IIRC this doesn't break tests because its only hit on reconnection, or something.
Basically, when a process needs to fetch missing updates for the `typing` stream it needs to query the writer instance via HTTP (as we don't write typing notifications to the DB), the problem was that the endpoint (`streams`) was only registered on master and specifically not on the typing writer worker.
Most of the stuff we do for replication commands can be done synchronously. There's no point spinning up background processes if we're not going to need them.
Handling of incoming typing stream updates from replication was not
hooked up on master, effecting set ups where typing was handled on a
different worker.
This is really only a problem if the master process is also handling
sync requests, which is unlikely for those that are at the stage of
moving typing off.
The other observable effect is that if a worker restarts or a
replication connect drops then the typing worker will issue a
`POSITION typing`, triggering master process to try and stream *all*
typing updates from position 0.
Fixes#7907
If we send out an event which refers to `prev_events` which other servers in
the federation are missing, then (after a round or two of backfill attempts),
they will end up asking us for `/state_ids` at a particular point in the DAG.
As per https://github.com/matrix-org/synapse/issues/7893, this is quite
expensive, and we tend to see lots of very similar requests around the same
time.
We can therefore handle this much more efficiently by using a cache, which (a)
ensures that if we see the same request from multiple servers (or even the same
server, multiple times), then they share the result, and (b) any other servers
that miss the initial excitement can also benefit from the work.
[It's interesting to note that `/state` has a cache for exactly this
reason. `/state` is now essentially unused and replaced with `/state_ids`, but
evidently when we replaced it we forgot to add a cache to the new endpoint.]
For inbound federation requests, if a given remote server makes too many
requests at once, we start stacking them up rather than processing them
immediatedly.
However, that means that there is a fair chance that the requesting server will
disconnect before we start processing the request. In that case, if it was a
read-only request (ie, a GET request), there is absolutely no point in
building a response (and some requests are quite expensive to handle).
Even in the case of a POST request, one of two things will happen:
* Most likely, the requesting server will retry the request and we'll get the
information anyway.
* Even if it doesn't, the requesting server has to assume that we didn't get
the memo, and act accordingly.
In short, we're better off aborting the request at this point rather than
ploughing on with what might be a quite expensive request.
The [postgres setup docs](https://github.com/matrix-org/synapse/blob/develop/docs/postgres.md#set-up-database) recommend setting up your database with user `synapse_user`.
However, uncommenting the postgres defaults in the sample config leave you with user `synapse`.
This PR switches the sample config to recommend `synapse_user`. Took a me a second to figure this out, so assume this will beneficial to others.
It serves no purpose and updating everytime we write to the device inbox
stream means all such transactions will conflict, causing lots of
transaction failures and retries.
When we get behind on replication, we tend to stack up background processes
behind a linearizer. Bg processes are heavy (particularly with respect to
prometheus metrics) and linearizers aren't terribly efficient once the queue
gets long either.
A better approach is to maintain a queue of requests to be processed, and
nominate a single process to work its way through the queue.
Fixes: #7444
Erik Johnston