Skip to content
This repository has been archived by the owner on Apr 26, 2024. It is now read-only.

Prefill more stream change caches. #12372

Merged
merged 7 commits into from
Apr 5, 2022
Merged

Conversation

erikjohnston
Copy link
Member

@erikjohnston erikjohnston commented Apr 5, 2022

We also make it so that get_cache_dict handles large gaps in stream IDs better (for device_lists_stream table).

We also increase the device_lists_stream limit a bit, as cache misses are expensive for /sync.

This follows on from work in #12367

@erikjohnston erikjohnston requested a review from a team as a code owner April 5, 2022 08:48
@babolivier babolivier self-assigned this Apr 5, 2022
Comment on lines +2041 to +2046
sql = f"""
SELECT {entity_column}, {stream_column}
FROM {table}
ORDER BY {stream_column} DESC
LIMIT ?
"""
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This change means that if an entity has many updates to its name (and others don't) we will end up with a smaller cache than before, right? Are we fine with that (eg because the limit is high enough that we don't really have to care)?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Before we were looking for rows in the range [MAX - LIMIT, MAX], which will return less than LIMIT ¹, so I think this will always return more rows than before.

¹ ... except if we have rows with the same stream ID, in which case this breaks.

@@ -0,0 +1 @@
Reduce overhead of restarting synchrotrons.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Doesn't it actually create more overhead, since we need to pull thousands of rows out of the database at startup?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah, so the reason this works is that we do the work once at start up, rather than many times for lots of different requests. The key thing here is that the stream change caches do not get updated on reads, but on the writes (or when we prefill them on startup).

Also c.f. #12367 (comment) why prefilling is good in general

Copy link
Contributor

@babolivier babolivier Apr 5, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I get why prefilling is a good thing, I'm just curious about the changelog claiming it reduces the restart the overhead of restarting synchrotrons but it actually increases it.

Edit: I reread the comment you're linking to and realised that it does somewhat answer my question

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Ah, I think I see what you mean: yes it does cause the initial restart to take longer, but then the first 10 minutes of up time uses much less DB and CPU. So I mean more along the lines of "reduce resource usage for the first few minutes after restart". Not sure if there is a better way of wording the changelog?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not sure if there is a better way of wording the changelog?

Don't think there is no, at least not in the way that would be user-friendly.

@erikjohnston
Copy link
Member Author

OH BLAST. I've just realised that this doesn't work for streams where we can have multiple rows per stream ID. GAH.

@erikjohnston
Copy link
Member Author

OH BLAST. I've just realised that this doesn't work for streams where we can have multiple rows per stream ID. GAH.

I've fixed this by simply adding one to the returned minimum stream ID

@erikjohnston erikjohnston merged commit 66053b6 into develop Apr 5, 2022
@erikjohnston erikjohnston deleted the erikj/get_cache_dict branch April 5, 2022 13:26
babolivier added a commit to matrix-org/synapse-dinsic that referenced this pull request Apr 28, 2022
Synapse 1.57.0 (2022-04-19)
===========================

This version includes a [change](matrix-org/synapse#12209) to the way transaction IDs are managed for application services. If your deployment uses a dedicated worker for application service traffic, **it must be stopped** when the database is upgraded (which normally happens when the main process is upgraded), to ensure the change is made safely without any risk of reusing transaction IDs.

See the [upgrade notes](https://github.com/matrix-org/synapse/blob/v1.57.0rc1/docs/upgrade.md#upgrading-to-v1570) for more details.

No significant changes since 1.57.0rc1.

