Sync response for device_lists is not immediate when running with workers

[hyossing]

Description

Hello,

Synapse does not create sync response for device_lists when logout in master and worker mode.

In a single synapse mode, sync response comes for 'device_lists' when logout.

Steps to reproduce

1. Set Master and Worker environment
   attached zip file is my test environment.
   synapse-worker-deployment.zip

2. login two devices with same id using element client.

• http://localhost:8008

3. logout from one of the two devices.

• check sync response from the non-logout device.

Log

Log For Master & Worker

• Master
  log_master.txt

2022-01-12 01:53:24,316 - synapse.replication.tcp.resource - 172 - DEBUG - replication_notifier-46 - Getting stream: device_lists: 19 -> 20

Master sends replication data to worker.

• Worker
  log_worker.txt

2022-01-12 01:53:24,326 - synapse.replication.tcp.handler - 510 - DEBUG - process-replication-data-86 - Received rdata device_lists (master) -> 20

Worker receives replication data from master. But, nothing happened.

Log For Single Process

• Single Process
  log_single_process.txt
  For the single process case, sync response is created for device list right after logout.

2022-01-12 02:08:31,602 - synapse.util.metrics - 152 - DEBUG - GET-55 - Entering block _generate_sync_entry_for_device_list


### Version information
- **Version**: 1.49.2
- **Install method**: docker-container
- **Platform**: mac os

[DMRobertson]

Hi @hyossing, can you clarify what you mean by "logout"? I'm guessing you've got element-web opened in two tabs in the same browser, and you close one tab while leaving the other open?

[hyossing]

Hello, @DMRobertson

1. Logout means, click logout menu in element-web client. So, synapse receives POST ..../logout request.

• If general(single process) case, synapse create sync response for device_lists to notify logout(changed) device.

2. No two tabs.

• It means, two devices. Login to synapse from different browser or physical machine.

And,

I'm sorry about the synapse-worker-deployment.zip.

Some configuration is wrong. It would not work.

• haproxy.cfg

I updated some configuration.
synapse-worker-deployment_20220113.zip

Thanks you.

[DMRobertson]

Thanks @hyossing. One quick point of clarification, to make sure I understand the problem:

For the single process case, sync response is created for device list right after logout.

Synapse does not create sync response for device_lists when logout in master and worker mode.

In master-worker mode, do you see an entry in device_lists eventually, e.g. in a later incremental sync? Or does it never appear at all?

[hyossing]

Dear @DMRobertson

It appears eventually in a later incremental sync as you said.

I found one more strange behavior.

The sync response comes right after logout if there is a room.

But, If there is no room(leave and forget the test room), then the sync response comes later.

Thanks you.

[DMRobertson]

It appears eventually in a later incremental sync as you said.

This is to be expected when we're running in worker mode. (There are delays as information propagates from one worker to another). We

Is this causing a particular problem for your use case?

[hyossing]

It is a problem for me.

There is a particular behavior of our own client when device list changes.

We expected the client will receive the same event right after logout in worker mode.

• Before that, we only run single mode.

Currently, we are considering to change our policy for the changed-device-list.

• Allow delayed device-list change information.

[DMRobertson]

Sorry for the delay in getting back to you. On reflection, I think I may have understood your problem. It sound a bit like #11457: something happens that clients should be told about, but we don't send a /sync response immediately, and have to wait for another event to actually send down the information.

Is the problem reproducible every time, or does it only happen sometimes?

[hyossing]

@DMRobertson
Sorry for the delay to reply.

The problem reproducible every time for the no room and worker case as I explained above.

In my opinion, the particular behavior is a trivial issue. We can handle by changing our policy a little bit.

BR
// Hyosung Lim

[richvdh]

It sounds like this is expected behaviour and the reporter has found a workaround. Closing.