Fix connection leak of session with dropped keyspace

[ikehara]

This PR fixes the connection leak problem of reconnection failure.

When a keyspace is dropped and the origin server is restarted, cql-proxy attempts to reconnect. However, the new connection fails to set the keyspace, leading to a memory leak each time cql-proxy tries to connect to the server.

2024-06-29T17:23:48.232+0900	INFO	proxycore/connpool.go:193	pool connection attempting to reconnect after delay	{"host": "127.0.0.1:9042", "delay": "10s"}

The root cause of this issue is that the connection cleanup process is not functioning as expected.
Specifically, the conn variable is always nil because the statement return nil, err overwrites the conn variable.

[absurdfarce]

Thanks for the report @ikehara!

After looking at this a bit I'm not sure there's a scenario which winds up leaking memory here. It's absolutely true that reconnect attempts in the scenario you describe (a dropped keyspace + origin server restart) will continue to try to reconnect via this operation. This will always fail (as the set keyspace op here can't possibly succeed against a dropped keyspace). But this shouldn't result in a memory leak; note that the output of the call to connect() is stored in a temp variable "c" and that this value is only copied to "conn" if "err" is nil. "err" is very definitely not nil in this case so nothing should be overwritten. "c" should get garbage collected normally when the block in question goes out of scope.

All of that said, there is a potentially interesting problem here. In the case under consideration the connection pool will fail to recreate new connections (for reasons discussed above) and will essentially fail to re-connect forever. This is due to the fact that the keyspace to set appears to be a client-level config and clients are currently created when a client connects to the proxy. We very likely need to tweak that logic a bit to handle the case you describe @ikehara.

[ikehara]

Thank you for the quick response.
Sorry, my previous message was a bit confusing. I meant a resource leak.
I'm afraid a connection (file descriptor of net.conn struct) is not closed by the GC, so a resource leak might occur if the Close method is not called. Even if it were closed by its finalizer, the number of TCP connections is limited and might run out before garbage collection happens, depending on the amount of memory.

[mpenick]

@absurdfarce It's because the existing logic returns the error instead of setting err which causes the defer not to close the connection.

[mpenick]

Nice find @ikehara