Skip to content

Conversation

edgurgel
Copy link
Member

@edgurgel edgurgel commented Oct 2, 2025

What kind of change does this PR introduce?

Support fastlane for phoenix presence_diff messages.

What is the current behavior?

presence_diff does not go through fastlane due to vanilla Phoenix.Presence using Phoenix.ChannelServer.dispatch/3 as dispatcher.

What is the new behavior?

Use a Phoenix fork with the option to pass a custom dispatcher to Phoenix.Presence. If we can get this PR merged we will just need to update Phoenix and ditch the fork

Another option is to copy and paste Phoenix.Presence and change the line that does the PubSub.local_broadcast

Additional context

Add any other context or screenshots.

@coveralls
Copy link

coveralls commented Oct 2, 2025

Coverage Status

coverage: 85.477% (-0.03%) from 85.506%
when pulling cce558e on feat/phoenix-presence-dispatcher
into 07de665 on main.

@edgurgel edgurgel force-pushed the feat/phoenix-presence-dispatcher branch from 3eed801 to c624928 Compare October 6, 2025 20:46
@edgurgel edgurgel merged commit ecac071 into main Oct 6, 2025
5 of 7 checks passed
@edgurgel edgurgel deleted the feat/phoenix-presence-dispatcher branch October 6, 2025 23:21
@edgurgel edgurgel changed the title Fastlane for phoenix presence_diff feat: fastlane for phoenix presence_diff Oct 7, 2025
Fudster added a commit to KBVE/realtime that referenced this pull request Oct 7, 2025
* fix: runtime setup error (supabase#1520)

* fix: use primary instead of replica on rename_settings_field (supabase#1521)

* feat: upgrade cowboy & ranch (supabase#1523)

* fix: Fix GenRpc to not try to connect to nodes that are not alive (supabase#1525)

* fix: enable presence on track message (supabase#1527)

currently the user would need to have enabled from the beginning of the channel. this will enable users to enable presence later in the flow by sending a track message which will enable presence messages for them

* fix: set cowboy active_n=100 as cowboy 2.12.0 (supabase#1530)

cowboy 2.13.0 set the default active_n=1

* fix: provide error_code metadata on RealtimeChannel.Logging (supabase#1531)

* feat: disable UTF8 validation on websocket frames (supabase#1532)

Currently all text frames as handled only with JSON which already requires UTF-8

* fix: move DB setup to happen after Connect.init (supabase#1533)

This change reduces the impact of slow DB setup impacting other tenants
trying to connect at the same time that landed on the same partition

* fix: handle wal bloat (supabase#1528)

Verify that replication connection is able to reconnect when faced with WAL bloat issues

* feat: replay realtime.messages (supabase#1526)

A new index was created on inserted_at DESC, topic WHERE private IS TRUE AND extension = "broadast"

The hardcoded limit is 25 for now.

* feat: gen_rpc pub sub adapter (supabase#1529)

Add a PubSub adapter that uses gen_rpc to send messages to other nodes.

It uses :gen_rpc.abcast/3 instead of :erlang.send/2

The adapter works very similarly to the PG2 adapter. It consists of
multiple workers that forward to the local node using PubSub.local_broadcast.

The way to choose the worker to be used is based on the sending process
just like PG2 adapter does

The number of workers is controlled by `:pool_size` or `:broadcast_pool_size`.
This distinction exists because Phoenix.PubSub uses `:pool_size` to
define how many partitions the PubSub registry will use. It's possible
to control them separately by using `:broadcast_pool_size`

* fix: ensure message id doesn't raise on non-map payloads (supabase#1534)

* fix: match error on Connect (supabase#1536)



---------

Co-authored-by: Eduardo Gurgel Pinho <[email protected]>

* feat: websocket max heap size configuration (supabase#1538)

* fix: set max process heap size to 500MB instead of 8GB
* feat: set websocket transport max heap size

WEBSOCKET_MAX_HEAP_SIZE can be used to configure it

* fix: update gen_rpc to fix gen_rpc_dispatcher issues (supabase#1537)

Issues:

* Single gen_rpc_dispatcher that can be a bottleneck if the connecting takes some time
* Many calls can land on the dispatcher but the node might be gone already. If we don't validate the node it might keep trying to connect until it times out instead of quickly giving up due to not being an actively connected node.

