You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@spark.apache.org by ru...@apache.org on 2026/03/26 06:09:17 UTC
(spark) branch master updated: [SPARK-56085][PYTHON][TEST] Add ASV micro-benchmarks for SQL_GROUPED_AGG_ARROW_UDF and SQL_GROUPED_AGG_ARROW_ITER_UDF
This is an automated email from the ASF dual-hosted git repository.
ruifengz pushed a commit to branch master
in repository https://gitbox.apache.org/repos/asf/spark.git
The following commit(s) were added to refs/heads/master by this push:
new cc85ba1d4c4a [SPARK-56085][PYTHON][TEST] Add ASV micro-benchmarks for SQL_GROUPED_AGG_ARROW_UDF and SQL_GROUPED_AGG_ARROW_ITER_UDF
cc85ba1d4c4a is described below
commit cc85ba1d4c4a08dc92bbae4c87c3134d7a04bcc5
Author: Yicong Huang <17...@users.noreply.github.com>
AuthorDate: Thu Mar 26 14:09:03 2026 +0800
[SPARK-56085][PYTHON][TEST] Add ASV micro-benchmarks for SQL_GROUPED_AGG_ARROW_UDF and SQL_GROUPED_AGG_ARROW_ITER_UDF
### What changes were proposed in this pull request?
Add ASV micro-benchmarks for `SQL_GROUPED_AGG_ARROW_UDF` and `SQL_GROUPED_AGG_ARROW_ITER_UDF` eval types to `bench_eval_type.py`.
### Why are the changes needed?
Part of [SPARK-55724](https://issues.apache.org/jira/browse/SPARK-55724). Establishes performance baselines before refactoring these eval types.
### Does this PR introduce _any_ user-facing change?
No
### How was this patch tested?
`COLUMNS=120 asv run --python=same --bench "GroupedAggArrow" --attribute "repeat=(3,5,5.0)"`:
**GroupedAggArrowUDFTimeBench** (`SQL_GROUPED_AGG_ARROW_UDF`):
```
================ ============ ================
-- udf
---------------- -----------------------------
scenario sum_udf mean_multi_udf
================ ============ ================
few_groups_sm 9.68±0.2ms 10.1±0.2ms
few_groups_lg 41.5±0.1ms 43.4±0.8ms
many_groups_sm 277±3ms 266±1ms
many_groups_lg 125±1ms 133±0.7ms
wide_cols 88.4±0.5ms 92.5±0.7ms
================ ============ ================
```
**GroupedAggArrowUDFPeakmemBench** (`SQL_GROUPED_AGG_ARROW_UDF`):
```
================ ========= ================
-- udf
---------------- --------------------------
scenario sum_udf mean_multi_udf
================ ========= ================
few_groups_sm 1.15G 1.15G
few_groups_lg 1.16G 1.16G
many_groups_sm 1.15G 1.15G
many_groups_lg 1.15G 1.15G
wide_cols 1.16G 1.15G
================ ========= ================
```
**GroupedAggArrowIterUDFTimeBench** (`SQL_GROUPED_AGG_ARROW_ITER_UDF`):
```
================ ============= ================
-- udf
---------------- ------------------------------
scenario sum_udf mean_multi_udf
================ ============= ================
few_groups_sm 7.36±0.09ms 7.84±0.06ms
few_groups_lg 30.4±0.2ms 33.0±0.3ms
many_groups_sm 218±2ms 201±2ms
many_groups_lg 88.8±0.9ms 96.6±0.4ms
wide_cols 57.3±0.1ms 56.9±2ms
================ ============= ================
```
**GroupedAggArrowIterUDFPeakmemBench** (`SQL_GROUPED_AGG_ARROW_ITER_UDF`):
```
================ ========= ================
-- udf
---------------- --------------------------
scenario sum_udf mean_multi_udf
================ ========= ================
few_groups_sm 1.15G 1.15G
few_groups_lg 1.15G 1.15G
many_groups_sm 1.15G 1.15G
many_groups_lg 1.15G 1.15G
wide_cols 1.15G 1.15G
================ ========= ================
```
### Was this patch authored or co-authored using generative AI tooling?
No
Closes #54932 from Yicong-Huang/SPARK-56085/bench/grouped-agg-arrow-udf.
