anshulbaliga7 opened a new pull request, #13610:
URL: https://github.com/apache/gluten/pull/13610

   ## What changes are proposed in this pull request?
   
   The executor-side build cache for broadcast hash joins has been the default 
since it landed, and there is no remaining case where building the hash table 
per task is wanted, except one: when a whole stage dump is requested for the 
micro benchmark, the build side has to reach the native join through its input 
iterator so the dumped stage can be replayed (GLUTEN-12504).
   
   This removes `spark.gluten.velox.buildHashTableOncePerExecutor.enabled`. The 
`enableBroadcastBuildOncePerExecutor` accessor stays and now means "no dump is 
requested", so the pre-projection path and the canonicalization rule that 
depend on it are unchanged in behavior. The one-time warning that the option 
was being ignored during dumping goes too, with the companion object that held 
its flag, since there is no option left to ignore.
   
   Tests that looped over both values of the option now exercise the always-on 
path only, and the config suite no longer asserts the option can be disabled. 
`VeloxBroadcastBuildOnceBenchmark` compared the two modes and is removed. The
   micro benchmark doc no longer refers to the option.
   
   One suite to watch: `DynamicOffHeapSizingSuite` had the option switched off 
with a note that the native hash table build needs off-heap memory, and it runs 
with static off-heap disabled under dynamic sizing. With the option gone it 
builds once per executor like everything else. If it turns out to need the 
per-task build, that is a real case for keeping the option and worth deciding 
here.
   
   Fixes #13253
   
   ## How was this patch tested?
   
   Existing tests, adjusted as described. Built locally with `./build/mvn 
-Pbackends-velox -Pspark-3.5 -pl backends-velox -am -DskipTests install` and 
`test-compile` of `gluten-ut/spark35` against it.
   
   ## Was this patch authored or co-authored using generative AI tooling?
   
   Generated-by: Claude claude-opus-4-8


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to