve_state_groups event_ids %s", event_ids)
+
+        # map from state group id to the state in that state group (where
+        # 'state' is a map from state key to event id)
+        # dict[int, dict[(str, str), str]]
+        state_groups_ids = yield self.store.get_state_groups_ids(
+            room_id, event_ids
+        )
+
+        if len(state_groups_ids) == 0:
+            defer.returnValue(_StateCacheEntry(
+                state={},
+                state_group=None,
+            ))
+        elif len(state_groups_ids) == 1:
+            name, state_list = list(state_groups_ids.items()).pop()
+
+            prev_group, delta_ids = yield self.store.get_state_group_delta(name)
+
+            defer.returnValue(_StateCacheEntry(
+                state=state_list,
+                state_group=name,
+                prev_group=prev_group,
+                delta_ids=delta_ids,
+            ))
+
+        room_version = yield self.store.get_room_version(room_id)
+
+        result = yield self._state_resolution_handler.resolve_state_groups(
+            room_id, room_version, state_groups_ids, None,
+            self._state_map_factory,
+        )
+        defer.returnValue(result)
+
+    def _state_map_factory(self, ev_ids):
+        return self.store.get_events(
+            ev_ids, get_prev_content=False, check_redacted=False,
+        )
+
+    def resolve_events(self, room_version, state_sets, event):
+        logger.info(
+            "Resolving state for %s with %d groups", event.room_id, len(state_sets)
+        )
+        state_set_ids = [{
+            (ev.type, ev.state_key): ev.event_id
+            for ev in st
+        } for st in state_sets]
+
+        state_map = {
+            ev.event_id: ev
+            for st in state_sets
+            for ev in st
+        }
+
+        with Measure(self.clock, "state._resolve_events"):
+            new_state = resolve_events_with_state_map(
+                room_version, state_set_ids, state_map,
+            )
+
+        new_state = {
+            key: state_map[ev_id] for key, ev_id in iteritems(new_state)
+        }
+
+        return new_state
+
+
+class StateResolutionHandler(object):
+    """Responsible for doing state conflict resolution.
+
+    Note that the storage layer depends on this handler, so all functions must
+    be storage-independent.
+    """
+    def __init__(self, hs):
+        self.clock = hs.get_clock()
+
+        # dict of set of event_ids -> _StateCacheEntry.
+        self._state_cache = None
+        self.resolve_linearizer = Linearizer(name="state_resolve_lock")
+
+    def start_caching(self):
+        logger.debug("start_caching")
+
+        self._state_cache = ExpiringCache(
+            cache_name="state_cache",
+            clock=self.clock,
+            max_len=SIZE_OF_CACHE,
+            expiry_ms=EVICTION_TIMEOUT_SECONDS * 1000,
+            iterable=True,
+            reset_expiry_on_get=True,
+        )
+
+        self._state_cache.start()
+
+    @defer.inlineCallbacks
+    @log_function
+    def resolve_state_groups(
+        self, room_id, room_version, state_groups_ids, event_map, state_map_factory,
+    ):
+        """Resolves conflicts between a set of state groups
+
+        Always generates a new state group (unless we hit the cache), so should
+        not be called for a single state group
+
+        Args:
+            room_id (str): room we are resolving for (used for logging)
+            room_version (str): version of the room
+            state_groups_ids (dict[int, dict[(str, str), str]]):
+                 map from state group id to the state in that state group
+                (where 'state' is a map from state key to event id)
+
+            event_map(dict[str,FrozenEvent]|None):
+                a dict from event_id to event, for any events that we happen to
+                have in flight (eg, those currently being persisted). This will be
+                used as a starting point fof finding the state we need; any missing
+                events will be requested via state_map_factory.
+
+                If None, all events will be fetched via state_map_factory.
