apache · adriangb · Sep 7, 2026 · Sep 7, 2026 · Sep 7, 2026 · Sep 8, 2026
diff --git a/benchmarks/bench.sh b/benchmarks/bench.sh
@@ -106,7 +106,7 @@ wide_schema:            Small-projection queries on a wide synthetic dataset (10
                           (runs both 'wide' and 'narrow' subgroups: narrow is an internal baseline; the wide-vs-narrow ratio is the signal)
 predicate_eval:         Conjunctive (AND) filter-evaluation micro-benchmarks; each subgroup is a different predicate pattern, to test how an
                           adaptive predicate-ordering system behaves across them (see https://github.com/apache/datafusion/issues/11262)
-                          (subgroups via BENCH_SUBGROUP: costsel, cost, selectivity, cardinality, width, scale, neutral, correlation, drift)
+                          (subgroups via BENCH_SUBGROUP: costsel, cost, selectivity, cardinality, width, scale, neutral, correlation, drift, nulls)
                           (toggle a system under test with its native DATAFUSION_* env var; size data with PRED_ROWS, string width with PRED_FILL)
 parquet_row_filter_skip: Per-RG fully-matched RowFilter skip on Parquet (apache/datafusion#23696); clustered string key + low-selectivity
                           range filter + pushdown, so most row groups are fully matched and the per-row RowFilter is skipped on them
@@ -876,7 +876,11 @@ run_push_down_topk() {
 # micro-benchmarks where each subgroup is a different predicate pattern, used to
 # test how an adaptive predicate-ordering system behaves across them (see
 # https://github.com/apache/datafusion/issues/11262). Data is generated inline
-# by the suite's load SQL, so there is no data step.
+# by the suite's load SQL, so there is no data step (drift q82 and q83 share 16
+# small Parquet files written into sql_benchmarks/predicate_eval/scratch/, which
+# is gitignored; q82 reads them with target_partitions=1, so the selectivity flip
+# lands halfway through one stream, and q83 with 16, so each stream gets one whole
+# file and a fixed profile).
 #
 # By default the suite measures DataFusion's built-in left-deep AND short-circuit
 # and sets no engine config of its own. To evaluate a system under test, export
@@ -885,7 +889,7 @@ run_push_down_topk() {
 #   DATAFUSION_EXECUTION_ADAPTIVE_FILTER_REORDERING=true ./bench.sh run predicate_eval
 # Suite-specific knobs (string-substituted into the load SQL, not engine config):
 #   BENCH_SUBGROUP   run one subgroup (costsel, cost, selectivity, cardinality,
-#                    width, scale, neutral, correlation, drift)
+#                    width, scale, neutral, correlation, drift, nulls)
 #   PRED_ROWS        synthetic row count (default 1_000_000; the scale subgroup
 #                    overrides this per query)
 #   PRED_FILL        filler chars per marker = string-column width knob

diff --git a/benchmarks/sql_benchmarks/README.md b/benchmarks/sql_benchmarks/README.md
@@ -43,7 +43,7 @@ in the community:
 | `tpcds`               | TPC‑DS queries                                                     |
 | `tpch`                | TPC‑H queries                                                      |
 | `wide_schema`         | Small-projection queries on a wide (1024-col, 256-file) synthetic dataset; runs `wide` + `narrow` subgroups for comparison |
-| `predicate_eval`      | Conjunctive (AND) filter-evaluation micro-benchmarks; each subgroup is a different predicate pattern, to test how an adaptive predicate-ordering system behaves across them ([#11262](https://github.com/apache/datafusion/issues/11262)). Subgroups (`--subgroup`): `costsel`, `cost`, `selectivity`, `cardinality`, `width`, `scale`, `neutral`, `correlation`, `drift`. Configure the system under test through its DataFusion settings. |
+| `predicate_eval`      | Conjunctive (AND) filter-evaluation micro-benchmarks; each subgroup is a different predicate pattern, to test how an adaptive predicate-ordering system behaves across them ([#11262](https://github.com/apache/datafusion/issues/11262)). Subgroups (`--subgroup`): `costsel`, `cost`, `selectivity`, `cardinality`, `width`, `scale`, `neutral`, `correlation`, `drift`, `nulls`. The suite sets no engine config of its own, so by default it measures DataFusion's built-in left-deep `AND` short-circuit; point it at a system under test by exporting that system's own DataFusion setting (the harness builds its `SessionConfig` with `SessionConfig::from_env`). Every query is a `count(*)`, and the counts are checked in under `predicate_eval/results/`, so `--result-mode validate` also checks that a reordering under test still returns the same rows; the checked-in counts were persisted at the suite defaults (`PRED_ROWS=1000000`, `PRED_FILL=30`), so validation assumes those (the `scale` and `width` subgroups pin their own values per query and validate at any setting). |
 | `parquet_row_filter_skip` | Micro-benchmark for the per-row-group fully-matched RowFilter skip on Parquet scans ([#23696](https://github.com/apache/datafusion/issues/23696)). Subgroups (`--subgroup`): `skip` (clustered key, most row groups fully matched by statistics so the per-row filter is skipped), `control` (scrambled key, no row group is ever fully matched). Size the data with `PRED_ROWS` and the row-group size with `RG_SIZE`. |
 