Synapse 1.57.0rc1 (2022-04-12)
==============================

Features
--------

- Send device list changes to application services as specified by [MSC3202](matrix-org/matrix-spec-proposals#3202), using unstable prefixes. The `msc3202_transaction_extensions` experimental homeserver config option must be enabled and `org.matrix.msc3202: true` must be present in the application service registration file for device list changes to be sent. The "left" field is currently always empty. ([\#11881](matrix-org/synapse#11881))
- Optimise fetching large quantities of missing room state over federation. ([\#12040](matrix-org/synapse#12040))
- Offload the `update_client_ip` background job from the main process to the background worker, when using Redis-based replication. ([\#12251](matrix-org/synapse#12251))
- Move `update_client_ip` background job from the main process to the background worker. ([\#12252](matrix-org/synapse#12252))
- Add a module callback to react to new 3PID (email address, phone number) associations. ([\#12302](matrix-org/synapse#12302))
- Add a configuration option to remove a specific set of rooms from sync responses. ([\#12310](matrix-org/synapse#12310))
- Add a module callback to react to account data changes. ([\#12327](matrix-org/synapse#12327))
- Allow setting user admin status using the module API. Contributed by Famedly. ([\#12341](matrix-org/synapse#12341))
- Reduce overhead of restarting synchrotrons. ([\#12367](matrix-org/synapse#12367), [\#12372](matrix-org/synapse#12372))
- Update `/messages` to use historic pagination tokens if no `from` query parameter is given. ([\#12370](matrix-org/synapse#12370))
- Add a module API for reading and writing global account data. ([\#12391](matrix-org/synapse#12391))
- Support the stable `v1` endpoint for `/relations`, per [MSC2675](matrix-org/matrix-spec-proposals#2675). ([\#12403](matrix-org/synapse#12403))
- Include bundled aggregations in search results
  ([MSC3666](matrix-org/matrix-spec-proposals#3666)). ([\#12436](matrix-org/synapse#12436))

Bugfixes
--------

- Fix a long-standing bug where updates to the server notices user profile (display name/avatar URL) in the configuration would not be applied to pre-existing rooms. Contributed by Jorge Florian. ([\#12115](matrix-org/synapse#12115))
- Fix a long-standing bug where events from ignored users were still considered for bundled aggregations. ([\#12235](matrix-org/synapse#12235), [\#12338](matrix-org/synapse#12338))
- Fix non-member state events not resolving for historical events when used in [MSC2716](matrix-org/matrix-spec-proposals#2716) `/batch_send` `state_events_at_start`. ([\#12329](matrix-org/synapse#12329))
- Fix a long-standing bug affecting URL previews that would generate a 500 response instead of a 403 if the previewed URL includes a port that isn't allowed by the relevant blacklist. ([\#12333](matrix-org/synapse#12333))
- Default to `private` room visibility rather than `public` when a client does not specify one, according to spec. ([\#12350](matrix-org/synapse#12350))
- Fix a spec compliance issue where requests to the `/publicRooms` federation API would specify `limit` as a string. ([\#12364](matrix-org/synapse#12364), [\#12410](matrix-org/synapse#12410))
- Fix a bug introduced in Synapse 1.49.0 which caused the `synapse_event_persisted_position` metric to have invalid values. ([\#12390](matrix-org/synapse#12390))

Updates to the Docker image
---------------------------

- Bundle locked versions of dependencies into the Docker image. ([\#12385](matrix-org/synapse#12385), [\#12439](matrix-org/synapse#12439))
- Fix up healthcheck generation for workers docker image. ([\#12405](matrix-org/synapse#12405))

Improved Documentation
----------------------

- Clarify documentation for running SyTest against Synapse, including use of Postgres and worker mode. ([\#12271](matrix-org/synapse#12271))
- Document the behaviour of `LoggingTransaction.call_after` and `LoggingTransaction.call_on_exception` methods when transactions are retried. ([\#12315](matrix-org/synapse#12315))
- Update dead links in `check-newsfragment.sh` to point to the correct documentation URL. ([\#12331](matrix-org/synapse#12331))
- Upgrade the version of `mdbook` in CI to 0.4.17. ([\#12339](matrix-org/synapse#12339))
- Updates to the Room DAG concepts development document to clarify that we mark events as outliers because we don't have any state for them. ([\#12345](matrix-org/synapse#12345))
- Update the link to Redis pub/sub documentation in the workers documentation. ([\#12369](matrix-org/synapse#12369))
- Remove documentation for converting a legacy structured logging configuration to the new format. ([\#12392](matrix-org/synapse#12392))

