Fediversity/matrix/synapse/workers.md

Table of Contents

• Introduction to a worker-based setup
• Redis
• Synapse
• Worker overview
• Types of workers
• Defining a worker
• Generic worker
• Media worker
• Worker logging
• Systemd

Introduction to a worker-based setup

Very busy servers are brought down because a single thread can't keep up with the load. So you want to create several threads for different types of work.

See this Matrix blog for some background information.

The traditional Synapse setup is one monolithic piece of software that does everything. Joining a very busy room makes a bottleneck, as the server will spend all its cycles on synchronizing that room.

You can split the server into workers, that are basically Synapse servers themselves. Redirect specific tasks to them and you have several different servers doing all kinds of tasks at the same time. A busy room will no longer freeze the rest.

Workers communicate with each other via UNIX sockets and Redis. We choose UNIX sockets because they're much more efficient than network sockets. Of course, if you scale to more than one machine, you will need network sockets instead.

Important note

While the use of workers can drastically improve speed, the law of diminished returns applies. Splitting off more and more workers will not further improve speed after a certain point. Plus: you need to understand what the most resource-consuming tasks are before you can start to plan how many workers for what tasks you need.

In this document we'll basically create a worker for every task, and several workers for a few heavy tasks, as an example. You mileage may not only vary, it will.

Tuning the rest of the machine and network also counts, especially PostgreSQL. A well-tuned PostgreSQL can make a really big difference and should probably be considered even before configuring workers.

With workers, PostgreSQL's configuration should be changed accordingly: see Tuning PostgreSQL for a Matrix Synapse server for hints and examples.

Redis

Workers need Redis as part of their communication, so our first step is to install Redis.

apt install redis-server

For less overhead we use a UNIX socket instead of a network connection to localhost. Disable the TCP listener and enable the socket in /etc/redis/redis.conf:

port 0

unixsocket /run/redis/redis-server.sock
unixsocketperm 770

Our matrix user (matrix-synapse) has to be able to read from and write to that socket, which is created by Redis and owned by redis:redis, so we add user matrix-synapse to the group redis. You may come up with a finer-grained permission solution, but for our example this will do.

adduser matrix-synapse redis

Restart Redis for these changes to take effect. Check for error messages in the logs, if port 6379 is no longer active, and if the socketfile /run/redis/redis-server.sock exists.

Now point Synapse at Redis in conf.d/redis.yaml:

redis:
  enabled: true
  path: /run/redis/redis-server.sock

Restart Synapse and check if it can connect to Redis via the socket, you should find log entries like this:

synapse.replication.tcp.redis - 292 - INFO - sentinel - Connecting to redis server UNIXAddress('/run/redis/redis-server.sock')
synapse.util.httpresourcetree - 56 - INFO - sentinel - Attaching <synapse.replication.http.ReplicationRestResource object at 0x7f95f850d150> to path b'/_synapse/replication'
synapse.replication.tcp.redis - 126 - INFO - sentinel - Connected to redis
synapse.replication.tcp.redis - 138 - INFO - subscribe-replication-0 - Sending redis SUBSCRIBE for ['matrix.example.com/USER_IP', 'matrix.example.com']
synapse.replication.tcp.redis - 141 - INFO - subscribe-replication-0 - Successfully subscribed to redis stream, sending REPLICATE command
synapse.replication.tcp.redis - 146 - INFO - subscribe-replication-0 - REPLICATE successfully sent

Synapse

Workers communicate with each other over sockets, that are all placed in one directory. These sockets are owned by matrix-synapse:matrix-synapse, so make sure nginx can write to them: add user www-data to group matrix-synapse and restart nginx.

Then, make sure systemd creates the directory for the sockets as soon as Synapse starts:

systemctl edit matrix-synapse

Now override parts of the Service stanza to add these two lines:

[Service]
RuntimeDirectory=matrix-synapse
RuntimeDirectoryPreserve=yes

The directory /run/matrix-synapse will be created as soon as Synapse starts, and will not be removed on restart or stop, because that would create problems with workers who suddenly lose their sockets.

Then we change Synapse from listening on localhost:8008 to listening on a socket. We'll do most of our workers work in conf.d/listeners.yaml, so let's put the new listener configuration for the main proccess there.

Remove the localhost:8008 stanza, and configure these two sockets:

listeners:
  - path: /run/matrix-synapse/inbound_main.sock
    mode: 0660
    type: http
    resources:
      - names:
        - client
        - consent
        - federation

  - path: /run/matrix-synapse/replication_main.sock
    mode: 0660
    type: http
    resources:
      - names:
        - replication 

This means Synapse will create two sockets under /run/matrix-synapse: one for incoming traffic that is forwarded by nginx (inbound_main.sock), and one for communicating with all the other workers (replication_main.sock).