* fix: improve ErlSysMon logging for processes (supabase#1540)

Include initial_call, ancestors, registered_name, message_queue_len and total_heap_size

Also bump long_schedule and long_gc

* fix: make pubsub adapter configurable (supabase#1539)

* fix: specify that only private channels are allowed when replaying (supabase#1543)

messages

* fix: rate limit connect module (supabase#1541)

On bad connection, we rate limit the Connect module so we prevent abuses and too much logging of errors

* build: automatically cancel old tests/build on new push (supabase#1545)

Currently, whenever you push any commit to your branch, the old builds are still running and a new build is started. Once a new commit is added, the old test results no longer matter and it's just a waste of CI resources. Also reduces confusion with multiple builds running in parallel for the same branch/possibly blocking any merges.

With this little change, we ensure that whenever a new commit is added, the previous build is immediately canceled/stopped and only the build (latest commit) runs.

* fix: move message queue data to off-heap for gen_rpc pub sub workers (supabase#1548)

* fix: rate limit Connect.lookup_or_start_connection on error only (supabase#1549)

* fix: increase connect error rate window to 30 seconds (supabase#1550)

* fix: set a lower fullsweep_after flag for GenRpcPubSub workers (supabase#1551)

* fix: hardcode presence limit (supabase#1552)

* fix: further decrease limit on presence events (supabase#1553)

* fix: bump up realtime (supabase#1554)

* fix: lower rate limit to 100 events per second (supabase#1556)

* fix: move connect rate limit to socket (supabase#1555)

* fix: reduce max_frame_size to 5MB
* fix: fullsweep_after=100 on gen rpc pub sub workers

---------

Co-authored-by: Eduardo Gurgel Pinho <[email protected]>

* fix: collect global metrics without tenant tagging (supabase#1557)

* feat: presence payload size (supabase#1559)

* Also tweak buckets to account all the way to 3000KB
* Start tagging the payload size metrics with message_type. message_type can be presence, broadcast or postgres_changes

* fix: use GenRpc for Realtime.Latency pings (supabase#1560)

* Fastlane for phoenix presence_diff (supabase#1558)

It uses a fork of Phoenix for time being

* fix: count presence_diff events on MessageDispatcher
* fix: remove traces from console during development

---------

Co-authored-by: Filipe Cabaço <[email protected]>
Co-authored-by: Eduardo Gurgel <[email protected]>
Co-authored-by: Kevin Grüneberg <[email protected]>
Co-authored-by: Bradley Haljendi <[email protected]>
Fudster added a commit to KBVE/realtime that referenced this pull request Oct 7, 2025
* 🔄 Sync with upstream changes (#2)

* chore: fix couple of flaky tests (supabase#1517)

* fix: Improve runtime setup logic (supabase#1511)

Cleanup runtime.exs logic to be more organized and easier to mantain

* fix: runtime setup error (supabase#1520)

---------

Co-authored-by: Eduardo Gurgel <[email protected]>
Co-authored-by: Filipe Cabaço <[email protected]>

* 🔄 Sync with upstream changes (#4)

* fix: runtime setup error (supabase#1520)

* fix: use primary instead of replica on rename_settings_field (supabase#1521)

---------

Co-authored-by: Filipe Cabaço <[email protected]>
Co-authored-by: Eduardo Gurgel <[email protected]>
Co-authored-by: Bradley Haljendi <[email protected]>

* 🔄 Sync with upstream changes (#6)

* fix: runtime setup error (supabase#1520)

* fix: use primary instead of replica on rename_settings_field (supabase#1521)