Deprecations and Removals
-------------------------

- Remove the unused and unstable `/aggregations` endpoint which was removed from [MSC2675](matrix-org/matrix-spec-proposals#2675). ([\#12293](matrix-org/synapse#12293))

Internal Changes
----------------

- Remove lingering unstable references to MSC2403 (knocking). ([\#12165](matrix-org/synapse#12165))
- Avoid trying to calculate the state at outlier events. ([\#12191](matrix-org/synapse#12191), [\#12316](matrix-org/synapse#12316), [\#12330](matrix-org/synapse#12330), [\#12332](matrix-org/synapse#12332), [\#12409](matrix-org/synapse#12409))
- Omit sending "offline" presence updates to application services after they are initially configured. ([\#12193](matrix-org/synapse#12193))
- Switch to using a sequence to generate AS transaction IDs. Contributed by Nick @ Beeper. If running synapse with a dedicated appservice worker, this MUST be stopped before upgrading the main process and database. ([\#12209](matrix-org/synapse#12209))
- Add missing type hints for storage. ([\#12267](matrix-org/synapse#12267))
- Add missing type definitions for scripts in docker folder. Contributed by Jorge Florian. ([\#12280](matrix-org/synapse#12280))
- Move [MSC2654](matrix-org/matrix-spec-proposals#2654) support behind an experimental configuration flag. ([\#12295](matrix-org/synapse#12295))
- Update docstrings to explain how to decipher live and historic pagination tokens. ([\#12317](matrix-org/synapse#12317))
- Add ground work for speeding up device list updates for users in large numbers of rooms. ([\#12321](matrix-org/synapse#12321))
- Fix typechecker problems exposed by signedjson 1.1.2. ([\#12326](matrix-org/synapse#12326))
- Remove the `tox` packaging job: it will be redundant once #11537 lands. ([\#12334](matrix-org/synapse#12334))
- Ignore `.envrc` for `direnv` users. ([\#12335](matrix-org/synapse#12335))
- Remove the (broadly unused, dev-only) dockerfile for pg tests. ([\#12336](matrix-org/synapse#12336))
- Remove redundant `get_success` calls in test code. ([\#12346](matrix-org/synapse#12346))
- Add type annotations for `tests/unittest.py`. ([\#12347](matrix-org/synapse#12347))
- Move single-use methods out of `TestCase`. ([\#12348](matrix-org/synapse#12348))
- Remove broken and unused development scripts. ([\#12349](matrix-org/synapse#12349), [\#12351](matrix-org/synapse#12351), [\#12355](matrix-org/synapse#12355))
- Convert `Linearizer` tests from `inlineCallbacks` to async. ([\#12353](matrix-org/synapse#12353))
- Update docstrings for `ReadWriteLock` tests. ([\#12354](matrix-org/synapse#12354))
- Refactor `Linearizer`, convert methods to async and use an async context manager. ([\#12357](matrix-org/synapse#12357))
- Fix a long-standing bug where `Linearizer`s could get stuck if a cancellation were to happen at the wrong time. ([\#12358](matrix-org/synapse#12358))
- Make `StreamToken.from_string` and `RoomStreamToken.parse` propagate cancellations instead of replacing them with `SynapseError`s. ([\#12366](matrix-org/synapse#12366))
- Add type hints to tests files. ([\#12371](matrix-org/synapse#12371))
- Allow specifying the Postgres database's port when running unit tests with Postgres. ([\#12376](matrix-org/synapse#12376))
- Remove temporary pin of signedjson<=1.1.1 that was added in Synapse 1.56.0. ([\#12379](matrix-org/synapse#12379))
- Add opentracing spans to calls to external cache. ([\#12380](matrix-org/synapse#12380))
- Lay groundwork for using `poetry` to manage Synapse's dependencies. ([\#12381](matrix-org/synapse#12381), [\#12407](matrix-org/synapse#12407), [\#12412](matrix-org/synapse#12412), [\#12418](matrix-org/synapse#12418))
- Make missing `importlib_metadata` dependency explicit. ([\#12384](matrix-org/synapse#12384), [\#12400](matrix-org/synapse#12400))
- Update type annotations for compatiblity with prometheus_client 0.14. ([\#12389](matrix-org/synapse#12389))
- Remove support for the unstable identifiers specified in [MSC3288](matrix-org/matrix-spec-proposals#3288). ([\#12398](matrix-org/synapse#12398))
- Add missing type hints to configuration classes. ([\#12402](matrix-org/synapse#12402))
- Add files used to build the Docker image used for complement testing into the Synapse repository. ([\#12404](matrix-org/synapse#12404))
- Do not include groups in the sync response when disabled. ([\#12408](matrix-org/synapse#12408))
- Improve type hints related to HTTP query parameters. ([\#12415](matrix-org/synapse#12415))
- Stop maintaining a list of lint targets. ([\#12420](matrix-org/synapse#12420))
- Make `synapse._scripts` pass type checks. ([\#12421](matrix-org/synapse#12421), [\#12422](matrix-org/synapse#12422))
- Add some type hints to datastore. ([\#12423](matrix-org/synapse#12423))
- Enable certificate checking during complement tests. ([\#12435](matrix-org/synapse#12435))
- Explicitly specify the `tls` extra for Twisted dependency. ([\#12444](matrix-org/synapse#12444))
PiotrKozimor added a commit to globekeeper/synapse that referenced this pull request May 5, 2022
Synapse 1.57.0 (2022-04-19)
===========================