+
+        Returns:
+            Deferred[_StateCacheEntry]: resolved state
+        """
+        logger.debug(
+            "resolve_state_groups state_groups %s",
+            state_groups_ids.keys()
+        )
+
+        group_names = frozenset(state_groups_ids.keys())
+
+        with (yield self.resolve_linearizer.queue(group_names)):
+            if self._state_cache is not None:
+                cache = self._state_cache.get(group_names, None)
+                if cache:
+                    defer.returnValue(cache)
+
+            logger.info(
+                "Resolving state for %s with %d groups", room_id, len(state_groups_ids)
+            )
+
+            # start by assuming we won't have any conflicted state, and build up the new
+            # state map by iterating through the state groups. If we discover a conflict,
+            # we give up and instead use `resolve_events_with_factory`.
+            #
+            # XXX: is this actually worthwhile, or should we just let
+            # resolve_events_with_factory do it?
+            new_state = {}
+            conflicted_state = False
+            for st in itervalues(state_groups_ids):
+                for key, e_id in iteritems(st):
+                    if key in new_state:
+                        conflicted_state = True
+                        break
+                    new_state[key] = e_id
+                if conflicted_state:
+                    break
+
+            if conflicted_state:
+                logger.info("Resolving conflicted state for %r", room_id)
+                with Measure(self.clock, "state._resolve_events"):
+                    new_state = yield resolve_events_with_factory(
+                        room_version,
+                        list(itervalues(state_groups_ids)),
+                        event_map=event_map,
+                        state_map_factory=state_map_factory,
+                    )
+
+            # if the new state matches any of the input state groups, we can
+            # use that state group again. Otherwise we will generate a state_id
+            # which will be used as a cache key for future resolutions, but
+            # not get persisted.
+
+            with Measure(self.clock, "state.create_group_ids"):
+                cache = _make_state_cache_entry(new_state, state_groups_ids)
+
+            if self._state_cache is not None:
+                self._state_cache[group_names] = cache
+
+            defer.returnValue(cache)
+
+
+def _make_state_cache_entry(
+    new_state,
+    state_groups_ids,
+):
+    """Given a resolved state, and a set of input state groups, pick one to base
+    a new state group on (if any), and return an appropriately-constructed
+    _StateCacheEntry.
+
+    Args:
+        new_state (dict[(str, str), str]): resolved state map (mapping from
+           (type, state_key) to event_id)
+
+        state_groups_ids (dict[int, dict[(str, str), str]]):
+                 map from state group id to the state in that state group
+                (where 'state' is a map from state key to event id)
+
+    Returns:
+        _StateCacheEntry
+    """
+    # if the new state matches any of the input state groups, we can
+    # use that state group again. Otherwise we will generate a state_id
+    # which will be used as a cache key for future resolutions, but
+    # not get persisted.
+
+    # first look for exact matches
+    new_state_event_ids = set(itervalues(new_state))
+    for sg, state in iteritems(state_groups_ids):
+        if len(new_state_event_ids) != len(state):
+            continue
+
+        old_state_event_ids = set(itervalues(state))
+        if new_state_event_ids == old_state_event_ids:
+            # got an exact match.
+            return _StateCacheEntry(
+                state=new_state,
+                state_group=sg,
+            )
+
+    # TODO: We want to create a state group for this set of events, to
+    # increase cache hits, but we need to make sure that it doesn't
+    # end up as a prev_group without being added to the database
+
+    # failing that, look for the closest match.
+    prev_group = None
+    delta_ids = None
+
+    for old_group, old_state in iteritems(state_groups_ids):
+        n_delta_ids = {
+            k: v
+            for k, v in iteritems(new_state)
+            if old_state.get(k) != v
+        }
+        if not delta_ids or len(n_delta_ids) < len(delta_ids):
+            prev_group = old_group
+            delta_ids = n_delta_ids
+
+    return _StateCacheEntry(
+        state=new_state,
+        state_group=None,
+        prev_group=prev_group,
+        delta_ids=delta_ids,
+    )
+
+
+def resolve_events_with_state_map(room_version, state_sets, state_map):
+    """
+    Args:
+        room_version(str): Version of the room
+        state_sets(list): List of dicts of (type, state_key) -> event_id,
+            which are the different state groups to resolve.