If you restart Synapse now, it won't do anything anymore, because nginx is still forwarding its traffic to localhost:8008. We'll get to nginx later, but for now you should change:

proxy_forward http://localhost:8008;

to

proxy_forward http://unix:/run/matrix-synapse/inbound_main.sock;

If you've done this, restart Synapse and nginx, and check if the sockets are created and have the correct permissions.

Synapse should work normally again, we've switched from network sockets to UNIX sockets, and added Redis. Now we'll create the actual workers.

Worker overview

Every worker is, in fact, a Synapse server, only with a limited set of tasks. Some tasks can be handled by a number of workers, others only by one. Every worker starts as a normal Synapse process, reading all the normal configuration files, and then a bit of configuration for the specific worker itself.

Workers need to communicate with each other and the main process, they do that via the replication sockets under /run/matrix-synapse and Redis.

Most worker also need a way to be fed traffic by nginx: they have an inbound socket for that, in the same directory.

Finally, all those replicating workers need to be registered in the main process: all workers and their replication sockets are listed in the instance_map.

Types of workers

We'll make separate workers for almost every task, and several for the heaviest tasks: synchronising. An overview of what endpoints are to be forwarded to a worker is in Synapse's documentation.

We'll create the following workers:

• login
• federation_sender
• mediaworker
• userdir
• pusher
• push_rules
• typing
• todevice
• accountdata
• presence
• receipts
• initial_sync: 1 and 2
• normal_sync: 1, 2 and 3

Some of them are stream_writers, and the documentation about stream_witers says:

Note: The same worker can handle multiple streams, but unless otherwise documented, each stream can only have a single writer.

So, stream writers must have unique tasks: you can't have two or more workers writing to the same stream. Stream writers have to be listed in stream_writers:

stream_writers:
  account_data:
    - accountdata
  presence:
    - presence
  receipts:
    - receipts
  to_device:
    - todevice
  typing:
    - typing
  push_rules:
    - push_rules

As you can see, we've given the stream workers the name of the stream they're writing to. We could combine all those streams into one worker, which would probably be enough for most instances.

We could define a worker with the name streamwriter and list it under all streams instead of a single worker for every stream.

Finally, we have to list all these workers under instance_map: their name and their replication socket:

instance_map:
  main:
    path: "/run/matrix-synapse/replication_main.sock"
  login:
    path: "/run/matrix-synapse/replication_login.sock"
  federation_sender:
    path: "/run/matrix-synapse/replication_federation_sender.sock"
  mediaworker:
    path: "/run/matrix-synapse/replication_mediaworker.sock"
...
  normal_sync1:
    path: "unix:/run/matrix-synapse/replication_normal_sync1.sock"
  normal_sync2:
    path: "unix:/run/matrix-synapse/replication_normal_sync2.sock"
  normal_sync3:
    path: "unix:/run/matrix-synapse/replication_normal_sync3.sock"

Defining a worker

Every working starts with the normal configuration files, and then loads its own. We put those files under /etc/matrix-synapse/workers. You have to create that directory, and make sure Synapse can read them. Being profesionally paranoid, we restrict access to that directory and the files in it:

mkdir /etc/matrix-synapse/workers
chown matrix-synapse:matrix-synapse /etc/matrix-synapse/workers
chmod 750 /etc/matrix-synapse-workers

We'll fill this directory with yaml files; one for each worker.

Generic worker

Workers look very much the same, very little configuration is needed. This is what you need:

• name
• replication socket (not every worker needs this)
• inbound socket (not every worker needs this)
• log configuration

One worker we use handles the login actions, this is how it's configured:

worker_app: "synapse.app.generic_worker"
worker_name: "login"
worker_log_config: "/etc/matrix-synapse/logconf.d/login.yaml"

worker_listeners:
  - path: "/run/matrix-synapse/inbound_login.sock"
    type: http
    resources:
      - names:
        - client
        - consent
        - federation

  - path: "/run/matrix-synapse/replication_login.sock"
    type: http
    resources:
      - names: [replication]

First listener is the socket where nginx sends all traffic related to logins to. You have to configure nginx to do that, we'll get to that later.

First line defines the type of worker. In the past there were quite a few different types, but most of them have been phased out in favour of one generic worker.

The worker_log_config defines how and where the worker logs. Of course you'll need to configure that too, see further.

The first listener is the inbound socket, that nginx uses to forward login related traffic to. Make sure nginx can write to this socket. The resources vary between workers.

The second listener is used for communication with the other workers and the main thread. The only resource it needs is replication. This socket needs to be listed in the instance_map in the main thread.

