synapse/docs/workers.md

Scaling synapse via workers

For small instances it recommended to run Synapse in the default monolith mode. For larger instances where performance is a concern it can be helpful to split out functionality into multiple separate python processes. These processes are called 'workers', and are (eventually) intended to scale horizontally independently.

Synapse's worker support is under active development and subject to change as we attempt to rapidly scale ever larger Synapse instances. However we are documenting it here to help admins needing a highly scalable Synapse instance similar to the one running matrix.org.

All processes continue to share the same database instance, and as such, workers only work with PostgreSQL-based Synapse deployments. SQLite should only be used for demo purposes and any admin considering workers should already be running PostgreSQL.

Main process/worker communication

The processes communicate with each other via a Synapse-specific protocol called 'replication' (analogous to MySQL- or Postgres-style database replication) which feeds streams of newly written data between processes so they can be kept in sync with the database state.

When configured to do so, Synapse uses a Redis pub/sub channel to send the replication stream between all configured Synapse processes. Additionally, processes may make HTTP requests to each other, primarily for operations which need to wait for a reply ─ such as sending an event.

Redis support was added in v1.13.0 with it becoming the recommended method in v1.18.0. It replaced the old direct TCP connections (which is deprecated as of v1.18.0) to the main process. With Redis, rather than all the workers connecting to the main process, all the workers and the main process connect to Redis, which relays replication commands between processes. This can give a significant cpu saving on the main process and will be a prerequisite for upcoming performance improvements.

See the Architectural diagram section at the end for a visualisation of what this looks like.

Setting up workers

A Redis server is required to manage the communication between the processes. The Redis server should be installed following the normal procedure for your distribution (e.g. apt install redis-server on Debian). It is safe to use an existing Redis deployment if you have one.

Once installed, check that Redis is running and accessible from the host running Synapse, for example by executing echo PING | nc -q1 localhost 6379 and seeing a response of +PONG.

The appropriate dependencies must also be installed for Synapse. If using a virtualenv, these can be installed with:

pip install matrix-synapse[redis]

Note that these dependencies are included when synapse is installed with pip install matrix-synapse[all]. They are also included in the debian packages from matrix.org and in the docker images at https://hub.docker.com/r/matrixdotorg/synapse/.

To make effective use of the workers, you will need to configure an HTTP reverse-proxy such as nginx or haproxy, which will direct incoming requests to the correct worker, or to the main synapse instance. See reverse_proxy.md for information on setting up a reverse proxy.

When using workers, each worker process has its own configuration file which contains settings specific to that worker, such as the HTTP listener that it provides (if any), logging configuration, etc.

Normally, the worker processes are configured to read from a shared configuration file as well as the worker-specific configuration files. This makes it easier to keep common configuration settings synchronised across all the processes.

The main process is somewhat special in this respect: it does not normally need its own configuration file and can take all of its configuration from the shared configuration file.

Shared configuration

Normally, only a couple of changes are needed to make an existing configuration file suitable for use with workers. First, you need to enable an "HTTP replication listener" for the main process; and secondly, you need to enable redis-based replication. For example:

# extend the existing `listeners` section. This defines the ports that the
# main process will listen on.
listeners:
  # The HTTP replication port
  - port: 9093
    bind_address: '127.0.0.1'
    type: http
    resources:
     - names: [replication]

redis:
    enabled: true

See the sample config for the full documentation of each option.

Under no circumstances should the replication listener be exposed to the public internet; it has no authentication and is unencrypted.

Worker configuration

In the config file for each worker, you must specify the type of worker application (worker_app), and you should specify a unqiue name for the worker (worker_name). The currently available worker applications are listed below. You must also specify the HTTP replication endpoint that it should talk to on the main synapse process. worker_replication_host should specify the host of the main synapse and worker_replication_http_port should point to the HTTP replication port. If the worker will handle HTTP requests then the worker_listeners option should be set with a http listener, in the same way as the listeners option in the shared config.

For example:

worker_app: synapse.app.generic_worker
worker_name: worker1

# The replication listener on the main synapse process.
worker_replication_host: 127.0.0.1
worker_replication_http_port: 9093

worker_listeners:
 - type: http
   port: 8083
   resources:
     - names:
       - client
       - federation

worker_log_config: /home/matrix/synapse/config/worker1_log_config.yaml

...is a full configuration for a generic worker instance, which will expose a plain HTTP endpoint on port 8083 separately serving various endpoints, e.g. /sync, which are listed below.

Obviously you should configure your reverse-proxy to route the relevant endpoints to the worker (localhost:8083 in the above example).

Running Synapse with workers

Finally, you need to start your worker processes. This can be done with either synctl or your distribution's preferred service manager such as systemd. We recommend the use of systemd where available: for information on setting up systemd to start synapse workers, see systemd-with-workers. To use synctl, see synctl_workers.md.

