matrix.org hotfixes: Back out in-flight state cache changes (#12117)

4 files changed, 25 insertions, 548 deletions

diff --git a/changelog.d/10870.misc b/changelog.d/10870.misc
deleted file mode 100644
index 3af049b969..0000000000
--- a/changelog.d/10870.misc
+++ /dev/null
@@ -1 +0,0 @@
-Deduplicate in-flight requests in `_get_state_for_groups`.
diff --git a/changelog.d/11608.misc b/changelog.d/11608.misc
deleted file mode 100644
index 3af049b969..0000000000
--- a/changelog.d/11608.misc
+++ /dev/null
@@ -1 +0,0 @@
-Deduplicate in-flight requests in `_get_state_for_groups`.
diff --git a/synapse/storage/databases/state/store.py b/synapse/storage/databases/state/store.py
index b8016f679a..7614d76ac6 100644
--- a/synapse/storage/databases/state/store.py
+++ b/synapse/storage/databases/state/store.py
@@ -13,23 +13,11 @@
 # limitations under the License.
 
 import logging
-from typing import (
-    TYPE_CHECKING,
-    Collection,
-    Dict,
-    Iterable,
-    Optional,
-    Sequence,
-    Set,
-    Tuple,
-)
+from typing import TYPE_CHECKING, Collection, Dict, Iterable, List, Optional, Set, Tuple
 
 import attr
 
-from twisted.internet import defer
-
 from synapse.api.constants import EventTypes
-from synapse.logging.context import make_deferred_yieldable, run_in_background
 from synapse.storage._base import SQLBaseStore
 from synapse.storage.database import (
     DatabasePool,
@@ -41,12 +29,6 @@ from synapse.storage.state import StateFilter
 from synapse.storage.types import Cursor
 from synapse.storage.util.sequence import build_sequence_generator
 from synapse.types import MutableStateMap, StateKey, StateMap
-from synapse.util import unwrapFirstError
-from synapse.util.async_helpers import (
-    AbstractObservableDeferred,
-    ObservableDeferred,
-    yieldable_gather_results,
-)
 from synapse.util.caches.descriptors import cached
 from synapse.util.caches.dictionary_cache import DictionaryCache
 
@@ -55,8 +37,8 @@ if TYPE_CHECKING:
 
 logger = logging.getLogger(__name__)
 
+
 MAX_STATE_DELTA_HOPS = 100
-MAX_INFLIGHT_REQUESTS_PER_GROUP = 5
 
 
 @attr.s(slots=True, frozen=True, auto_attribs=True)
@@ -124,12 +106,6 @@ class StateGroupDataStore(StateBackgroundUpdateStore, SQLBaseStore):
             500000,
         )
 
-        # Current ongoing get_state_for_groups in-flight requests
-        # {group ID -> {StateFilter -> ObservableDeferred}}
-        self._state_group_inflight_requests: Dict[
-            int, Dict[StateFilter, AbstractObservableDeferred[StateMap[str]]]
-        ] = {}
-
         def get_max_state_group_txn(txn: Cursor) -> int:
             txn.execute("SELECT COALESCE(max(id), 0) FROM state_groups")
             return txn.fetchone()[0]  # type: ignore
@@ -181,7 +157,7 @@ class StateGroupDataStore(StateBackgroundUpdateStore, SQLBaseStore):
         )
 
     async def _get_state_groups_from_groups(
-        self, groups: Sequence[int], state_filter: StateFilter
+        self, groups: List[int], state_filter: StateFilter
     ) -> Dict[int, StateMap[str]]:
         """Returns the state groups for a given set of groups from the
         database, filtering on types of state events.
@@ -252,165 +228,6 @@ class StateGroupDataStore(StateBackgroundUpdateStore, SQLBaseStore):
 
