Split state groups into a separate data store (#6296)

This commit is contained in:
Erik Johnston 2019-12-20 10:48:24 +00:00 committed by GitHub
parent fa780e9721
commit 75d8f26ac8
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
28 changed files with 1159 additions and 1168 deletions

1
changelog.d/6245.misc Normal file
View file

@ -0,0 +1 @@
Split out state storage into separate data store.

View file

@ -51,11 +51,12 @@ from synapse.storage.data_stores.main.registration import (
from synapse.storage.data_stores.main.room import RoomBackgroundUpdateStore from synapse.storage.data_stores.main.room import RoomBackgroundUpdateStore
from synapse.storage.data_stores.main.roommember import RoomMemberBackgroundUpdateStore from synapse.storage.data_stores.main.roommember import RoomMemberBackgroundUpdateStore
from synapse.storage.data_stores.main.search import SearchBackgroundUpdateStore from synapse.storage.data_stores.main.search import SearchBackgroundUpdateStore
from synapse.storage.data_stores.main.state import StateBackgroundUpdateStore from synapse.storage.data_stores.main.state import MainStateBackgroundUpdateStore
from synapse.storage.data_stores.main.stats import StatsStore from synapse.storage.data_stores.main.stats import StatsStore
from synapse.storage.data_stores.main.user_directory import ( from synapse.storage.data_stores.main.user_directory import (
UserDirectoryBackgroundUpdateStore, UserDirectoryBackgroundUpdateStore,
) )
from synapse.storage.data_stores.state.bg_updates import StateBackgroundUpdateStore
from synapse.storage.database import Database, make_conn from synapse.storage.database import Database, make_conn
from synapse.storage.engines import create_engine from synapse.storage.engines import create_engine
from synapse.storage.prepare_database import prepare_database from synapse.storage.prepare_database import prepare_database
@ -138,6 +139,7 @@ class Store(
RoomMemberBackgroundUpdateStore, RoomMemberBackgroundUpdateStore,
SearchBackgroundUpdateStore, SearchBackgroundUpdateStore,
StateBackgroundUpdateStore, StateBackgroundUpdateStore,
MainStateBackgroundUpdateStore,
UserDirectoryBackgroundUpdateStore, UserDirectoryBackgroundUpdateStore,
StatsStore, StatsStore,
): ):
@ -496,9 +498,7 @@ class Porter(object):
def run(self): def run(self):
try: try:
self.sqlite_store = yield self.build_db_store( self.sqlite_store = yield self.build_db_store(
DatabaseConnectionConfig( DatabaseConnectionConfig("master-sqlite", self.sqlite_config)
"master", self.sqlite_config, data_stores=["main"]
)
) )
# Check if all background updates are done, abort if not. # Check if all background updates are done, abort if not.

View file

@ -34,10 +34,12 @@ class DatabaseConnectionConfig:
module name, and `args` for the args to give to the database module name, and `args` for the args to give to the database
connector. connector.
data_stores: The list of data stores that should be provisioned on the data_stores: The list of data stores that should be provisioned on the
database. database. Defaults to all data stores.
""" """
def __init__(self, name: str, db_config: dict, data_stores: List[str]): def __init__(
self, name: str, db_config: dict, data_stores: List[str] = ["main", "state"]
):
if db_config["name"] not in ("sqlite3", "psycopg2"): if db_config["name"] not in ("sqlite3", "psycopg2"):
raise ConfigError("Unsupported database type %r" % (db_config["name"],)) raise ConfigError("Unsupported database type %r" % (db_config["name"],))
@ -62,9 +64,7 @@ class DatabaseConfig(Config):
if database_config is None: if database_config is None:
database_config = {"name": "sqlite3", "args": {}} database_config = {"name": "sqlite3", "args": {}}
self.databases = [ self.databases = [DatabaseConnectionConfig("master", database_config)]
DatabaseConnectionConfig("master", database_config, data_stores=["main"])
]
self.set_databasepath(config.get("database_path")) self.set_databasepath(config.get("database_path"))

View file

@ -15,6 +15,7 @@
import logging import logging
from synapse.storage.data_stores.state import StateGroupDataStore
from synapse.storage.database import Database, make_conn from synapse.storage.database import Database, make_conn
from synapse.storage.engines import create_engine from synapse.storage.engines import create_engine
from synapse.storage.prepare_database import prepare_database from synapse.storage.prepare_database import prepare_database
@ -55,6 +56,10 @@ class DataStores(object):
logger.info("Starting 'main' data store") logger.info("Starting 'main' data store")
self.main = main_store_class(database, db_conn, hs) self.main = main_store_class(database, db_conn, hs)
if "state" in database_config.data_stores:
logger.info("Starting 'state' data store")
self.state = StateGroupDataStore(database, db_conn, hs)
db_conn.commit() db_conn.commit()
self.databases.append(database) self.databases.append(database)

View file

@ -1757,163 +1757,6 @@ class EventsStore(
return state_groups return state_groups
def purge_unreferenced_state_groups(
self, room_id: str, state_groups_to_delete
) -> defer.Deferred:
"""Deletes no longer referenced state groups and de-deltas any state
groups that reference them.
Args:
room_id: The room the state groups belong to (must all be in the
same room).
state_groups_to_delete (Collection[int]): Set of all state groups
to delete.
"""
return self.db.runInteraction(
"purge_unreferenced_state_groups",
self._purge_unreferenced_state_groups,
room_id,
state_groups_to_delete,
)
def _purge_unreferenced_state_groups(self, txn, room_id, state_groups_to_delete):
logger.info(
"[purge] found %i state groups to delete", len(state_groups_to_delete)
)
rows = self.db.simple_select_many_txn(
txn,
table="state_group_edges",
column="prev_state_group",
iterable=state_groups_to_delete,
keyvalues={},
retcols=("state_group",),
)
remaining_state_groups = set(
row["state_group"]
for row in rows
if row["state_group"] not in state_groups_to_delete
)
logger.info(
"[purge] de-delta-ing %i remaining state groups",
len(remaining_state_groups),
)
# Now we turn the state groups that reference to-be-deleted state
# groups to non delta versions.
for sg in remaining_state_groups:
logger.info("[purge] de-delta-ing remaining state group %s", sg)
curr_state = self._get_state_groups_from_groups_txn(txn, [sg])
curr_state = curr_state[sg]
self.db.simple_delete_txn(
txn, table="state_groups_state", keyvalues={"state_group": sg}
)
self.db.simple_delete_txn(
txn, table="state_group_edges", keyvalues={"state_group": sg}
)
self.db.simple_insert_many_txn(
txn,
table="state_groups_state",
values=[
{
"state_group": sg,
"room_id": room_id,
"type": key[0],
"state_key": key[1],
"event_id": state_id,
}
for key, state_id in iteritems(curr_state)
],
)
logger.info("[purge] removing redundant state groups")
txn.executemany(
"DELETE FROM state_groups_state WHERE state_group = ?",
((sg,) for sg in state_groups_to_delete),
)
txn.executemany(
"DELETE FROM state_groups WHERE id = ?",
((sg,) for sg in state_groups_to_delete),
)
@defer.inlineCallbacks
def get_previous_state_groups(self, state_groups):
"""Fetch the previous groups of the given state groups.
Args:
state_groups (Iterable[int])
Returns:
Deferred[dict[int, int]]: mapping from state group to previous
state group.
"""
rows = yield self.db.simple_select_many_batch(
table="state_group_edges",
column="prev_state_group",
iterable=state_groups,
keyvalues={},
retcols=("prev_state_group", "state_group"),
desc="get_previous_state_groups",
)
return {row["state_group"]: row["prev_state_group"] for row in rows}
def purge_room_state(self, room_id, state_groups_to_delete):
"""Deletes all record of a room from state tables
Args:
room_id (str):
state_groups_to_delete (list[int]): State groups to delete
"""
return self.db.runInteraction(
"purge_room_state",
self._purge_room_state_txn,
room_id,
state_groups_to_delete,
)
def _purge_room_state_txn(self, txn, room_id, state_groups_to_delete):
# first we have to delete the state groups states
logger.info("[purge] removing %s from state_groups_state", room_id)
self.db.simple_delete_many_txn(
txn,
table="state_groups_state",
column="state_group",
iterable=state_groups_to_delete,
keyvalues={},
)
# ... and the state group edges
logger.info("[purge] removing %s from state_group_edges", room_id)
self.db.simple_delete_many_txn(
txn,
table="state_group_edges",
column="state_group",
iterable=state_groups_to_delete,
keyvalues={},
)
# ... and the state groups
logger.info("[purge] removing %s from state_groups", room_id)
self.db.simple_delete_many_txn(
txn,
table="state_groups",
column="id",
iterable=state_groups_to_delete,
keyvalues={},
)
async def is_event_after(self, event_id1, event_id2): async def is_event_after(self, event_id1, event_id2):
"""Returns True if event_id1 is after event_id2 in the stream """Returns True if event_id1 is after event_id2 in the stream
""" """

View file

@ -20,7 +20,6 @@ DROP INDEX IF EXISTS events_room_id; -- Prefix of events_room_stream
DROP INDEX IF EXISTS events_order; -- Prefix of events_order_topo_stream_room DROP INDEX IF EXISTS events_order; -- Prefix of events_order_topo_stream_room
DROP INDEX IF EXISTS events_topological_ordering; -- Prefix of events_order_topo_stream_room DROP INDEX IF EXISTS events_topological_ordering; -- Prefix of events_order_topo_stream_room
DROP INDEX IF EXISTS events_stream_ordering; -- Duplicate of PRIMARY KEY DROP INDEX IF EXISTS events_stream_ordering; -- Duplicate of PRIMARY KEY
DROP INDEX IF EXISTS state_groups_id; -- Duplicate of PRIMARY KEY
DROP INDEX IF EXISTS event_to_state_groups_id; -- Duplicate of PRIMARY KEY DROP INDEX IF EXISTS event_to_state_groups_id; -- Duplicate of PRIMARY KEY
DROP INDEX IF EXISTS event_push_actions_room_id_event_id_user_id_profile_tag; -- Duplicate of UNIQUE CONSTRAINT DROP INDEX IF EXISTS event_push_actions_room_id_event_id_user_id_profile_tag; -- Duplicate of UNIQUE CONSTRAINT