This version includes a [change](matrix-org#12209) to the way transaction IDs are managed for application services. If your deployment uses a dedicated worker for application service traffic, **it must be stopped** when the database is upgraded (which normally happens when the main process is upgraded), to ensure the change is made safely without any risk of reusing transaction IDs.

See the [upgrade notes](https://github.com/matrix-org/synapse/blob/v1.57.0rc1/docs/upgrade.md#upgrading-to-v1570) for more details.

No significant changes since 1.57.0rc1.

Synapse 1.57.0rc1 (2022-04-12)
==============================

Features
--------

- Send device list changes to application services as specified by [MSC3202](matrix-org/matrix-spec-proposals#3202), using unstable prefixes. The `msc3202_transaction_extensions` experimental homeserver config option must be enabled and `org.matrix.msc3202: true` must be present in the application service registration file for device list changes to be sent. The "left" field is currently always empty. ([\matrix-org#11881](matrix-org#11881))
- Optimise fetching large quantities of missing room state over federation. ([\matrix-org#12040](matrix-org#12040))
- Offload the `update_client_ip` background job from the main process to the background worker, when using Redis-based replication. ([\matrix-org#12251](matrix-org#12251))
- Move `update_client_ip` background job from the main process to the background worker. ([\matrix-org#12252](matrix-org#12252))
- Add a module callback to react to new 3PID (email address, phone number) associations. ([\matrix-org#12302](matrix-org#12302))
- Add a configuration option to remove a specific set of rooms from sync responses. ([\matrix-org#12310](matrix-org#12310))
- Add a module callback to react to account data changes. ([\matrix-org#12327](matrix-org#12327))
- Allow setting user admin status using the module API. Contributed by Famedly. ([\matrix-org#12341](matrix-org#12341))
- Reduce overhead of restarting synchrotrons. ([\matrix-org#12367](matrix-org#12367), [\matrix-org#12372](matrix-org#12372))
- Update `/messages` to use historic pagination tokens if no `from` query parameter is given. ([\matrix-org#12370](matrix-org#12370))
- Add a module API for reading and writing global account data. ([\matrix-org#12391](matrix-org#12391))
- Support the stable `v1` endpoint for `/relations`, per [MSC2675](matrix-org/matrix-spec-proposals#2675). ([\matrix-org#12403](matrix-org#12403))
- Include bundled aggregations in search results
  ([MSC3666](matrix-org/matrix-spec-proposals#3666)). ([\matrix-org#12436](matrix-org#12436))