---------

Co-authored-by: Filipe Cabaço <[email protected]>
Co-authored-by: Eduardo Gurgel <[email protected]>

* 🔄 Sync with upstream changes (#7)

* fix: runtime setup error (supabase#1520)

* fix: use primary instead of replica on rename_settings_field (supabase#1521)

* feat: upgrade cowboy & ranch (supabase#1523)

* fix: Fix GenRpc to not try to connect to nodes that are not alive (supabase#1525)

* fix: enable presence on track message (supabase#1527)

currently the user would need to have enabled from the beginning of the channel. this will enable users to enable presence later in the flow by sending a track message which will enable presence messages for them

* fix: set cowboy active_n=100 as cowboy 2.12.0 (supabase#1530)

cowboy 2.13.0 set the default active_n=1

* fix: provide error_code metadata on RealtimeChannel.Logging (supabase#1531)

* feat: disable UTF8 validation on websocket frames (supabase#1532)

Currently all text frames as handled only with JSON which already requires UTF-8

* fix: move DB setup to happen after Connect.init (supabase#1533)

This change reduces the impact of slow DB setup impacting other tenants
trying to connect at the same time that landed on the same partition

* fix: handle wal bloat (supabase#1528)

Verify that replication connection is able to reconnect when faced with WAL bloat issues

* feat: replay realtime.messages (supabase#1526)

A new index was created on inserted_at DESC, topic WHERE private IS TRUE AND extension = "broadast"

The hardcoded limit is 25 for now.

* feat: gen_rpc pub sub adapter (supabase#1529)

Add a PubSub adapter that uses gen_rpc to send messages to other nodes.

It uses :gen_rpc.abcast/3 instead of :erlang.send/2

The adapter works very similarly to the PG2 adapter. It consists of
multiple workers that forward to the local node using PubSub.local_broadcast.

The way to choose the worker to be used is based on the sending process
just like PG2 adapter does

The number of workers is controlled by `:pool_size` or `:broadcast_pool_size`.
This distinction exists because Phoenix.PubSub uses `:pool_size` to
define how many partitions the PubSub registry will use. It's possible
to control them separately by using `:broadcast_pool_size`

* fix: ensure message id doesn't raise on non-map payloads (supabase#1534)

* fix: match error on Connect (supabase#1536)



---------

Co-authored-by: Eduardo Gurgel Pinho <[email protected]>

* feat: websocket max heap size configuration (supabase#1538)

* fix: set max process heap size to 500MB instead of 8GB
* feat: set websocket transport max heap size

WEBSOCKET_MAX_HEAP_SIZE can be used to configure it

* fix: update gen_rpc to fix gen_rpc_dispatcher issues (supabase#1537)

Issues:

* Single gen_rpc_dispatcher that can be a bottleneck if the connecting takes some time
* Many calls can land on the dispatcher but the node might be gone already. If we don't validate the node it might keep trying to connect until it times out instead of quickly giving up due to not being an actively connected node.

* fix: improve ErlSysMon logging for processes (supabase#1540)

Include initial_call, ancestors, registered_name, message_queue_len and total_heap_size

Also bump long_schedule and long_gc

* fix: make pubsub adapter configurable (supabase#1539)

---------

Co-authored-by: Filipe Cabaço <[email protected]>
Co-authored-by: Eduardo Gurgel <[email protected]>
Co-authored-by: Bradley Haljendi <[email protected]>

* 🔄 Sync with upstream changes (#9)

* fix: runtime setup error (supabase#1520)

* fix: use primary instead of replica on rename_settings_field (supabase#1521)

* feat: upgrade cowboy & ranch (supabase#1523)

* fix: Fix GenRpc to not try to connect to nodes that are not alive (supabase#1525)

* fix: enable presence on track message (supabase#1527)

currently the user would need to have enabled from the beginning of the channel. this will enable users to enable presence later in the flow by sending a track message which will enable presence messages for them