View file

@ -975,40 +975,6 @@ CREATE TABLE state_events (
CREATE TABLE state_group_edges (
state_group bigint NOT NULL,
prev_state_group bigint NOT NULL
);
CREATE SEQUENCE state_group_id_seq
START WITH 1
INCREMENT BY 1
NO MINVALUE
NO MAXVALUE
CACHE 1;
CREATE TABLE state_groups (
id bigint NOT NULL,
room_id text NOT NULL,
event_id text NOT NULL
);
CREATE TABLE state_groups_state (
state_group bigint NOT NULL,
room_id text NOT NULL,
type text NOT NULL,
state_key text NOT NULL,
event_id text NOT NULL
);
CREATE TABLE stats_stream_pos ( CREATE TABLE stats_stream_pos (
lock character(1) DEFAULT 'X'::bpchar NOT NULL, lock character(1) DEFAULT 'X'::bpchar NOT NULL,
stream_id bigint, stream_id bigint,
@ -1482,12 +1448,6 @@ ALTER TABLE ONLY state_events
ADD CONSTRAINT state_events_event_id_key UNIQUE (event_id); ADD CONSTRAINT state_events_event_id_key UNIQUE (event_id);
ALTER TABLE ONLY state_groups
ADD CONSTRAINT state_groups_pkey PRIMARY KEY (id);
ALTER TABLE ONLY stats_stream_pos ALTER TABLE ONLY stats_stream_pos
ADD CONSTRAINT stats_stream_pos_lock_key UNIQUE (lock); ADD CONSTRAINT stats_stream_pos_lock_key UNIQUE (lock);
@ -1928,18 +1888,6 @@ CREATE UNIQUE INDEX room_stats_room_ts ON room_stats USING btree (room_id, ts);
CREATE INDEX state_group_edges_idx ON state_group_edges USING btree (state_group);
CREATE INDEX state_group_edges_prev_idx ON state_group_edges USING btree (prev_state_group);
CREATE INDEX state_groups_state_type_idx ON state_groups_state USING btree (state_group, type, state_key);
CREATE INDEX stream_ordering_to_exterm_idx ON stream_ordering_to_exterm USING btree (stream_ordering); CREATE INDEX stream_ordering_to_exterm_idx ON stream_ordering_to_exterm USING btree (stream_ordering);

View file

@ -42,8 +42,6 @@ CREATE INDEX ev_edges_id ON event_edges(event_id);
CREATE INDEX ev_edges_prev_id ON event_edges(prev_event_id); CREATE INDEX ev_edges_prev_id ON event_edges(prev_event_id);
CREATE TABLE room_depth( room_id TEXT NOT NULL, min_depth INTEGER NOT NULL, UNIQUE (room_id) ); CREATE TABLE room_depth( room_id TEXT NOT NULL, min_depth INTEGER NOT NULL, UNIQUE (room_id) );
CREATE INDEX room_depth_room ON room_depth(room_id); CREATE INDEX room_depth_room ON room_depth(room_id);
CREATE TABLE state_groups( id BIGINT PRIMARY KEY, room_id TEXT NOT NULL, event_id TEXT NOT NULL );
CREATE TABLE state_groups_state( state_group BIGINT NOT NULL, room_id TEXT NOT NULL, type TEXT NOT NULL, state_key TEXT NOT NULL, event_id TEXT NOT NULL );
CREATE TABLE event_to_state_groups( event_id TEXT NOT NULL, state_group BIGINT NOT NULL, UNIQUE (event_id) ); CREATE TABLE event_to_state_groups( event_id TEXT NOT NULL, state_group BIGINT NOT NULL, UNIQUE (event_id) );
CREATE TABLE local_media_repository ( media_id TEXT, media_type TEXT, media_length INTEGER, created_ts BIGINT, upload_name TEXT, user_id TEXT, quarantined_by TEXT, url_cache TEXT, last_access_ts BIGINT, UNIQUE (media_id) ); CREATE TABLE local_media_repository ( media_id TEXT, media_type TEXT, media_length INTEGER, created_ts BIGINT, upload_name TEXT, user_id TEXT, quarantined_by TEXT, url_cache TEXT, last_access_ts BIGINT, UNIQUE (media_id) );
CREATE TABLE local_media_repository_thumbnails ( media_id TEXT, thumbnail_width INTEGER, thumbnail_height INTEGER, thumbnail_type TEXT, thumbnail_method TEXT, thumbnail_length INTEGER, UNIQUE ( media_id, thumbnail_width, thumbnail_height, thumbnail_type ) ); CREATE TABLE local_media_repository_thumbnails ( media_id TEXT, thumbnail_width INTEGER, thumbnail_height INTEGER, thumbnail_type TEXT, thumbnail_method TEXT, thumbnail_length INTEGER, UNIQUE ( media_id, thumbnail_width, thumbnail_height, thumbnail_type ) );
@ -120,9 +118,6 @@ CREATE TABLE device_max_stream_id ( stream_id BIGINT NOT NULL );
CREATE TABLE public_room_list_stream ( stream_id BIGINT NOT NULL, room_id TEXT NOT NULL, visibility BOOLEAN NOT NULL , appservice_id TEXT, network_id TEXT); CREATE TABLE public_room_list_stream ( stream_id BIGINT NOT NULL, room_id TEXT NOT NULL, visibility BOOLEAN NOT NULL , appservice_id TEXT, network_id TEXT);
CREATE INDEX public_room_list_stream_idx on public_room_list_stream( stream_id ); CREATE INDEX public_room_list_stream_idx on public_room_list_stream( stream_id );
CREATE INDEX public_room_list_stream_rm_idx on public_room_list_stream( room_id, stream_id ); CREATE INDEX public_room_list_stream_rm_idx on public_room_list_stream( room_id, stream_id );
CREATE TABLE state_group_edges( state_group BIGINT NOT NULL, prev_state_group BIGINT NOT NULL );
CREATE INDEX state_group_edges_idx ON state_group_edges(state_group);
CREATE INDEX state_group_edges_prev_idx ON state_group_edges(prev_state_group);
CREATE TABLE stream_ordering_to_exterm ( stream_ordering BIGINT NOT NULL, room_id TEXT NOT NULL, event_id TEXT NOT NULL ); CREATE TABLE stream_ordering_to_exterm ( stream_ordering BIGINT NOT NULL, room_id TEXT NOT NULL, event_id TEXT NOT NULL );
CREATE INDEX stream_ordering_to_exterm_idx on stream_ordering_to_exterm( stream_ordering ); CREATE INDEX stream_ordering_to_exterm_idx on stream_ordering_to_exterm( stream_ordering );
CREATE INDEX stream_ordering_to_exterm_rm_idx on stream_ordering_to_exterm( room_id, stream_ordering ); CREATE INDEX stream_ordering_to_exterm_rm_idx on stream_ordering_to_exterm( room_id, stream_ordering );
@ -254,6 +249,5 @@ CREATE INDEX user_ips_last_seen_only ON user_ips (last_seen);
CREATE INDEX users_creation_ts ON users (creation_ts); CREATE INDEX users_creation_ts ON users (creation_ts);
CREATE INDEX event_to_state_groups_sg_index ON event_to_state_groups (state_group); CREATE INDEX event_to_state_groups_sg_index ON event_to_state_groups (state_group);
CREATE UNIQUE INDEX device_lists_remote_cache_unique_id ON device_lists_remote_cache (user_id, device_id); CREATE UNIQUE INDEX device_lists_remote_cache_unique_id ON device_lists_remote_cache (user_id, device_id);
CREATE INDEX state_groups_state_type_idx ON state_groups_state(state_group, type, state_key);
CREATE UNIQUE INDEX device_lists_remote_extremeties_unique_idx ON device_lists_remote_extremeties (user_id); CREATE UNIQUE INDEX device_lists_remote_extremeties_unique_idx ON device_lists_remote_extremeties (user_id);
CREATE UNIQUE INDEX user_ips_user_token_ip_unique_index ON user_ips (user_id, access_token, ip); CREATE UNIQUE INDEX user_ips_user_token_ip_unique_index ON user_ips (user_id, access_token, ip);

