How does watch API scale w/ PG? If there are 10000 RPCs to the watch endpoint simultaneously streaming events, are there 10000 queries running periodically?

Right now each watch opens its own connection on the backend, so you wouldn't want to open that many. The recommendation is to open a watch for all possible consumers and then distribute to individual clients on the application side. But we've definitely floated ideas to make this simpler / easier, especially with Spanner where there are a limited number of watch slots. Might be worth writing up a GH issue to discuss some future options to simplify this.

there's also been discussion of providing easier mechanisms to move events from the watch API to something like kafka, such as through a connector

Thank you both! Is there a GH issue around a "connector"? We may be interested in contributing to something like that

not yet - my understanding is that there's been a couple of starts on this idea (one using Bento and one that i started at my previous company but haven't pushed up yet) but nothing written down yet

@alechenninger we'd also welcome collab on it too

@alechenninger I added a tracking issue here: https://github.com/authzed/spicedb/issues/2015 and we'd welcome your input on the topic