         return state_filter.filter_state(state_dict_ids), not missing_types
 
-    def _get_state_for_group_gather_inflight_requests(
-        self, group: int, state_filter_left_over: StateFilter
-    ) -> Tuple[Sequence[AbstractObservableDeferred[StateMap[str]]], StateFilter]:
-        """
-        Attempts to gather in-flight requests and re-use them to retrieve state
-        for the given state group, filtered with the given state filter.
-
-        If there are more than MAX_INFLIGHT_REQUESTS_PER_GROUP in-flight requests,
-        and there *still* isn't enough information to complete the request by solely
-        reusing others, a full state filter will be requested to ensure that subsequent
-        requests can reuse this request.
-
-        Used as part of _get_state_for_group_using_inflight_cache.
-
-        Returns:
-            Tuple of two values:
-                A sequence of ObservableDeferreds to observe
-                A StateFilter representing what else needs to be requested to fulfill the request
-        """
-
-        inflight_requests = self._state_group_inflight_requests.get(group)
-        if inflight_requests is None:
-            # no requests for this group, need to retrieve it all ourselves
-            return (), state_filter_left_over
-
-        # The list of ongoing requests which will help narrow the current request.
-        reusable_requests = []
-        for (request_state_filter, request_deferred) in inflight_requests.items():
-            new_state_filter_left_over = state_filter_left_over.approx_difference(
-                request_state_filter
-            )
-            if new_state_filter_left_over == state_filter_left_over:
-                # Reusing this request would not gain us anything, so don't bother.
-                continue
-
-            reusable_requests.append(request_deferred)
-            state_filter_left_over = new_state_filter_left_over
-            if state_filter_left_over == StateFilter.none():
-                # we have managed to collect enough of the in-flight requests
-                # to cover our StateFilter and give us the state we need.
-                break
-
-        if (
-            state_filter_left_over != StateFilter.none()
-            and len(inflight_requests) >= MAX_INFLIGHT_REQUESTS_PER_GROUP
-        ):
-            # There are too many requests for this group.
-            # To prevent even more from building up, we request the whole
-            # state filter to guarantee that we can be reused by any subsequent
-            # requests for this state group.
-            return (), StateFilter.all()
-
-        return reusable_requests, state_filter_left_over
-
-    async def _get_state_for_group_fire_request(
-        self, group: int, state_filter: StateFilter
-    ) -> StateMap[str]:
-        """
-        Fires off a request to get the state at a state group,
-        potentially filtering by type and/or state key.
-
-        This request will be tracked in the in-flight request cache and automatically
-        removed when it is finished.
-
-        Used as part of _get_state_for_group_using_inflight_cache.
-
-        Args:
-            group: ID of the state group for which we want to get state
-            state_filter: the state filter used to fetch state from the database
-        """
-        cache_sequence_nm = self._state_group_cache.sequence
-        cache_sequence_m = self._state_group_members_cache.sequence
-
-        # Help the cache hit ratio by expanding the filter a bit
-        db_state_filter = state_filter.return_expanded()
-
-        async def _the_request() -> StateMap[str]:
-            group_to_state_dict = await self._get_state_groups_from_groups(