* fix: set cowboy active_n=100 as cowboy 2.12.0 (supabase#1530)

cowboy 2.13.0 set the default active_n=1

* fix: provide error_code metadata on RealtimeChannel.Logging (supabase#1531)

* feat: disable UTF8 validation on websocket frames (supabase#1532)

Currently all text frames as handled only with JSON which already requires UTF-8

* fix: move DB setup to happen after Connect.init (supabase#1533)

This change reduces the impact of slow DB setup impacting other tenants
trying to connect at the same time that landed on the same partition

* fix: handle wal bloat (supabase#1528)

Verify that replication connection is able to reconnect when faced with WAL bloat issues

* feat: replay realtime.messages (supabase#1526)

A new index was created on inserted_at DESC, topic WHERE private IS TRUE AND extension = "broadast"

The hardcoded limit is 25 for now.

* feat: gen_rpc pub sub adapter (supabase#1529)

Add a PubSub adapter that uses gen_rpc to send messages to other nodes.

It uses :gen_rpc.abcast/3 instead of :erlang.send/2

The adapter works very similarly to the PG2 adapter. It consists of
multiple workers that forward to the local node using PubSub.local_broadcast.

The way to choose the worker to be used is based on the sending process
just like PG2 adapter does

The number of workers is controlled by `:pool_size` or `:broadcast_pool_size`.
This distinction exists because Phoenix.PubSub uses `:pool_size` to
define how many partitions the PubSub registry will use. It's possible
to control them separately by using `:broadcast_pool_size`

* fix: ensure message id doesn't raise on non-map payloads (supabase#1534)

* fix: match error on Connect (supabase#1536)



---------

Co-authored-by: Eduardo Gurgel Pinho <[email protected]>

* feat: websocket max heap size configuration (supabase#1538)

* fix: set max process heap size to 500MB instead of 8GB
* feat: set websocket transport max heap size

WEBSOCKET_MAX_HEAP_SIZE can be used to configure it

* fix: update gen_rpc to fix gen_rpc_dispatcher issues (supabase#1537)

Issues:

* Single gen_rpc_dispatcher that can be a bottleneck if the connecting takes some time
* Many calls can land on the dispatcher but the node might be gone already. If we don't validate the node it might keep trying to connect until it times out instead of quickly giving up due to not being an actively connected node.

* fix: improve ErlSysMon logging for processes (supabase#1540)

Include initial_call, ancestors, registered_name, message_queue_len and total_heap_size

Also bump long_schedule and long_gc

* fix: make pubsub adapter configurable (supabase#1539)

---------

Co-authored-by: Filipe Cabaço <[email protected]>
Co-authored-by: Eduardo Gurgel <[email protected]>
Co-authored-by: Bradley Haljendi <[email protected]>

* 🔄 Sync with upstream changes (#11)

* fix: runtime setup error (supabase#1520)

* fix: use primary instead of replica on rename_settings_field (supabase#1521)

* feat: upgrade cowboy & ranch (supabase#1523)

* fix: Fix GenRpc to not try to connect to nodes that are not alive (supabase#1525)

* fix: enable presence on track message (supabase#1527)

currently the user would need to have enabled from the beginning of the channel. this will enable users to enable presence later in the flow by sending a track message which will enable presence messages for them

* fix: set cowboy active_n=100 as cowboy 2.12.0 (supabase#1530)

cowboy 2.13.0 set the default active_n=1

* fix: provide error_code metadata on RealtimeChannel.Logging (supabase#1531)

* feat: disable UTF8 validation on websocket frames (supabase#1532)

Currently all text frames as handled only with JSON which already requires UTF-8

* fix: move DB setup to happen after Connect.init (supabase#1533)

This change reduces the impact of slow DB setup impacting other tenants
trying to connect at the same time that landed on the same partition

* fix: handle wal bloat (supabase#1528)

Verify that replication connection is able to reconnect when faced with WAL bloat issues

* feat: replay realtime.messages (supabase#1526)

A new index was created on inserted_at DESC, topic WHERE private IS TRUE AND extension = "broadast"

The hardcoded limit is 25 for now.

