diff --git a/synapse/federation/sender/__init__.py b/synapse/federation/sender/__init__.py
index 1bcc353d18..1dc041752b 100644
--- a/synapse/federation/sender/__init__.py
+++ b/synapse/federation/sender/__init__.py
@@ -104,7 +104,26 @@ class FederationSender(object):
self._processing_pending_presence = False
+ # map from room_id to a set of PerDestinationQueues which we believe are
+ # awaiting a call to flush_read_receipts_for_room. The presence of an entry
+ # here for a given room means that we are rate-limiting RR flushes to that room,
+ # and that there is a pending call to _flush_rrs_for_room in the system.
+ self._queues_awaiting_rr_flush_by_room = {
+ } # type: dict[str, set[PerDestinationQueue]]
+
+ self._rr_txn_interval_per_room_ms = (
+ 1000.0 / hs.get_config().federation_rr_transactions_per_room_per_second
+ )
+
def _get_per_destination_queue(self, destination):
+ """Get or create a PerDestinationQueue for the given destination
+
+ Args:
+ destination (str): server_name of remote server
+
+ Returns:
+ PerDestinationQueue
+ """
queue = self._per_destination_queues.get(destination)
if not queue:
queue = PerDestinationQueue(self.hs, self._transaction_manager, destination)
@@ -250,33 +269,91 @@ class FederationSender(object):
Args:
receipt (synapse.types.ReadReceipt): receipt to be sent
"""
+
+ # Some background on the rate-limiting going on here.
+ #
+ # It turns out that if we attempt to send out RRs as soon as we get them from
+ # a client, then we end up trying to do several hundred Hz of federation
+ # transactions. (The number of transactions scales as O(N^2) on the size of a
+ # room, since in a large room we have both more RRs coming in, and more servers
+ # to send them to.)
+ #
+ # This leads to a lot of CPU load, and we end up getting behind. The solution
+ # currently adopted is as follows:
+ #
+ # The first receipt in a given room is sent out immediately, at time T0. Any
+ # further receipts are, in theory, batched up for N seconds, where N is calculated
+ # based on the number of servers in the room to achieve a transaction frequency
+ # of around 50Hz. So, for example, if there were 100 servers in the room, then
+ # N would be 100 / 50Hz = 2 seconds.
+ #
+ # Then, after T+N, we flush out any receipts that have accumulated, and restart
+ # the timer to flush out more receipts at T+2N, etc. If no receipts accumulate,
+ # we stop the cycle and go back to the start.
+ #
+ # However, in practice, it is often possible to flush out receipts earlier: in
+ # particular, if we are sending a transaction to a given server anyway (for
+ # example, because we have a PDU or a RR in another room to send), then we may
+ # as well send out all of the pending RRs for that server. So it may be that
+ # by the time we get to T+N, we don't actually have any RRs left to send out.
+ # Nevertheless we continue to buffer up RRs for the room in question until we
+ # reach the point that no RRs arrive between timer ticks.
+ #
+ # For even more background, see https://github.com/matrix-org/synapse/issues/4730.
+
+ room_id = receipt.room_id
+
# Work out which remote servers should be poked and poke them.
- domains = yield self.state.get_current_hosts_in_room(receipt.room_id)
+ domains = yield self.state.get_current_hosts_in_room(room_id)
domains = [d for d in domains if d != self.server_name]
if not domains:
return
- logger.debug("Sending receipt to: %r", domains)
+ queues_pending_flush = self._queues_awaiting_rr_flush_by_room.get(
+ room_id
+ )
- content = {
- receipt.room_id: {
- receipt.receipt_type: {
- receipt.user_id: {
- "event_ids": receipt.event_ids,
- "data": receipt.data,
- },
- },
- },
- }
- key = (receipt.room_id, receipt.receipt_type, receipt.user_id)
+ # if there is no flush yet scheduled, we will send out these receipts with
+ # immediate flushes, and schedule the next flush for this room.
+ if queues_pending_flush is not None:
+ logger.debug("Queuing receipt for: %r", domains)
+ else:
+ logger.debug("Sending receipt to: %r", domains)
+ self._schedule_rr_flush_for_room(room_id, len(domains))
for domain in domains:
- self.build_and_send_edu(
- destination=domain,
- edu_type="m.receipt",
- content=content,
- key=key,
- )
+ queue = self._get_per_destination_queue(domain)
+ queue.queue_read_receipt(receipt)
+
+ # if there is already a RR flush pending for this room, then make sure this
+ # destination is registered for the flush
+ if queues_pending_flush is not None:
+ queues_pending_flush.add(queue)
+ else:
+ queue.flush_read_receipts_for_room(room_id)
+
+ def _schedule_rr_flush_for_room(self, room_id, n_domains):
+ # that is going to cause approximately len(domains) transactions, so now back
+ # off for that multiplied by RR_TXN_INTERVAL_PER_ROOM
+ backoff_ms = self._rr_txn_interval_per_room_ms * n_domains
+
+ logger.debug("Scheduling RR flush in %s in %d ms", room_id, backoff_ms)
+ self.clock.call_later(backoff_ms, self._flush_rrs_for_room, room_id)
+ self._queues_awaiting_rr_flush_by_room[room_id] = set()
+
+ def _flush_rrs_for_room(self, room_id):
+ queues = self._queues_awaiting_rr_flush_by_room.pop(room_id)
+ logger.debug("Flushing RRs in %s to %s", room_id, queues)
+
+ if not queues:
+ # no more RRs arrived for this room; we are done.
+ return
+
+ # schedule the next flush
+ self._schedule_rr_flush_for_room(room_id, len(queues))
+
+ for queue in queues:
+ queue.flush_read_receipts_for_room(room_id)
@logcontext.preserve_fn # the caller should not yield on this
@defer.inlineCallbacks
|