 # Running Benchmarks

diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q30.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q30.benchmark
@@ -1,7 +1,6 @@
 subgroup cardinality
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=cardinality
-QPAD=30
+QUERY=cardinality/q30
 DATASET=ints
 NAME=cardinality_q30_k2
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q31.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q31.benchmark
@@ -1,7 +1,6 @@
 subgroup cardinality
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=cardinality
-QPAD=31
+QUERY=cardinality/q31
 DATASET=ints
 NAME=cardinality_q31_k4
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q32.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q32.benchmark
@@ -1,7 +1,6 @@
 subgroup cardinality
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=cardinality
-QPAD=32
+QUERY=cardinality/q32
 DATASET=ints
 NAME=cardinality_q32_k8
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q33.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q33.benchmark
@@ -1,7 +1,6 @@
 subgroup cardinality
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=cardinality
-QPAD=33
+QUERY=cardinality/q33
 DATASET=ints
 NAME=cardinality_q33_k16
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q34.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cardinality/q34.benchmark
@@ -0,0 +1,6 @@
+subgroup cardinality
+
+template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
+QUERY=cardinality/q34
+DATASET=ints_wide
+NAME=cardinality_q34_k8_wide64
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q70.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q70.benchmark
@@ -1,7 +1,6 @@
 subgroup correlation
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=correlation
-QPAD=70
+QUERY=correlation/q70
 DATASET=corr
 NAME=correlation_q70_independent
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q71.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q71.benchmark
@@ -1,7 +1,6 @@
 subgroup correlation
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=correlation
-QPAD=71
+QUERY=correlation/q71
 DATASET=corr
 NAME=correlation_q71_positive
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q72.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q72.benchmark
@@ -1,7 +1,6 @@
 subgroup correlation
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=correlation
-QPAD=72
+QUERY=correlation/q72
 DATASET=corr
 NAME=correlation_q72_anti
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q73.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/correlation/q73.benchmark
@@ -1,7 +1,6 @@
 subgroup correlation
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=correlation
-QPAD=73
+QUERY=correlation/q73
 DATASET=corrproxy
 NAME=correlation_q73_redundant_proxy
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cost/q10.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cost/q10.benchmark
@@ -1,7 +1,6 @@
 subgroup cost
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=cost
-QPAD=10
+QUERY=cost/q10
 DATASET=mixed
 NAME=cost_q10_expensive_first
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cost/q11.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/cost/q11.benchmark
@@ -1,7 +1,6 @@
 subgroup cost
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=cost
-QPAD=11
+QUERY=cost/q11
 DATASET=mixed
 NAME=cost_q11_cheap_first
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q01.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q01.benchmark
@@ -1,7 +1,6 @@
 subgroup costsel
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=costsel
-QPAD=01
+QUERY=costsel/q01
 DATASET=markers
 NAME=costsel_q01_regexp_selective_last
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q02.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q02.benchmark
@@ -1,7 +1,6 @@
 subgroup costsel
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=costsel
-QPAD=02
+QUERY=costsel/q02
 DATASET=markers
 NAME=costsel_q02_regexp_selective_first
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q03.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q03.benchmark
@@ -1,7 +1,6 @@
 subgroup costsel
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=costsel
-QPAD=03
+QUERY=costsel/q03
 DATASET=mixed
 NAME=costsel_q03_cheap_unselective_then_expensive_selective
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q04.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/costsel/q04.benchmark
@@ -0,0 +1,6 @@
+subgroup costsel
+
+template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
+QUERY=costsel/q04
+DATASET=mixed
+NAME=costsel_q04_expensive_selective_then_cheap_unselective
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q80.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q80.benchmark
@@ -1,7 +1,6 @@
 subgroup drift
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=drift
-QPAD=80
+QUERY=drift/q80
 DATASET=drift
 NAME=drift_q80_a_then_b
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q81.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q81.benchmark
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q82.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q82.benchmark
@@ -0,0 +1,10 @@
+subgroup drift
+
+template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
+QUERY=drift/q82
+DATASET=drift_split
+NAME=drift_q82_late_flip
+
+# One stream over the whole directory: the flip lands mid-scan.
