[EPIC] Improved performance in H2O.ai benchmarks

[alamb]

Is your feature request related to a problem or challenge?

The basic aggregate functions like COUNT and SUM in DataFusion are very fast (see Apache DataFusion is now the fastest single node engine for querying Apache Parquet files)

However, many of the other aggregate functions are not particularly fast, and this shows up specifically on some of the H20 benchmarks

We saw this in the results in the 2024 DataFusion SIGMOD paper

(BTW we have made median faster)

@MrPowers has also observed similar results on discord (link):

DataFusion was added to the h2o benchmarks (which are now maintained by duckdb) and DataFusion performs quite well for most of the "basic" groupby queries. It performs poorly for some of the advanced questions on the 50GB dataset. Here are the results:
https://duckdblabs.github.io/db-benchmark/

See his version of the benchmarks here
https://github.com/MrPowers/mrpowers-benchmarks

Testing

• Add H2O.ai Database-like Ops benchmark to dfbench #7209

Functions

• Improve performance of corr function #13549
• Improve performance of median function #13550

Other improvements

• Improve join performance for h2o queries #13765

Describe the solution you'd like

DataFusion has two APIs ways to implement Aggregate functions like SUM and COUNT

• Easy (but slow) way: Accumulator (api docs)
• Fast (but complicated way): GroupsAccumulator (api docs)

The basic aggregates are implemented using GroupsAccumulator and are part of DataFusions performance

This ticket tracks the effort to improve the performance of these for these "more advanced" aggregate functions, likely by implementing GroupsAccumulator

Describe alternatives you've considered

For each function listed above, ideally we would:

1. Add a new benchmark. Either add a specific one for H20 benchmarks or add a query to the ClickBench extended benchmark Documentation Here in one PR
2. Implement GroupsAccumulator for the relevant aggregate function in a second PR (along with tests for correctness). We would use the benchmark to verify the performance

Here is a pretty good example of how @eejbyfeldt did this for STDDEV:

• Implement groups accumulator for stddev and variance #12095

Additional context

No response

[alamb]

For posterity, here is a link to the discord chat: https://discord.com/channels/885562378132000778/1309883046886903870/1309887744595595324

[MrPowers]

Would like to note that the DataFusion performance really starts to lag when the dataset size grows.

Take a look at this query: select id2, id4, power(corr(v1, v2), 2) as r2 from x group by id2, id4.

When the dataset is 10 million rows, then Polars takes 3 seconds and DataFusion takes 3.6 seconds, so pretty similar.

When the dataset is 100 million rows, then Polars takes 126 seconds and DataFusion takes 2,100 seconds.

[alamb]

When the dataset is 100 million rows, then Polars takes 126 seconds and DataFusion takes 2,100 seconds.

What version are you working with?

@Rachelint has some ideas of how to improve this:

• Sketch for aggregation intermediate results blocked management #11943
• Manage group values and states by blocks in aggregation #11931

[Dandandan]

Hm this seems something quadratic in nature?

When the dataset is 100 million rows, then Polars takes 126 seconds and DataFusion takes 2,100 seconds.

What version are you working with?

@Rachelint has some ideas of how to improve this:

• Sketch for aggregation intermediate results blocked management #11943
• Manage group values and states by blocks in aggregation #11931

Does it fully explain the dramatic difference? @MrPowers how do you generate the 10M vs 100M rows?

[alamb]

I would also expect this to help (but it was merged and depends on when it is merged)

• Skipping partial aggregation when it is not helping for high cardinality aggregates #11627
  I think the right thing to do is to get the query / dataset and profile it

[MrPowers]

@Dandandan - thanks to the great work by @SemyonSinchenko, it's easy to generate these datasets with falsa.

Here's the command to generate the 10 million row dataset: falsa groupby --path-prefix=~/data --size SMALL --data-format PARQUET. Just use MEDIUM to generate the 100 million row dataset.

[Rachelint]

@Dandandan - thanks to the great work by @SemyonSinchenko, it's easy to generate these datasets with falsa.

Here's the command to generate the 10 million row dataset: falsa groupby --path-prefix=~/data --size SMALL --data-format PARQUET. Just use MEDIUM to generate the 100 million row dataset.

Thanks, I will profile and see what happen about the so long time cost in datafusion.

[Rachelint]

When the dataset is 100 million rows, then Polars takes 126 seconds and DataFusion takes 2,100 seconds.

What version are you working with?

@Rachelint has some ideas of how to improve this:

* [Sketch for aggregation intermediate results blocked management #11943](https://github.com/apache/datafusion/pull/11943)

* [Manage group values and states by blocks in aggregation #11931](https://github.com/apache/datafusion/issues/11931)

🤔 I guess it may be caused by the similar reason of what we encountered during benchmarking in #11827

[alamb]

When the dataset is 100 million rows, then Polars takes 126 seconds and DataFusion takes 2,100 seconds.

What version are you working with?
@Rachelint has some ideas of how to improve this:

* [Sketch for aggregation intermediate results blocked management #11943](https://github.com/apache/datafusion/pull/11943)

* [Manage group values and states by blocks in aggregation #11931](https://github.com/apache/datafusion/issues/11931)

🤔 I guess it may be caused by the similar reason of what we encountered during benchmarking in #11827

Specifically that power and corr need to support convert_to_state?

[Rachelint]

When the dataset is 100 million rows, then Polars takes 126 seconds and DataFusion takes 2,100 seconds.

What version are you working with?
@Rachelint has some ideas of how to improve this:

* [Sketch for aggregation intermediate results blocked management #11943](https://github.com/apache/datafusion/pull/11943)

* [Manage group values and states by blocks in aggregation #11931](https://github.com/apache/datafusion/issues/11931)

🤔 I guess it may be caused by the similar reason of what we encountered during benchmarking in #11827

Specifically that power and corr need to support convert_to_state?

I am not sure, but I think it maybe really related to GroupAccumulatorAdapter as #11827?
I am running and profiling it to find the answer.

[2010YOUY01]

I rerun the H2O Q9 with GroupsAccumulator for corr(). See #13581
h2o dataset is in parquet format

Result
----
main, h2o_10m: 0.8s
main, h2o_100m: 12s
pr, h2o_10m: 0.2s
pr, h2o_100m: 4s

I didn't reproduce the drastic slowdown in main branch🤔

When the dataset is 10 million rows, then Polars takes 3 seconds and DataFusion takes 3.6 seconds, so pretty similar.

When the dataset is 100 million rows, then Polars takes 126 seconds and DataFusion takes 2,100 seconds.

[MrPowers]

The h2o benchmarks are run on a Intel(R) Xeon(R) Platinum 8375C CPU @ 2.90GHz machine with 128 cores and 250 GB of RAM.

DataFusion groupby queries perform well on the 100 million row dataset (~5GB of data in a CSV file):

Some don't run with the 1 billion row dataset (~50GB of data in an uncompressed CSV file):

I am using a M3 Macbook with 16 GB of RAM. How much RAM does your machine have? Perhaps DataFusion only struggles with query 9 when the machine doesn't have lots of extra RAM.