One day, alertmanager could not send alerts in my production. So i made a dump file. The dump file shows that Dispather blocks on cleaning up empty aggrGroups while aggrGroup blocks on ag.done in aggrGroup.stop(). After analysing the code, aggrGroup.stop() should terminate the aggrGroup.run() loop, and then aggrGroup.stop() exits normally. But in fact, it does not work as we hope.

I think that there is a memory visibility error. aggrGroup.run() runs as a goroutine which initializes ag.done, and aggrGroup.stop() runs in another goroutine which reads ag.done channel itself and waiting value from this channel. This is a data race! If those two goroutine runs on different cpu cores, aggrGroup.stop() may be read from a nil channel even though aggrGroup.run() has initialized the channel. So i fix this error by initializing ag.done in newAggrGroup().

There is may be another error. If aggrGroup.stop() runs before aggrGroup.run(), aggrGroup.stop() may be also waiting from a nil channel because ag.done not be initialized in aggrGroup.run(). Of course, it seems that aggrGroup.stop() won't run before aggrGroup.run(). But i don't think that initializing shared variables in one goroutine is a good practice.