-                (group,), state_filter=db_state_filter
-            )
-
-            # Now let's update the caches
-            self._insert_into_cache(
-                group_to_state_dict,
-                db_state_filter,
-                cache_seq_num_members=cache_sequence_m,
-                cache_seq_num_non_members=cache_sequence_nm,
-            )
-
-            # Remove ourselves from the in-flight cache
-            group_request_dict = self._state_group_inflight_requests[group]
-            del group_request_dict[db_state_filter]
-            if not group_request_dict:
-                # If there are no more requests in-flight for this group,
-                # clean up the cache by removing the empty dictionary
-                del self._state_group_inflight_requests[group]
-
-            return group_to_state_dict[group]
-
-        # We don't immediately await the result, so must use run_in_background
-        # But we DO await the result before the current log context (request)
-        # finishes, so don't need to run it as a background process.
-        request_deferred = run_in_background(_the_request)
-        observable_deferred = ObservableDeferred(request_deferred, consumeErrors=True)
-
-        # Insert the ObservableDeferred into the cache
-        group_request_dict = self._state_group_inflight_requests.setdefault(group, {})
-        group_request_dict[db_state_filter] = observable_deferred
-
-        return await make_deferred_yieldable(observable_deferred.observe())
-
-    async def _get_state_for_group_using_inflight_cache(
-        self, group: int, state_filter: StateFilter
-    ) -> MutableStateMap[str]:
-        """
-        Gets the state at a state group, potentially filtering by type and/or
-        state key.
-
-        1. Calls _get_state_for_group_gather_inflight_requests to gather any
-           ongoing requests which might overlap with the current request.
-        2. Fires a new request, using _get_state_for_group_fire_request,
-           for any state which cannot be gathered from ongoing requests.
-
-        Args:
-            group: ID of the state group for which we want to get state
-            state_filter: the state filter used to fetch state from the database
-        Returns:
-            state map
-        """
-
-        # first, figure out whether we can re-use any in-flight requests
-        # (and if so, what would be left over)
-        (
-            reusable_requests,
-            state_filter_left_over,
-        ) = self._get_state_for_group_gather_inflight_requests(group, state_filter)
-
-        if state_filter_left_over != StateFilter.none():
-            # Fetch remaining state
-            remaining = await self._get_state_for_group_fire_request(
-                group, state_filter_left_over
-            )
-            assembled_state: MutableStateMap[str] = dict(remaining)
-        else:
-            assembled_state = {}
-
-        gathered = await make_deferred_yieldable(
-            defer.gatherResults(
-                (r.observe() for r in reusable_requests), consumeErrors=True
-            )
-        ).addErrback(unwrapFirstError)
-
-        # assemble our result.
-        for result_piece in gathered:
-            assembled_state.update(result_piece)
-
-        # Filter out any state that may be more than what we asked for.
-        return state_filter.filter_state(assembled_state)
-
     async def _get_state_for_groups(
         self, groups: Iterable[int], state_filter: Optional[StateFilter] = None
     ) -> Dict[int, MutableStateMap[str]]:
@@ -452,17 +269,31 @@ class StateGroupDataStore(StateBackgroundUpdateStore, SQLBaseStore):
         if not incomplete_groups:
             return state
 
