Regression in federation sender instance calls to get joined profile

[Fizzadar]

v1.61.1 release dramatically reduced federation sender calls to the _get_joined_profile_from_event_id cache (see: #12955 (comment)). This seems to have returned in 1.63.1, the jump below is when we upgraded from 1.61.1 -> 1.63.1:

[reivilibre]

It would appear that matrix.org has also seen something like that:

I don't currently know of any relevant changes that led to this.

This is currently at roughly 1% of DB time on the federation sender, so also not sure it's urgent.

[Fizzadar]

Out of interest - how many federation senders are running on matrix.org? I wonder if something is weird about our setup; we have 3x of them each doing ~1k/s on this cache.

[richvdh]

I've spent a while looking into this.

On matrix.org there is a notable uptick in miss rate on 12th July (coinciding with a deployment of these changes, roughly 1.62.0+ to 1.63.0rc1-), followed by a second uptick on 22nd July (coinciding with a deployment with these changes, roughly equivalent to v1.63.0 to v1.64.0rc1). However, both increases are an order of magnitude smaller than those reported by @Fizzadar (300 requests/second vs 2000.) So: bad news for @Fizzadar in that it looks like this gets even worse for 1.64.0.

The federation senders primarily use this lookup to determine which servers they should send events to. Normally, it's expected that the event creators should precache the list of servers in redis, so we'd only expect to see requests to _get_joined_profile_from_event_id where a server list was not found in the redis cache, and indeed we can see a corresponding uptick in cache misses on the get_joined_hosts external cache:

I've not been able to find any obvious reason why we might be seeing more cache misses here. In particular, I'm not seeing any change to the rate of cache population, and given it's a redis cache, I think it's unlikely there will have been any changes to the expiry characteristics. What's particularly odd is that the event_to_prev_state_group hit rate is unchanged, and the two are supposedly populated at pretty much the same time.

It might be instructive to add some logging to report which events are seeing a get_joined_hosts cache miss, and see if we can find any pattern.

@Fizzadar it would be interesting to know if you also see a corresponding pattern in cache misses for the get_joined_hosts external cache. Note that you'll probably have to update the grafana dashboard as I just added a "External cache miss rate" panel.

[richvdh]

Out of interest - how many federation senders are running on matrix.org? I wonder if something is weird about our setup; we have 3x of them each doing ~1k/s on this cache.

sorry, missed this. matrix.org apparently has 20 federation senders, handling a total of around 100 outgoing federation PDUs per second. But note that the code that hits get_joined_hosts is done before the work is sharded (ie, each of the federation senders does the lookup for every event).

[Fizzadar]

@richvdh sorry for the delay - some interesting charts from our end. Misses (bump is 1.64 rollout):

Hits for the same period (so we're ~99% hit rate):

And the joined profile cache hits:

---

What seems crazy to me is we're talking 1/s missed cache hit on get_joined_hosts resulting in ~3-4k requests back to the joined profile query. As far as I know our synapse has no changes within the federation logic. This feels like we're missing the cache for one or two huge rooms (#matrix:matrix.org perhaps?) which is resulting in the profile request.

Not had a chance yet to poke into the changes you linked which could be interesting, I wonder why this is getting missed...

[erikjohnston]

Next steps: as per Rich's #13444 (comment) let's add some log lines to figure out where this is coming from.