View file

@ -17,8 +17,7 @@ import logging
from collections import namedtuple from collections import namedtuple
from typing import Iterable, Tuple from typing import Iterable, Tuple
from six import iteritems, itervalues from six import iteritems
from six.moves import range
from twisted.internet import defer from twisted.internet import defer
@ -29,11 +28,9 @@ from synapse.events.snapshot import EventContext
from synapse.storage._base import SQLBaseStore from synapse.storage._base import SQLBaseStore
from synapse.storage.data_stores.main.events_worker import EventsWorkerStore from synapse.storage.data_stores.main.events_worker import EventsWorkerStore
from synapse.storage.database import Database from synapse.storage.database import Database
from synapse.storage.engines import PostgresEngine
from synapse.storage.state import StateFilter from synapse.storage.state import StateFilter
from synapse.util.caches import get_cache_factor_for, intern_string from synapse.util.caches import intern_string
from synapse.util.caches.descriptors import cached, cachedList from synapse.util.caches.descriptors import cached, cachedList
from synapse.util.caches.dictionary_cache import DictionaryCache
from synapse.util.stringutils import to_ascii from synapse.util.stringutils import to_ascii
logger = logging.getLogger(__name__) logger = logging.getLogger(__name__)
@ -55,207 +52,14 @@ class _GetStateGroupDelta(
return len(self.delta_ids) if self.delta_ids else 0 return len(self.delta_ids) if self.delta_ids else 0
class StateGroupBackgroundUpdateStore(SQLBaseStore):
"""Defines functions related to state groups needed to run the state backgroud
updates.
"""
def _count_state_group_hops_txn(self, txn, state_group):
"""Given a state group, count how many hops there are in the tree.
This is used to ensure the delta chains don't get too long.
"""
if isinstance(self.database_engine, PostgresEngine):
sql = """
WITH RECURSIVE state(state_group) AS (
VALUES(?::bigint)
UNION ALL
SELECT prev_state_group FROM state_group_edges e, state s
WHERE s.state_group = e.state_group
)
SELECT count(*) FROM state;
"""
txn.execute(sql, (state_group,))
row = txn.fetchone()
if row and row[0]:
return row[0]
else:
return 0
else:
# We don't use WITH RECURSIVE on sqlite3 as there are distributions
# that ship with an sqlite3 version that doesn't support it (e.g. wheezy)
next_group = state_group
count = 0
while next_group:
next_group = self.db.simple_select_one_onecol_txn(
txn,
table="state_group_edges",
keyvalues={"state_group": next_group},
retcol="prev_state_group",
allow_none=True,
)
if next_group:
count += 1
return count
def _get_state_groups_from_groups_txn(
self, txn, groups, state_filter=StateFilter.all()
):
results = {group: {} for group in groups}
where_clause, where_args = state_filter.make_sql_filter_clause()
# Unless the filter clause is empty, we're going to append it after an
# existing where clause
if where_clause:
where_clause = " AND (%s)" % (where_clause,)
if isinstance(self.database_engine, PostgresEngine):
# Temporarily disable sequential scans in this transaction. This is
# a temporary hack until we can add the right indices in
txn.execute("SET LOCAL enable_seqscan=off")
# The below query walks the state_group tree so that the "state"
# table includes all state_groups in the tree. It then joins
# against `state_groups_state` to fetch the latest state.
# It assumes that previous state groups are always numerically
# lesser.
# The PARTITION is used to get the event_id in the greatest state
# group for the given type, state_key.
# This may return multiple rows per (type, state_key), but last_value
# should be the same.
sql = """
WITH RECURSIVE state(state_group) AS (
VALUES(?::bigint)
UNION ALL
SELECT prev_state_group FROM state_group_edges e, state s
WHERE s.state_group = e.state_group
)
SELECT DISTINCT type, state_key, last_value(event_id) OVER (
PARTITION BY type, state_key ORDER BY state_group ASC
ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING
) AS event_id FROM state_groups_state
WHERE state_group IN (
SELECT state_group FROM state
)
"""
for group in groups:
args = [group]
args.extend(where_args)
txn.execute(sql + where_clause, args)
for row in txn:
typ, state_key, event_id = row
key = (typ, state_key)
results[group][key] = event_id
else:
max_entries_returned = state_filter.max_entries_returned()
# We don't use WITH RECURSIVE on sqlite3 as there are distributions
# that ship with an sqlite3 version that doesn't support it (e.g. wheezy)
for group in groups:
next_group = group
while next_group:
# We did this before by getting the list of group ids, and
# then passing that list to sqlite to get latest event for
# each (type, state_key). However, that was terribly slow
# without the right indices (which we can't add until
# after we finish deduping state, which requires this func)
args = [next_group]
args.extend(where_args)
txn.execute(
"SELECT type, state_key, event_id FROM state_groups_state"
" WHERE state_group = ? " + where_clause,
args,
)
results[group].update(
((typ, state_key), event_id)
for typ, state_key, event_id in txn
if (typ, state_key) not in results[group]
)
# If the number of entries in the (type,state_key)->event_id dict
# matches the number of (type,state_keys) types we were searching
# for, then we must have found them all, so no need to go walk
# further down the tree... UNLESS our types filter contained
# wildcards (i.e. Nones) in which case we have to do an exhaustive
# search
if (
max_entries_returned is not None
and len(results[group]) == max_entries_returned
):
break
next_group = self.db.simple_select_one_onecol_txn(
txn,
table="state_group_edges",
keyvalues={"state_group": next_group},
retcol="prev_state_group",
allow_none=True,
)
return results
# this inherits from EventsWorkerStore because it calls self.get_events # this inherits from EventsWorkerStore because it calls self.get_events
class StateGroupWorkerStore( class StateGroupWorkerStore(EventsWorkerStore, SQLBaseStore):
EventsWorkerStore, StateGroupBackgroundUpdateStore, SQLBaseStore
):
"""The parts of StateGroupStore that can be called from workers. """The parts of StateGroupStore that can be called from workers.
""" """
STATE_GROUP_DEDUPLICATION_UPDATE_NAME = "state_group_state_deduplication"
STATE_GROUP_INDEX_UPDATE_NAME = "state_group_state_type_index"
CURRENT_STATE_INDEX_UPDATE_NAME = "current_state_members_idx"
def __init__(self, database: Database, db_conn, hs): def __init__(self, database: Database, db_conn, hs):
super(StateGroupWorkerStore, self).__init__(database, db_conn, hs) super(StateGroupWorkerStore, self).__init__(database, db_conn, hs)
# Originally the state store used a single DictionaryCache to cache the
# event IDs for the state types in a given state group to avoid hammering
# on the state_group* tables.
#
# The point of using a DictionaryCache is that it can cache a subset
# of the state events for a given state group (i.e. a subset of the keys for a
# given dict which is an entry in the cache for a given state group ID).
#
# However, this poses problems when performing complicated queries
# on the store - for instance: "give me all the state for this group, but
# limit members to this subset of users", as DictionaryCache's API isn't
# rich enough to say "please cache any of these fields, apart from this subset".
# This is problematic when lazy loading members, which requires this behaviour,
# as without it the cache has no choice but to speculatively load all
# state events for the group, which negates the efficiency being sought.
#
# Rather than overcomplicating DictionaryCache's API, we instead split the
# state_group_cache into two halves - one for tracking non-member events,
# and the other for tracking member_events. This means that lazy loading
# queries can be made in a cache-friendly manner by querying both caches
# separately and then merging the result. So for the example above, you
# would query the members cache for a specific subset of state keys
# (which DictionaryCache will handle efficiently and fine) and the non-members
# cache for all state (which DictionaryCache will similarly handle fine)
# and then just merge the results together.
#
# We size the non-members cache to be smaller than the members cache as the
# vast majority of state in Matrix (today) is member events.
self._state_group_cache = DictionaryCache(
"*stateGroupCache*",
# TODO: this hasn't been tuned yet
50000 * get_cache_factor_for("stateGroupCache"),
)
self._state_group_members_cache = DictionaryCache(
"*stateGroupMembersCache*",
500000 * get_cache_factor_for("stateGroupMembersCache"),
)
@defer.inlineCallbacks @defer.inlineCallbacks
def get_room_version(self, room_id): def get_room_version(self, room_id):
"""Get the room_version of a given room """Get the room_version of a given room
@ -431,229 +235,6 @@ class StateGroupWorkerStore(
return event.content.get("canonical_alias") return event.content.get("canonical_alias")
@cached(max_entries=10000, iterable=True)
def get_state_group_delta(self, state_group):
"""Given a state group try to return a previous group and a delta between
the old and the new.
Returns:
(prev_group, delta_ids), where both may be None.
"""
def _get_state_group_delta_txn(txn):
prev_group = self.db.simple_select_one_onecol_txn(
txn,
table="state_group_edges",
keyvalues={"state_group": state_group},
retcol="prev_state_group",
allow_none=True,
)
if not prev_group:
return _GetStateGroupDelta(None, None)
delta_ids = self.db.simple_select_list_txn(
txn,
table="state_groups_state",
keyvalues={"state_group": state_group},
retcols=("type", "state_key", "event_id"),
)
return _GetStateGroupDelta(
prev_group,
{(row["type"], row["state_key"]): row["event_id"] for row in delta_ids},
)
return self.db.runInteraction(
"get_state_group_delta", _get_state_group_delta_txn
)
@defer.inlineCallbacks
def get_state_groups_ids(self, _room_id, event_ids):
