feat: do not fallback to Spark for COUNT(distinct)

[comphead]

Which issue does this PR close?

Just experiment, related to #2292
Closes #.

Rationale for this change

What changes are included in this PR?

How are these changes tested?

[comphead]

Might be fixed by #2407
@andygrove WDYT do you have anything in your mind to double test the count(distinct) ?

Running CometFuzzAggregateSuite

[mbutrovich]

I think it was fixed by several issues. I think #2407 is one, but I think the native shuffle rewrite and bumping the Arrow Java version contributed as well. It's encouraging to see!

[comphead]

I'm adding more tests to make sure it is working now

[andygrove]

Don't we need to pass the aggExpr.isDistinct value into the protobuf plan?

[comphead]

this is a good point, I was thinking the same but IMO Spark doesn't call the count distinct on partial phase.

                +-----------------------------+
                |         Driver              |
                |   COUNT(DISTINCT name)      |
                +-------------+---------------+
                              |
                              v
     +-------------------+          +-------------------+
     |   Executor 1      |          |   Executor 2      |
     | Partitions P0,P1  |          | Partitions P2,P3  |
     | Local distinct:   |          | Local distinct:   |
     | {Alice,Bob,Eve}   |          | {Mallory,Eve,Bob,Trent}|
     +---------+---------+          +---------+---------+
               |                              |
               |          Shuffle             |
               v                              v
        +--------------+               +--------------+
        | Reducer R0   |               | Reducer R1   |
        | {Alice,Bob,Eve}              | {Mallory,Trent} |
        +--------------+               +--------------+
               \                              /
                \                            /
                 \                          /
                  +-----------+-------------+
                              v
                    Driver Final Merge
                      DISTINCT = 5

The local distinct is made by HashAggregate so when count distinct get called as aggExpr it might not be needing the flag as data already deduped on reducers. Checking the Final stage though

[andygrove]

Thanks for explaining that. We should add tests for other distinct aggregates as well, such as sum and avg. I'm not sure if there are others?

[andygrove]

Spark has tests for the following aggregates with DISTINCT:

• count
• sum
• avg
• first
• last
• corr
• var_pop
• var_samp

[andygrove]

For this PR, we could just remove the fallback for COUNT?

[comphead]

yep, will do once

[codecov-commenter]

Codecov Report

❌ Patch coverage is 75.00000% with 1 line in your changes missing coverage. Please review.
✅ Project coverage is 58.48%. Comparing base (f09f8af) to head (9d3a40f).
⚠️ Report is 845 commits behind head on main.

Files with missing lines	Patch %	Lines
.../scala/org/apache/comet/serde/QueryPlanSerde.scala	75.00%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@             Coverage Diff              @@
##               main    #2429      +/-   ##
============================================
+ Coverage     56.12%   58.48%   +2.35%     
- Complexity      976     1440     +464     
============================================
  Files           119      146      +27     
  Lines         11743    13519    +1776     
  Branches       2251     2352     +101     
============================================
+ Hits           6591     7906    +1315     
- Misses         4012     4379     +367     
- Partials       1140     1234      +94     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[comphead]

depends on #2258

[andygrove]

could you also add tests for count distinct with multiple columns e.g. COUNT(DISTINCT col1, col2, col3)

[comphead]

Thanks @andygrove I'll add them separately as well

[comphead]

Fuzz tests fall to Spark

- count distinct (native_comet, native shuffle) *** FAILED *** (418 milliseconds)
  Expected only Comet native operators, but found HashAggregate.
  plan: HashAggregate(keys=[], functions=[count(distinct c0#92074)], output=[count(DISTINCT c0)#92169L])
  +- Exchange SinglePartition, ENSURE_REQUIREMENTS, [plan_id=716408]
     +- HashAggregate(keys=[], functions=[partial_count(distinct c0#92074)], output=[count#92173L])
        +- CometHashAggregate [c0#92074], [c0#92074]
           +- CometExchange hashpartitioning(c0#92074, 10), ENSURE_REQUIREMENTS, CometNativeShuffle, [plan_id=716405]
              +- CometHashAggregate [c0#92074], [c0#92074]
                 +- CometScan [native_comet] parquet [c0#92074] Batched: true, DataFilters: [], Format: CometParquet, Location: InMemoryFileIndex(1 paths)

checking

[mbutrovich]

Can we add a case to CometAggregateBenchmark to tout our gains? :)

[comphead]

Can we add a case to CometAggregateBenchmark to tout our gains? :)

Added

[mbutrovich]

LGTM. I guess Comet already knew how to serde everything and it was just a matter of removing the fallback and beefing up tests. Thanks @comphead!

I also ran the benchmark locally: ~1.4x on most cases! 🚀

CometAggregateBenchmark.txt

[andygrove]

I found that we fall back to Spark when there are multiple count distinct e.g.

  test("mulitple count distinct with group column") {
    val df = spark.read.parquet(filename)
    df.createOrReplaceTempView("t1")
    val sql = s"SELECT c1, count(distinct c2), count(distinct c3) FROM t1 group by c1"
    val (_, cometPlan) = checkSparkAnswer(sql)
    if (usingDataSourceExec) {
      assert(1 == collectNativeScans(cometPlan).length)
    }
  }

This isn't an issue for this PR, but thought I should make a note of this.

Comet cannot accelerate HashAggregateExec because: Aggregate expression with filter is not supported

[andygrove]

LGTM. Thanks @comphead!

[comphead]

I found that we fall back to Spark when there are multiple count distinct e.g.

  test("mulitple count distinct with group column") {
    val df = spark.read.parquet(filename)
    df.createOrReplaceTempView("t1")
    val sql = s"SELECT c1, count(distinct c2), count(distinct c3) FROM t1 group by c1"
    val (_, cometPlan) = checkSparkAnswer(sql)
    if (usingDataSourceExec) {
      assert(1 == collectNativeScans(cometPlan).length)
    }
  }

This isn't an issue for this PR, but thought I should make a note of this.

Comet cannot accelerate HashAggregateExec because: Aggregate expression with filter is not supported

Filed #2456