This is done by creating a custom `RedisFactory` subclass that
periodically pings all connections in its pool.
We also ensure that the `replyTimeout` param is non-null, so that we
timeout waiting for the reply to those pings (and thus triggering a
reconnect).
I was trying to make it so that we didn't have to start a background task when handling RDATA, but that is a bigger job (due to all the code in `generic_worker`). However I still think not pulling the event from the DB may help reduce some DB usage due to replication, even if most workers will simply go and pull that event from the DB later anyway.
Co-authored-by: Patrick Cloke <clokep@users.noreply.github.com>
#8567 started a span for every background process. This is good as it means all Synapse code that gets run should be in a span (unless in the sentinel logging context), but it means we generate about 15x the number of spans as we did previously.
This PR attempts to reduce that number by a) not starting one for send commands to Redis, and b) deferring starting background processes until after we're sure they're necessary.
I don't really know how much this will help.
Currently background proccesses stream the events stream use the "minimum persisted position" (i.e. `get_current_token()`) rather than the vector clock style tokens. This is broadly fine as it doesn't matter if the background processes lag a small amount. However, in extreme cases (i.e. SyTests) where we only write to one event persister the background processes will never make progress.
This PR changes it so that the `MultiWriterIDGenerator` keeps the current position of a given instance as up to date as possible (i.e using the latest token it sees if its not in the process of persisting anything), and then periodically announces that over replication. This then allows the "minimum persisted position" to advance, albeit with a small lag.
The idea is to remove some of the places we pass around `int`, where it can represent one of two things:
1. the position of an event in the stream; or
2. a token that partitions the stream, used as part of the stream tokens.
The valid operations are then:
1. did a position happen before or after a token;
2. get all events that happened before or after a token; and
3. get all events between two tokens.
(Note that we don't want to allow other operations as we want to change the tokens to be vector clocks rather than simple ints)
slots use less memory (and attribute access is faster) while slightly
limiting the flexibility of the class attributes. This focuses on objects
which are instantiated "often" and for short periods of time.
This is *not* ready for production yet. Caveats:
1. We should write some tests...
2. The stream token that we use for events can get stalled at the minimum position of all writers. This means that new events may not be processed and e.g. sent down sync streams if a writer isn't writing or is slow.
The idea here is that we pass the `max_stream_id` to everything, and only use the stream ID of the particular event to figure out *when* the max stream position has caught up to the event and we can notify people about it.
This is to maintain the distinction between the position of an item in the stream (i.e. event A has stream ID 513) and a token that can be used to partition the stream (i.e. give me all events after stream ID 352). This distinction becomes important when the tokens are more complicated than a single number, which they will be once we start tracking the position of multiple writers in the tokens.
The valid operations here are:
1. Is a position before or after a token
2. Fetching all events between two tokens
3. Merging multiple tokens to get the "max", i.e. `C = max(A, B)` means that for all positions P where P is before A *or* before B, then P is before C.
Future PR will change the token type to a dedicated type.
`pusher_pool.on_new_notifications` expected a min and max stream ID, however that was not what we were passing in. Instead, let's just pass it the current max stream ID and have it track the last stream ID it got passed.
I believe that it mostly worked as we called the function for every event. However, it would break for events that got persisted out of order, i.e, that were persisted but the max stream ID wasn't incremented as not all preceding events had finished persisting, and push for that event would be delayed until another event got pushed to the effected users.
This is *not* ready for production yet. Caveats:
1. We should write some tests...
2. The stream token that we use for events can get stalled at the minimum position of all writers. This means that new events may not be processed and e.g. sent down sync streams if a writer isn't writing or is slow.
This fixes a bug where having multiple callers waiting on the same
stream and position will cause it to try and compare two deferreds,
which fails (due to the sorted list having an entry of `Tuple[int,
Deferred]`).
It's just a thin wrapper around two ID gens to make `get_current_token`
and `get_next` return tuples. This can easily be replaced by calling the
appropriate methods on the underlying ID gens directly.
The function is used for two purposes: 1) for subscribers of streams to
get a token they can use to get further updates with, and 2) for
replication to track position of the writers of the stream.
For streams with a single writer the two scenarios produce the same
result, however the situation becomes complicated for streams with
multiple writers. The current `MultiWriterIdGenerator` does not
correctly handle the first case (which is not an issue as its only used
for the `caches` stream which nothing subscribes to outside of
replication).
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
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
The CI appears to use the latest version of isort, which is a problem when isort gets a major version bump. Rather than try to pin the version, I've done the necessary to make isort5 happy with synapse.
The aim here is to make it easier to reason about when streams are limited and when they're not, by moving the logic into the database functions themselves. This should mean we can kill of `db_query_to_update_function` function.
* Ensure account data stream IDs are unique.
The account data stream is shared between three tables, and the maximum
allocated ID was tracked in a dedicated table. Updating the max ID
happened outside the transaction that allocated the ID, leading to a
race where if the server was restarted then the same ID could be
allocated but the max ID failed to be updated, leading it to be reused.
The ID generators have support for tracking across multiple tables, so
we may as well use that instead of a dedicated table.
* Fix bug in account data replication stream.
If the same stream ID was used in both global and room account data then
the getting updates for the replication stream would fail due to
`heapq.merge(..)` trying to compare a `str` with a `None`. (This is
because you'd have two rows like `(534, '!room')` and `(534, None)` from
the room and global account data tables).
Fix is just to order by stream ID, since we don't rely on the ordering
beyond that. The bug where stream IDs can be reused should be fixed now,
so this case shouldn't happen going forward.
Fixes#7617
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.
