1 files changed, 191 insertions, 0 deletions
diff --git a/docs/message_retention_policies.md b/docs/message_retention_policies.md
new file mode 100644
index 0000000000..78055b2f64
--- /dev/null
+++ b/docs/message_retention_policies.md
@@ -0,0 +1,191 @@
+# Message retention policies
+
+Synapse admins can enable support for message retention policies on
+their homeserver. Message retention policies exist at a room level,
+follow the semantics described in
+[MSC1763](https://github.com/matrix-org/matrix-doc/blob/matthew/msc1763/proposals/1763-configurable-retention-periods.md),
+and allow server and room admins to configure how long messages should
+be kept in a homeserver's database before being purged from it.
+
+A message retention policy is mainly defined by its `max_lifetime`
+parameter, which defines how long a message can be kept around after
+it's been sent in the room. If a room doesn't have a message retention
+policy, and there's no default one for a given server, then no message
+sent in that room is ever purged on that server.
+
+MSC1763 also specifies semantics for a `min_lifetime` parameter which
+defines the amount of time after which an event _can_ get purged (after
+it's been sent to the room), but Synapse doesn't currently support it
+beyond registering it.
+
+Both `max_lifetime` and `min_lifetime` are optional parameters.
+
+Note that message retention policies don't apply to state events.
+
+Once an event reaches its expiry date (defined as the time it was sent
+plus the value for `max_lifetime` in the room), two things happen:
+
+* Synapse stops serving the event to clients via any endpoint.
+* The message gets picked up by the next purge job (see the "Purge jobs"
+  section) and is removed from Synapse's database.
+
+Since purge jobs don't run continuously, this means that an event might
+stay in a server's database for longer than the value for `max_lifetime`
+in the room would allow, though hidden from clients.
+
+Similarly, if a server (with support for message retention policies
+enabled) receives from another server an event that should have been
+purged according to its room's policy, then the receiving server will
+process and store that event until it's picked up by the next purge job,
+though it will always hide it from clients.
+
+
+## Room configuration
+
+To configure a room's message retention policy, a room's admin or
+moderator needs to send a state event in that room with the type
+`m.room.retention` and the following content:
+
+```json
+{
+    "max_lifetime": ...
+}
+```
+
+In this event's content, the `max_lifetime` parameter has the same
+meaning as previously described, and needs to be expressed in
+milliseconds. The event's content can also include a `min_lifetime`
+parameter, which has the same meaning and limited support as previously
+described.
+
+Note that over every server in the room, only the ones with support for
+message retention policies will actually remove expired events. While
+we plan to eventually enable this support by default in Synapse, this
+isn't currently the case.
+
+
+## Server configuration
+
+Support for this feature can be enabled and configured in the
+`retention` section of the Synapse configuration file (see the
+[sample file](https://github.com/matrix-org/synapse/blob/v1.7.3/docs/sample_config.yaml#L332-L393)).
+
+To enable support for message retentions policies, set the setting
+`enabled` in this section to `true`.
+
+
+### Default policy
+
+A default message retention policy is a policy defined in Synapse's
+configuration that is used by Synapse for every room that doesn't have a
+message retention policy configured in its state. This allows server
+admins to ensure that messages are never kept indefinitely in a server's
+database. 
+
+A default policy can be defined as such, in the `retention` section of
+the configuration file:
+
+```yaml
+  default_policy:
+    min_lifetime: 1d
+    max_lifetime: 1y
+```
+
+Here, `min_lifetime` and `max_lifetime` have the same meaning and level
+of support as previously described. They can be expressed either as a
+duration (using the units `s` (seconds), `m` (minutes), `h` (hours),
+`d` (days), `w` (weeks) and `y` (years)) or as a number of milliseconds.
+
+
+### Purge jobs
+
+Purge jobs are the jobs that Synapse run in the background to purge
+expired events from the database. They are only run if support for
+message retention policies is enabled in the server's configuration. If
+no configuration for purge jobs is configured by the server admin,
+Synapse will run one daily that will handle every room with a message
+retention policy (or, if the server has a default policy configured,
+every room it knows), which should be enough in most cases.
+
+Some server admins might want a finer control on when events are removed
+depending on an event's room's policy. This can be done by setting the
+`purge_jobs` sub-section in the `retention` section of the configuration
+file. An example of such configuration could be:
+
+```yaml
+  purge_jobs:
+    - longest_max_lifetime: 3d
+      interval: 12h
+    - shortest_max_lifetime: 3d
+      longest_max_lifetime: 1w
+      interval: 1d
+    - shortest_max_lifetime: 1w
+      interval: 2d
+```
+
+In this example, we define two jobs:
+
+* one that runs twice a day (every 12 hours) and purges events in rooms
+  which policy's `max_lifetime` is lower or equal to 3 days.
+* one that runs once a day and purges events in rooms which policy's
+  `max_lifetime` is between 3 days and a week.
+* one that runs once every 2 days and purges events in rooms which
+  policy's `max_lifetime` is greater than a week.
+
+Note that this example is tailored to show different configurations and
+features slightly more jobs than it's probably necessary (in practice, a
+server admin would probably consider it better to replace the two last
+jobs with one that runs once a day and handles rooms which which
+policy's `max_lifetime` is greater than 3 days).
+
+Keep in mind, when configuring these jobs, that a purge job can become
+quite heavy on the server if it targets many rooms, therefore prefer
+having jobs with a low interval that target a limited set of rooms. Also
+make sure to include a job with no minimum and one with no maximum to
+make sure your configuration handles every policy.
+
+As previously mentioned in this documentation, while a purge job that
+runs e.g. every day means that an expired event might stay in the
+database for up to a day after its expiry, Synapse hides expired events
+from clients as soon as they expire, so the event is not visible to
+local users between its expiry date and the moment it gets purged from
+the server's database.
+
+
+### Lifetime limits
+
+**Note: this feature is mainly useful within a closed federation or on
+servers that don't federate, because there currently is no way to
+enforce these limits in an open federation.**
+
+Server admins can restrict the values their local users are allowed to
+use for both `min_lifetime` and `max_lifetime`. These limits can be
+defined as such in the `retention` section of the configuration file:
+
+```yaml
+  allowed_lifetime_min: 1d
+  allowed_lifetime_max: 1y
+```
+
+Here, `allowed_lifetime_min` is the lowest value a local user can set
+for both `min_lifetime` and `max_lifetime`, and `allowed_lifetime_max`
+is the highest value. Both parameters are optional (e.g. setting
+`allowed_lifetime_min` but not `allowed_lifetime_max` only enforces a
+minimum and no maximum).
+
+Like other settings in this section, these parameters can be expressed
+either as a duration or as a number of milliseconds.
+
+
+## Note on reclaiming disk space
+
+While purge jobs actually delete data from the database, the disk space
+used by the database might not decrease immediately on the database's
+host. However, even though the database engine won't free up the disk
+space, it will start writing new data into where the purged data was.
+
+If you want to reclaim the freed disk space anyway and return it to the
+operating system, the server admin needs to run `VACUUM FULL;` on the
+database (see the related
+[PostgreSQL documentation](https://www.postgresql.org/docs/current/sql-vacuum.html)).
+