When a lot of locks are waiting for a single lock, notifying all locks
independently with `call_later` on each release is really costly and
incurs some kind of async contention, where the CPU is spinning a lot
for not much.
The included test is taking around 30s before the change, and 0.5s
after.
It was found following failing tests with
https://github.com/element-hq/synapse/pull/16827.
This PR aims to fix#16895, caused by a regression in #7 and not fixed
by #16903. The PR #16903 only fixes a starvation issue, where the CPU
isn't released. There is a second issue, where the execution is blocked.
This theory is supported by the flame graphs provided in #16895 and the
fact that I see the CPU usage reducing and far below the limit.
Since the changes in #7, the method `check_state_independent_auth_rules`
is called with the additional parameter `batched_auth_events`:
6fa13b4f92/synapse/handlers/federation_event.py (L1741-L1743)
It makes the execution enter this if clause, introduced with #151956fa13b4f92/synapse/event_auth.py (L178-L189)
There are two issues in the above code snippet.
First, there is the blocking issue. I'm not entirely sure if this is a
deadlock, starvation, or something different. In the beginning, I
thought the copy operation was responsible. It wasn't. Then I
investigated the nested `store.get_events` inside the function `update`.
This was also not causing the blocking issue. Only when I replaced the
set difference operation (`-` ) with a list comprehension, the blocking
was resolved. Creating and comparing sets with a very large amount of
events seems to be problematic.
This is how the flamegraph looks now while persisting outliers. As you
can see, the execution no longer locks up in the above function.
Second, the copying here doesn't serve any purpose, because only a
shallow copy is created. This means the same objects from the original
dict are referenced. This fails the intention of protecting these
objects from mutation. The review of the original PR
https://github.com/matrix-org/synapse/pull/15195 had an extensive
discussion about this matter.
Various approaches to copying the auth_events were attempted:
1) Implementing a deepcopy caused issues due to
builtins.EventInternalMetadata not being pickleable.
2) Creating a dict with new objects akin to a deepcopy.
3) Creating a dict with new objects containing only necessary
attributes.
Concluding, there is no easy way to create an actual copy of the
objects. Opting for a deepcopy can significantly strain memory and CPU
resources, making it an inefficient choice. I don't see why the copy is
necessary in the first place. Therefore I'm proposing to remove it
altogether.
After these changes, I was able to successfully join these rooms,
without the main worker locking up:
- #synapse:matrix.org
- #element-android:matrix.org
- #element-web:matrix.org
- #ecips:matrix.org
- #ipfs-chatter:ipfs.io
- #python:matrix.org
- #matrix:matrix.org
Since Synapse 1.76.0, any module which registers a `on_new_event`
callback would brick the ability to join remote rooms.
This is because this callback tried to get the full state of the room,
which would end up in a deadlock.
Related:
https://github.com/matrix-org/synapse-auto-accept-invite/issues/18
The following module would brick the ability to join remote rooms:
```python
from typing import Any, Dict, Literal, Union
import logging
from synapse.module_api import ModuleApi, EventBase
logger = logging.getLogger(__name__)
class MyModule:
def __init__(self, config: None, api: ModuleApi):
self._api = api
self._config = config
self._api.register_third_party_rules_callbacks(
on_new_event=self.on_new_event,
)
async def on_new_event(self, event: EventBase, _state_map: Any) -> None:
logger.info(f"Received new event: {event}")
@staticmethod
def parse_config(_config: Dict[str, Any]) -> None:
return None
```
This is technically a breaking change, as we are now passing partial
state on the `on_new_event` callback.
However, this callback was broken for federated rooms since 1.76.0, and
local rooms have full state anyway, so it's unlikely that it would
change anything.
This adds a counter `synapse_emails_sent_total` for emails sent. They
are broken down by `type`, which are `password_reset`, `registration`,
`add_threepid`, `notification` (matching the methods of `Mailer`).
We do this by adding support to the LRU cache for "extra indices" based
on the cached value. This allows us to efficiently map from room ID to
the cached events and only invalidate those.
List of users not to send out device list updates for when they register
new devices. This is useful to handle bot accounts.