-        async def get_from_cache(group: int, state_filter: StateFilter) -> None:
-            state[group] = await self._get_state_for_group_using_inflight_cache(
-                group, state_filter
-            )
+        cache_sequence_nm = self._state_group_cache.sequence
+        cache_sequence_m = self._state_group_members_cache.sequence
 
-        await yieldable_gather_results(
-            get_from_cache,
-            incomplete_groups,
-            state_filter,
+        # Help the cache hit ratio by expanding the filter a bit
+        db_state_filter = state_filter.return_expanded()
+
+        group_to_state_dict = await self._get_state_groups_from_groups(
+            list(incomplete_groups), state_filter=db_state_filter
         )
 
+        # Now lets update the caches
+        self._insert_into_cache(
+            group_to_state_dict,
+            db_state_filter,
+            cache_seq_num_members=cache_sequence_m,
+            cache_seq_num_non_members=cache_sequence_nm,
+        )
+
+        # And finally update the result dict, by filtering out any extra
+        # stuff we pulled out of the database.
+        for group, group_state_dict in group_to_state_dict.items():
+            # We just replace any existing entries, as we will have loaded
+            # everything we need from the database anyway.
+            state[group] = state_filter.filter_state(group_state_dict)
+
         return state
 
     def _get_state_for_groups_using_cache(
diff --git a/tests/storage/databases/test_state_store.py b/tests/storage/databases/test_state_store.py
deleted file mode 100644
index 076b660809..0000000000
--- a/tests/storage/databases/test_state_store.py
+++ /dev/null
@@ -1,352 +0,0 @@
-# Copyright 2022 The Matrix.org Foundation C.I.C.
-#
-# 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 typing
-from typing import Dict, List, Sequence, Tuple
-from unittest.mock import patch
-
-from twisted.internet.defer import Deferred, ensureDeferred
-from twisted.test.proto_helpers import MemoryReactor
-
-from synapse.api.constants import EventTypes
-from synapse.storage.databases.state.store import MAX_INFLIGHT_REQUESTS_PER_GROUP
-from synapse.storage.state import StateFilter
-from synapse.types import StateMap
-from synapse.util import Clock
-
-from tests.unittest import HomeserverTestCase
-
-if typing.TYPE_CHECKING:
-    from synapse.server import HomeServer
-
-# StateFilter for ALL non-m.room.member state events
-ALL_NON_MEMBERS_STATE_FILTER = StateFilter.freeze(
-    types={EventTypes.Member: set()},
-    include_others=True,
-)
-
-FAKE_STATE = {
-    (EventTypes.Member, "@alice:test"): "join",
-    (EventTypes.Member, "@bob:test"): "leave",
-    (EventTypes.Member, "@charlie:test"): "invite",
-    ("test.type", "a"): "AAA",
-    ("test.type", "b"): "BBB",
-    ("other.event.type", "state.key"): "123",
-}
-
-
-class StateGroupInflightCachingTestCase(HomeserverTestCase):
-    def prepare(
-        self, reactor: MemoryReactor, clock: Clock, homeserver: "HomeServer"
-    ) -> None:
-        self.state_storage = homeserver.get_storage().state
-        self.state_datastore = homeserver.get_datastores().state
-        # Patch out the `_get_state_groups_from_groups`.
-        # This is useful because it lets us pretend we have a slow database.
-        get_state_groups_patch = patch.object(
-            self.state_datastore,
-            "_get_state_groups_from_groups",
-            self._fake_get_state_groups_from_groups,
-        )
-        get_state_groups_patch.start()
-
-        self.addCleanup(get_state_groups_patch.stop)
-        self.get_state_group_calls: List[
-            Tuple[Tuple[int, ...], StateFilter, Deferred[Dict[int, StateMap[str]]]]
-        ] = []
-
-    def _fake_get_state_groups_from_groups(
-        self, groups: Sequence[int], state_filter: StateFilter
-    ) -> "Deferred[Dict[int, StateMap[str]]]":
-        d: Deferred[Dict[int, StateMap[str]]] = Deferred()
-        self.get_state_group_calls.append((tuple(groups), state_filter, d))
-        return d
-
-    def _complete_request_fake(
-        self,
-        groups: Tuple[int, ...],
-        state_filter: StateFilter,
-        d: "Deferred[Dict[int, StateMap[str]]]",
-    ) -> None:
-        """
-        Assemble a fake database response and complete the database request.
-        """
-
-        # Return a filtered copy of the fake state
-        d.callback({group: state_filter.filter_state(FAKE_STATE) for group in groups})
-
-    def test_duplicate_requests_deduplicated(self) -> None:
-        """
-        Tests that duplicate requests for state are deduplicated.
-
-        This test:
-        - requests some state (state group 42, 'all' state filter)
-        - requests it again, before the first request finishes
-        - checks to see that only one database query was made
-        - completes the database query
-        - checks that both requests see the same retrieved state
-        """
-        req1 = ensureDeferred(
