Design a queue-less replicator

At the moment we Enqueue and Dequeue events in a SQL database table. Each event tracks a replication action.

We've seen this design isn't great for several reasons:

The metadata of each event might get outdated or incorrect when an event gets dequeued
It doesn't scale well, especially because duplicate events might get created #3940 (closed)
and more...

Goal

The database already has all information to make a decision on which repos need replication. There are still a few challenges to make the data queuryable and indexable. This issue aims to define a design that allows us to do that. There has been discussion on-going on what a possible solution could be, see #3940 (comment 871401396).

The outcome of this issue wouldn't be any code change, only a theoretical design, maybe in the form of an RFC.

To upload designs, you'll need to enable LFS and have an admin enable hashed storage. More information