This is undocumented as its mostly a hack to test on matrix.org.
Note: This will still send out device list updates if the device is
later updated, e.g. end to end keys are added.
This basically reverts a change that was in
https://github.com/element-hq/synapse/pull/16833, where we reduced the
batching.
The smaller batching can cause performance issues on busy servers and
databases.
During the migration the automated script to update the copyright
headers accidentally got rid of some of the existing copyright lines.
Reinstate them.
Prior to this PR, if a request to create a public (public as in
published to the rooms directory) room violated the room list
publication rules set in the
[config](https://matrix-org.github.io/synapse/latest/usage/configuration/config_documentation.html#room_list_publication_rules),
the request to create the room was denied and the room was not created.
This PR changes the behavior such that when a request to create a room
published to the directory violates room list publication rules, the
room is still created but the room is not published to the directory.
The current query supports passing in a list of users, which generates a
query using `user_id = ANY(..)`. This is generates a less efficient
query plan that is notably slower than a simple `user_id = ?` condition.
Note: The new function is mostly a copy and paste and then a
simplification of the existing function.
The crux of the change is to try and make the queries simpler and pull
out fewer rows. Before, there were quite a few joins against subqueries,
which caused postgres to pull out more rows than necessary.
Instead, let's simplify the query and do some of the filtering out in
Python instead, letting Postgres do better optimizations now that it
doesn't have to deal with joins against subqueries.
Review note: this is a complete rewrite of the function, so not sure how
useful the diff is.
---------
Co-authored-by: Andrew Morgan <1342360+anoadragon453@users.noreply.github.com>
In previous versions of authlib using `client_secret_basic` without a
`client_secret` would result in an invalid auth header. Since authlib
1.3 it throws an exception.
The configuration may be accepted in by very lax servers, so we don't
want to deny it outright. Instead, let's default the
`client_auth_method` to `none`, which does the right thing. If the
config specifies `client_auth_method` and no `client_secret` then that
is going to be bogus and we should reject it
Sometimes we fail to fetch events during backfill due to missing state,
and we often end up querying the same bad events periodically (as people
backpaginate). In such cases its likely we will continue to fail to get
the state, and therefore we should try *before* loading the state that
we have from the DB (as otherwise it's wasted DB and memory).
---------
Co-authored-by: reivilibre <oliverw@matrix.org>
There are two changes here:
1. Only pull out the required state when handling the request.
2. Change the get filtered state return type to check that we're only
querying state that was requested
---------
Co-authored-by: reivilibre <oliverw@matrix.org>
Instead of persisting outliers in a bunch of batches, let's just do them
all at once.
This is fine because all `_auth_and_persist_outliers_inner` is doing is
checking the auth rules for each event, which requires the events to be
topologically sorted by the auth graph.
The idea here being that the directory server shouldn't advertise rooms
to a requesting server is the requesting server would not be allowed to
join or participate in the room.
<!--
Fixes: # <!-- -->
<!--
Supersedes: # <!-- -->
<!--
Follows: # <!-- -->
<!--
Part of: # <!-- -->
Base: `develop` <!-- git-stack-base-branch:develop -->
<!--
This pull request is commit-by-commit review friendly. <!-- -->
<!--
This pull request is intended for commit-by-commit review. <!-- -->
Original commit schedule, with full messages:
<ol>
<li>
Pass `from_federation_origin` down into room list retrieval code
</li>
<li>
Don't cache /publicRooms response for inbound federated requests
</li>
<li>
fixup! Don't cache /publicRooms response for inbound federated requests
</li>
<li>
Cap the number of /publicRooms entries to 100
</li>
<li>
Simplify code now that you can't request unlimited rooms
</li>
<li>
Filter out rooms from federated requests that don't have the correct ACL
</li>
<li>
Request a handful more when filtering ACLs so that we can try to avoid
shortchanging the requester
</li>
</ol>
---------
Signed-off-by: Olivier Wilkinson (reivilibre) <oliverw@matrix.org>
There are a couple of things we need to be careful of here:
1. The current python code does no validation when loading from the DB,
so we need to be careful to ignore such errors (at least on jki.re there
are some old events with internal metadata fields of the wrong type).
