Arvados

I recently ran a toy websocket client against pirca (running 2.6.3), just subscribed to everything and dumping the records to JSON. It seemed to work, so websockets aren't completely broken.

Brett Smith wrote in #note-6:

I recently ran a toy websocket client against pirca (running 2.6.3), just subscribed to everything and dumping the records to JSON. It seemed to work, so websockets aren't completely broken.

Tom and I chatted about this a couple weeks ago, I suspect there's a bug in Workbench 2 where it sometimes it sends an empty subscription and thus doesn't receive any events. Nobody has gone in to track it down yet.

Unfortunately, I don't yet know how to reproduce this.

See graphs on #20930

Our listener pings create the conditions for a deadlock issue.

Specifically: our event loop can deadlock if enough (~32) server notifications arrive after the event loop decides to call Ping (e.g., while listener.Ping() is waiting for a response from the server, or in the time.Sleep() invoked by testSlowPing).

(*ListenerConn)listenerConnLoop() doesn't see the server's ping response until it finishes sending a previous notification through its internal queue to (*Listener)listenerConnLoop(), which is blocked on sending to our Notify channel, which is blocked on waiting for the Ping response.

(The lib/pq example uses "go listener.Ping()" so it doesn't deadlock.)