+init
+set datafusion.execution.target_partitions = 1;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q83.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/drift/q83.benchmark
@@ -0,0 +1,10 @@
+subgroup drift
+
+template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
+QUERY=drift/q83
+DATASET=drift_split
+NAME=drift_q83_per_partition_skew
+
+# One file per stream: each stream sees a single profile.
+init
+set datafusion.execution.target_partitions = 16;
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/neutral/q60.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/neutral/q60.benchmark
@@ -1,7 +1,6 @@
 subgroup neutral
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=neutral
-QPAD=60
+QUERY=neutral/q60
 DATASET=ints
 NAME=neutral_q60_cheap_uniform
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/neutral/q61.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/neutral/q61.benchmark
@@ -1,7 +1,6 @@
 subgroup neutral
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=neutral
-QPAD=61
+QUERY=neutral/q61
 DATASET=markers
 NAME=neutral_q61_expensive_uniform
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/nulls/q90.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/nulls/q90.benchmark
@@ -0,0 +1,6 @@
+subgroup nulls
+
+template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
+QUERY=nulls/q90
+DATASET=ints
+NAME=nulls_q90_nullable_selective_first
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/nulls/q91.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/nulls/q91.benchmark
@@ -0,0 +1,6 @@
+subgroup nulls
+
+template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
+QUERY=nulls/q91
+DATASET=ints
+NAME=nulls_q91_nullable_selective_last
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q50.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q50.benchmark
@@ -1,8 +1,7 @@
 subgroup scale
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=scale
-QPAD=50
+QUERY=costsel/q03
 DATASET=mixed
 PRED_ROWS=5000
 NAME=scale_q50_5k
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q51.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q51.benchmark
@@ -1,8 +1,7 @@
 subgroup scale
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=scale
-QPAD=51
+QUERY=costsel/q03
 DATASET=mixed
 PRED_ROWS=100000
 NAME=scale_q51_100k
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q52.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q52.benchmark
@@ -1,8 +1,7 @@
 subgroup scale
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=scale
-QPAD=52
+QUERY=costsel/q03
 DATASET=mixed
 PRED_ROWS=5000000
 NAME=scale_q52_5m
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q53.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/scale/q53.benchmark
@@ -1,8 +1,7 @@
 subgroup scale
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=scale
-QPAD=53
+QUERY=costsel/q03
 DATASET=mixed
 PRED_ROWS=50000000
 NAME=scale_q53_50m
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/selectivity/q20.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/selectivity/q20.benchmark
@@ -1,7 +1,6 @@
 subgroup selectivity
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=selectivity
-QPAD=20
+QUERY=selectivity/q20
 DATASET=ints
 NAME=selectivity_q20_unselective_first
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/selectivity/q21.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/selectivity/q21.benchmark
@@ -1,7 +1,6 @@
 subgroup selectivity
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=selectivity
-QPAD=21
+QUERY=selectivity/q21
 DATASET=ints
 NAME=selectivity_q21_selective_first
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q40.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q40.benchmark
@@ -1,8 +1,7 @@
 subgroup width
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=width
-QPAD=40
+QUERY=costsel/q01
 DATASET=markers
 PRED_FILL=2
 NAME=width_q40_narrow
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q41.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q41.benchmark
@@ -1,8 +1,7 @@
 subgroup width
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=width
-QPAD=41
+QUERY=costsel/q01
 DATASET=markers
 PRED_FILL=30
 NAME=width_q41_wide
diff --git a/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q42.benchmark b/benchmarks/sql_benchmarks/predicate_eval/benchmarks/width/q42.benchmark
@@ -1,8 +1,7 @@
 subgroup width
 
 template sql_benchmarks/predicate_eval/predicate_eval.benchmark.template
-SUBGROUP=width
-QPAD=42
+QUERY=costsel/q01
 DATASET=markers
 PRED_FILL=170
 NAME=width_q42_xwide
diff --git a/benchmarks/sql_benchmarks/predicate_eval/load/corr.sql b/benchmarks/sql_benchmarks/predicate_eval/load/corr.sql
@@ -1,15 +1,6 @@
--- Correlation dataset: a base column plus derived columns that control the
--- *conditional* selectivity of one predicate given another (its selectivity
--- among the rows that already passed the other).
