Parallelize background tree service sync #166

New issue

Closed

opened 2023-03-23 12:44:15 +00:00 by fyrchik · 1 comment

fyrchik commented

2023-03-23 12:44:15 +00:00

Owner

Currently we iterate over all container nodes starting from the same height.
Let's say i-th node returned log up to h_i height.
Then we take min(h_i) and store it in the database as the next point.

There are 2 opportunities for synchronization here:

Process container nodes in parallel. This should allow us merging operations streams from GetOpLog() to process each height exactly once.
Perform TreeApply() in parallel. This is a bit more difficult, but it allows to speed up synchronization by a lot. So basically we will have 3 components:

Workers which receive GetOpLog operations.
Worker which merges multiple stream into one stream of strictly increasing log operations
Workers which apply operations to the local storage.

I suggest starting with (1) and do (2) separately.

Currently we iterate over all container nodes starting from the same height. Let's say i-th node returned log up to h_i height. Then we take min(h_i) and store it in the database as the next point. There are 2 opportunities for synchronization here: 1. Process container nodes in parallel. This should allow us merging operations streams from GetOpLog() to process each height exactly once. 2. Perform TreeApply() in parallel. This is a bit more difficult, but it allows to speed up synchronization by a lot. So basically we will have 3 components: - Workers which receive `GetOpLog` operations. - Worker which merges multiple stream into one stream of strictly increasing log operations - Workers which apply operations to the local storage. I suggest starting with (1) and do (2) separately.