This PR attempts to eliminate unnecessary presence sending work when your local server joins a room, or when a remote server joins a room your server is participating in by processing state deltas in chunks rather than individually.
---
When your server joins a room for the first time, it requests the historical state as well. This chunk of new state is passed to the presence handler which, after filtering that state down to only membership joins, will send presence updates to homeservers for each join processed.
It turns out that we were being a bit naive and processing each event individually, and sending out presence updates for every one of those joins. Even if many different joins were users on the same server (hello IRC bridges), we'd send presence to that same homeserver for every remote user join we saw.
This PR attempts to deduplicate all of that by processing the entire batch of state deltas at once, instead of only doing each join individually. We process the joins and note down which servers need which presence:
* If it was a local user join, send that user's latest presence to all servers in the room
* If it was a remote user join, send the presence for all local users in the room to that homeserver
We deduplicate by inserting all of those pending updates into a dictionary of the form:
```
{
server_name1: {presence_update1, ...},
server_name2: {presence_update1, presence_update2, ...}
}
```
Only after building this dict do we then start sending out presence updates.
This is a small bug that I noticed while working on #8956.
We have a for-loop which attempts to strip all presence changes for each user except for the final one, as we don't really care about older presence:
9e19c6aab4/synapse/handlers/presence.py (L368-L371)
`new_states_dict` stores this stripped copy of latest presence state for each user, before it is... put into a new variable `new_state`, which is just overridden by the subsequent for loop.
I believe this was instead meant to override `new_states`. Without doing so, it effectively meant:
1. The for loop had no effect.
2. We were still processing old presence state for users.
- Update black version to the latest
- Run black auto formatting over the codebase
- Run autoformatting according to [`docs/code_style.md
`](80d6dc9783/docs/code_style.md)
- Update `code_style.md` docs around installing black to use the correct version
Cached functions accept an `on_invalidate` function, which we failed to add to the type signature. It's rarely used in the files that we have typed, which is why we haven't noticed it before.
This removes `SourcePaginationConfig` and `get_pagination_rows`. The reasoning behind this is that these generic classes/functions erased the types of the IDs it used (i.e. instead of passing around `StreamToken` it'd pass in e.g. `token.room_key`, which don't have uniform types).
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.
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.
* Fix presence timeouts when synchrotron restarts.
Handling timeouts would fail if there was an external process that had
timed out, e.g. a synchrotron restarting. This was due to a couple of
variable name typoes.
Fixes#3715.
Hopefully this will fix the occasional failures we were seeing in the room directory.
The problem was that events are not necessarily persisted (and `current_state_delta_stream` updated) in the same order as their stream_id. So for instance current_state_delta 9 might be persisted *before* current_state_delta 8. Then, when the room stats saw stream_id 9, it assumed it had done everything up to 9, and never came back to do stream_id 8.
We can solve this easily by only processing up to the stream_id where we know all events have been persisted.
Python will return a tuple whether there are parentheses around the returned values or not.
I'm just sick of my editor complaining about this all over the place :)
If, for some reason, presence updates take a while to persist then it
can trigger clients to tightloop calling `/sync` due to the presence
handler returning updates but not advancing the stream token.
Fixes#5503.
* Fix background updates to handle redactions/rejections
In background updates based on current state delta stream we need to
handle that we may not have all the events (or at least that
`get_events` may raise an exception).
Primarily this fixes a bug in the handling of remote users joining a
room where the server sent out the presence for all local users in the
room to all servers in the room.
We also change to using the state delta stream, rather than the
distributor, as it will make it easier to split processing out of the
master process (as well as being more flexible).
Finally, when sending presence states to newly joined servers we filter
out old presence states to reduce the number sent. Initially we filter
out states that are offline and have a last active more than a week ago,
though this can be changed down the line.
Fixes#3962
In worker mode, on the federation sender, when we receive an edu for sending
over the replication socket, it is parsed into an Edu object. There is no point
extracting the contents of it so that we can then immediately build another Edu.
There were a bunch of places where we fire off a process to happen in the
background, but don't have any exception handling on it - instead relying on
the unhandled error being logged when the relevent deferred gets
garbage-collected.
This is unsatisfactory for a number of reasons:
- logging on garbage collection is best-effort and may happen some time after
the error, if at all
- it can be hard to figure out where the error actually happened.
- it is logged as a scary CRITICAL error which (a) I always forget to grep for
and (b) it's not really CRITICAL if a background process we don't care about
fails.
So this is an attempt to add exception handling to everything we fire off into
the background.
This is mainly done by moving the calculation of where to send presence
updates from the presence handler to the transaction queue, so we only
need to send the presence event (and not the destinations) across the
replication connection. Before we were duplicating by sending the full
state across once per destination.
The TCP replication protocol streams deltas of who has started or
stopped syncing. This is different from the HTTP API which periodically
sends the full list of users who are syncing. This commit adds support
for the new TCP style of sending deltas.
Andrew Morgan