---
---   x       uniform [0,100)
---   x_pos   = x         (perfectly positively correlated: `x<k AND x_pos<k`
---                        passes ~k%, not the ~k%^2 an independence assumption
---                        predicts)
---   x_anti  = 99 - x    (anti-correlated: `x<k AND x_anti<k` is empty for k<=50)
---   ind     independent control column, uniform [0,100)
---
--- PRED_ROWS sizes the table.
+-- Correlation dataset: `x` uniform on [0,100); `x_pos` = x (perfectly positively
+-- correlated); `x_anti` = 99 - x (anti-correlated); `ind` uniform on [0,100) and
+-- independent of x. PRED_ROWS sizes the table.
 CREATE TABLE t AS
 SELECT
   (value * 7)  % 100        AS x,

diff --git a/benchmarks/sql_benchmarks/predicate_eval/load/corrproxy.sql b/benchmarks/sql_benchmarks/predicate_eval/load/corrproxy.sql
@@ -1,30 +1,20 @@
--- Correlated-proxy dataset: a cheap integer predicate that is a perfect proxy
--- for three string predicates, plus one independent string predicate.
---
+-- Correlated-proxy dataset:
 --   c0          = 1 for ~30% of rows (cheap proxy)
---   s1, s2, s3  each contain a marker exactly where c0 = 1     (correlated)
---   s4          contains a marker for an independent ~30%      (independent)
---
--- The four string columns are deliberately *identical in shape*: same width,
--- the same single marker at the same offset, each matched by an equally cheap
--- regex with the same ~30% marginal selectivity. Marginally the four regex
--- predicates are therefore indistinguishable -- same cost, same selectivity, in
--- every position -- so neither a marginal cost/selectivity estimator nor
--- runtime timing can prefer one over another. Only their *conditional*
--- behaviour behind the proxy differs: after `c0 = 1`, the s1/s2/s3 regexes keep
--- every survivor (each re-tests the proxy's condition) while the s4 regex still
--- discards ~70%. Only joint statistics can see that; an independence assumption
--- prices all four regexes identically in every position.
+--   s1, s2, s3  each carry a marker exactly where c0 = 1  (correlated)
+--   s4          carries a marker for an independent ~30%  (independent)
+-- PRED_FILL sets the filler width each side of the marker, PRED_ROWS the row count.
 --
--- PRED_FILL sets the filler width on each side of the marker (a non-matching
--- `regexp_like` must scan the whole value), and PRED_ROWS sizes the table.
+-- The four string columns are identical in shape -- same width, same marker
+-- offset, same regex cost, same ~30% marginal selectivity -- so marginally they
+-- are indistinguishable in every position. Only their joint distribution with the
+-- proxy differs: after `c0 = 1` the s1/s2/s3 regexes keep every survivor while s4
+-- still discards ~70%. Ranking them therefore takes joint statistics; an
+-- independence assumption prices all four regexes identically everywhere.
 CREATE TABLE t AS
 WITH base AS (
   SELECT
-    -- The cheap proxy and the independent control share one definition each, so
-    -- the perfect-proxy / independence invariants can't drift apart silently.
-    (value * 7)  % 100 < 30 AS proxy,   -- ~30%, drives c0 and s1/s2/s3
-    (value * 13) % 100 < 30 AS indep    -- ~30%, independent of proxy, drives s4
+    (value * 7)  % 100 < 30 AS proxy,
+    (value * 13) % 100 < 30 AS indep
   FROM generate_series(1, ${PRED_ROWS:-1000000})
 )
 SELECT