-            self.state_datastore._get_state_for_group_using_inflight_cache(
-                42, StateFilter.all()
-            )
-        )
-        self.pump(by=0.1)
-
-        # This should have gone to the database
-        self.assertEqual(len(self.get_state_group_calls), 1)
-        self.assertFalse(req1.called)
-
-        req2 = ensureDeferred(
-            self.state_datastore._get_state_for_group_using_inflight_cache(
-                42, StateFilter.all()
-            )
-        )
-        self.pump(by=0.1)
-
-        # No more calls should have gone to the database
-        self.assertEqual(len(self.get_state_group_calls), 1)
-        self.assertFalse(req1.called)
-        self.assertFalse(req2.called)
-
-        groups, sf, d = self.get_state_group_calls[0]
-        self.assertEqual(groups, (42,))
-        self.assertEqual(sf, StateFilter.all())
-
-        # Now we can complete the request
-        self._complete_request_fake(groups, sf, d)
-
-        self.assertEqual(self.get_success(req1), FAKE_STATE)
-        self.assertEqual(self.get_success(req2), FAKE_STATE)
-
-    def test_smaller_request_deduplicated(self) -> None:
-        """
-        Tests that duplicate requests for state are deduplicated.
-
-        This test:
-        - requests some state (state group 42, 'all' state filter)
-        - requests a subset of that state, before the first request finishes
-        - checks to see that only one database query was made
-        - completes the database query
-        - checks that both requests see the correct retrieved state
-        """
-        req1 = ensureDeferred(
-            self.state_datastore._get_state_for_group_using_inflight_cache(
-                42, StateFilter.from_types((("test.type", None),))
-            )
-        )
-        self.pump(by=0.1)
-
-        # This should have gone to the database
-        self.assertEqual(len(self.get_state_group_calls), 1)
-        self.assertFalse(req1.called)
-
-        req2 = ensureDeferred(
-            self.state_datastore._get_state_for_group_using_inflight_cache(
-                42, StateFilter.from_types((("test.type", "b"),))
-            )
-        )
-        self.pump(by=0.1)
-
-        # No more calls should have gone to the database, because the second
-        # request was already in the in-flight cache!
-        self.assertEqual(len(self.get_state_group_calls), 1)
-        self.assertFalse(req1.called)
-        self.assertFalse(req2.called)
-
-        groups, sf, d = self.get_state_group_calls[0]
-        self.assertEqual(groups, (42,))
-        # The state filter is expanded internally for increased cache hit rate,
-        # so we the database sees a wider state filter than requested.
-        self.assertEqual(sf, ALL_NON_MEMBERS_STATE_FILTER)
-
-        # Now we can complete the request
-        self._complete_request_fake(groups, sf, d)
-
-        self.assertEqual(
-            self.get_success(req1),
-            {("test.type", "a"): "AAA", ("test.type", "b"): "BBB"},
-        )
-        self.assertEqual(self.get_success(req2), {("test.type", "b"): "BBB"})
-
-    def test_partially_overlapping_request_deduplicated(self) -> None:
-        """
-        Tests that partially-overlapping requests are partially deduplicated.
-
-        This test:
-        - requests a single type of wildcard state
-          (This is internally expanded to be all non-member state)
-        - requests the entire state in parallel
-        - checks to see that two database queries were made, but that the second
-          one is only for member state.
-        - completes the database queries
-        - checks that both requests have the correct result.
-        """
-
-        req1 = ensureDeferred(
-            self.state_datastore._get_state_for_group_using_inflight_cache(
-                42, StateFilter.from_types((("test.type", None),))
-            )
-        )
-        self.pump(by=0.1)
-
-        # This should have gone to the database
-        self.assertEqual(len(self.get_state_group_calls), 1)
-        self.assertFalse(req1.called)
-
-        req2 = ensureDeferred(
-            self.state_datastore._get_state_for_group_using_inflight_cache(
-                42, StateFilter.all()
-            )
-        )
-        self.pump(by=0.1)
-
-        # Because it only partially overlaps, this also went to the database
-        self.assertEqual(len(self.get_state_group_calls), 2)
-        self.assertFalse(req1.called)
-        self.assertFalse(req2.called)
-
-        # First request:
-        groups, sf, d = self.get_state_group_calls[0]
-        self.assertEqual(groups, (42,))
-        # The state filter is expanded internally for increased cache hit rate,
-        # so we the database sees a wider state filter than requested.
-        self.assertEqual(sf, ALL_NON_MEMBERS_STATE_FILTER)
-        self._complete_request_fake(groups, sf, d)
-
-        # Second request:
-        groups, sf, d = self.get_state_group_calls[1]
-        self.assertEqual(groups, (42,))
