Rebuild state serially about omid

[fpj]

Isn't it sufficient to make sure that we execute operations in the order of ledger entry id? If operations are logged in order, then replaying this way should be sufficient. We can still read in parallel to speed up the read, but we need to make sure that we follow the order of entries.

Building the sequence and replaying might require a pretty large buffer in some cases, no? It might not be such a good idea to do it.

[maysamyabandeh]

I agree with you Daniel. The required memory for this purpose is in order of 10% of the physical memory and it is returned to the system right after recovery. Moreover that this approach greatly simplifies the recovery procedure, as it emulates a natural in-order processing of a sequence of events.

[fpj]

I still don't understand the need of buffering before replaying. If we follow the order of entry ids, then I don't see why it should be incorrect.

I also don't understand where you're pulling 10% of memory from, Maysam.

[dgomezferro]

I think we all agree that we should just replay in the order of entry ids, but for that we have to buffer, no?

1. We want to start reading the log from the tail to avoid reading the whole log and stop early.
2. We want to replay the log in the order of entry ids (from the start to the tail).

So we cannot execute anything until we reach the stop condition, and then we execute all operations in order, so we should either buffer the read entries or reread them from BK.

[fpj]

Correct, I keep forgetting that we are reading backwards.

[francisco-perez-sorrosal]

Current version is not based on BK