Bugfixes
--------

- Fix a long-standing bug where updates to the server notices user profile (display name/avatar URL) in the configuration would not be applied to pre-existing rooms. Contributed by Jorge Florian. ([\matrix-org#12115](matrix-org#12115))
- Fix a long-standing bug where events from ignored users were still considered for bundled aggregations. ([\matrix-org#12235](matrix-org#12235), [\matrix-org#12338](matrix-org#12338))
- Fix non-member state events not resolving for historical events when used in [MSC2716](matrix-org/matrix-spec-proposals#2716) `/batch_send` `state_events_at_start`. ([\matrix-org#12329](matrix-org#12329))
- Fix a long-standing bug affecting URL previews that would generate a 500 response instead of a 403 if the previewed URL includes a port that isn't allowed by the relevant blacklist. ([\matrix-org#12333](matrix-org#12333))
- Default to `private` room visibility rather than `public` when a client does not specify one, according to spec. ([\matrix-org#12350](matrix-org#12350))
- Fix a spec compliance issue where requests to the `/publicRooms` federation API would specify `limit` as a string. ([\matrix-org#12364](matrix-org#12364), [\matrix-org#12410](matrix-org#12410))
- Fix a bug introduced in Synapse 1.49.0 which caused the `synapse_event_persisted_position` metric to have invalid values. ([\matrix-org#12390](matrix-org#12390))

Updates to the Docker image
---------------------------

- Bundle locked versions of dependencies into the Docker image. ([\matrix-org#12385](matrix-org#12385), [\matrix-org#12439](matrix-org#12439))
- Fix up healthcheck generation for workers docker image. ([\matrix-org#12405](matrix-org#12405))

Improved Documentation
----------------------

- Clarify documentation for running SyTest against Synapse, including use of Postgres and worker mode. ([\matrix-org#12271](matrix-org#12271))
- Document the behaviour of `LoggingTransaction.call_after` and `LoggingTransaction.call_on_exception` methods when transactions are retried. ([\matrix-org#12315](matrix-org#12315))
- Update dead links in `check-newsfragment.sh` to point to the correct documentation URL. ([\matrix-org#12331](matrix-org#12331))
- Upgrade the version of `mdbook` in CI to 0.4.17. ([\matrix-org#12339](matrix-org#12339))
- Updates to the Room DAG concepts development document to clarify that we mark events as outliers because we don't have any state for them. ([\matrix-org#12345](matrix-org#12345))
- Update the link to Redis pub/sub documentation in the workers documentation. ([\matrix-org#12369](matrix-org#12369))
- Remove documentation for converting a legacy structured logging configuration to the new format. ([\matrix-org#12392](matrix-org#12392))

Deprecations and Removals
-------------------------

- Remove the unused and unstable `/aggregations` endpoint which was removed from [MSC2675](matrix-org/matrix-spec-proposals#2675). ([\matrix-org#12293](matrix-org#12293))

