8d3885d357
This replaces the manual threadpool management in ReceiveCommits with a more general approach. That support was added for slow NoSQL backends, which may also have poor performance characteristics when updating lots of NoteDb refs in parallel. This approach substantially rewrites the BatchUpdate loop to stage all NoteDb update operations in memory in separate threads, each with its own copy of the ReviewDb and change repo, and then aggregates the results together in the main thread with a single ObjectInserter and BatchRefUpdate. One nice thing about this approach is it doesn't require any changes to BatchUpdate.Op implementations: each Op's methods are executed sequentially, in one thread at a time, with proper barriers (an executor) to ensure later methods see the results of writes in the background thread. Use an in-memory implementation of ObjectInserter in each of the background threads' NoteDbUpdateManagers to buffer writes completely in memory until it's time for the caller to flush. This wastes a small amount of memory for the buffer, but these are just NoteDb objects, so they should be quite small. This implementation using immutable result types is preferable to trying to share a Repository/ObjectReader/ ObjectInserter across threads, which requires manual locking. That is not only painful but also produces deadlocks when mixing repo-level locks and SQL-level locks (e.g. H2's transaction implementation) across threads. Change-Id: I40545a4d48fcfa892bd3e4c0cd9b72ab7fac9436 |
||
|---|---|---|
| .. | ||
| src | ||
| BUCK | ||