+        state_map(dict): a dict from event_id to event, for all events in
+            state_sets.
+
+    Returns
+        dict[(str, str), str]:
+            a map from (type, state_key) to event_id.
+    """
+    if room_version in (RoomVersions.V1, RoomVersions.VDH_TEST,):
+        return v1.resolve_events_with_state_map(
+            state_sets, state_map,
+        )
+    else:
+        # This should only happen if we added a version but forgot to add it to
+        # the list above.
+        raise Exception(
+            "No state resolution algorithm defined for version %r" % (room_version,)
+        )
+
+
+def resolve_events_with_factory(room_version, state_sets, event_map, state_map_factory):
+    """
+    Args:
+        room_version(str): Version of the room
+
+        state_sets(list): List of dicts of (type, state_key) -> event_id,
+            which are the different state groups to resolve.
+
+        event_map(dict[str,FrozenEvent]|None):
+            a dict from event_id to event, for any events that we happen to
+            have in flight (eg, those currently being persisted). This will be
+            used as a starting point fof finding the state we need; any missing
+            events will be requested via state_map_factory.
+
+            If None, all events will be fetched via state_map_factory.
+
+        state_map_factory(func): will be called
+            with a list of event_ids that are needed, and should return with
+            a Deferred of dict of event_id to event.
+
+    Returns
+        Deferred[dict[(str, str), str]]:
+            a map from (type, state_key) to event_id.
+    """
+    if room_version in (RoomVersions.V1, RoomVersions.VDH_TEST,):
+        return v1.resolve_events_with_factory(
+            state_sets, event_map, state_map_factory,
+        )
+    else:
+        # This should only happen if we added a version but forgot to add it to
+        # the list above.
+        raise Exception(
+            "No state resolution algorithm defined for version %r" % (room_version,)
+        )
diff --git a/synapse/state/v1.py b/synapse/state/v1.py
new file mode 100644
index 0000000000..3a1f7054a1
--- /dev/null
+++ b/synapse/state/v1.py
@@ -0,0 +1,321 @@
+# -*- coding: utf-8 -*-
+# Copyright 2018 New Vector 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 hashlib
+import logging
+
+from six import iteritems, iterkeys, itervalues
+
+from twisted.internet import defer
+
+from synapse import event_auth
+from synapse.api.constants import EventTypes
+from synapse.api.errors import AuthError
+
+logger = logging.getLogger(__name__)
+
+
+POWER_KEY = (EventTypes.PowerLevels, "")
+
+
+def resolve_events_with_state_map(state_sets, state_map):
+    """
+    Args:
+        state_sets(list): List of dicts of (type, state_key) -> event_id,
+            which are the different state groups to resolve.
+        state_map(dict): a dict from event_id to event, for all events in
+            state_sets.
+
+    Returns
+        dict[(str, str), str]:
+            a map from (type, state_key) to event_id.
+    """
+    if len(state_sets) == 1:
+        return state_sets[0]
+
+    unconflicted_state, conflicted_state = _seperate(
+        state_sets,
+    )
+
+    auth_events = _create_auth_events_from_maps(
+        unconflicted_state, conflicted_state, state_map
+    )
+
+    return _resolve_with_state(
+        unconflicted_state, conflicted_state, auth_events, state_map
+    )
+
+
+@defer.inlineCallbacks
+def resolve_events_with_factory(state_sets, event_map, state_map_factory):
+    """
+    Args:
+        state_sets(list): List of dicts of (type, state_key) -> event_id,
+            which are the different state groups to resolve.
+
+        event_map(dict[str,FrozenEvent]|None):
+            a dict from event_id to event, for any events that we happen to
+            have in flight (eg, those currently being persisted). This will be
+            used as a starting point fof finding the state we need; any missing
+            events will be requested via state_map_factory.
+
+            If None, all events will be fetched via state_map_factory.
+
+        state_map_factory(func): will be called
+            with a list of event_ids that are needed, and should return with
+            a Deferred of dict of event_id to event.