Available worker applications

synapse.app.generic_worker

This worker can handle API requests matching the following regular expressions:

# Sync requests
^/_matrix/client/(v2_alpha|r0)/sync$
^/_matrix/client/(api/v1|v2_alpha|r0)/events$
^/_matrix/client/(api/v1|r0)/initialSync$
^/_matrix/client/(api/v1|r0)/rooms/[^/]+/initialSync$

# Federation requests
^/_matrix/federation/v1/event/
^/_matrix/federation/v1/state/
^/_matrix/federation/v1/state_ids/
^/_matrix/federation/v1/backfill/
^/_matrix/federation/v1/get_missing_events/
^/_matrix/federation/v1/publicRooms
^/_matrix/federation/v1/query/
^/_matrix/federation/v1/make_join/
^/_matrix/federation/v1/make_leave/
^/_matrix/federation/v1/send_join/
^/_matrix/federation/v2/send_join/
^/_matrix/federation/v1/send_leave/
^/_matrix/federation/v2/send_leave/
^/_matrix/federation/v1/invite/
^/_matrix/federation/v2/invite/
^/_matrix/federation/v1/query_auth/
^/_matrix/federation/v1/event_auth/
^/_matrix/federation/v1/exchange_third_party_invite/
^/_matrix/federation/v1/user/devices/
^/_matrix/federation/v1/get_groups_publicised$
^/_matrix/key/v2/query

# Inbound federation transaction request
^/_matrix/federation/v1/send/

# Client API requests
^/_matrix/client/(api/v1|r0|unstable)/publicRooms$
^/_matrix/client/(api/v1|r0|unstable)/rooms/.*/joined_members$
^/_matrix/client/(api/v1|r0|unstable)/rooms/.*/context/.*$
^/_matrix/client/(api/v1|r0|unstable)/rooms/.*/members$
^/_matrix/client/(api/v1|r0|unstable)/rooms/.*/state$
^/_matrix/client/(api/v1|r0|unstable)/account/3pid$
^/_matrix/client/(api/v1|r0|unstable)/keys/query$
^/_matrix/client/(api/v1|r0|unstable)/keys/changes$
^/_matrix/client/versions$
^/_matrix/client/(api/v1|r0|unstable)/voip/turnServer$
^/_matrix/client/(api/v1|r0|unstable)/joined_groups$
^/_matrix/client/(api/v1|r0|unstable)/publicised_groups$
^/_matrix/client/(api/v1|r0|unstable)/publicised_groups/
^/_synapse/client/password_reset/email/submit_token$

# Registration/login requests
^/_matrix/client/(api/v1|r0|unstable)/login$
^/_matrix/client/(r0|unstable)/register$
^/_matrix/client/(r0|unstable)/auth/.*/fallback/web$

# Event sending requests
^/_matrix/client/(api/v1|r0|unstable)/rooms/.*/send
^/_matrix/client/(api/v1|r0|unstable)/rooms/.*/state/
^/_matrix/client/(api/v1|r0|unstable)/rooms/.*/(join|invite|leave|ban|unban|kick)$
^/_matrix/client/(api/v1|r0|unstable)/join/
^/_matrix/client/(api/v1|r0|unstable)/profile/

Additionally, the following REST endpoints can be handled for GET requests:

^/_matrix/federation/v1/groups/

Pagination requests can also be handled, but all requests for a given room must be routed to the same instance. Additionally, care must be taken to ensure that the purge history admin API is not used while pagination requests for the room are in flight:

^/_matrix/client/(api/v1|r0|unstable)/rooms/.*/messages$

Note that a HTTP listener with client and federation resources must be configured in the worker_listeners option in the worker config.

Load balancing

It is possible to run multiple instances of this worker app, with incoming requests being load-balanced between them by the reverse-proxy. However, different endpoints have different characteristics and so admins may wish to run multiple groups of workers handling different endpoints so that load balancing can be done in different ways.

For /sync and /initialSync requests it will be more efficient if all requests from a particular user are routed to a single instance. Extracting a user ID from the access token or Authorization header is currently left as an exercise for the reader. Admins may additionally wish to separate out /sync requests that have a since query parameter from those that don't (and /initialSync), as requests that don't are known as "initial sync" that happens when a user logs in on a new device and can be very resource intensive, so isolating these requests will stop them from interfering with other users ongoing syncs.

Federation and client requests can be balanced via simple round robin.

The inbound federation transaction request ^/_matrix/federation/v1/send/ should be balanced by source IP so that transactions from the same remote server go to the same process.

Registration/login requests can be handled separately purely to help ensure that unexpected load doesn't affect new logins and sign ups.

Finally, event sending requests can be balanced by the room ID in the URI (or the full URI, or even just round robin), the room ID is the path component after /rooms/. If there is a large bridge connected that is sending or may send lots of events, then a dedicated set of workers can be provisioned to limit the effects of bursts of events from that bridge on events sent by normal users.

Stream writers

Additionally, there is experimental support for moving writing of specific streams (such as events) off of the main process to a particular worker. (This is only supported with Redis-based replication.)

Currently support streams are events and typing.