2. We want to be memory efficient, as we often have many hundreds of
thousands of events in the cache at a time.
---------
Co-authored-by: Quentin Gliech <quenting@element.io>
We remove these fields as they're just duplicating data the event
already stores, and (for reasons 🤫) I'd like to simplify
the class to only store simple types.
I'm not entirely convinced that we shouldn't instead add helper methods
to the event class to generate stream tokens, but I don't really think
that's where they belong either
This is an extra response parameter just added to MSC3981. In the
current impl, the recursion depth is always 3, so this just returns a
static 3 if the recurse parameter is supplied.
Previously, the response status of `HTMLResource` was hardcoded as
`200`. However, for proper redirection after the user verifies their
email, we require the status to be `302`. This PR addresses that issue
by using `code` as response status.
Closes:
- https://github.com/matrix-org/synapse/issues/10397
- #10397
An administrator should know whether he wants to set a password or not.
There are many uses cases where a blank password is required.
- Use of only some users with SSO.
- Use of bots with password, users with SSO
This reverts two commits:
0bb8e418a4
"Fix postgres schema after dropping old tables (#16730)"
and
51e4e35653
"Add a Postgres `REPLICA IDENTITY` to tables that do not have an implicit one. This should allow use of Postgres logical replication. (take 2, now with no added deadlocks!) (#16658)"
and also amends the changelog.
* Add `ALTER TABLE ... REPLICA IDENTITY ...` for individual tables
We can't combine them into one file as it makes it likely to hit a deadlock
if Synapse is running, as it only takes one other transaction to access two
tables in a different order to the schema delta.
* Add notes
* Newsfile
Signed-off-by: Olivier Wilkinson (reivilibre) <oliverw@matrix.org>
* Re-introduce REPLICA IDENTITY test
---------
Signed-off-by: Olivier Wilkinson (reivilibre) <oliverw@matrix.org>
Implement MSC3860 to follow redirects for federated media downloads.
Note that the Client-Server API doesn't support this (yet) since the media
repository in Synapse doesn't have a way of supporting redirects.
* Describe `insert_client_ip`
* Pull out client_ips and MAU tracking to BaseAuth
* Define HAS_AUTHLIB once in tests
sick of copypasting
* Track ips and token usage when delegating auth
* Test that we track MAU and user_ips
* Don't track `__oidc_admin`
Keeping track of a lower bound of stream ID where we've deleted everything below makes the queries much faster. Otherwise, every time we scan for rows to delete we'd re-scan across all the rows that have previously deleted (until the next table VACUUM).
If a worker reconnects to Redis we send out the current positions of all our streams. However, if we're also trying to send out a backlog of RDATA at the same time then we can end up sending a `POSITION` with the current token *before* we've sent all the RDATA before the current token.
This doesn't cause actual bugs as the receiving servers see the POSITION, fetch the relevant rows from the DB, and then ignore the old RDATA as they come in. However, this is inefficient so it'd be better if we didn't send out-of-order positions
* Fix the CI query that did not detect all cases of missing primary keys
* Add more missing REPLICA IDENTITY entries
* Newsfile
Signed-off-by: Olivier Wilkinson (reivilibre) <oliverw@matrix.org>
---------
Signed-off-by: Olivier Wilkinson (reivilibre) <oliverw@matrix.org>
* Add Postgres replica identities to tables that don't have an implicit one
Fixes#16224
* Newsfile
Signed-off-by: Olivier Wilkinson (reivilibre) <oliverw@matrix.org>
* Move the delta to version 83 as we missed the boat for 82
* Add a test that all tables have a REPLICA IDENTITY
* Extend the test to include when indices are deleted
* isort
* black
* Fully qualify `oid` as it is a 'hidden attribute' in Postgres 11
* Update tests/storage/test_database.py
Co-authored-by: Patrick Cloke <clokep@users.noreply.github.com>
* Add missed tables
---------
Signed-off-by: Olivier Wilkinson (reivilibre) <oliverw@matrix.org>
Co-authored-by: Patrick Cloke <clokep@users.noreply.github.com>
If simple_{insert,upsert,update}_many_txn is called without any data
to modify then return instead of executing the query.