Authored-by: Yicong Huang <17...@users.noreply.github.com>
Signed-off-by: Ruifeng Zheng <ru...@apache.org>
---
python/benchmarks/bench_eval_type.py | 182 +++++++++++++++++++++++++++++++++++
1 file changed, 182 insertions(+)
diff --git a/python/benchmarks/bench_eval_type.py b/python/benchmarks/bench_eval_type.py
index 7a2a5ddc0357..a489d5d15ce5 100644
--- a/python/benchmarks/bench_eval_type.py
+++ b/python/benchmarks/bench_eval_type.py
@@ -559,6 +559,188 @@ def _make_grouped_batches(
return groups, return_type
+# -- SQL_GROUPED_AGG_ARROW_UDF ------------------------------------------------
+# UDF receives pa.Array columns per group, returns scalar.
+
+
+def _grouped_agg_arrow_sum(col):
+ """Sum a single Arrow column."""
+ import pyarrow.compute as pc
+
+ return pc.sum(col).as_py()
+
+
+def _grouped_agg_arrow_mean_multi(col0, col1):
+ """Mean of two Arrow columns combined."""
+ import pyarrow.compute as pc
+
+ return (pc.mean(col0).as_py() or 0) + (pc.mean(col1).as_py() or 0)
+
+
+_GROUPED_AGG_ARROW_UDFS = {
+ "sum_udf": _grouped_agg_arrow_sum,
+ "mean_multi_udf": _grouped_agg_arrow_mean_multi,
+}
+
+
+# -- SQL_GROUPED_AGG_ARROW_ITER_UDF ------------------------------------------
+# UDF receives Iterator[pa.Array] (or Iterator[Tuple[pa.Array, ...]]) per group,
+# returns scalar.
+
+
+def _grouped_agg_arrow_iter_sum(batch_iter):
+ """Sum across batches via iterator."""
+ import pyarrow.compute as pc
+
+ total = 0
+ for col in batch_iter:
+ total += pc.sum(col).as_py() or 0
+ return total
+
+
+def _grouped_agg_arrow_iter_mean_multi(batch_iter):
+ """Mean across batches of tuples via iterator."""
+ import pyarrow.compute as pc
+
+ total = 0.0
+ for col0, col1 in batch_iter:
+ total += (pc.mean(col0).as_py() or 0) + (pc.mean(col1).as_py() or 0)
+ return total
+
+
+_GROUPED_AGG_ARROW_ITER_UDFS = {
+ "sum_udf": _grouped_agg_arrow_iter_sum,
+ "mean_multi_udf": _grouped_agg_arrow_iter_mean_multi,
+}
+
+
+def _make_agg_arrow_groups(
+ num_groups: int,
+ rows_per_group: int,
+ n_cols: int,
+) -> list[tuple[pa.RecordBatch, ...]]:
+ """Create grouped data for agg Arrow UDFs (plain batches, not struct-wrapped)."""
+ type_cycle = [
+ lambda r: pa.array(np.random.randint(0, 1000, r, dtype=np.int64)),
+ lambda r: pa.array(np.random.rand(r)),
+ lambda r: pa.array(np.random.randint(0, 100, r, dtype=np.int64)),
+ lambda r: pa.array(np.random.rand(r)),
+ ]
+
+ groups = []
+ for _ in range(num_groups):
+ arrays = [type_cycle[i % len(type_cycle)](rows_per_group) for i in range(n_cols)]
+ names = [f"col_{i}" for i in range(n_cols)]
+ batch = pa.RecordBatch.from_arrays(arrays, names=names)
+ groups.append((batch,))
+
+ return groups
+
+
+def _build_grouped_agg_arrow_scenarios():
+ """Build scenarios for SQL_GROUPED_AGG_ARROW_UDF / AGG_ARROW_ITER_UDF.
+
+ Returns dict mapping name to (groups, n_cols, arg_offsets_map) where
+ arg_offsets_map provides offsets per UDF name.
+ """
+ scenarios = {}
+
+ for name, (num_groups, rows_per_group, n_cols) in {
+ "few_groups_sm": (50, 5_000, 5),
+ "few_groups_lg": (50, 50_000, 5),
+ "many_groups_sm": (2_000, 500, 5),
+ "many_groups_lg": (500, 10_000, 5),
+ "wide_cols": (200, 5_000, 20),
+ }.items():
+ groups = _make_agg_arrow_groups(num_groups, rows_per_group, n_cols)
+ scenarios[name] = (groups, n_cols)
+
+ return scenarios
+
+
+_GROUPED_AGG_ARROW_SCENARIOS = _build_grouped_agg_arrow_scenarios()
+
+
+class _GroupedAggArrowBenchMixin:
+ """Provides _write_scenario for SQL_GROUPED_AGG_ARROW_UDF."""