"""Get the event IDs of all the state for the state groups for the given events
Args:
_room_id (str): id of the room for these events
event_ids (iterable[str]): ids of the events
Returns:
Deferred[dict[int, dict[tuple[str, str], str]]]:
dict of state_group_id -> (dict of (type, state_key) -> event id)
"""
if not event_ids:
return {}
event_to_groups = yield self._get_state_group_for_events(event_ids)
groups = set(itervalues(event_to_groups))
group_to_state = yield self._get_state_for_groups(groups)
return group_to_state
@defer.inlineCallbacks
def get_state_ids_for_group(self, state_group):
"""Get the event IDs of all the state in the given state group
Args:
state_group (int)
Returns:
Deferred[dict]: Resolves to a map of (type, state_key) -> event_id
"""
group_to_state = yield self._get_state_for_groups((state_group,))
return group_to_state[state_group]
@defer.inlineCallbacks
def get_state_groups(self, room_id, event_ids):
""" Get the state groups for the given list of event_ids
Returns:
Deferred[dict[int, list[EventBase]]]:
dict of state_group_id -> list of state events.
"""
if not event_ids:
return {}
group_to_ids = yield self.get_state_groups_ids(room_id, event_ids)
state_event_map = yield self.get_events(
[
ev_id
for group_ids in itervalues(group_to_ids)
for ev_id in itervalues(group_ids)
],
get_prev_content=False,
)
return {
group: [
state_event_map[v]
for v in itervalues(event_id_map)
if v in state_event_map
]
for group, event_id_map in iteritems(group_to_ids)
}
@defer.inlineCallbacks
def _get_state_groups_from_groups(self, groups, state_filter):
"""Returns the state groups for a given set of groups, filtering on
types of state events.
Args:
groups(list[int]): list of state group IDs to query
state_filter (StateFilter): The state filter used to fetch state
from the database.
Returns:
Deferred[dict[int, dict[tuple[str, str], str]]]:
dict of state_group_id -> (dict of (type, state_key) -> event id)
"""
results = {}
chunks = [groups[i : i + 100] for i in range(0, len(groups), 100)]
for chunk in chunks:
res = yield self.db.runInteraction(
"_get_state_groups_from_groups",
self._get_state_groups_from_groups_txn,
chunk,
state_filter,
)
results.update(res)
return results
@defer.inlineCallbacks
def get_state_for_events(self, event_ids, state_filter=StateFilter.all()):
"""Given a list of event_ids and type tuples, return a list of state
dicts for each event.
Args:
event_ids (list[string])
state_filter (StateFilter): The state filter used to fetch state
from the database.
Returns:
deferred: A dict of (event_id) -> (type, state_key) -> [state_events]
"""
event_to_groups = yield self._get_state_group_for_events(event_ids)
groups = set(itervalues(event_to_groups))
group_to_state = yield self._get_state_for_groups(groups, state_filter)
state_event_map = yield self.get_events(
[ev_id for sd in itervalues(group_to_state) for ev_id in itervalues(sd)],
get_prev_content=False,
)
event_to_state = {
event_id: {
k: state_event_map[v]
for k, v in iteritems(group_to_state[group])
if v in state_event_map
}
for event_id, group in iteritems(event_to_groups)
}
return {event: event_to_state[event] for event in event_ids}
@defer.inlineCallbacks
def get_state_ids_for_events(self, event_ids, state_filter=StateFilter.all()):
"""
Get the state dicts corresponding to a list of events, containing the event_ids
of the state events (as opposed to the events themselves)
Args:
event_ids(list(str)): events whose state should be returned
state_filter (StateFilter): The state filter used to fetch state
from the database.
Returns:
A deferred dict from event_id -> (type, state_key) -> event_id
"""
event_to_groups = yield self._get_state_group_for_events(event_ids)
groups = set(itervalues(event_to_groups))
group_to_state = yield self._get_state_for_groups(groups, state_filter)
event_to_state = {
event_id: group_to_state[group]
for event_id, group in iteritems(event_to_groups)
}
return {event: event_to_state[event] for event in event_ids}
@defer.inlineCallbacks
def get_state_for_event(self, event_id, state_filter=StateFilter.all()):
"""
Get the state dict corresponding to a particular event
Args:
event_id(str): event whose state should be returned
state_filter (StateFilter): The state filter used to fetch state
from the database.
Returns:
A deferred dict from (type, state_key) -> state_event
"""
state_map = yield self.get_state_for_events([event_id], state_filter)
return state_map[event_id]
@defer.inlineCallbacks
def get_state_ids_for_event(self, event_id, state_filter=StateFilter.all()):
"""
Get the state dict corresponding to a particular event
Args:
event_id(str): event whose state should be returned
state_filter (StateFilter): The state filter used to fetch state
from the database.
Returns:
A deferred dict from (type, state_key) -> state_event
"""
state_map = yield self.get_state_ids_for_events([event_id], state_filter)
return state_map[event_id]
@cached(max_entries=50000) @cached(max_entries=50000)
def _get_state_group_for_event(self, event_id): def _get_state_group_for_event(self, event_id):
return self.db.simple_select_one_onecol( return self.db.simple_select_one_onecol(
@ -684,329 +265,6 @@ class StateGroupWorkerStore(
return {row["event_id"]: row["state_group"] for row in rows} return {row["event_id"]: row["state_group"] for row in rows}
def _get_state_for_group_using_cache(self, cache, group, state_filter):
"""Checks if group is in cache. See `_get_state_for_groups`
Args:
cache(DictionaryCache): the state group cache to use
group(int): The state group to lookup
state_filter (StateFilter): The state filter used to fetch state
from the database.
Returns 2-tuple (`state_dict`, `got_all`).
`got_all` is a bool indicating if we successfully retrieved all
requests state from the cache, if False we need to query the DB for the
missing state.
"""
is_all, known_absent, state_dict_ids = cache.get(group)
if is_all or state_filter.is_full():
# Either we have everything or want everything, either way
# `is_all` tells us whether we've gotten everything.
return state_filter.filter_state(state_dict_ids), is_all
# tracks whether any of our requested types are missing from the cache
missing_types = False
if state_filter.has_wildcards():
# We don't know if we fetched all the state keys for the types in
# the filter that are wildcards, so we have to assume that we may
# have missed some.
missing_types = True
else:
# There aren't any wild cards, so `concrete_types()` returns the
# complete list of event types we're wanting.
for key in state_filter.concrete_types():
if key not in state_dict_ids and key not in known_absent:
missing_types = True
break
return state_filter.filter_state(state_dict_ids), not missing_types
@defer.inlineCallbacks
def _get_state_for_groups(self, groups, state_filter=StateFilter.all()):
"""Gets the state at each of a list of state groups, optionally
filtering by type/state_key
Args:
groups (iterable[int]): list of state groups for which we want
to get the state.
state_filter (StateFilter): The state filter used to fetch state
from the database.
Returns:
Deferred[dict[int, dict[tuple[str, str], str]]]:
dict of state_group_id -> (dict of (type, state_key) -> event id)
"""
member_filter, non_member_filter = state_filter.get_member_split()
# Now we look them up in the member and non-member caches
(
non_member_state,
incomplete_groups_nm,
) = yield self._get_state_for_groups_using_cache(
groups, self._state_group_cache, state_filter=non_member_filter
)
(
member_state,
incomplete_groups_m,
) = yield self._get_state_for_groups_using_cache(
groups, self._state_group_members_cache, state_filter=member_filter
)
state = dict(non_member_state)
for group in groups:
state[group].update(member_state[group])
# Now fetch any missing groups from the database
incomplete_groups = incomplete_groups_m | incomplete_groups_nm
if not incomplete_groups:
return state
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()
group_to_state_dict = yield 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 iteritems(group_to_state_dict):
# 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(self, groups, cache, state_filter):
"""Gets the state at each of a list of state groups, optionally
filtering by type/state_key, querying from a specific cache.
Args:
groups (iterable[int]): list of state groups for which we want
to get the state.
cache (DictionaryCache): the cache of group ids to state dicts which
we will pass through - either the normal state cache or the specific
members state cache.
state_filter (StateFilter): The state filter used to fetch state
from the database.
Returns:
tuple[dict[int, dict[tuple[str, str], str]], set[int]]: Tuple of
dict of state_group_id -> (dict of (type, state_key) -> event id)
of entries in the cache, and the state group ids either missing
from the cache or incomplete.
"""
results = {}
incomplete_groups = set()
for group in set(groups):
state_dict_ids, got_all = self._get_state_for_group_using_cache(
cache, group, state_filter
)
results[group] = state_dict_ids
if not got_all:
incomplete_groups.add(group)
return results, incomplete_groups
def _insert_into_cache(
self,
group_to_state_dict,
state_filter,
cache_seq_num_members,
cache_seq_num_non_members,
):
"""Inserts results from querying the database into the relevant cache.
Args:
group_to_state_dict (dict): The new entries pulled from database.
Map from state group to state dict
state_filter (StateFilter): The state filter used to fetch state
from the database.
cache_seq_num_members (int): Sequence number of member cache since
last lookup in cache
cache_seq_num_non_members (int): Sequence number of member cache since
last lookup in cache
"""
# We need to work out which types we've fetched from the DB for the
# member vs non-member caches. This should be as accurate as possible,
# but can be an underestimate (e.g. when we have wild cards)
member_filter, non_member_filter = state_filter.get_member_split()
if member_filter.is_full():
