Peter Toth created SPARK-62098:
----------------------------------

             Summary: Apply per-partition runtime bloom filters in the scan of 
the application side
                 Key: SPARK-62098
                 URL: https://issues.apache.org/jira/browse/SPARK-62098
             Project: Spark
          Issue Type: Sub-task
          Components: SQL
    Affects Versions: 5.0.0
            Reporter: Peter Toth


SPARK-60541 checks each partition of the application side against the bloom 
filter of the matching partition of the creation side, in a semi join. Under a 
storage-partitioned join no shuffle lies between the scan and that join, so the 
semi join saves little.

This puts the per-partition bloom filters into the scan of the application side 
instead, as a storage filter (SPARK-60465):
* Each split of the scan gets the bloom filter of the creation side's partition 
with the split's partition key.
* A split whose partition key has no bloom filter is not read.
* The scan keeps the partitioning it reports, so it adds no shuffle and cuts no 
tasks.

The semi join goes, and so do:
* the planning that judged whether it was free 
({{ChoosePerPartitionRuntimeFilters}});
* {{might_contain}} reading its bloom filter from a column;
* a hash existence join reusing its bloom filter build column.

A scan that takes no storage filter gets no per-partition bloom filter.




--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

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

Reply via email to