The idea here is that if an instance persists an event via the replication HTTP API it can return before we receive that event over replication, which can lead to races where code assumes that persisting an event immediately updates various caches (e.g. current state of the room).
Most of Synapse doesn't hit such races, so we don't do the waiting automagically, instead we do so where necessary to avoid unnecessary delays. We may decide to change our minds here if it turns out there are a lot of subtle races going on.
People probably want to look at this commit by commit.
For in memory streams when fetching updates on workers we need to query the source of the stream, which currently is hard coded to be master. This PR threads through the source instance we received via `POSITION` through to the update function in each stream, which can then be passed to the replication client for in memory streams.
We move the processing of typing and federation replication traffic into their handlers so that `Stream.current_token()` points to a valid token. This allows us to remove `get_streams_to_replicate()` and `stream_positions()`.
The aim here is to move the command handling out of the TCP protocol classes and to also merge the client and server command handling (so that we can reuse them for redis protocol). This PR simply moves the client paths to the new `ReplicationCommandHandler`, a future PR will move the server paths too.
* Remove `conn_id` usage for UserSyncCommand.
Each tcp replication connection is assigned a "conn_id", which is used
to give an ID to a remotely connected worker. In a redis world, there
will no longer be a one to one mapping between connection and instance,
so instead we need to replace such usages with an ID generated by the
remote instances and included in the replicaiton commands.
This really only effects UserSyncCommand.
* Add CLEAR_USER_SYNCS command that is sent on shutdown.
This should help with the case where a synchrotron gets restarted
gracefully, rather than rely on 5 minute timeout.
This changes the replication protocol so that the server does not send down `RDATA` for rows that happened before the client connected. Instead, the server will send a `POSITION` and clients then query the database (or master out of band) to get up to date.
* Port synapse.replication.tcp to async/await
* Newsfile
* Correctly document type of on_<FOO> functions as async
* Don't be overenthusiastic with the asyncing....
If the client failed to process incoming commands during the initial set
up of the replication connection it would immediately disconnect and
reconnect, resulting in a tightloop.
This can happen, for example, when subscribing to a stream that has a
row that is too long in the backlog.
The fix here is to not consider the connection successfully set up until
the client has succesfully subscribed and caught up with the streams.
This ensures that the retry logic timers aren't reset until then,
meaning that if an error does happen during start up the client will
continue backing off before retrying again.
Run the handlers for replication commands as background processes. This should
improve the visibility in our metrics, and reduce the number of "running db
transaction from sentinel context" warnings.
Ideally it means converting the things that fire off deferreds into the night
into things that actually return a Deferred when they are done. I've made a bit
of a stab at this, but it will probably be leaky.
on_notifier_poke no longer runs synchonously, so we have to do a different hack
to make sure that the replication data has been sent. Let's actually listen for
its arrival.
