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.
Make sure that the AccountDataStream presents complete updates, in the right
order.
This is much the same fix as #7337 and #7358, but applied to a different stream.
This allows us to have the logic on both master and workers, which is necessary to move event persistence off master.
We also combine the instantiation of ID generators from DataStore and slave stores to the base worker stores. This allows us to select which process writes events independently of the master/worker splits.
Before all streams were only written to from master, so only master needed to respond to `REPLICATE` commands.
Before all instances wrote to the cache invalidation stream, but didn't respond to `REPLICATE`. This was a bug, which could lead to missed rows from cache invalidation stream if an instance is restarted, however all the caches would be empty in that case so it wasn't a problem.
Proactively send out `POSITION` commands (as if we had just received a `REPLICATE`) when we connect to Redis. This is important as other instances won't notice we've connected to issue a `REPLICATE` command (unlike for direct TCP connections). This is only currently an issue if master process reconnects without restarting (if it restarts then it won't have written anything and so other instances probably won't have missed anything).
* release-v1.13.0:
Don't UPGRADE database rows
RST indenting
Put rollback instructions in upgrade notes
Fix changelog typo
Oh yeah, RST
Absolute URL it is then
Fix upgrade notes link
Provide summary of upgrade issues in changelog. Fix )
Move next version notes from changelog to upgrade notes
Changelog fixes
1.13.0rc1
Documentation on setting up redis (#7446)
Rework UI Auth session validation for registration (#7455)
Fix errors from malformed log line (#7454)
Drop support for redis.dbid (#7450)
We forgot to set the password on the subscriber connection, as well as
not calling super methods for overridden connectionMade/connectionLost
functions.
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()`.
This is primarily for allowing us to send those commands from workers, but for now simply allows us to ignore echoed RDATA/POSITION commands that we sent (we get echoes of sent commands when using redis). Currently we log a WARNING on the master process every time we receive an echoed RDATA.
For direct TCP connections we need the master to relay REMOTE_SERVER_UP
commands to the other connections so that all instances get notified
about it. The old implementation just relayed to all connections,
assuming that sending back to the original sender of the command was
safe. This is not true for redis, where commands sent get echoed back to
the sender, which was causing master to effectively infinite loop
sending and then re-receiving REMOTE_SERVER_UP commands that it sent.
The fix is to ensure that we only relay to *other* connections and not
to the connection we received the notification from.
Fixes#7334.
* Factor out functions for injecting events into database
I want to add some more flexibility to the tools for injecting events into the
database, and I don't want to clutter up HomeserverTestCase with them, so let's
factor them out to a new file.
* Rework TestReplicationDataHandler
This wasn't very easy to work with: the mock wrapping was largely superfluous,
and it's useful to be able to inspect the received rows, and clear out the
received list.
* Fix AssertionErrors being thrown by EventsStream
Part of the problem was that there was an off-by-one error in the assertion,
but also the limit logic was too simple. Fix it all up and add some tests.
Figuring out how to correctly limit updates from this stream without dropping
entries is far more complicated than just counting the number of rows being
returned. We need to consider each query separately and, if any one query hits
the limit, truncate the results from the others.
I think this also fixes some potentially long-standing bugs where events or
state changes could get missed if we hit the limit on either query.
Long story short: if we're handling presence on the current worker, we shouldn't be sending USER_SYNC commands over replication.
In an attempt to figure out what is going on here, I ended up refactoring some bits of the presencehandler code, so the first 4 commits here are non-functional refactors to move this code slightly closer to sanity. (There's still plenty to do here :/). Suggest reviewing individual commits.
Fixes (I hope) #7257.
Other parts of the code (such as the StreamChangeCache) assume that there will
not be multiple changes with the same stream id.
This code was introduced in #7024, and I hope this fixes#7206.
The general idea here is to get rid of the type: ignore annotations on all of the current_token and update_function assignments, which would have caught #7290.
After a bit of experimentation, it seems like the least-awful way to do this is to pass the offending functions in as parameters to the Stream constructor. Unfortunately that means that the concrete implementations no longer have the same constructor signature as Stream itself, which means that it gets hard to correctly annotate STREAMS_MAP.
I've also introduced a couple of new types, to take out some duplication.
Some of the query functions return generators rather than lists, so we can't
index into the result. Happily we already have a copy of the results.
(think this was introduced in #7024)
`REPLICATE` is now a valid command, and it's nice if you can issue it from the
console without remembering to call it `REPLICATE ` with a trailing space.
Separate `SimpleCommand` from `Command`, so that things which don't want to use
the `data` property don't have to, and thus fix the warnings PyCharm was giving
me about not calling `__init__` in the base class.
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.
This broke in a recent PR (#7024) and is no longer useful due to all
replication clients implicitly subscribing to all streams, so let's
just remove it.
* 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.
* Add 'device_lists_outbound_pokes' as extra table.
This makes sure we check all the relevant tables to get the current max
stream ID.
Currently not doing so isn't problematic as the max stream ID in
`device_lists_outbound_pokes` is the same as in `device_lists_stream`,
however that will change.
* Change device lists stream to have one row per id.
This will make it possible to process the streams more incrementally,
avoiding having to process large chunks at once.
* Change device list replication to match new semantics.
Instead of sending down batches of user ID/host tuples, send down a row
per entity (user ID or host).
* Newsfile
* Remove handling of multiple rows per ID
* Fix worker handling
* Comments from review
This is a bit fiddly because it all has to be done on one fell swoop:
* Wherever we create a new event, pass in the room version (and check it matches the format version)
* When we prune an event, use the room version of the unpruned event to create the pruned version.
* When we pass an event over the replication protocol, pass the room version over alongside it, and use it when deserialising the event again.
This makes sure we check all the relevant tables to get the current max
stream ID.
Currently not doing so isn't problematic as the max stream ID in
`device_lists_outbound_pokes` is the same as in `device_lists_stream`,
however that will change.
When we get an invite over federation, store the room version in the rooms table.
The general idea here is that, when we pull the invite out again, we'll want to know what room_version it belongs to (so that we can later redact it if need be). So we need to store it somewhere...
* Port synapse.replication.tcp to async/await
* Newsfile
* Correctly document type of on_<FOO> functions as async
* Don't be overenthusiastic with the asyncing....
Currently we rely on `current_state_events` to figure out what rooms a
user was in and their last membership event in there. However, if the
server leaves the room then the table may be cleaned up and that
information is lost. So lets add a table that separately holds that
information.
When the `/keys/query` API is hit on client_reader worker Synapse may
decide that it needs to resync some remote deivces. Usually this happens
on master, and then gets cached. However, that fails on workers and so
it falls back to fetching devices from remotes directly, which may in
turn fail if the remote is down.
Erik Johnston