Of course, if you need to scale up to the point where you need more than one machine, these listeners can no longer use UNIX sockets, but will have to use the network. This creates extra overhead, so you want to use sockets whenever possible.

Media worker

The media worker is slightly different than the generic one. It doesn't use the synapse.app.generic_worker, but a specialised one: synapse.app.media_repository. To prevent the main process from handling media itself, you have to explicitly tell it to leave that to the worker, by adding this to the configuration (in our setup conf.d/listeners.yaml):

enable_media_repo: false
media_instance_running_background_jobs: mediaworker

The worker mediaworker looks like this:

worker_app: "synapse.app.media_repository"
worker_name: "mediaworker"
worker_log_config: "/etc/matrix-synapse/logconf.d/media.yaml"

worker_listeners:
  - path: "/run/matrix-synapse/inbound_mediaworker.sock"
    type: http
    resources:
      - names:
        - media
        - federation

  - path: "/run/matrix-synapse/replication_mediaworker.sock"
    type: http
    resources:
      - names: [replication]

If you use more than one mediaworker, know that they must all run on the same machine; scaling it over more than one machine will not work.

Worker logging

As stated before, you configure the logging of workers in a separate yaml file. As with the definitions of the workers themselves, you need a directory for that. We'll use /etc/matrix-synapse/logconf.d for that; make it and fix the permissions.

There's a lot you can configure for logging, but for now we'll give every worker the same layout. Here's the configuration for the login worker:

version: 1
formatters:
  precise:
    format: '%(asctime)s - %(name)s - %(lineno)d - %(levelname)s - %(request)s - %(message)s'
handlers:
  file:
    class: logging.handlers.TimedRotatingFileHandler
    formatter: precise
    filename: /var/log/matrix-synapse/login.log
    when: midnight
    backupCount: 3
    encoding: utf8

  buffer:
    class: synapse.logging.handlers.PeriodicallyFlushingMemoryHandler
    target: file
    capacity: 10
    flushLevel: 30
    period: 5

loggers:
  synapse.metrics:
    level: WARN
    handlers: [buffer]
  synapse.replication.tcp:
    level: WARN
    handlers: [buffer]
  synapse.util.caches.lrucache:
    level: WARN
    handlers: [buffer]
  twisted:
    level: WARN
    handlers: [buffer]
  synapse:
    level: INFO
    handlers: [buffer]

root:
  level: INFO
  handlers: [buffer]

The only thing you need to change if the filename to which the logs are written. You could create only one configuration and use that in every worker, but that would mean all logs will end up in the same file, which may not be what you want.

See the Python documentation for all the ins and outs of logging.

Systemd

You want Synapse and its workers managed by systemd. First of all we define a target: a group of services that belong together.

systemctl edit --force --full matrix-synapse.target

Feed it with this bit:

[Unit]
Description=Matrix Synapse with all its workers
After=network.target

[Install]
WantedBy=multi-user.target

First add matrix-synapse.service to this target by overriding the WantedBy in the unit file (systemctl edit matrix-synapse.service):

[Install]
WantedBy=matrix.target

The same WantedBy need to go in the unit files for every worker. For the workers we're using a template instead of separate unit files for every single one. Create the template:

systemctl edit --full --force matrix-synapse-worker@

Fill it with this content:

[Unit]
Description=Synapse %i
AssertPathExists=/etc/matrix-synapse/workers/%i.yaml

# This service should be restarted when the synapse target is restarted.
PartOf=matrix-synapse.target
ReloadPropagatedFrom=matrix-synapse.target

# if this is started at the same time as the main, let the main process start
# first, to initialise the database schema.
After=matrix-synapse.service

[Service]
Type=notify
NotifyAccess=main
User=matrix-synapse
WorkingDirectory=/var/lib/matrix-synapse
EnvironmentFile=-/etc/default/matrix-synapse
ExecStart=/opt/venvs/matrix-synapse/bin/python -m synapse.app.generic_worker --config-path=/etc/matrix-synapse/homeserver.yaml --config-path=/etc/matrix-synapse/conf.d/ --config-path=/etc/matrix-synapse/workers/%i.yaml
ExecReload=/bin/kill -HUP $MAINPID
Restart=always
RestartSec=3
SyslogIdentifier=matrix-synapse-%i

[Install]
WantedBy=matrix-synapse.target

Every worker needs to be enabled and started individually. Quickest way to do that, is to run a loop in the directory:

cd /etc/matrix-synapse/workers
for worker in `ls *yaml`; do systemctl enable --now matrix-synapse-worker@$worker; done

After a reboot, Synapse and all its workers should be started.