This matches the behavior of simple_{select,delete}_many_txn.
Fetch information needed for push rule evaluation in parallel.
Ideally this would use query pipelining, but this is not
available in psycopg2.
Due to the database thread pool this may result in little
to no parallelization.
Previously only Twisted's EPollReactor was compatible with the
reactor timing metric, notably not working when asyncio was used.
After this change, the following configurations support the reactor
timing metric:
* poll, epoll, or select reactors
* asyncio reactor with a poll, epoll, select, /dev/poll, or kqueue event loop.
The event persistence code used to handle multiple rooms
at a time, but was simplified to only ever be called with a
single room at a time (different rooms are now handled in
parallel). The code is still generic to multiple rooms causing
a lot of work that is unnecessary (e.g. unnecessary loops, and
partitioning data by room).
This strips out the ability to handle multiple rooms at once, greatly
simplifying the code.
Just to standardize on the normal helpers, it might also have
a slight perf improvement on PostgreSQL which will now use
`ANY (?)` instead of `IN (?, ?, ...)`.
This is mostly useful for federated rooms where some users
would get stuck in the invite or knock state when the room
was purged from their homeserver.
This adds a module API which allows a module to update a user's
presence state/status message. This is useful for controlling presence
from an external system.
To fully control presence from the module the presence.enabled config
parameter gains a new state of "untracked" which disables internal tracking
of presence changes via user actions, etc. Only updates from the module will
be persisted and sent down sync properly).
* Fix bug where a new writer advances their token too quickly
When starting a new writer (for e.g. persisting events), the
`MultiWriterIdGenerator` doesn't have a minimum token for it as there
are no rows matching that new writer in the DB.
This results in the the first stream ID it acquired being announced as
persisted *before* it actually finishes persisting, if another writer
gets and persists a subsequent stream ID. This is due to the logic of
setting the minimum persisted position to the minimum known position of
across all writers, and the new writer starts off not being considered.
* Fix sending out POSITIONs when our token advances without update
Broke in #14820
* For replication HTTP requests, only wait for minimal position
This could happen if the last rows in the account data stream were inserted into `account_data`. After a restart the max account ID would be calculated without looking at the `account_data` table, and so have an old ID.
If using the script remotely, there's no particularly convincing reason
to disable certificate verification, as this makes the connection
interceptible.
If on the other hand, the script is used locally (the most common use
case), you can simply target the HTTP listener and avoid TLS altogether.
This is what the script already attempts to do if passed a homeserver
configuration YAML file.
This splits thinsg into two queries, but most of the time we won't have
new event backwards extremities so this shouldn't actually add an extra
RTT for the majority of cases.
Note this removes the check for events with no prev events, but that was
part of MSC2716 work that has since been removed.
Synapse was incorrectly implemented with a knock_state_events
property on some APIs (instead of knock_room_state). This was
correct in Synapse 1.70.0, but *both* fields were sent to also be
compatible with Synapse versions expecting the wrong field.
Enough time has passed that only the correct field needs to be
included/handled.
This converts the media servlet URLs in the same way as
(most) of the rest of Synapse. This will give more flexibility
in the versions each endpoint exists under.
This avoids calling cursor_to_dict and then immediately
unpacking the values in the dict for other users. By not
creating the intermediate dictionary we can avoid allocating
the dictionary and strings for the keys, which should generally
be more performant.
Additionally this improves type hints by avoid Dict[str, Any]
dictionaries coming out of the database layer.
Co-authored-by: David Robertson <davidr@element.io>
Co-authored-by: Patrick Cloke <patrickc@matrix.org>
Co-authored-by: Erik Johnston <erik@matrix.org>
Assert that the return type of callables wrapped in @cached
and @cachedList are cachable (aka immutable).
This is because if a worker reaches ~100% CPU then everything starts
lagging and we hit the log line a lot. When at error we invoke sentry
and that has a lot of overhead, which then puts even more pressure on
the worker.
