CASSGO-41 Deadlock in refreshDebouncer when reconnection fails

[kevinkyyro]

What version of Cassandra are you using?

astra-classic

What version of Gocql are you using?

v1.6.0

What version of Go are you using?

1.21

What did you do?

Connection errors, I think due to overload, lead to frequent reconnection attempts and failures

What did you expect to see?

Should retry until connection succeeds

What did you see instead?

Deadlock

498297 goroutine 1324045437 [chan send, 113 minutes]:
498298 github.com/gocql/gocql.(*refreshDebouncer).stop(0xc0b826a7c0)
498299         /go/pkg/mod/github.com/gocql/gocql@v1.6.0/host_source.go:848 +0x8c
498300 github.com/gocql/gocql.(*Session).Close(0xc03efb0c00)
498301         /go/pkg/mod/github.com/gocql/gocql@v1.6.0/session.go:494 +0x105
498302 github.com/gocql/gocql.NewSession({{0xc24be58930, 0x3, 0x3}, {0x2ef55cf, 0x5}, 0x4, 0x12a05f200, 0x12a05f200, 0x0, 0x755a, ...})
498303         /go/pkg/mod/github.com/gocql/gocql@v1.6.0/session.go:180 +0x98d
498304 github.com/gocql/gocql.(*ClusterConfig).CreateSession(...)
498305         /go/pkg/mod/github.com/gocql/gocql@v1.6.0/cluster.go:289

It looks like a race condition between (*refreshDebouncer).stop() and (*refreshDebouncer).flusher()

1. stop() acquires d.mu and sets d.stopped to true
2. flusher() exits the select at the top of the loop and blocks on acquiring d.mu
3. stop() releases d.mu and tries to write to d.quit
4. flusher() acquires d.mu and returns because d.stopped is true
5. stop() is deadlocked because d.quit is unbuffered and the reader has stopped

[joao-r-reis]

I can work on a fix for this but it's likely that it will only get merged when this driver is donated to the ASF (see #1749 (comment) and #1751 )

[joao-r-reis]

By the way @kevinkyyro , this won't really help with gocql not reconnecting, it will just stop the panic from happening. I believe your issue with the reconnection is related to gocql-astra adding a single contact point which causes gocql to fail after 1 connection attempt (see: datastax/gocql-astra#24 )

In any case I've opened a PR to fix this deadlock issue: #1767