Frank McSherry

I saw it! The intent looks good; I'll try and take a read through later tonight or tomorrow, if that is ok.

Afaiu, this is "solved" with option (4.). We might want to improve the ability of multi-process replicas to reconcile efficiently, but the issue is no longer that they will be...

One sorta-resolution of that issue could be that in the dataflow implementation of postgres sources, it can hotwire in a bespoke use of timely's `partition` operator. That might be easier...

This seems like it could/should be fixable with a `Uuid` (or maybe `SeqNo`) from the controller as part of replica initialization. Each instance should only form a mesh with other...

It would be great to hoist it up into the API (e.g. return a `Result`) as I think there is potentially a recovery option: advance the dataflow locally and produce...

I think if you put a sleep in to replica hydration just after it determines what commands to send (and so picks a frontier), you could probably repro the issue...

I have something that might reduce the noise, but there would still be a "bug" in that we don't know the full space of reasons this triggers. In particular, I...

Adding a few specific use cases for RTR, to add context: 1. Many users want to know "how up to date are my views of my sources". This has led...

Imo we should: 1. Make sure we can measure interactive query times. We have introspection views for peek latencies, and tbh these haven't usually been very high. They are often...

As discussed in a meeting, one thing we might do is introduce "source frontiers" like we have "materialization frontiers", reporting for each source the frontier through which we have received...