Huang Kuan Hao created HDDS-16281:
-------------------------------------
Summary: Reduce per-key overhead in Recon reprocess tasks
Key: HDDS-16281
URL: https://issues.apache.org/jira/browse/HDDS-16281
Project: Apache Ozone
Issue Type: Improvement
Reporter: Huang Kuan Hao
On the Recon reprocess (full rebuild) loop over the whole keyspace, the OBS
task does a raw RocksDB point-read per key to fetch the bucket, even though
consecutive keys share a bucket (the incremental process() path caches via
lookupBucketCached):
String bucketDBKey = omMetadataManager.getBucketKey(volumeName, bucketName);
OmBucketInfo omBucketInfo =
omMetadataManager.getBucketTable().getSkipCache(bucketDBKey);
and each worker re-fetches its own accumulation map per key by boxing the
thread id:
Map<..> localMap = allLocalMaps.computeIfAbsent(
Thread.currentThread().getId(), k -> new ConcurrentHashMap<>());
Fix: (a) add a per-worker bucket cache (reprocess is multi-threaded, so
per-worker / thread-safe); (b) resolve the worker map once via a ThreadLocal,
keeping a concurrent registry of worker maps for the final merge
(allLocalMaps.values() is iterated to flush). Behavior unchanged.
Files: NSSummaryTaskWithOBS.java:122,315; ContainerKeyMapperHelper.java:149;
FileSizeCountTaskHelper.java:167; NSSummaryTaskWithFSO.java:282
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
