Mika Naylor created FLINK-41519:
-----------------------------------
Summary: Performance regression in RocksDB MapState iterator and
isEmpty benchmarks since Sep 24
Key: FLINK-41519
URL: https://issues.apache.org/jira/browse/FLINK-41519
Project: Flink
Issue Type: Bug
Components: Runtime / State Backends
Affects Versions: 2.4.0
Reporter: Mika Naylor
Since around Sep 27 the daily flink-benchmarks runs show a ~4–6% regression in
RocksDB MapState benchmarks that open an iterator:
||Benchmark||Baseline (ops/ms)||Current (ops/ms)||Δ||
|mapEntries.ROCKSDB|294.5|280.9|-4.6%|
|mapIsEmpty.ROCKSDB|38.0|36.3|-4.6%|
|mapIsEmpty.ROCKSDB_CHANGELOG|38.3|36.0|-5.9%|
|mapIterator.ROCKSDB_CHANGELOG|295.0|283.2|-4.0%|
*Cause:* FLINK-40584 (ce32f1ef, 2d30c035).
{{RocksDBOperationUtils#getRocksIteratorBoundedByPrefix}} creates and frees a
private {{ReadOptions}} copy and an upper-bound {{Slice}} on every
{{{}isEmpty(){}}}, {{clear()}} and map iterator cache load. That is about 8
extra JNI calls and 2 native allocations per call.
I was able to replicate the regression by reverting and un-reverting the
commits in question.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)