Merge branch 'matthew/gin_work_mem' into matthew/hit_the_gin

This commit is contained in:
Richard van der Hoff 2018-02-13 16:45:36 +00:00
commit ddb6a79b68
116 changed files with 4226 additions and 1834 deletions

View file

@ -13,19 +13,26 @@
# See the License for the specific language governing permissions and
# limitations under the License.
from collections import namedtuple
import logging
import re
import sys
import ujson as json
from twisted.internet import defer
from .background_updates import BackgroundUpdateStore
from synapse.api.errors import SynapseError
from synapse.storage.engines import PostgresEngine, Sqlite3Engine
import logging
import re
import ujson as json
logger = logging.getLogger(__name__)
SearchEntry = namedtuple('SearchEntry', [
'key', 'value', 'event_id', 'room_id', 'stream_ordering',
'origin_server_ts',
])
class SearchStore(BackgroundUpdateStore):
@ -60,16 +67,17 @@ class SearchStore(BackgroundUpdateStore):
@defer.inlineCallbacks
def _background_reindex_search(self, progress, batch_size):
# we work through the events table from highest stream id to lowest
target_min_stream_id = progress["target_min_stream_id_inclusive"]
max_stream_id = progress["max_stream_id_exclusive"]
rows_inserted = progress.get("rows_inserted", 0)
INSERT_CLUMP_SIZE = 1000
TYPES = ["m.room.name", "m.room.message", "m.room.topic"]
def reindex_search_txn(txn):
sql = (
"SELECT stream_ordering, event_id, room_id, type, content FROM events"
"SELECT stream_ordering, event_id, room_id, type, content, "
" origin_server_ts FROM events"
" WHERE ? <= stream_ordering AND stream_ordering < ?"
" AND (%s)"
" ORDER BY stream_ordering DESC"
@ -78,6 +86,10 @@ class SearchStore(BackgroundUpdateStore):
txn.execute(sql, (target_min_stream_id, max_stream_id, batch_size))
# we could stream straight from the results into
# store_search_entries_txn with a generator function, but that
# would mean having two cursors open on the database at once.
# Instead we just build a list of results.
rows = self.cursor_to_dict(txn)
if not rows:
return 0
@ -90,6 +102,8 @@ class SearchStore(BackgroundUpdateStore):
event_id = row["event_id"]
room_id = row["room_id"]
etype = row["type"]
stream_ordering = row["stream_ordering"]
origin_server_ts = row["origin_server_ts"]
try:
content = json.loads(row["content"])
except Exception:
@ -104,6 +118,8 @@ class SearchStore(BackgroundUpdateStore):
elif etype == "m.room.name":
key = "content.name"
value = content["name"]
else:
raise Exception("unexpected event type %s" % etype)
except (KeyError, AttributeError):
# If the event is missing a necessary field then
# skip over it.
@ -114,25 +130,16 @@ class SearchStore(BackgroundUpdateStore):
# then skip over it
continue
event_search_rows.append((event_id, room_id, key, value))
event_search_rows.append(SearchEntry(
key=key,
value=value,
event_id=event_id,
room_id=room_id,
stream_ordering=stream_ordering,
origin_server_ts=origin_server_ts,
))
if isinstance(self.database_engine, PostgresEngine):
sql = (
"INSERT INTO event_search (event_id, room_id, key, vector)"
" VALUES (?,?,?,to_tsvector('english', ?))"
)
elif isinstance(self.database_engine, Sqlite3Engine):
sql = (
"INSERT INTO event_search (event_id, room_id, key, value)"
" VALUES (?,?,?,?)"
)
else:
# This should be unreachable.
raise Exception("Unrecognized database engine")
for index in range(0, len(event_search_rows), INSERT_CLUMP_SIZE):
clump = event_search_rows[index:index + INSERT_CLUMP_SIZE]
txn.executemany(sql, clump)
self.store_search_entries_txn(txn, event_search_rows)
progress = {
"target_min_stream_id_inclusive": target_min_stream_id,
@ -276,6 +283,92 @@ class SearchStore(BackgroundUpdateStore):
defer.returnValue(num_rows)
def store_event_search_txn(self, txn, event, key, value):
"""Add event to the search table
Args:
txn (cursor):
event (EventBase):
key (str):
value (str):
"""
self.store_search_entries_txn(
txn,
(SearchEntry(
key=key,
value=value,
event_id=event.event_id,
room_id=event.room_id,
stream_ordering=event.internal_metadata.stream_ordering,
origin_server_ts=event.origin_server_ts,
),),
)
def store_search_entries_txn(self, txn, entries):
"""Add entries to the search table
Args:
txn (cursor):
entries (iterable[SearchEntry]):
entries to be added to the table
"""
if isinstance(self.database_engine, PostgresEngine):
sql = (
"INSERT INTO event_search"
" (event_id, room_id, key, vector, stream_ordering, origin_server_ts)"
" VALUES (?,?,?,to_tsvector('english', ?),?,?)"
)
args = ((
entry.event_id, entry.room_id, entry.key, entry.value,
entry.stream_ordering, entry.origin_server_ts,
) for entry in entries)
# inserts to a GIN index are normally batched up into a pending
# list, and then all committed together once the list gets to a
# certain size. The trouble with that is that postgres (pre-9.5)
# uses work_mem to determine the length of the list, and work_mem
# is typically very large.
#
# We therefore reduce work_mem while we do the insert.
#
# (postgres 9.5 uses the separate gin_pending_list_limit setting,
# so doesn't suffer the same problem, but changing work_mem will
# be harmless)
txn.execute("SET work_mem='256kB'")
try:
txn.executemany(sql, args)
except Exception:
# we need to reset work_mem, but doing so may throw a new
# exception and we want to preserve the original
t, v, tb = sys.exc_info()
try:
txn.execute("RESET work_mem")
except Exception as e:
logger.warn(
"exception resetting work_mem during exception "
"handling: %r",
e,
)
raise t, v, tb
else:
txn.execute("RESET work_mem")
elif isinstance(self.database_engine, Sqlite3Engine):
sql = (
"INSERT INTO event_search (event_id, room_id, key, value)"
" VALUES (?,?,?,?)"
)
args = ((
entry.event_id, entry.room_id, entry.key, entry.value,
) for entry in entries)
txn.executemany(sql, args)
else:
# This should be unreachable.
raise Exception("Unrecognized database engine")
@defer.inlineCallbacks
def search_msgs(self, room_ids, search_term, keys):
"""Performs a full text search over events with given keys.