diff options
Diffstat (limited to 'synapse/replication/tcp/streams/events.py')
| -rw-r--r-- | synapse/replication/tcp/streams/events.py | 133 |
1 files changed, 108 insertions, 25 deletions
diff --git a/synapse/replication/tcp/streams/events.py b/synapse/replication/tcp/streams/events.py index d97669c886..f370390331 100644 --- a/synapse/replication/tcp/streams/events.py +++ b/synapse/replication/tcp/streams/events.py @@ -13,13 +13,14 @@ # 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 heapq +from collections import Iterable +from typing import List, Tuple, Type import attr -from twisted.internet import defer - -from ._base import Stream +from ._base import Stream, StreamUpdateResult, Token, current_token_without_instance """Handling of the 'events' replication stream @@ -63,7 +64,8 @@ class BaseEventsStreamRow(object): Specifies how to identify, serialize and deserialize the different types. """ - TypeId = None # Unique string that ids the type. Must be overriden in sub classes. + # Unique string that ids the type. Must be overriden in sub classes. + TypeId = None # type: str @classmethod def from_data(cls, data): @@ -99,9 +101,12 @@ class EventsStreamCurrentStateRow(BaseEventsStreamRow): event_id = attr.ib() # str, optional -TypeToRow = { - Row.TypeId: Row for Row in (EventsStreamEventRow, EventsStreamCurrentStateRow) -} +_EventRows = ( + EventsStreamEventRow, + EventsStreamCurrentStateRow, +) # type: Tuple[Type[BaseEventsStreamRow], ...] + +TypeToRow = {Row.TypeId: Row for Row in _EventRows} class EventsStream(Stream): @@ -112,29 +117,107 @@ class EventsStream(Stream): def __init__(self, hs): self._store = hs.get_datastore() - self.current_token = self._store.get_current_events_token - - super(EventsStream, self).__init__(hs) - - @defer.inlineCallbacks - def update_function(self, from_token, current_token, limit=None): - event_rows = yield self._store.get_all_new_forward_event_rows( - from_token, current_token, limit - ) - event_updates = ( - (row[0], EventsStreamEventRow.TypeId, row[1:]) for row in event_rows + super().__init__( + hs.get_instance_name(), + current_token_without_instance(self._store.get_current_events_token), + self._update_function, ) - state_rows = yield self._store.get_all_updated_current_state_deltas( - from_token, current_token, limit - ) - state_updates = ( - (row[0], EventsStreamCurrentStateRow.TypeId, row[1:]) for row in state_rows + async def _update_function( + self, + instance_name: str, + from_token: Token, + current_token: Token, + target_row_count: int, + ) -> StreamUpdateResult: + + # the events stream merges together three separate sources: + # * new events + # * current_state changes + # * events which were previously outliers, but have now been de-outliered. + # + # The merge operation is complicated by the fact that we only have a single + # "stream token" which is supposed to indicate how far we have got through + # all three streams. It's therefore no good to return rows 1-1000 from the + # "new events" table if the state_deltas are limited to rows 1-100 by the + # target_row_count. + # + # In other words: we must pick a new upper limit, and must return *all* rows + # up to that point for each of the three sources. + # + # Start by trying to split the target_row_count up. We expect to have a + # negligible number of ex-outliers, and a rough approximation based on recent + # traffic on sw1v.org shows that there are approximately the same number of + # event rows between a given pair of stream ids as there are state + # updates, so let's split our target_row_count among those two types. The target + # is only an approximation - it doesn't matter if we end up going a bit over it. + + target_row_count //= 2 + + # now we fetch up to that many rows from the events table + + event_rows = await self._store.get_all_new_forward_event_rows( + from_token, current_token, target_row_count + ) # type: List[Tuple] + + # we rely on get_all_new_forward_event_rows strictly honouring the limit, so + # that we know it is safe to just take upper_limit = event_rows[-1][0]. + assert ( + len(event_rows) <= target_row_count + ), "get_all_new_forward_event_rows did not honour row limit" + + # if we hit the limit on event_updates, there's no point in going beyond the + # last stream_id in the batch for the other sources. + + if len(event_rows) == target_row_count: + limited = True + upper_limit = event_rows[-1][0] # type: int + else: + limited = False + upper_limit = current_token + + # next up is the state delta table. + ( + state_rows, + upper_limit, + state_rows_limited, + ) = await self._store.get_all_updated_current_state_deltas( + from_token, upper_limit, target_row_count ) - all_updates = heapq.merge(event_updates, state_updates) + limited = limited or state_rows_limited + + # finally, fetch the ex-outliers rows. We assume there are few enough of these + # not to bother with the limit. + + ex_outliers_rows = await self._store.get_ex_outlier_stream_rows( + from_token, upper_limit + ) # type: List[Tuple] + + # we now need to turn the raw database rows returned into tuples suitable + # for the replication protocol (basically, we add an identifier to + # distinguish the row type). At the same time, we can limit the event_rows + # to the max stream_id from state_rows. - return all_updates + event_updates = ( + (stream_id, (EventsStreamEventRow.TypeId, rest)) + for (stream_id, *rest) in event_rows + if stream_id <= upper_limit + ) # type: Iterable[Tuple[int, Tuple]] + + state_updates = ( + (stream_id, (EventsStreamCurrentStateRow.TypeId, rest)) + for (stream_id, *rest) in state_rows + ) # type: Iterable[Tuple[int, Tuple]] + + ex_outliers_updates = ( + (stream_id, (EventsStreamEventRow.TypeId, rest)) + for (stream_id, *rest) in ex_outliers_rows + ) # type: Iterable[Tuple[int, Tuple]] + + # we need to return a sorted list, so merge them together. + updates = list(heapq.merge(event_updates, state_updates, ex_outliers_updates)) + return updates, upper_limit, limited @classmethod def parse_row(cls, row): |