+
+    Returns
+        Deferred[dict[(str, str), str]]:
+            a map from (type, state_key) to event_id.
+    """
+    if len(state_sets) == 1:
+        defer.returnValue(state_sets[0])
+
+    unconflicted_state, conflicted_state = _seperate(
+        state_sets,
+    )
+
+    needed_events = set(
+        event_id
+        for event_ids in itervalues(conflicted_state)
+        for event_id in event_ids
+    )
+    if event_map is not None:
+        needed_events -= set(iterkeys(event_map))
+
+    logger.info("Asking for %d conflicted events", len(needed_events))
+
+    # dict[str, FrozenEvent]: a map from state event id to event. Only includes
+    # the state events which are in conflict (and those in event_map)
+    state_map = yield state_map_factory(needed_events)
+    if event_map is not None:
+        state_map.update(event_map)
+
+    # get the ids of the auth events which allow us to authenticate the
+    # conflicted state, picking only from the unconflicting state.
+    #
+    # dict[(str, str), str]: a map from state key to event id
+    auth_events = _create_auth_events_from_maps(
+        unconflicted_state, conflicted_state, state_map
+    )
+
+    new_needed_events = set(itervalues(auth_events))
+    new_needed_events -= needed_events
+    if event_map is not None:
+        new_needed_events -= set(iterkeys(event_map))
+
+    logger.info("Asking for %d auth events", len(new_needed_events))
+
+    state_map_new = yield state_map_factory(new_needed_events)
+    state_map.update(state_map_new)
+
+    defer.returnValue(_resolve_with_state(
+        unconflicted_state, conflicted_state, auth_events, state_map
+    ))
+
+
+def _seperate(state_sets):
+    """Takes the state_sets and figures out which keys are conflicted and
+    which aren't. i.e., which have multiple different event_ids associated
+    with them in different state sets.
+
+    Args:
+        state_sets(iterable[dict[(str, str), str]]):
+            List of dicts of (type, state_key) -> event_id, which are the
+            different state groups to resolve.
+
+    Returns:
+        (dict[(str, str), str], dict[(str, str), set[str]]):
+            A tuple of (unconflicted_state, conflicted_state), where:
+
+            unconflicted_state is a dict mapping (type, state_key)->event_id
+            for unconflicted state keys.
+
+            conflicted_state is a dict mapping (type, state_key) to a set of
+            event ids for conflicted state keys.
+    """
+    state_set_iterator = iter(state_sets)
+    unconflicted_state = dict(next(state_set_iterator))
+    conflicted_state = {}
+
+    for state_set in state_set_iterator:
+        for key, value in iteritems(state_set):
+            # Check if there is an unconflicted entry for the state key.
+            unconflicted_value = unconflicted_state.get(key)
+            if unconflicted_value is None:
+                # There isn't an unconflicted entry so check if there is a
+                # conflicted entry.
+                ls = conflicted_state.get(key)
+                if ls is None:
+                    # There wasn't a conflicted entry so haven't seen this key before.
+                    # Therefore it isn't conflicted yet.
+                    unconflicted_state[key] = value
+                else:
+                    # This key is already conflicted, add our value to the conflict set.
+                    ls.add(value)
+            elif unconflicted_value != value:
+                # If the unconflicted value is not the same as our value then we
+                # have a new conflict. So move the key from the unconflicted_state
+                # to the conflicted state.
