Using the new `TaskScheduler` meant that we'ed create lots of new
metrics (due to adding task ID to the desc of background process),
resulting in requests for metrics taking an increasing amount of CPU.
Adds three new configuration variables:
* destination_min_retry_interval is identical to before (10mn).
* destination_retry_multiplier is now 2 instead of 5, the maximum value will
be reached slower.
* destination_max_retry_interval is one day instead of (essentially) infinity.
Capping this will cause destinations to continue to be retried sometimes instead
of being lost forever. The previous value was 2 ^ 62 milliseconds.
The cached decorators always return a Deferred, which was not
properly propagated. It was close enough when wrapping coroutines,
but failed if a bare function was wrapped.
This change fixes two memory leaks during `trial` test runs.
Garbage collection is disabled during each test case and a gen-0 GC is
run at the end of each test. However, when the gen-0 GC is run, the
`TestCase` object usually still holds references to the `HomeServer`
used during the test. As a result, the `HomeServer` gets promoted to
gen-1 and then never garbage collected.
Fix this by periodically running full GCs.
Additionally, fix `HomeServer`s leaking after tests that touch inbound
federation due to `FederationRateLimiter`s adding themselves to a global
set, by turning the set into a `WeakSet`.
Resolves#15622.
Signed-off-by: Sean Quah <seanq@matrix.org>
Additionally:
* Consistently use `freeze()` in test
---------
Co-authored-by: Patrick Cloke <clokep@users.noreply.github.com>
Co-authored-by: 6543 <6543@obermui.de>
When there are many synchronous requests waiting on a
`_PerHostRatelimiter`, each request will be started recursively just
after the previous request has completed. Under the right conditions,
this leads to stack exhaustion.
A common way for requests to become synchronous is when the remote
client disconnects early, because the homeserver is overloaded and slow
to respond.
Avoid stack exhaustion under these conditions by deferring subsequent
requests until the next reactor tick.
Fixes#14480.
Signed-off-by: Sean Quah <seanq@matrix.org>
PKCE can protect against certain attacks and is enabled by default. Support
can be controlled manually by setting the pkce_method of each oidc_providers
entry to 'auto' (default), 'always', or 'never'.
This is required by Twitter OAuth 2.0 support.
Then adapts calling code to retry when needed so it doesn't 500
to clients.
Signed-off-by: Mathieu Velten <mathieuv@matrix.org>
Co-authored-by: Sean Quah <8349537+squahtx@users.noreply.github.com>
An empty cache does not mean the entity has no changed, if
it is earlier than the earliest known stream position return that
the entity *has* changed since the cache cannot accurately
answer that query.
The internal methods of the StreamChangeCache were inconsistently
treating the earliest known stream position as valid. It is now treated as
invalid, meaning the cache cannot determine if an entity at the earliest
known stream position has changed or not.
Avoid an n+1 query problem and fetch the bundled aggregations for
m.annotation relations in a single query instead of a query per event.
This applies similar logic for as was previously done for edits in
8b309adb43 (#11660) and threads
in b65acead42 (#11752).
Remove type hints from comments which have been added
as Python type hints. This helps avoid drift between comments
and reality, as well as removing redundant information.
Also adds some missing type hints which were simple to fill in.
This code was factored out to a method, but also left in-place.
Calling this twice in a row makes no sense: the first call will reduce
the size appropriately, but the loop will immediately exit since the
cache size was already reduced.
Patrick Cloke