# We fetched all member events
member_types = None
else:
# `concrete_types()` will only return a subset when there are wild
# cards in the filter, but that's fine.
member_types = member_filter.concrete_types()
if non_member_filter.is_full():
# We fetched all non member events
non_member_types = None
else:
non_member_types = non_member_filter.concrete_types()
for group, group_state_dict in iteritems(group_to_state_dict):
state_dict_members = {}
state_dict_non_members = {}
for k, v in iteritems(group_state_dict):
if k[0] == EventTypes.Member:
state_dict_members[k] = v
else:
state_dict_non_members[k] = v
self._state_group_members_cache.update(
cache_seq_num_members,
key=group,
value=state_dict_members,
fetched_keys=member_types,
)
self._state_group_cache.update(
cache_seq_num_non_members,
key=group,
value=state_dict_non_members,
fetched_keys=non_member_types,
)
def store_state_group(
self, event_id, room_id, prev_group, delta_ids, current_state_ids
):
"""Store a new set of state, returning a newly assigned state group.
Args:
event_id (str): The event ID for which the state was calculated
room_id (str)
prev_group (int|None): A previous state group for the room, optional.
delta_ids (dict|None): The delta between state at `prev_group` and
`current_state_ids`, if `prev_group` was given. Same format as
`current_state_ids`.
current_state_ids (dict): The state to store. Map of (type, state_key)
to event_id.
Returns:
Deferred[int]: The state group ID
"""
def _store_state_group_txn(txn):
if current_state_ids is None:
# AFAIK, this can never happen
raise Exception("current_state_ids cannot be None")
state_group = self.database_engine.get_next_state_group_id(txn)
self.db.simple_insert_txn(
txn,
table="state_groups",
values={"id": state_group, "room_id": room_id, "event_id": event_id},
)
# We persist as a delta if we can, while also ensuring the chain
# of deltas isn't tooo long, as otherwise read performance degrades.
if prev_group:
is_in_db = self.db.simple_select_one_onecol_txn(
txn,
table="state_groups",
keyvalues={"id": prev_group},
retcol="id",
allow_none=True,
)
if not is_in_db:
raise Exception(
"Trying to persist state with unpersisted prev_group: %r"
% (prev_group,)
)
potential_hops = self._count_state_group_hops_txn(txn, prev_group)
if prev_group and potential_hops < MAX_STATE_DELTA_HOPS:
self.db.simple_insert_txn(
txn,
table="state_group_edges",
values={"state_group": state_group, "prev_state_group": prev_group},
)
self.db.simple_insert_many_txn(
txn,
table="state_groups_state",
values=[
{
"state_group": state_group,
"room_id": room_id,
"type": key[0],
"state_key": key[1],
"event_id": state_id,
}
for key, state_id in iteritems(delta_ids)
],
)
else:
self.db.simple_insert_many_txn(
txn,
table="state_groups_state",
values=[
{
"state_group": state_group,
"room_id": room_id,
"type": key[0],
"state_key": key[1],
"event_id": state_id,
}
for key, state_id in iteritems(current_state_ids)
],
)
# Prefill the state group caches with this group.
# It's fine to use the sequence like this as the state group map
# is immutable. (If the map wasn't immutable then this prefill could
# race with another update)
current_member_state_ids = {
s: ev
for (s, ev) in iteritems(current_state_ids)
if s[0] == EventTypes.Member
}
txn.call_after(
self._state_group_members_cache.update,
self._state_group_members_cache.sequence,
key=state_group,
value=dict(current_member_state_ids),
)
current_non_member_state_ids = {
s: ev
for (s, ev) in iteritems(current_state_ids)
if s[0] != EventTypes.Member
}
txn.call_after(
self._state_group_cache.update,
self._state_group_cache.sequence,
key=state_group,
value=dict(current_non_member_state_ids),
)
return state_group
return self.db.runInteraction("store_state_group", _store_state_group_txn)
@defer.inlineCallbacks @defer.inlineCallbacks
def get_referenced_state_groups(self, state_groups): def get_referenced_state_groups(self, state_groups):
"""Check if the state groups are referenced by events. """Check if the state groups are referenced by events.
@ -1031,22 +289,14 @@ class StateGroupWorkerStore(
return set(row["state_group"] for row in rows) return set(row["state_group"] for row in rows)
class StateBackgroundUpdateStore(StateGroupBackgroundUpdateStore): class MainStateBackgroundUpdateStore(SQLBaseStore):
STATE_GROUP_DEDUPLICATION_UPDATE_NAME = "state_group_state_deduplication"
STATE_GROUP_INDEX_UPDATE_NAME = "state_group_state_type_index"
CURRENT_STATE_INDEX_UPDATE_NAME = "current_state_members_idx" CURRENT_STATE_INDEX_UPDATE_NAME = "current_state_members_idx"
EVENT_STATE_GROUP_INDEX_UPDATE_NAME = "event_to_state_groups_sg_index" EVENT_STATE_GROUP_INDEX_UPDATE_NAME = "event_to_state_groups_sg_index"
def __init__(self, database: Database, db_conn, hs): def __init__(self, database: Database, db_conn, hs):
super(StateBackgroundUpdateStore, self).__init__(database, db_conn, hs) super(MainStateBackgroundUpdateStore, self).__init__(database, db_conn, hs)
self.db.updates.register_background_update_handler(
self.STATE_GROUP_DEDUPLICATION_UPDATE_NAME,
self._background_deduplicate_state,
)
self.db.updates.register_background_update_handler(
self.STATE_GROUP_INDEX_UPDATE_NAME, self._background_index_state
)
self.db.updates.register_background_index_update( self.db.updates.register_background_index_update(
self.CURRENT_STATE_INDEX_UPDATE_NAME, self.CURRENT_STATE_INDEX_UPDATE_NAME,
index_name="current_state_events_member_index", index_name="current_state_events_member_index",
@ -1061,181 +311,8 @@ class StateBackgroundUpdateStore(StateGroupBackgroundUpdateStore):
columns=["state_group"], columns=["state_group"],
) )
@defer.inlineCallbacks
def _background_deduplicate_state(self, progress, batch_size):
"""This background update will slowly deduplicate state by reencoding
them as deltas.
"""
last_state_group = progress.get("last_state_group", 0)
rows_inserted = progress.get("rows_inserted", 0)
max_group = progress.get("max_group", None)
BATCH_SIZE_SCALE_FACTOR = 100 class StateStore(StateGroupWorkerStore, MainStateBackgroundUpdateStore):
batch_size = max(1, int(batch_size / BATCH_SIZE_SCALE_FACTOR))
if max_group is None:
rows = yield self.db.execute(
"_background_deduplicate_state",
None,
"SELECT coalesce(max(id), 0) FROM state_groups",
)
max_group = rows[0][0]
def reindex_txn(txn):
new_last_state_group = last_state_group
for count in range(batch_size):
txn.execute(
"SELECT id, room_id FROM state_groups"
" WHERE ? < id AND id <= ?"
" ORDER BY id ASC"
" LIMIT 1",
(new_last_state_group, max_group),
)
row = txn.fetchone()
if row:
state_group, room_id = row
if not row or not state_group:
return True, count
txn.execute(
"SELECT state_group FROM state_group_edges"
" WHERE state_group = ?",
(state_group,),
)
# If we reach a point where we've already started inserting
# edges we should stop.
if txn.fetchall():
return True, count
txn.execute(
"SELECT coalesce(max(id), 0) FROM state_groups"
" WHERE id < ? AND room_id = ?",
(state_group, room_id),
)
(prev_group,) = txn.fetchone()
new_last_state_group = state_group
if prev_group:
potential_hops = self._count_state_group_hops_txn(txn, prev_group)
if potential_hops >= MAX_STATE_DELTA_HOPS:
# We want to ensure chains are at most this long,#
# otherwise read performance degrades.
continue
prev_state = self._get_state_groups_from_groups_txn(
txn, [prev_group]
)
prev_state = prev_state[prev_group]
curr_state = self._get_state_groups_from_groups_txn(
txn, [state_group]
)
curr_state = curr_state[state_group]
if not set(prev_state.keys()) - set(curr_state.keys()):
# We can only do a delta if the current has a strict super set
# of keys
delta_state = {
key: value
for key, value in iteritems(curr_state)
if prev_state.get(key, None) != value
}
self.db.simple_delete_txn(
txn,
table="state_group_edges",
keyvalues={"state_group": state_group},
)
self.db.simple_insert_txn(
txn,
table="state_group_edges",
values={
"state_group": state_group,
"prev_state_group": prev_group,
},
)
self.db.simple_delete_txn(
txn,
table="state_groups_state",
keyvalues={"state_group": state_group},
)
self.db.simple_insert_many_txn(
txn,
table="state_groups_state",
values=[
{
"state_group": state_group,
"room_id": room_id,
"type": key[0],
"state_key": key[1],
"event_id": state_id,
}
for key, state_id in iteritems(delta_state)
],
)
progress = {
"last_state_group": state_group,
"rows_inserted": rows_inserted + batch_size,
"max_group": max_group,
}
self.db.updates._background_update_progress_txn(
txn, self.STATE_GROUP_DEDUPLICATION_UPDATE_NAME, progress
)
return False, batch_size
finished, result = yield self.db.runInteraction(
self.STATE_GROUP_DEDUPLICATION_UPDATE_NAME, reindex_txn
)
if finished:
yield self.db.updates._end_background_update(
self.STATE_GROUP_DEDUPLICATION_UPDATE_NAME
)
return result * BATCH_SIZE_SCALE_FACTOR
@defer.inlineCallbacks
def _background_index_state(self, progress, batch_size):
def reindex_txn(conn):
conn.rollback()
if isinstance(self.database_engine, PostgresEngine):
# postgres insists on autocommit for the index
conn.set_session(autocommit=True)
try:
txn = conn.cursor()
txn.execute(
"CREATE INDEX CONCURRENTLY state_groups_state_type_idx"
" ON state_groups_state(state_group, type, state_key)"
)
txn.execute("DROP INDEX IF EXISTS state_groups_state_id")
finally:
conn.set_session(autocommit=False)
else:
txn = conn.cursor()
txn.execute(
"CREATE INDEX state_groups_state_type_idx"
" ON state_groups_state(state_group, type, state_key)"
)
txn.execute("DROP INDEX IF EXISTS state_groups_state_id")
yield self.db.runWithConnection(reindex_txn)
yield self.db.updates._end_background_update(self.STATE_GROUP_INDEX_UPDATE_NAME)
return 1
class StateStore(StateGroupWorkerStore, StateBackgroundUpdateStore):
""" Keeps track of the state at a given event. """ Keeps track of the state at a given event.
This is done by the concept of `state groups`. Every event is a assigned This is done by the concept of `state groups`. Every event is a assigned