-        # The state filter is narrowed to only request membership state, because
-        # the remainder of the state is already being queried in the first request!
-        self.assertEqual(
-            sf, StateFilter.freeze({EventTypes.Member: None}, include_others=False)
-        )
-        self._complete_request_fake(groups, sf, d)
-
-        # Check the results are correct
-        self.assertEqual(
-            self.get_success(req1),
-            {("test.type", "a"): "AAA", ("test.type", "b"): "BBB"},
-        )
-        self.assertEqual(self.get_success(req2), FAKE_STATE)
-
-    def test_in_flight_requests_stop_being_in_flight(self) -> None:
-        """
-        Tests that in-flight request deduplication doesn't somehow 'hold on'
-        to completed requests: once they're done, they're taken out of the
-        in-flight cache.
-        """
-        req1 = ensureDeferred(
-            self.state_datastore._get_state_for_group_using_inflight_cache(
-                42, StateFilter.all()
-            )
-        )
-        self.pump(by=0.1)
-
-        # This should have gone to the database
-        self.assertEqual(len(self.get_state_group_calls), 1)
-        self.assertFalse(req1.called)
-
-        # Complete the request right away.
-        self._complete_request_fake(*self.get_state_group_calls[0])
-        self.assertTrue(req1.called)
-
-        # Send off another request
-        req2 = ensureDeferred(
-            self.state_datastore._get_state_for_group_using_inflight_cache(
-                42, StateFilter.all()
-            )
-        )
-        self.pump(by=0.1)
-
-        # It should have gone to the database again, because the previous request
-        # isn't in-flight and therefore isn't available for deduplication.
-        self.assertEqual(len(self.get_state_group_calls), 2)
-        self.assertFalse(req2.called)
-
-        # Complete the request right away.
-        self._complete_request_fake(*self.get_state_group_calls[1])
-        self.assertTrue(req2.called)
-        groups, sf, d = self.get_state_group_calls[0]
-
-        self.assertEqual(self.get_success(req1), FAKE_STATE)
-        self.assertEqual(self.get_success(req2), FAKE_STATE)
-
-    def test_inflight_requests_capped(self) -> None:
-        """
-        Tests that the number of in-flight requests is capped to 5.
-
-        - requests several pieces of state separately
-          (5 to hit the limit, 1 to 'shunt out', another that comes after the
-          group has been 'shunted out')
-        - checks to see that the torrent of requests is shunted out by
-          rewriting one of the filters as the 'all' state filter
-        - requests after that one do not cause any additional queries
-        """
-        # 5 at the time of writing.
-        CAP_COUNT = MAX_INFLIGHT_REQUESTS_PER_GROUP
-
-        reqs = []
-
-        # Request 7 different keys (1 to 7) of the `some.state` type.
-        for req_id in range(CAP_COUNT + 2):
-            reqs.append(
-                ensureDeferred(
-                    self.state_datastore._get_state_for_group_using_inflight_cache(
-                        42,
-                        StateFilter.freeze(
-                            {"some.state": {str(req_id + 1)}}, include_others=False
-                        ),
-                    )
-                )
-            )
-        self.pump(by=0.1)
-
-        # There should only be 6 calls to the database, not 7.
-        self.assertEqual(len(self.get_state_group_calls), CAP_COUNT + 1)
-
-        # Assert that the first 5 are exact requests for the individual pieces
-        # wanted
-        for req_id in range(CAP_COUNT):
-            groups, sf, d = self.get_state_group_calls[req_id]
-            self.assertEqual(
-                sf,
-                StateFilter.freeze(
-                    {"some.state": {str(req_id + 1)}}, include_others=False
-                ),
-            )
-
-        # The 6th request should be the 'all' state filter
-        groups, sf, d = self.get_state_group_calls[CAP_COUNT]
-        self.assertEqual(sf, StateFilter.all())
-
-        # Complete the queries and check which requests complete as a result
-        for req_id in range(CAP_COUNT):
-            # This request should not have been completed yet
-            self.assertFalse(reqs[req_id].called)
-
-            groups, sf, d = self.get_state_group_calls[req_id]
-            self._complete_request_fake(groups, sf, d)
-
-            # This should have only completed this one request
-            self.assertTrue(reqs[req_id].called)
-
-        # Now complete the final query; the last 2 requests should complete
-        # as a result
-        self.assertFalse(reqs[CAP_COUNT].called)
-        self.assertFalse(reqs[CAP_COUNT + 1].called)
-        groups, sf, d = self.get_state_group_calls[CAP_COUNT]
-        self._complete_request_fake(groups, sf, d)
-        self.assertTrue(reqs[CAP_COUNT].called)
-        self.assertTrue(reqs[CAP_COUNT + 1].called)