+                conflicted_state[key] = {value, unconflicted_value}
+                unconflicted_state.pop(key, None)
+
+    return unconflicted_state, conflicted_state
+
+
+def _create_auth_events_from_maps(unconflicted_state, conflicted_state, state_map):
+    auth_events = {}
+    for event_ids in itervalues(conflicted_state):
+        for event_id in event_ids:
+            if event_id in state_map:
+                keys = event_auth.auth_types_for_event(state_map[event_id])
+                for key in keys:
+                    if key not in auth_events:
+                        event_id = unconflicted_state.get(key, None)
+                        if event_id:
+                            auth_events[key] = event_id
+    return auth_events
+
+
+def _resolve_with_state(unconflicted_state_ids, conflicted_state_ids, auth_event_ids,
+                        state_map):
+    conflicted_state = {}
+    for key, event_ids in iteritems(conflicted_state_ids):
+        events = [state_map[ev_id] for ev_id in event_ids if ev_id in state_map]
+        if len(events) > 1:
+            conflicted_state[key] = events
+        elif len(events) == 1:
+            unconflicted_state_ids[key] = events[0].event_id
+
+    auth_events = {
+        key: state_map[ev_id]
+        for key, ev_id in iteritems(auth_event_ids)
+        if ev_id in state_map
+    }
+
+    try:
+        resolved_state = _resolve_state_events(
+            conflicted_state, auth_events
+        )
+    except Exception:
+        logger.exception("Failed to resolve state")
+        raise
+
+    new_state = unconflicted_state_ids
+    for key, event in iteritems(resolved_state):
+        new_state[key] = event.event_id
+
+    return new_state
+
+
+def _resolve_state_events(conflicted_state, auth_events):
+    """ This is where we actually decide which of the conflicted state to
+    use.
+
+    We resolve conflicts in the following order:
+        1. power levels
+        2. join rules
+        3. memberships
+        4. other events.
+    """
+    resolved_state = {}
+    if POWER_KEY in conflicted_state:
+        events = conflicted_state[POWER_KEY]
+        logger.debug("Resolving conflicted power levels %r", events)
+        resolved_state[POWER_KEY] = _resolve_auth_events(
+            events, auth_events)
+
+    auth_events.update(resolved_state)
+
+    for key, events in iteritems(conflicted_state):
+        if key[0] == EventTypes.JoinRules:
+            logger.debug("Resolving conflicted join rules %r", events)
+            resolved_state[key] = _resolve_auth_events(
+                events,
+                auth_events
+            )
+
+    auth_events.update(resolved_state)
+
+    for key, events in iteritems(conflicted_state):
+        if key[0] == EventTypes.Member:
+            logger.debug("Resolving conflicted member lists %r", events)
+            resolved_state[key] = _resolve_auth_events(
+                events,
+                auth_events
+            )
+
+    auth_events.update(resolved_state)
+
+    for key, events in iteritems(conflicted_state):
+        if key not in resolved_state:
+            logger.debug("Resolving conflicted state %r:%r", key, events)
+            resolved_state[key] = _resolve_normal_events(
+                events, auth_events
+            )
+
+    return resolved_state
+
+
+def _resolve_auth_events(events, auth_events):
+    reverse = [i for i in reversed(_ordered_events(events))]
+
+    auth_keys = set(
+        key
+        for event in events
+        for key in event_auth.auth_types_for_event(event)
+    )
+
+    new_auth_events = {}
+    for key in auth_keys:
+        auth_event = auth_events.get(key, None)
+        if auth_event:
+            new_auth_events[key] = auth_event
+
+    auth_events = new_auth_events
+
+    prev_event = reverse[0]
+    for event in reverse[1:]:
+        auth_events[(prev_event.type, prev_event.state_key)] = prev_event
+        try:
+            # The signatures have already been checked at this point
+            event_auth.check(event, auth_events, do_sig_check=False, do_size_check=False)
+            prev_event = event
+        except AuthError:
+            return prev_event
+
+    return event
+
+
+def _resolve_normal_events(events, auth_events):
+    for event in _ordered_events(events):
+        try:
+            # The signatures have already been checked at this point
+            event_auth.check(event, auth_events, do_sig_check=False, do_size_check=False)
+            return event
+        except AuthError:
+            pass
+
+    # Use the last event (the one with the least depth) if they all fail
+    # the auth check.
+    return event
+
+
+def _ordered_events(events):
+    def key_func(e):
+        return -int(e.depth), hashlib.sha1(e.event_id.encode('ascii')).hexdigest()
+
+    return sorted(events, key=key_func)