View file

@ -0,0 +1,16 @@
# -*- coding: utf-8 -*-
# Copyright 2019 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.
from synapse.storage.data_stores.state.store import StateGroupDataStore # noqa: F401

View file

@ -0,0 +1,374 @@
# -*- coding: utf-8 -*-
# Copyright 2014-2016 OpenMarket Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
from six import iteritems
from twisted.internet import defer
from synapse.storage._base import SQLBaseStore
from synapse.storage.database import Database
from synapse.storage.engines import PostgresEngine
from synapse.storage.state import StateFilter
logger = logging.getLogger(__name__)
MAX_STATE_DELTA_HOPS = 100
class StateGroupBackgroundUpdateStore(SQLBaseStore):
"""Defines functions related to state groups needed to run the state backgroud
updates.
"""
def _count_state_group_hops_txn(self, txn, state_group):
"""Given a state group, count how many hops there are in the tree.
This is used to ensure the delta chains don't get too long.
"""
if isinstance(self.database_engine, PostgresEngine):
sql = """
WITH RECURSIVE state(state_group) AS (
VALUES(?::bigint)
UNION ALL
SELECT prev_state_group FROM state_group_edges e, state s
WHERE s.state_group = e.state_group
)
SELECT count(*) FROM state;
"""
txn.execute(sql, (state_group,))
row = txn.fetchone()
if row and row[0]:
return row[0]
else:
return 0
else:
# We don't use WITH RECURSIVE on sqlite3 as there are distributions
# that ship with an sqlite3 version that doesn't support it (e.g. wheezy)
next_group = state_group
count = 0
while next_group:
next_group = self.db.simple_select_one_onecol_txn(
txn,
table="state_group_edges",
keyvalues={"state_group": next_group},
retcol="prev_state_group",
allow_none=True,
)
if next_group:
count += 1
return count
def _get_state_groups_from_groups_txn(
self, txn, groups, state_filter=StateFilter.all()
):
results = {group: {} for group in groups}
where_clause, where_args = state_filter.make_sql_filter_clause()
# Unless the filter clause is empty, we're going to append it after an
# existing where clause
if where_clause:
where_clause = " AND (%s)" % (where_clause,)
if isinstance(self.database_engine, PostgresEngine):
# Temporarily disable sequential scans in this transaction. This is
# a temporary hack until we can add the right indices in
txn.execute("SET LOCAL enable_seqscan=off")
# The below query walks the state_group tree so that the "state"
# table includes all state_groups in the tree. It then joins
# against `state_groups_state` to fetch the latest state.
# It assumes that previous state groups are always numerically
# lesser.
# The PARTITION is used to get the event_id in the greatest state
# group for the given type, state_key.
# This may return multiple rows per (type, state_key), but last_value
# should be the same.
sql = """
WITH RECURSIVE state(state_group) AS (
VALUES(?::bigint)
UNION ALL
SELECT prev_state_group FROM state_group_edges e, state s
WHERE s.state_group = e.state_group
)
SELECT DISTINCT type, state_key, last_value(event_id) OVER (
PARTITION BY type, state_key ORDER BY state_group ASC
ROWS BETWEEN UNBOUNDED PRECEDING AND UNBOUNDED FOLLOWING
) AS event_id FROM state_groups_state
WHERE state_group IN (
SELECT state_group FROM state
)
"""
for group in groups:
args = [group]
args.extend(where_args)
txn.execute(sql + where_clause, args)
for row in txn:
typ, state_key, event_id = row
key = (typ, state_key)
results[group][key] = event_id
else:
max_entries_returned = state_filter.max_entries_returned()
# We don't use WITH RECURSIVE on sqlite3 as there are distributions
# that ship with an sqlite3 version that doesn't support it (e.g. wheezy)
for group in groups:
next_group = group
while next_group:
# We did this before by getting the list of group ids, and
# then passing that list to sqlite to get latest event for
# each (type, state_key). However, that was terribly slow
# without the right indices (which we can't add until
# after we finish deduping state, which requires this func)
args = [next_group]
args.extend(where_args)
txn.execute(
"SELECT type, state_key, event_id FROM state_groups_state"
" WHERE state_group = ? " + where_clause,
args,
)
results[group].update(
((typ, state_key), event_id)
for typ, state_key, event_id in txn
if (typ, state_key) not in results[group]
)
# If the number of entries in the (type,state_key)->event_id dict
# matches the number of (type,state_keys) types we were searching
# for, then we must have found them all, so no need to go walk
# further down the tree... UNLESS our types filter contained
# wildcards (i.e. Nones) in which case we have to do an exhaustive
# search
if (
max_entries_returned is not None
and len(results[group]) == max_entries_returned
):
break
next_group = self.db.simple_select_one_onecol_txn(
txn,
table="state_group_edges",
keyvalues={"state_group": next_group},
retcol="prev_state_group",
allow_none=True,
)
return results
class StateBackgroundUpdateStore(StateGroupBackgroundUpdateStore):
STATE_GROUP_DEDUPLICATION_UPDATE_NAME = "state_group_state_deduplication"
STATE_GROUP_INDEX_UPDATE_NAME = "state_group_state_type_index"
STATE_GROUPS_ROOM_INDEX_UPDATE_NAME = "state_groups_room_id_idx"
def __init__(self, database: Database, db_conn, hs):
super(StateBackgroundUpdateStore, self).__init__(database, db_conn, hs)
self.db.updates.register_background_update_handler(
self.STATE_GROUP_DEDUPLICATION_UPDATE_NAME,
self._background_deduplicate_state,
)
self.db.updates.register_background_update_handler(
self.STATE_GROUP_INDEX_UPDATE_NAME, self._background_index_state
)
self.db.updates.register_background_index_update(
self.STATE_GROUPS_ROOM_INDEX_UPDATE_NAME,
index_name="state_groups_room_id_idx",
table="state_groups",
columns=["room_id"],
)
@defer.inlineCallbacks
def _background_deduplicate_state(self, progress, batch_size):
"""This background update will slowly deduplicate state by reencoding
them as deltas.
"""
last_state_group = progress.get("last_state_group", 0)
rows_inserted = progress.get("rows_inserted", 0)
max_group = progress.get("max_group", None)
BATCH_SIZE_SCALE_FACTOR = 100
batch_size = max(1, int(batch_size / BATCH_SIZE_SCALE_FACTOR))
if max_group is None:
rows = yield self.db.execute(
"_background_deduplicate_state",
None,
"SELECT coalesce(max(id), 0) FROM state_groups",
)
max_group = rows[0][0]
def reindex_txn(txn):
new_last_state_group = last_state_group
for count in range(batch_size):
txn.execute(
"SELECT id, room_id FROM state_groups"
" WHERE ? < id AND id <= ?"
" ORDER BY id ASC"
" LIMIT 1",
(new_last_state_group, max_group),
)
row = txn.fetchone()
if row:
state_group, room_id = row
if not row or not state_group:
return True, count
txn.execute(
"SELECT state_group FROM state_group_edges"
" WHERE state_group = ?",
(state_group,),
)
# If we reach a point where we've already started inserting
# edges we should stop.
if txn.fetchall():
return True, count
txn.execute(
"SELECT coalesce(max(id), 0) FROM state_groups"
" WHERE id < ? AND room_id = ?",
(state_group, room_id),
)
(prev_group,) = txn.fetchone()
new_last_state_group = state_group
if prev_group:
potential_hops = self._count_state_group_hops_txn(txn, prev_group)
if potential_hops >= MAX_STATE_DELTA_HOPS:
# We want to ensure chains are at most this long,#
# otherwise read performance degrades.
continue
prev_state = self._get_state_groups_from_groups_txn(
txn, [prev_group]
)
prev_state = prev_state[prev_group]
curr_state = self._get_state_groups_from_groups_txn(
txn, [state_group]
)
curr_state = curr_state[state_group]
if not set(prev_state.keys()) - set(curr_state.keys()):
# We can only do a delta if the current has a strict super set
# of keys
delta_state = {
key: value
for key, value in iteritems(curr_state)
if prev_state.get(key, None) != value
}
self.db.simple_delete_txn(
txn,
table="state_group_edges",
keyvalues={"state_group": state_group},
)
self.db.simple_insert_txn(
txn,
table="state_group_edges",
values={
"state_group": state_group,
"prev_state_group": prev_group,
},
)
self.db.simple_delete_txn(
txn,
table="state_groups_state",
keyvalues={"state_group": state_group},
)
self.db.simple_insert_many_txn(
txn,
table="state_groups_state",
values=[
{
"state_group": state_group,
"room_id": room_id,
"type": key[0],
"state_key": key[1],
"event_id": state_id,
}
for key, state_id in iteritems(delta_state)
],
)
progress = {
"last_state_group": state_group,
"rows_inserted": rows_inserted + batch_size,
"max_group": max_group,
}
self.db.updates._background_update_progress_txn(
txn, self.STATE_GROUP_DEDUPLICATION_UPDATE_NAME, progress
)
return False, batch_size
finished, result = yield self.db.runInteraction(
self.STATE_GROUP_DEDUPLICATION_UPDATE_NAME, reindex_txn
)
if finished:
yield self.db.updates._end_background_update(
self.STATE_GROUP_DEDUPLICATION_UPDATE_NAME
)
return result * BATCH_SIZE_SCALE_FACTOR
@defer.inlineCallbacks
def _background_index_state(self, progress, batch_size):
def reindex_txn(conn):
conn.rollback()
if isinstance(self.database_engine, PostgresEngine):
# postgres insists on autocommit for the index
conn.set_session(autocommit=True)
try:
txn = conn.cursor()
txn.execute(
"CREATE INDEX CONCURRENTLY state_groups_state_type_idx"
" ON state_groups_state(state_group, type, state_key)"
)
txn.execute("DROP INDEX IF EXISTS state_groups_state_id")
finally:
conn.set_session(autocommit=False)
else:
txn = conn.cursor()
txn.execute(
"CREATE INDEX state_groups_state_type_idx"
" ON state_groups_state(state_group, type, state_key)"
)
txn.execute("DROP INDEX IF EXISTS state_groups_state_id")
yield self.db.runWithConnection(reindex_txn)
yield self.db.updates._end_background_update(self.STATE_GROUP_INDEX_UPDATE_NAME)
return 1

View file

@ -0,0 +1,19 @@
/* Copyright 2016 OpenMarket Ltd
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
-- The following indices are redundant, other indices are equivalent or
-- supersets
DROP INDEX IF EXISTS state_groups_id; -- Duplicate of PRIMARY KEY

View file

@ -0,0 +1,17 @@
/* Copyright 2019 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.
*/
INSERT INTO background_updates (update_name, progress_json) VALUES
('state_groups_room_id_idx', '{}');

View file

@ -0,0 +1,37 @@
/* Copyright 2019 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.
*/
CREATE TABLE state_groups (
id BIGINT PRIMARY KEY,
room_id TEXT NOT NULL,
event_id TEXT NOT NULL
);
CREATE TABLE state_groups_state (
state_group BIGINT NOT NULL,
room_id TEXT NOT NULL,
type TEXT NOT NULL,
state_key TEXT NOT NULL,
event_id TEXT NOT NULL
);
CREATE TABLE state_group_edges (
state_group BIGINT NOT NULL,
prev_state_group BIGINT NOT NULL
);
CREATE INDEX state_group_edges_idx ON state_group_edges (state_group);
CREATE INDEX state_group_edges_prev_idx ON state_group_edges (prev_state_group);
CREATE INDEX state_groups_state_type_idx ON state_groups_state (state_group, type, state_key);

View file

@ -0,0 +1,21 @@
/* Copyright 2019 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.
*/
CREATE SEQUENCE state_group_id_seq
START WITH 1
INCREMENT BY 1
NO MINVALUE
NO MAXVALUE
CACHE 1;

