Failures in RedisPubSubTest

[tsegismont]

Version

Vert.x 5 and Vert.x 4

Context

In CI, RedisPubSubTest has failures in different branches (the main branch, 4.x and 5.0):

https://github.com/vert-x3/vertx-redis-client/actions/runs/16385215052/job/46303737074#step:5:374
https://github.com/vert-x3/vertx-redis-client/actions/runs/16385215052/job/46303737074#step:5:374
https://github.com/vert-x3/vertx-redis-client/actions/runs/16385215052/job/46303737074#step:5:374
https://github.com/vert-x3/vertx-redis-client/actions/runs/16385215052/job/46303737074#step:5:374

[tsegismont]

@Ladicek does that ring a bell?

[Ladicek]

I don't recall seeing that, all tests pass locally and passed in CI last time I made a PR 🤔

[Ladicek]

This seems quite flaky. I cannot reproduce the issue in my fork.

[Ladicek]

OK, took me a while, but finally reproduced the issue in my fork. Investigation will likely take a while :-)

[Ladicek]

Added the following reproduction steps and debug logs Ladicek@3553565 and in the failing run, the NetSocket just doesn't deliver the expected message. It does deliver the subscription, which is weird...

[Ladicek]

The PUBLISH command returns that the message was delivered to 0 consumers, which is consistent with the behavior...

[Ladicek]

Found the issue and have the fix. The problem is that when the SUBSCRIBE command returns, there's no guarantee that the subscription has actually completed. That is only guaranteed after the subscription connection receives the subscribe message. The fix is to simply wait until the subscribe message is received by the subscription connection and only then run PUBLISH on the publish connection.