* feat: gen_rpc pub sub adapter (supabase#1529)

Add a PubSub adapter that uses gen_rpc to send messages to other nodes.

It uses :gen_rpc.abcast/3 instead of :erlang.send/2

The adapter works very similarly to the PG2 adapter. It consists of
multiple workers that forward to the local node using PubSub.local_broadcast.

The way to choose the worker to be used is based on the sending process
just like PG2 adapter does

The number of workers is controlled by `:pool_size` or `:broadcast_pool_size`.
This distinction exists because Phoenix.PubSub uses `:pool_size` to
define how many partitions the PubSub registry will use. It's possible
to control them separately by using `:broadcast_pool_size`

* fix: ensure message id doesn't raise on non-map payloads (supabase#1534)

* fix: match error on Connect (supabase#1536)



---------

Co-authored-by: Eduardo Gurgel Pinho <[email protected]>

* feat: websocket max heap size configuration (supabase#1538)

* fix: set max process heap size to 500MB instead of 8GB
* feat: set websocket transport max heap size

WEBSOCKET_MAX_HEAP_SIZE can be used to configure it

* fix: update gen_rpc to fix gen_rpc_dispatcher issues (supabase#1537)

Issues:

* Single gen_rpc_dispatcher that can be a bottleneck if the connecting takes some time
* Many calls can land on the dispatcher but the node might be gone already. If we don't validate the node it might keep trying to connect until it times out instead of quickly giving up due to not being an actively connected node.

* fix: improve ErlSysMon logging for processes (supabase#1540)

Include initial_call, ancestors, registered_name, message_queue_len and total_heap_size

Also bump long_schedule and long_gc

* fix: make pubsub adapter configurable (supabase#1539)

* fix: specify that only private channels are allowed when replaying (supabase#1543)

messages

* fix: rate limit connect module (supabase#1541)

On bad connection, we rate limit the Connect module so we prevent abuses and too much logging of errors

---------

Co-authored-by: Filipe Cabaço <[email protected]>
Co-authored-by: Eduardo Gurgel <[email protected]>
Co-authored-by: Bradley Haljendi <[email protected]>

* 🔄 Sync with upstream changes (#13)

* fix: runtime setup error (supabase#1520)

* fix: use primary instead of replica on rename_settings_field (supabase#1521)

* feat: upgrade cowboy & ranch (supabase#1523)

* fix: Fix GenRpc to not try to connect to nodes that are not alive (supabase#1525)

* fix: enable presence on track message (supabase#1527)

currently the user would need to have enabled from the beginning of the channel. this will enable users to enable presence later in the flow by sending a track message which will enable presence messages for them

* fix: set cowboy active_n=100 as cowboy 2.12.0 (supabase#1530)

cowboy 2.13.0 set the default active_n=1

* fix: provide error_code metadata on RealtimeChannel.Logging (supabase#1531)

* feat: disable UTF8 validation on websocket frames (supabase#1532)

Currently all text frames as handled only with JSON which already requires UTF-8

* fix: move DB setup to happen after Connect.init (supabase#1533)

This change reduces the impact of slow DB setup impacting other tenants
trying to connect at the same time that landed on the same partition

* fix: handle wal bloat (supabase#1528)

Verify that replication connection is able to reconnect when faced with WAL bloat issues

* feat: replay realtime.messages (supabase#1526)

A new index was created on inserted_at DESC, topic WHERE private IS TRUE AND extension = "broadast"

The hardcoded limit is 25 for now.

* feat: gen_rpc pub sub adapter (supabase#1529)

Add a PubSub adapter that uses gen_rpc to send messages to other nodes.

It uses :gen_rpc.abcast/3 instead of :erlang.send/2

The adapter works very similarly to the PG2 adapter. It consists of
multiple workers that forward to the local node using PubSub.local_broadcast.

