We have seen a failure mode here where if there are many in flight
unfinished IDs then marking an ID as finished takes a lot of CPU (as
calling deque.remove iterates over the list)
Introduced in #9104
This wasn't picked up by the tests as this is all fine the first time you run Synapse (after upgrading), but then when you restart the wrong value is pulled from `stream_positions`.
We asserted that the IDs returned by postgres sequence was greater than
any we had seen, however this is technically racey as we may update the
current positions out of order.
We now assert that the sequences are correct on startup, so the
assertion is no longer really required, so we remove them.
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.
We call `_update_stream_positions_table_txn` a lot, which is an UPSERT
that can conflict in `REPEATABLE READ` isolation level. Instead of doing
a transaction consisting of a single query we may as well run it outside
of a transaction.
This is so we can tell what is going on when things are taking a while to start up.
The main change here is to ensure that transactions that are created during startup get correctly logged like normal transactions.
For negative streams we have to negate the internal stream ID before
querying the DB.
The effect of this bug was to query far too many rows, slowing start up
time, but we would correctly filter the results afterwards so there was
no ill effect.
* Fix table scan of events on worker startup.
This happened because we assumed "new" writers had an initial stream
position of 0, so the replication code tried to fetch all events written
by the instance between 0 and the current position.
Instead, set the initial position of new writers to the current
persisted up to position, on the assumption that new writers won't have
written anything before that point.
* Consider old writers coming back as "new".
Otherwise we'd try and fetch entries between the old stale token and the
current position, even though it won't have written any rows.
Co-authored-by: Andrew Morgan <1342360+anoadragon453@users.noreply.github.com>
Co-authored-by: Andrew Morgan <1342360+anoadragon453@users.noreply.github.com>
On startup `MultiWriteIdGenerator` fetches the maximum stream ID for
each instance from the table and uses that as its initial "current
position" for each writer. This is problematic as a) it involves either
a scan of events table or an index (neither of which is ideal), and b)
if rows are being persisted out of order elsewhere while the process
restarts then using the maximum stream ID is not correct. This could
theoretically lead to race conditions where e.g. events that are
persisted out of order are not sent down sync streams.
We fix this by creating a new table that tracks the current positions of
each writer to the stream, and update it each time we finish persisting
a new entry. This is a relatively small overhead when persisting events.
However for the cache invalidation stream this is a much bigger relative
overhead, so instead we note that for invalidation we don't actually
care about reliability over restarts (as there's no caches to
invalidate) and simply don't bother reading and writing to the new table
in that particular case.
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 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.
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).
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.
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 :)
... and update some docstrings to correctly reflect the types being used.
get_new_device_msgs_for_remote can return a long under some circumstances,
which was being stored in last_device_list_stream_id_by_dest, and was then
upsetting things on the next loop.