View file

@ -0,0 +1,640 @@
# -*- coding: utf-8 -*-
# Copyright 2014-2016 OpenMarket Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import logging
from collections import namedtuple
from six import iteritems
from six.moves import range
from twisted.internet import defer
from synapse.api.constants import EventTypes
from synapse.storage._base import SQLBaseStore
from synapse.storage.data_stores.state.bg_updates import StateBackgroundUpdateStore
from synapse.storage.database import Database
from synapse.storage.state import StateFilter
from synapse.util.caches import get_cache_factor_for
from synapse.util.caches.descriptors import cached
from synapse.util.caches.dictionary_cache import DictionaryCache
logger = logging.getLogger(__name__)
MAX_STATE_DELTA_HOPS = 100
class _GetStateGroupDelta(
namedtuple("_GetStateGroupDelta", ("prev_group", "delta_ids"))
):
"""Return type of get_state_group_delta that implements __len__, which lets
us use the itrable flag when caching
"""
__slots__ = []
def __len__(self):
return len(self.delta_ids) if self.delta_ids else 0
class StateGroupDataStore(StateBackgroundUpdateStore, SQLBaseStore):
"""A data store for fetching/storing state groups.
"""
def __init__(self, database: Database, db_conn, hs):
super(StateGroupDataStore, self).__init__(database, db_conn, hs)
# Originally the state store used a single DictionaryCache to cache the
# event IDs for the state types in a given state group to avoid hammering
# on the state_group* tables.
#
# The point of using a DictionaryCache is that it can cache a subset
# of the state events for a given state group (i.e. a subset of the keys for a
# given dict which is an entry in the cache for a given state group ID).
#
# However, this poses problems when performing complicated queries
# on the store - for instance: "give me all the state for this group, but
# limit members to this subset of users", as DictionaryCache's API isn't
# rich enough to say "please cache any of these fields, apart from this subset".
# This is problematic when lazy loading members, which requires this behaviour,
# as without it the cache has no choice but to speculatively load all
# state events for the group, which negates the efficiency being sought.
#
# Rather than overcomplicating DictionaryCache's API, we instead split the
# state_group_cache into two halves - one for tracking non-member events,
# and the other for tracking member_events. This means that lazy loading
# queries can be made in a cache-friendly manner by querying both caches
# separately and then merging the result. So for the example above, you
# would query the members cache for a specific subset of state keys
# (which DictionaryCache will handle efficiently and fine) and the non-members
# cache for all state (which DictionaryCache will similarly handle fine)
# and then just merge the results together.
#
# We size the non-members cache to be smaller than the members cache as the
# vast majority of state in Matrix (today) is member events.
self._state_group_cache = DictionaryCache(
"*stateGroupCache*",
# TODO: this hasn't been tuned yet
50000 * get_cache_factor_for("stateGroupCache"),
)
self._state_group_members_cache = DictionaryCache(
"*stateGroupMembersCache*",
500000 * get_cache_factor_for("stateGroupMembersCache"),
)
@cached(max_entries=10000, iterable=True)
def get_state_group_delta(self, state_group):
"""Given a state group try to return a previous group and a delta between
the old and the new.
Returns:
(prev_group, delta_ids), where both may be None.
"""
def _get_state_group_delta_txn(txn):
prev_group = self.db.simple_select_one_onecol_txn(
txn,
table="state_group_edges",
keyvalues={"state_group": state_group},
retcol="prev_state_group",
allow_none=True,
)
if not prev_group:
return _GetStateGroupDelta(None, None)
delta_ids = self.db.simple_select_list_txn(
txn,
table="state_groups_state",
keyvalues={"state_group": state_group},
retcols=("type", "state_key", "event_id"),
)
return _GetStateGroupDelta(
prev_group,
{(row["type"], row["state_key"]): row["event_id"] for row in delta_ids},
)
return self.db.runInteraction(
"get_state_group_delta", _get_state_group_delta_txn
)
@defer.inlineCallbacks
def _get_state_groups_from_groups(self, groups, state_filter):
"""Returns the state groups for a given set of groups, filtering on
types of state events.
Args:
groups(list[int]): list of state group IDs to query
state_filter (StateFilter): The state filter used to fetch state
from the database.
Returns:
Deferred[dict[int, dict[tuple[str, str], str]]]:
dict of state_group_id -> (dict of (type, state_key) -> event id)
"""
results = {}
chunks = [groups[i : i + 100] for i in range(0, len(groups), 100)]
for chunk in chunks:
res = yield self.db.runInteraction(
"_get_state_groups_from_groups",
self._get_state_groups_from_groups_txn,
chunk,
state_filter,
)
results.update(res)
return results
def _get_state_for_group_using_cache(self, cache, group, state_filter):
"""Checks if group is in cache. See `_get_state_for_groups`
Args:
cache(DictionaryCache): the state group cache to use
group(int): The state group to lookup
state_filter (StateFilter): The state filter used to fetch state
from the database.
Returns 2-tuple (`state_dict`, `got_all`).
`got_all` is a bool indicating if we successfully retrieved all
requests state from the cache, if False we need to query the DB for the
missing state.
"""
is_all, known_absent, state_dict_ids = cache.get(group)
if is_all or state_filter.is_full():
# Either we have everything or want everything, either way
# `is_all` tells us whether we've gotten everything.
return state_filter.filter_state(state_dict_ids), is_all
# tracks whether any of our requested types are missing from the cache
missing_types = False
if state_filter.has_wildcards():
# We don't know if we fetched all the state keys for the types in
# the filter that are wildcards, so we have to assume that we may
# have missed some.
missing_types = True
else:
# There aren't any wild cards, so `concrete_types()` returns the
# complete list of event types we're wanting.
for key in state_filter.concrete_types():
if key not in state_dict_ids and key not in known_absent:
missing_types = True
break
return state_filter.filter_state(state_dict_ids), not missing_types
@defer.inlineCallbacks
def _get_state_for_groups(self, groups, state_filter=StateFilter.all()):
"""Gets the state at each of a list of state groups, optionally
filtering by type/state_key
Args:
groups (iterable[int]): list of state groups for which we want
to get the state.
state_filter (StateFilter): The state filter used to fetch state
from the database.
Returns:
Deferred[dict[int, dict[tuple[str, str], str]]]:
dict of state_group_id -> (dict of (type, state_key) -> event id)
"""
member_filter, non_member_filter = state_filter.get_member_split()
# Now we look them up in the member and non-member caches
(
non_member_state,
incomplete_groups_nm,
) = yield self._get_state_for_groups_using_cache(
groups, self._state_group_cache, state_filter=non_member_filter
)
(
member_state,
incomplete_groups_m,
) = yield self._get_state_for_groups_using_cache(
groups, self._state_group_members_cache, state_filter=member_filter
)
state = dict(non_member_state)
for group in groups:
state[group].update(member_state[group])
# Now fetch any missing groups from the database
incomplete_groups = incomplete_groups_m | incomplete_groups_nm
if not incomplete_groups:
return state
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()
group_to_state_dict = yield 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 iteritems(group_to_state_dict):
# 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(self, groups, cache, state_filter):
"""Gets the state at each of a list of state groups, optionally
filtering by type/state_key, querying from a specific cache.
Args:
groups (iterable[int]): list of state groups for which we want
to get the state.
cache (DictionaryCache): the cache of group ids to state dicts which
we will pass through - either the normal state cache or the specific
members state cache.
state_filter (StateFilter): The state filter used to fetch state
from the database.
Returns:
tuple[dict[int, dict[tuple[str, str], str]], set[int]]: Tuple of
dict of state_group_id -> (dict of (type, state_key) -> event id)
of entries in the cache, and the state group ids either missing
from the cache or incomplete.
"""
results = {}
incomplete_groups = set()
for group in set(groups):
state_dict_ids, got_all = self._get_state_for_group_using_cache(
cache, group, state_filter
)
results[group] = state_dict_ids
if not got_all:
incomplete_groups.add(group)
return results, incomplete_groups
def _insert_into_cache(
self,
group_to_state_dict,
state_filter,
cache_seq_num_members,
cache_seq_num_non_members,
):
"""Inserts results from querying the database into the relevant cache.
Args:
group_to_state_dict (dict): The new entries pulled from database.
Map from state group to state dict
state_filter (StateFilter): The state filter used to fetch state
from the database.
cache_seq_num_members (int): Sequence number of member cache since
last lookup in cache
cache_seq_num_non_members (int): Sequence number of member cache since
last lookup in cache
"""
# We need to work out which types we've fetched from the DB for the
# member vs non-member caches. This should be as accurate as possible,
# but can be an underestimate (e.g. when we have wild cards)
member_filter, non_member_filter = state_filter.get_member_split()
if member_filter.is_full():
# We fetched all member events
member_types = None
else:
# `concrete_types()` will only return a subset when there are wild
# cards in the filter, but that's fine.
member_types = member_filter.concrete_types()
if non_member_filter.is_full():
# We fetched all non member events
non_member_types = None
else:
non_member_types = non_member_filter.concrete_types()
for group, group_state_dict in iteritems(group_to_state_dict):
state_dict_members = {}
state_dict_non_members = {}
for k, v in iteritems(group_state_dict):
if k[0] == EventTypes.Member:
state_dict_members[k] = v
else:
state_dict_non_members[k] = v
self._state_group_members_cache.update(
cache_seq_num_members,
key=group,
value=state_dict_members,
fetched_keys=member_types,
)
self._state_group_cache.update(
cache_seq_num_non_members,
key=group,
value=state_dict_non_members,
fetched_keys=non_member_types,
)
def store_state_group(
self, event_id, room_id, prev_group, delta_ids, current_state_ids
):
"""Store a new set of state, returning a newly assigned state group.
Args:
event_id (str): The event ID for which the state was calculated
room_id (str)
prev_group (int|None): A previous state group for the room, optional.
delta_ids (dict|None): The delta between state at `prev_group` and
`current_state_ids`, if `prev_group` was given. Same format as
`current_state_ids`.
current_state_ids (dict): The state to store. Map of (type, state_key)
to event_id.
Returns:
Deferred[int]: The state group ID
"""
def _store_state_group_txn(txn):
if current_state_ids is None:
# AFAIK, this can never happen
raise Exception("current_state_ids cannot be None")
state_group = self.database_engine.get_next_state_group_id(txn)
self.db.simple_insert_txn(
txn,
table="state_groups",
values={"id": state_group, "room_id": room_id, "event_id": event_id},
)
# We persist as a delta if we can, while also ensuring the chain
# of deltas isn't tooo long, as otherwise read performance degrades.
if prev_group:
is_in_db = self.db.simple_select_one_onecol_txn(
txn,
table="state_groups",
keyvalues={"id": prev_group},
retcol="id",
allow_none=True,
)
if not is_in_db:
raise Exception(
"Trying to persist state with unpersisted prev_group: %r"
% (prev_group,)
)
potential_hops = self._count_state_group_hops_txn(txn, prev_group)
if prev_group and potential_hops < MAX_STATE_DELTA_HOPS:
self.db.simple_insert_txn(
txn,
table="state_group_edges",
values={"state_group": state_group, "prev_state_group": prev_group},
)
self.db.simple_insert_many_txn(
txn,
table="state_groups_state",
values=[
{
"state_group": state_group,
"room_id": room_id,
"type": key[0],
"state_key": key[1],
"event_id": state_id,
}
for key, state_id in iteritems(delta_ids)
],
)
else:
self.db.simple_insert_many_txn(
txn,
table="state_groups_state",
values=[
{
"state_group": state_group,
"room_id": room_id,
"type": key[0],
"state_key": key[1],
"event_id": state_id,
}
for key, state_id in iteritems(current_state_ids)
],
)
# Prefill the state group caches with this group.
# It's fine to use the sequence like this as the state group map
# is immutable. (If the map wasn't immutable then this prefill could
# race with another update)
current_member_state_ids = {
s: ev
for (s, ev) in iteritems(current_state_ids)
if s[0] == EventTypes.Member
}
txn.call_after(
self._state_group_members_cache.update,
self._state_group_members_cache.sequence,
key=state_group,
value=dict(current_member_state_ids),
)
current_non_member_state_ids = {
s: ev
for (s, ev) in iteritems(current_state_ids)
if s[0] != EventTypes.Member
}
txn.call_after(
self._state_group_cache.update,
self._state_group_cache.sequence,
key=state_group,
value=dict(current_non_member_state_ids),
)
return state_group
return self.db.runInteraction("store_state_group", _store_state_group_txn)
def purge_unreferenced_state_groups(
self, room_id: str, state_groups_to_delete
) -> defer.Deferred:
"""Deletes no longer referenced state groups and de-deltas any state
groups that reference them.
Args:
room_id: The room the state groups belong to (must all be in the
same room).
state_groups_to_delete (Collection[int]): Set of all state groups
to delete.
"""
return self.db.runInteraction(
"purge_unreferenced_state_groups",
self._purge_unreferenced_state_groups,
room_id,
state_groups_to_delete,
)
def _purge_unreferenced_state_groups(self, txn, room_id, state_groups_to_delete):
logger.info(
"[purge] found %i state groups to delete", len(state_groups_to_delete)
)
rows = self.db.simple_select_many_txn(
txn,
table="state_group_edges",
column="prev_state_group",
iterable=state_groups_to_delete,
keyvalues={},
retcols=("state_group",),
)
remaining_state_groups = set(
row["state_group"]
for row in rows
if row["state_group"] not in state_groups_to_delete
)
logger.info(
"[purge] de-delta-ing %i remaining state groups",
len(remaining_state_groups),
)
# Now we turn the state groups that reference to-be-deleted state
# groups to non delta versions.
for sg in remaining_state_groups:
logger.info("[purge] de-delta-ing remaining state group %s", sg)
curr_state = self._get_state_groups_from_groups_txn(txn, [sg])
curr_state = curr_state[sg]
self.db.simple_delete_txn(
txn, table="state_groups_state", keyvalues={"state_group": sg}
)
self.db.simple_delete_txn(
txn, table="state_group_edges", keyvalues={"state_group": sg}
)
self.db.simple_insert_many_txn(
txn,
table="state_groups_state",
values=[
{
"state_group": sg,
"room_id": room_id,
"type": key[0],
"state_key": key[1],
"event_id": state_id,
}
for key, state_id in iteritems(curr_state)
],
)
logger.info("[purge] removing redundant state groups")
txn.executemany(
"DELETE FROM state_groups_state WHERE state_group = ?",
((sg,) for sg in state_groups_to_delete),
)
txn.executemany(
"DELETE FROM state_groups WHERE id = ?",
((sg,) for sg in state_groups_to_delete),
)
@defer.inlineCallbacks
def get_previous_state_groups(self, state_groups):
"""Fetch the previous groups of the given state groups.
Args:
state_groups (Iterable[int])
Returns:
Deferred[dict[int, int]]: mapping from state group to previous
state group.
"""
rows = yield self.db.simple_select_many_batch(
table="state_group_edges",
column="prev_state_group",
iterable=state_groups,
keyvalues={},
retcols=("prev_state_group", "state_group"),
desc="get_previous_state_groups",
)
return {row["state_group"]: row["prev_state_group"] for row in rows}
def purge_room_state(self, room_id, state_groups_to_delete):
"""Deletes all record of a room from state tables
Args:
room_id (str):
state_groups_to_delete (list[int]): State groups to delete
"""
return self.db.runInteraction(
"purge_room_state",
self._purge_room_state_txn,
room_id,
state_groups_to_delete,
)
def _purge_room_state_txn(self, txn, room_id, state_groups_to_delete):
# first we have to delete the state groups states
logger.info("[purge] removing %s from state_groups_state", room_id)
self.db.simple_delete_many_txn(
txn,
table="state_groups_state",
column="state_group",
iterable=state_groups_to_delete,
keyvalues={},
)
# ... and the state group edges
logger.info("[purge] removing %s from state_group_edges", room_id)
self.db.simple_delete_many_txn(
txn,
table="state_group_edges",
column="state_group",
iterable=state_groups_to_delete,
keyvalues={},
)
# ... and the state groups
logger.info("[purge] removing %s from state_groups", room_id)
self.db.simple_delete_many_txn(
txn,
table="state_groups",
column="id",
iterable=state_groups_to_delete,
keyvalues={},
)

