Hi Colin, > I agree controlled shutdown is tricky to get just right. I suppose this is a > case where the RPCs we send out are not purely "fire and forget"; we have to > listen for the response. But that can be done in an event-based way. > Controlled shutdown is also probably the last thing we'll implement once we > have the basic lift and shift.
LeaderAndIsr requests are also not "fire and forget" either as the broker response can include per-partition errors which trigger ISR and leadership changes. > Topic deletion is another annoying thing. I wonder if we could just delete > immediately here and skip the complexity of implementing "deleting state." > Topic IDs will exist in IBP 3.4, in both ZK and KRaft mode, so it should be > OK, right? I don’t think this is OK because we still use topic names (not topic IDs) to persist records under log directories. If one of the brokers is unavailable when a topic is deleted and re-created with the same name that broker might come back later with conflicting state/records on that topic. Best, -- Igor
