Matthew Seal

@davidbrochart @kevin-bates FYI

I think we could support a client queue and feed queries down to the kernel 1 by 1 until an error came up to abort. Or alternatively the servers should...

> in a high-latency environment you could end up adding multiple extra seconds to simple executions. Not to derail the conversation from where / what might need to change --...

Ah ok yeah I agree. Kernels patterns weren't designed for sub 50ms requests in rapid succession and usually the application expectations above them carry that same assumption.

> Do we agree that accessing in some way the current file path/name/url/uuid is something we want to allow ? I think it's asked for often enough that having a...

Thanks for putting together such a detailed issue. There's a few different executions paths across repos that use the data directory listing to fetch files they care about. It would...

Hmm thanks for digging into it some. Ironically, if the APPDATA env variable is not set it would actually use `~/.jupyter` -- see [jupyter_core](https://github.com/jupyter/jupyter_core/blob/ed50a593d03684d93c0234372cd2947e500ef7b4/jupyter_core/paths.py#L87-L92).

Thanks for pulling that topic out. I think the most important thing is to make a test plan to try it out. Some of the downstream libraries have parallelism tests...

> Just creation of sockets and sending a message? And teardown, yeah. I've used parallel papermill or nbclient processes over a simple notebook as a way to test in the...

That's a good idea to add and should be an easy change. If you have interest in opening a PR for it I can review quickly, otherwise I'll tag it...