The way to choose the worker to be used is based on the sending process
just like PG2 adapter does

The number of workers is controlled by `:pool_size` or `:broadcast_pool_size`.
This distinction exists because Phoenix.PubSub uses `:pool_size` to
define how many partitions the PubSub registry will use. It's possible
to control them separately by using `:broadcast_pool_size`

* fix: ensure message id doesn't raise on non-map payloads (supabase#1534)

* fix: match error on Connect (supabase#1536)



---------

Co-authored-by: Eduardo Gurgel Pinho <[email protected]>

* feat: websocket max heap size configuration (supabase#1538)

* fix: set max process heap size to 500MB instead of 8GB
* feat: set websocket transport max heap size

WEBSOCKET_MAX_HEAP_SIZE can be used to configure it

* fix: update gen_rpc to fix gen_rpc_dispatcher issues (supabase#1537)

Issues:

* Single gen_rpc_dispatcher that can be a bottleneck if the connecting takes some time
* Many calls can land on the dispatcher but the node might be gone already. If we don't validate the node it might keep trying to connect until it times out instead of quickly giving up due to not being an actively connected node.

* fix: improve ErlSysMon logging for processes (supabase#1540)

Include initial_call, ancestors, registered_name, message_queue_len and total_heap_size

Also bump long_schedule and long_gc

* fix: make pubsub adapter configurable (supabase#1539)

* fix: specify that only private channels are allowed when replaying (supabase#1543)

messages

* fix: rate limit connect module (supabase#1541)

On bad connection, we rate limit the Connect module so we prevent abuses and too much logging of errors

* build: automatically cancel old tests/build on new push (supabase#1545)

Currently, whenever you push any commit to your branch, the old builds are still running and a new build is started. Once a new commit is added, the old test results no longer matter and it's just a waste of CI resources. Also reduces confusion with multiple builds running in parallel for the same branch/possibly blocking any merges.

With this little change, we ensure that whenever a new commit is added, the previous build is immediately canceled/stopped and only the build (latest commit) runs.

* fix: move message queue data to off-heap for gen_rpc pub sub workers (supabase#1548)

* fix: rate limit Connect.lookup_or_start_connection on error only (supabase#1549)

* fix: increase connect error rate window to 30 seconds (supabase#1550)

* fix: set a lower fullsweep_after flag for GenRpcPubSub workers (supabase#1551)

* fix: hardcode presence limit (supabase#1552)

* fix: further decrease limit on presence events (supabase#1553)

* fix: bump up realtime (supabase#1554)

* fix: lower rate limit to 100 events per second (supabase#1556)

* fix: move connect rate limit to socket (supabase#1555)

* fix: reduce max_frame_size to 5MB
* fix: fullsweep_after=100 on gen rpc pub sub workers

---------

Co-authored-by: Eduardo Gurgel Pinho <[email protected]>

* fix: collect global metrics without tenant tagging (supabase#1557)

* feat: presence payload size (supabase#1559)

* Also tweak buckets to account all the way to 3000KB
* Start tagging the payload size metrics with message_type. message_type can be presence, broadcast or postgres_changes

* fix: use GenRpc for Realtime.Latency pings (supabase#1560)

* Fastlane for phoenix presence_diff (supabase#1558)

It uses a fork of Phoenix for time being

* fix: count presence_diff events on MessageDispatcher
* fix: remove traces from console during development

---------

Co-authored-by: Filipe Cabaço <[email protected]>
Co-authored-by: Eduardo Gurgel <[email protected]>
Co-authored-by: Kevin Grüneberg <[email protected]>
Co-authored-by: Bradley Haljendi <[email protected]>

---------

Co-authored-by: Al @h0lybyte <[email protected]>
Co-authored-by: Eduardo Gurgel <[email protected]>
Co-authored-by: Filipe Cabaço <[email protected]>
Co-authored-by: Bradley Haljendi <[email protected]>
Co-authored-by: Kevin Grüneberg <[email protected]>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants