Some Synapse instances have been hammering their database after v1.66.0 -> v1.68.0 update

[jaywink]

Description

Some EMS hosted Synapse instances are hammering their database after upgrading from v1.66.0 to v1.68.0. The host concentrating here on is ecf6bc70-0bd7-11ec-8fb7-11c2f603e85f-live (EMS internal host ID, please check with EMS team for real hostnames).

The offensive query is:

SELECT c.state_key FROM current_state_events as c
            /* Get the depth of the event from the events table */
            INNER JOIN events AS e USING (event_id)
            WHERE c.type = ? AND c.room_id = ? AND membership = ?
            /* Sorted by lowest depth first */
            ORDER BY e.depth ASC

The background update running at the time was event_push_backfill_thread_id, if relevant.

Graphs:

IOPS increase at upgrade. The initial plateau at 4K was due to the database being locked to 4K IOPS. Now it has 10K and has consistently continued to hammer the database after ~7 hours since the upgrade.

Degraded event send times especially when constrained to 4K IOPS, which the host has been running with for a long time fine.

Stateres worst-case seems to reflect the database usage, just side effect of a busy db?

DB usage for background jobs had a rather massive spike for notify_interested_appservices_ephemeral right after upgrade.

Taking that away from the graph, we see db usage for background jobs higher since upgrade all around.

DB transactions:

Cache eviction seems to indicate we should raise the get_local_users_in_room cache as it is being evicted a lot by size. However, this has been the case pre-upgrade as well.

Appservice transactions have not changed during this time by a large factor (3 bridges):

A few other hosts manually found:

• 01bbd800-4670-11e9-8324-b54a9efc8abc-live

• db0718c0-2480-11e9-83c4-ad579ecfcc33-live

Day time based changes in traffic have been ruled out, all these issues started on upgrade with no other changes to the hosting or deployment stack. There are probably more hosts affected by the db usage increase.

---

Also discussed in backend internal.

Steps to reproduce

Uprgade from v1.66.0 to v1.68.0.

Homeserver

ecf6bc70-0bd7-11ec-8fb7-11c2f603e85f-live, 01bbd800-4670-11e9-8324-b54a9efc8abc-live, db0718c0-2480-11e9-83c4-ad579ecfcc33-live

Synapse Version

v1.68.0

Installation Method

Other (please mention below)

Platform

EMS flavour Docker images built from upstream images. Kubernetes cluster.

Relevant log output

-

Anything else that would be useful to know?

No response

[reivilibre]

That query was touched recently by #13575.

[DMRobertson]

$ git log --oneline -S get_users_in_room v1.66.0..v1.68.0
  e124b24ad 1.68.0rc1
  906cead9c Update docstrings to explain the impact of partial state (#13750)
  cdf7fb737 1.67.0rc1
  1a209efdb Update `get_users_in_room` mis-use to get hosts with dedicated `get_current_hosts_in_room` (#13605)
  d58615c82 Directly lookup local membership instead of getting all members in a room first (`get_users_in_room` mis-use) (#13608)
  c807b814a Use dedicated `get_local_users_in_room` to find local users when calculating `join_authorised_via_users_server` of a `/make_jo  in` request (#13606)

#13605 #13608 #13606

On the affected hosts we saw that postgres used a heap scan rather than an index scan for the aggressive query. Forcing an index-only scan (SET enable_bitmapscan to off) made the query quicker (500->300ms) for one trial set of parameters. Unclear if those are representitive. Also unclear if the index-only scan used more or less file IO.

But this might be barking up the wrong tree. Are we perhaps calling get_users_in_room in more circumstances now? Do we need to up its cache?

[anoadragon453]

Note that on matrix.org the cache factor specifically for get_users_in_room is between 7 and 20 across most workers. The global cache factor is typically either 1 or 2. This may help explain a discrepancy between matrix.org and these hosts.

[DMRobertson]

That query was touched recently by #13575.

By trying different versions we've confirmed that #13575 contributed to the performance regression, and we haven't seen any evidence of other queries being suddenly more expensive.

We did this by trying 51d732d and its parent 4f6de33 on a test server. (I tried to revert 51d732d on top of 1.68.0 for testing, but there were merge conflicts and I wasn't sure how to handle them---could use advice from @MadLittleMods on that one, and on how to proceed here more generally.

Anyway, here's the graph of database IO (bigger = more IO credits available) from the test server.

Seems convincing to me.

[MadLittleMods]

#13605 #13608 #13606

[...]

Are we perhaps calling get_users_in_room in more circumstances now? Do we need to up its cache?

From the PRs you linked, we use get_users_in_room a lot less now. It was being used as a crutch to find the current hosts in the room but we have a dedicated optimized function for that, get_current_hosts_in_room, which it was switched to. And in some cases, we didn't even need to use get_users_in_room or get_current_hosts_in_room at at all.

#13575 does make the get_users_in_room query more complicated by adding a join. Some ideas for removing the join are in that same thread

You can see from the matrix.org graphs that we call it a lot less (it's not even a top db transaction by rate anymore) and the total time is way down but it does take a bit more time now when it runs. (the switch happens with the dramatic change in the graphs around Sept. 6th)

https://grafana.matrix.org/d/000000012/synapse?orgId=1&from=1661891889156&to=1664483889157&viewPanel=11

---

The giant spike in notify_interested_services_ephemeral (from the graphs in the description) would account for a lot more get_users_in_room calls down the line though 🐠

---

Also as a note, we will always use get_users_in_room to calculate get_current_hosts_in_room if you're using SQLite but I think the EMS hosts are using Postgres from the comment above,

synapse/synapse/storage/databases/main/roommember.py

Lines 960 to 963 in ebd9e2d

	if users is None and isinstance(self.database_engine, Sqlite3Engine):
	# If we're using SQLite then let's just always use
	# `get_users_in_room` rather than funky SQL.
	users = await self.get_users_in_room(room_id)

[MadLittleMods]

Summarizing some more discussion in backend internal:

We don't know exactly what upstream callers are accounting for all of these get_users_in_room calls but everything kinda points to this being appservice related.

This also makes sense since one of the affected hosts is Element One which has a lot of bridged appservice users.

I suspect we have some more get_users_in_room mis-uses in the appservice code where we should be using get_local_users_in_room since the appservice only needs to worry about users on its own server. Fixing these mis-uses doesn't fix get_users_in_room performance itself but using get_local_users_in_room would be more performant than whatever we end up changing with get_users_in_room anyway.

@clokep pointed out a couple get_users_in_room mis-uses already:

• synapse/synapse/appservice/__init__.py

  Lines 175 to 177 in 7bc08f3

  	member_list = await store.get_users_in_room(
  	room_id, on_invalidate=cache_context.invalidate
  	)

• synapse/synapse/storage/databases/main/appservice.py

  Lines 160 to 163 in 21eeacc

  	users_in_room = await self.get_users_in_room(
  	room_id, on_invalidate=cache_context.invalidate
  	)
  	return list(filter(app_service.is_interested_in_user, users_in_room))

I'll work on a PR now to change these over ⏩