slachiewicz opened a new pull request, #3811:
URL: https://github.com/apache/thrift/pull/3811
Fixes THRIFT-5814.
`TestNoHangDuringStopFromClientNoDataSendDuringAcceptLoop` slept 10 ms after
starting the server and again after dialing it, then asserted that `Stop` took
at least the 50 ms stop timeout. `Stop` waits on the wait group `TSimpleServer`
fills in `innerAccept`, so whenever the accept and the goroutine start did not
fit inside those naps, the group was still empty, `Stop` returned in
microseconds, and the assertion failed. That is the roughly 1-in-100 flake in
the report.
The processor now closes a channel when it is entered, and the test waits
for that before calling `Stop`. `simple_server.go` calls `p.wg.Add(2)` in
`innerAccept` before starting the goroutine that reaches the processor, so
entering the processor is a sound barrier for what `Stop` waits on. Nothing
unblocks the read in the window between the signal and `ReadMessageBegin`:
`InterruptFunc` closes the listener, not the accepted connection — which is why
that connection now lives until cleanup rather than going out of scope.
### Reproduction
The flake is reachable on demand under a single P, which is how this was
verified rather than by waiting on chance:
```
$ GOMAXPROCS=1 go test -run
TestNoHangDuringStopFromClientNoDataSendDuringAcceptLoop -count=500 .
--- FAIL: TestNoHangDuringStopFromClientNoDataSendDuringAcceptLoop (0.03s)
simple_server_test.go:221: stop cost less time than server stop timeout,
server stop timeout:50ms,cost time:136.625µs
```
With this change the same command passes 500 of 500. The test also passes
200 runs under `-race` at `GOMAXPROCS=1`, and the full package passes under
`-race`.
Earlier attempts on the ticket changed the transport — unix domain socket,
extra client-side handshakes — which left the timing assumption in place. This
removes it.
*This change was created with AI assistance.*
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