View file

@ -183,7 +183,7 @@ class EventsPersistenceStorage(object):
# so we use separate variables here even though they point to the same # so we use separate variables here even though they point to the same
# store for now. # store for now.
self.main_store = stores.main self.main_store = stores.main
self.state_store = stores.main self.state_store = stores.state
self._clock = hs.get_clock() self._clock = hs.get_clock()
self.is_mine_id = hs.is_mine_id self.is_mine_id = hs.is_mine_id

View file

@ -42,7 +42,7 @@ class UpgradeDatabaseException(PrepareDatabaseException):
pass pass
def prepare_database(db_conn, database_engine, config, data_stores=["main"]): def prepare_database(db_conn, database_engine, config, data_stores=["main", "state"]):
"""Prepares a database for usage. Will either create all necessary tables """Prepares a database for usage. Will either create all necessary tables
or upgrade from an older schema version. or upgrade from an older schema version.

View file

@ -58,7 +58,7 @@ class PurgeEventsStorage(object):
sg_to_delete = yield self._find_unreferenced_groups(state_groups) sg_to_delete = yield self._find_unreferenced_groups(state_groups)
yield self.stores.main.purge_unreferenced_state_groups(room_id, sg_to_delete) yield self.stores.state.purge_unreferenced_state_groups(room_id, sg_to_delete)
@defer.inlineCallbacks @defer.inlineCallbacks
def _find_unreferenced_groups(self, state_groups): def _find_unreferenced_groups(self, state_groups):
@ -102,7 +102,7 @@ class PurgeEventsStorage(object):
# groups that are referenced. # groups that are referenced.
current_search -= referenced current_search -= referenced
edges = yield self.stores.main.get_previous_state_groups(current_search) edges = yield self.stores.state.get_previous_state_groups(current_search)
prevs = set(edges.values()) prevs = set(edges.values())
# We don't bother re-handling groups we've already seen # We don't bother re-handling groups we've already seen

View file

@ -342,7 +342,7 @@ class StateGroupStorage(object):
(prev_group, delta_ids) (prev_group, delta_ids)
""" """
return self.stores.main.get_state_group_delta(state_group) return self.stores.state.get_state_group_delta(state_group)
@defer.inlineCallbacks @defer.inlineCallbacks
def get_state_groups_ids(self, _room_id, event_ids): def get_state_groups_ids(self, _room_id, event_ids):
@ -362,7 +362,7 @@ class StateGroupStorage(object):
event_to_groups = yield self.stores.main._get_state_group_for_events(event_ids) event_to_groups = yield self.stores.main._get_state_group_for_events(event_ids)
groups = set(itervalues(event_to_groups)) groups = set(itervalues(event_to_groups))
group_to_state = yield self.stores.main._get_state_for_groups(groups) group_to_state = yield self.stores.state._get_state_for_groups(groups)
return group_to_state return group_to_state
@ -423,7 +423,7 @@ class StateGroupStorage(object):
dict of state_group_id -> (dict of (type, state_key) -> event id) dict of state_group_id -> (dict of (type, state_key) -> event id)
""" """
return self.stores.main._get_state_groups_from_groups(groups, state_filter) return self.stores.state._get_state_groups_from_groups(groups, state_filter)
@defer.inlineCallbacks @defer.inlineCallbacks
def get_state_for_events(self, event_ids, state_filter=StateFilter.all()): def get_state_for_events(self, event_ids, state_filter=StateFilter.all()):
@ -439,7 +439,7 @@ class StateGroupStorage(object):
event_to_groups = yield self.stores.main._get_state_group_for_events(event_ids) event_to_groups = yield self.stores.main._get_state_group_for_events(event_ids)
groups = set(itervalues(event_to_groups)) groups = set(itervalues(event_to_groups))
group_to_state = yield self.stores.main._get_state_for_groups( group_to_state = yield self.stores.state._get_state_for_groups(
groups, state_filter groups, state_filter
) )
@ -476,7 +476,7 @@ class StateGroupStorage(object):
event_to_groups = yield self.stores.main._get_state_group_for_events(event_ids) event_to_groups = yield self.stores.main._get_state_group_for_events(event_ids)
groups = set(itervalues(event_to_groups)) groups = set(itervalues(event_to_groups))
group_to_state = yield self.stores.main._get_state_for_groups( group_to_state = yield self.stores.state._get_state_for_groups(
groups, state_filter groups, state_filter
) )
@ -532,7 +532,7 @@ class StateGroupStorage(object):
Deferred[dict[int, dict[tuple[str, str], str]]]: Deferred[dict[int, dict[tuple[str, str], str]]]:
dict of state_group_id -> (dict of (type, state_key) -> event id) dict of state_group_id -> (dict of (type, state_key) -> event id)
""" """
return self.stores.main._get_state_for_groups(groups, state_filter) return self.stores.state._get_state_for_groups(groups, state_filter)
def store_state_group( def store_state_group(
self, event_id, room_id, prev_group, delta_ids, current_state_ids self, event_id, room_id, prev_group, delta_ids, current_state_ids
@ -552,6 +552,6 @@ class StateGroupStorage(object):
Returns: Returns:
Deferred[int]: The state group ID Deferred[int]: The state group ID
""" """
return self.stores.main.store_state_group( return self.stores.state.store_state_group(
event_id, room_id, prev_group, delta_ids, current_state_ids event_id, room_id, prev_group, delta_ids, current_state_ids
) )

View file

@ -35,7 +35,7 @@ class StateStoreTestCase(tests.unittest.TestCase):
self.store = hs.get_datastore() self.store = hs.get_datastore()
self.storage = hs.get_storage() self.storage = hs.get_storage()
self.state_datastore = self.store self.state_datastore = self.storage.state.stores.state
self.event_builder_factory = hs.get_event_builder_factory() self.event_builder_factory = hs.get_event_builder_factory()
self.event_creation_handler = hs.get_event_creation_handler() self.event_creation_handler = hs.get_event_creation_handler()

View file

@ -231,7 +231,7 @@ def setup_test_homeserver(
"args": {"database": ":memory:", "cp_min": 1, "cp_max": 1}, "args": {"database": ":memory:", "cp_min": 1, "cp_max": 1},
} }
database = DatabaseConnectionConfig("master", database_config, ["main"]) database = DatabaseConnectionConfig("master", database_config)
config.database.databases = [database] config.database.databases = [database]
db_engine = create_engine(database.config) db_engine = create_engine(database.config)