To enable this, the worker must have a HTTP replication listener configured, have a worker_name and be listed in the instance_map config. For example to move event persistence off to a dedicated worker, the shared configuration would include:

instance_map:
    event_persister1:
        host: localhost
        port: 8034

stream_writers:
    events: event_persister1

synapse.app.pusher

Handles sending push notifications to sygnal and email. Doesn't handle any REST endpoints itself, but you should set start_pushers: False in the shared configuration file to stop the main synapse sending push notifications.

Note this worker cannot be load-balanced: only one instance should be active.

synapse.app.appservice

Handles sending output traffic to Application Services. Doesn't handle any REST endpoints itself, but you should set notify_appservices: False in the shared configuration file to stop the main synapse sending appservice notifications.

Note this worker cannot be load-balanced: only one instance should be active.

synapse.app.federation_sender

Handles sending federation traffic to other servers. Doesn't handle any REST endpoints itself, but you should set send_federation: False in the shared configuration file to stop the main synapse sending this traffic.

If running multiple federation senders then you must list each instance in the federation_sender_instances option by their worker_name. All instances must be stopped and started when adding or removing instances. For example:

federation_sender_instances:
    - federation_sender1
    - federation_sender2

synapse.app.media_repository

Handles the media repository. It can handle all endpoints starting with:

/_matrix/media/

... and the following regular expressions matching media-specific administration APIs:

^/_synapse/admin/v1/purge_media_cache$
^/_synapse/admin/v1/room/.*/media.*$
^/_synapse/admin/v1/user/.*/media.*$
^/_synapse/admin/v1/media/.*$
^/_synapse/admin/v1/quarantine_media/.*$

You should also set enable_media_repo: False in the shared configuration file to stop the main synapse running background jobs related to managing the media repository.

In the media_repository worker configuration file, configure the http listener to expose the media resource. For example:

    worker_listeners:
     - type: http
       port: 8085
       resources:
         - names:
           - media

Note that if running multiple media repositories they must be on the same server and you must configure a single instance to run the background tasks, e.g.:

    media_instance_running_background_jobs: "media-repository-1"

synapse.app.user_dir

Handles searches in the user directory. It can handle REST endpoints matching the following regular expressions:

^/_matrix/client/(api/v1|r0|unstable)/user_directory/search$

When using this worker you must also set update_user_directory: False in the shared configuration file to stop the main synapse running background jobs related to updating the user directory.

synapse.app.frontend_proxy

Proxies some frequently-requested client endpoints to add caching and remove load from the main synapse. It can handle REST endpoints matching the following regular expressions:

^/_matrix/client/(api/v1|r0|unstable)/keys/upload

If use_presence is False in the homeserver config, it can also handle REST endpoints matching the following regular expressions:

^/_matrix/client/(api/v1|r0|unstable)/presence/[^/]+/status

This "stub" presence handler will pass through GET request but make the PUT effectively a no-op.

It will proxy any requests it cannot handle to the main synapse instance. It must therefore be configured with the location of the main instance, via the worker_main_http_uri setting in the frontend_proxy worker configuration file. For example:

worker_main_http_uri: http://127.0.0.1:8008

Historical apps

Note: Historically there used to be more apps, however they have been amalgamated into a single synapse.app.generic_worker app. The remaining apps are ones that do specific processing unrelated to requests, e.g. the pusher that handles sending out push notifications for new events. The intention is for all these to be folded into the generic_worker app and to use config to define which processes handle the various proccessing such as push notifications.

Migration from old config

There are two main independent changes that have been made: introducing Redis support and merging apps into synapse.app.generic_worker. Both these changes are backwards compatible and so no changes to the config are required, however server admins are encouraged to plan to migrate to Redis as the old style direct TCP replication config is deprecated.

To migrate to Redis add the redis config as above, and optionally remove the TCP replication listener from master and worker_replication_port from worker config.

To migrate apps to use synapse.app.generic_worker simply update the worker_app option in the worker configs, and where worker are started (e.g. in systemd service files, but not required for synctl).

Architectural diagram

The following shows an example setup using Redis and a reverse proxy:

                     Clients & Federation
                              |
                              v
                        +-----------+
                        |           |
                        |  Reverse  |
                        |  Proxy    |
                        |           |
                        +-----------+
                            | | |
                            | | | HTTP requests
        +-------------------+ | +-----------+
        |                 +---+             |
        |                 |                 |
        v                 v                 v
+--------------+  +--------------+  +--------------+  +--------------+
|   Main       |  |   Generic    |  |   Generic    |  |  Event       |
|   Process    |  |   Worker 1   |  |   Worker 2   |  |  Persister   |
+--------------+  +--------------+  +--------------+  +--------------+
      ^    ^          |   ^   |         |   ^   |          ^    ^
      |    |          |   |   |         |   |   |          |    |
      |    |          |   |   |  HTTP   |   |   |          |    |
      |    +----------+<--|---|---------+   |   |          |    |
      |                   |   +-------------|-->+----------+    |
      |                   |                 |                   |
      |                   |                 |                   |
      v                   v                 v                   v
====================================================================
                                                         Redis pub/sub channel