Internal Changes
----------------

- Remove lingering unstable references to MSC2403 (knocking). ([\matrix-org#12165](matrix-org#12165))
- Avoid trying to calculate the state at outlier events. ([\matrix-org#12191](matrix-org#12191), [\matrix-org#12316](matrix-org#12316), [\matrix-org#12330](matrix-org#12330), [\matrix-org#12332](matrix-org#12332), [\matrix-org#12409](matrix-org#12409))
- Omit sending "offline" presence updates to application services after they are initially configured. ([\matrix-org#12193](matrix-org#12193))
- Switch to using a sequence to generate AS transaction IDs. Contributed by Nick @ Beeper. If running synapse with a dedicated appservice worker, this MUST be stopped before upgrading the main process and database. ([\matrix-org#12209](matrix-org#12209))
- Add missing type hints for storage. ([\matrix-org#12267](matrix-org#12267))
- Add missing type definitions for scripts in docker folder. Contributed by Jorge Florian. ([\matrix-org#12280](matrix-org#12280))
- Move [MSC2654](matrix-org/matrix-spec-proposals#2654) support behind an experimental configuration flag. ([\matrix-org#12295](matrix-org#12295))
- Update docstrings to explain how to decipher live and historic pagination tokens. ([\matrix-org#12317](matrix-org#12317))
- Add ground work for speeding up device list updates for users in large numbers of rooms. ([\matrix-org#12321](matrix-org#12321))
- Fix typechecker problems exposed by signedjson 1.1.2. ([\matrix-org#12326](matrix-org#12326))
- Remove the `tox` packaging job: it will be redundant once matrix-org#11537 lands. ([\matrix-org#12334](matrix-org#12334))
- Ignore `.envrc` for `direnv` users. ([\matrix-org#12335](matrix-org#12335))
- Remove the (broadly unused, dev-only) dockerfile for pg tests. ([\matrix-org#12336](matrix-org#12336))
- Remove redundant `get_success` calls in test code. ([\matrix-org#12346](matrix-org#12346))
- Add type annotations for `tests/unittest.py`. ([\matrix-org#12347](matrix-org#12347))
- Move single-use methods out of `TestCase`. ([\matrix-org#12348](matrix-org#12348))
- Remove broken and unused development scripts. ([\matrix-org#12349](matrix-org#12349), [\matrix-org#12351](matrix-org#12351), [\matrix-org#12355](matrix-org#12355))
- Convert `Linearizer` tests from `inlineCallbacks` to async. ([\matrix-org#12353](matrix-org#12353))
- Update docstrings for `ReadWriteLock` tests. ([\matrix-org#12354](matrix-org#12354))
- Refactor `Linearizer`, convert methods to async and use an async context manager. ([\matrix-org#12357](matrix-org#12357))
- Fix a long-standing bug where `Linearizer`s could get stuck if a cancellation were to happen at the wrong time. ([\matrix-org#12358](matrix-org#12358))
- Make `StreamToken.from_string` and `RoomStreamToken.parse` propagate cancellations instead of replacing them with `SynapseError`s. ([\matrix-org#12366](matrix-org#12366))
- Add type hints to tests files. ([\matrix-org#12371](matrix-org#12371))
- Allow specifying the Postgres database's port when running unit tests with Postgres. ([\matrix-org#12376](matrix-org#12376))
- Remove temporary pin of signedjson<=1.1.1 that was added in Synapse 1.56.0. ([\matrix-org#12379](matrix-org#12379))
- Add opentracing spans to calls to external cache. ([\matrix-org#12380](matrix-org#12380))
- Lay groundwork for using `poetry` to manage Synapse's dependencies. ([\matrix-org#12381](matrix-org#12381), [\matrix-org#12407](matrix-org#12407), [\matrix-org#12412](matrix-org#12412), [\matrix-org#12418](matrix-org#12418))
- Make missing `importlib_metadata` dependency explicit. ([\matrix-org#12384](matrix-org#12384), [\matrix-org#12400](matrix-org#12400))
- Update type annotations for compatiblity with prometheus_client 0.14. ([\matrix-org#12389](matrix-org#12389))
- Remove support for the unstable identifiers specified in [MSC3288](matrix-org/matrix-spec-proposals#3288). ([\matrix-org#12398](matrix-org#12398))
- Add missing type hints to configuration classes. ([\matrix-org#12402](matrix-org#12402))
- Add files used to build the Docker image used for complement testing into the Synapse repository. ([\matrix-org#12404](matrix-org#12404))
- Do not include groups in the sync response when disabled. ([\matrix-org#12408](matrix-org#12408))
- Improve type hints related to HTTP query parameters. ([\matrix-org#12415](matrix-org#12415))
- Stop maintaining a list of lint targets. ([\matrix-org#12420](matrix-org#12420))
- Make `synapse._scripts` pass type checks. ([\matrix-org#12421](matrix-org#12421), [\matrix-org#12422](matrix-org#12422))
- Add some type hints to datastore. ([\matrix-org#12423](matrix-org#12423))
- Enable certificate checking during complement tests. ([\matrix-org#12435](matrix-org#12435))
- Explicitly specify the `tls` extra for Twisted dependency. ([\matrix-org#12444](matrix-org#12444))
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants