When we need to do a full replay, we currently do a synchronous clone of the source operator's state. This holds up processing other writes and backfills through that operator, which is sad. If we could find some kind of data-structure that allowed us to efficiently capture a snapshot of it which we could then use for the backfill, that'd be awesome!
When we need to do a full replay, we currently do a synchronous clone of the source operator's state. This holds up processing other writes and backfills through that operator, which is sad. If we could find some kind of data-structure that allowed us to efficiently capture a snapshot of it which we could then use for the backfill, that'd be awesome!