+
+ def _write_scenario(self, scenario, udf_name, buf):
+ groups, n_cols = self._scenarios[scenario]
+ udf_func = self._udfs[udf_name]
+
+ # sum_udf uses 1 arg, mean_multi_udf uses 2 args
+ if "multi" in udf_name:
+ arg_offsets = [0, 1]
+ else:
+ arg_offsets = [0]
+
+ return_type = DoubleType()
+
+ def write_command(b):
+ _build_udf_payload(udf_func, return_type, arg_offsets, b)
+
+ _write_worker_input(
+ PythonEvalType.SQL_GROUPED_AGG_ARROW_UDF,
+ write_command,
+ lambda b: _write_grouped_arrow_data(groups, num_dfs=1, buf=b),
+ buf,
+ )
+
+
+class GroupedAggArrowUDFTimeBench(_GroupedAggArrowBenchMixin, _TimeBenchBase):
+ _scenarios = _GROUPED_AGG_ARROW_SCENARIOS
+ _udfs = _GROUPED_AGG_ARROW_UDFS
+ params = [list(_GROUPED_AGG_ARROW_SCENARIOS), list(_GROUPED_AGG_ARROW_UDFS)]
+ param_names = ["scenario", "udf"]
+
+
+class GroupedAggArrowUDFPeakmemBench(_GroupedAggArrowBenchMixin, _PeakmemBenchBase):
+ _scenarios = _GROUPED_AGG_ARROW_SCENARIOS
+ _udfs = _GROUPED_AGG_ARROW_UDFS
+ params = [list(_GROUPED_AGG_ARROW_SCENARIOS), list(_GROUPED_AGG_ARROW_UDFS)]
+ param_names = ["scenario", "udf"]
+
+
+class _GroupedAggArrowIterBenchMixin:
+ """Provides _write_scenario for SQL_GROUPED_AGG_ARROW_ITER_UDF."""
+
+ def _write_scenario(self, scenario, udf_name, buf):
+ groups, n_cols = self._scenarios[scenario]
+ udf_func = self._udfs[udf_name]
+
+ # sum_udf uses 1 arg, mean_multi_udf uses 2 args
+ if "multi" in udf_name:
+ arg_offsets = [0, 1]
+ else:
+ arg_offsets = [0]
+
+ return_type = DoubleType()
+
+ def write_command(b):
+ _build_udf_payload(udf_func, return_type, arg_offsets, b)
+
+ _write_worker_input(
+ PythonEvalType.SQL_GROUPED_AGG_ARROW_ITER_UDF,
+ write_command,
+ lambda b: _write_grouped_arrow_data(groups, num_dfs=1, buf=b),
+ buf,
+ )
+
+
+class GroupedAggArrowIterUDFTimeBench(_GroupedAggArrowIterBenchMixin, _TimeBenchBase):
+ _scenarios = _GROUPED_AGG_ARROW_SCENARIOS
+ _udfs = _GROUPED_AGG_ARROW_ITER_UDFS
+ params = [list(_GROUPED_AGG_ARROW_SCENARIOS), list(_GROUPED_AGG_ARROW_ITER_UDFS)]
+ param_names = ["scenario", "udf"]
+
+
+class GroupedAggArrowIterUDFPeakmemBench(_GroupedAggArrowIterBenchMixin, _PeakmemBenchBase):
+ _scenarios = _GROUPED_AGG_ARROW_SCENARIOS
+ _udfs = _GROUPED_AGG_ARROW_ITER_UDFS
+ params = [list(_GROUPED_AGG_ARROW_SCENARIOS), list(_GROUPED_AGG_ARROW_ITER_UDFS)]
+ param_names = ["scenario", "udf"]
+
+
# -- SQL_GROUPED_MAP_ARROW_UDF ------------------------------------------------
# UDF receives (key: pa.RecordBatch, values: Iterator[pa.RecordBatch]),
# returns Iterator[pa.RecordBatch].
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscribe@spark.apache.org
For additional commands, e-mail: commits-help@spark.apache.org