Konstantin Orlov created IGNITE-23682:
-----------------------------------------
Summary: Sql. Avoid unnecessary task scheduling for simple plan
Key: IGNITE-23682
URL: https://issues.apache.org/jira/browse/IGNITE-23682
Project: Ignite
Issue Type: Improvement
Components: sql
Reporter: Konstantin Orlov
KvPlan unconditionally schedules completion of prefetch callback to another
thread pool:
{code:java}
// KeyValueGetPlan.java
public <RowT> AsyncCursor<InternalSqlRow> execute(...) {
...
if (firstPageReadyCallback != null) { // firstPageReadyCallback is never
null
Executor executor = task -> ctx.execute(task::run,
firstPageReadyCallback::onPrefetchComplete);
result.whenCompleteAsync((res, err) ->
firstPageReadyCallback.onPrefetchComplete(err), executor);
}
...
{code}
The problem is that for single statement read operations this callback is not
used, yet scheduling results in moderate performance penalty. Below are results
acquired on MBP M3 simply replacing {{result.whenCompleteAsync}} -->
{{{}result.whenComplete{}}}:
{code:java}
Before
Benchmark (clusterSize) (fsync) Mode Cnt Score Error Units
SelectBenchmark.sqlGet 1 false avgt 20 9.056 ± 0.407 us/op
After
Benchmark (clusterSize) (fsync) Mode Cnt Score Error Units
SelectBenchmark.sqlGet 1 false avgt 20 6.236 ± 0.448 us/op
{code}
Let's revise usage of prefetchCallback and get rid of unnecessary scheduling,
when possible.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
