Bench: Cover Boolean filtering and oversized dictionaries by Rich-T-kid · Pull Request #12054 · apache/arrow-rs

Rich-T-kid · 2026-11-03T15:25:22Z

Which issue does this PR close?

Part of #11885. Adds benchmark coverage for #11901 and #12008, and follows up on the Boolean filter coverage requested in #11918.

This consolidates and supersedes #11980 and #12041.

Why is this change needed?

The existing suites do not cover Boolean arrays as filter data or the cost of operating on small key ranges backed by oversized dictionaries. They also do not exercise the 16-values-per-selected-key boundary between dense and sparse dictionary mappings.

What changes are included in this PR?

Adds nullable and non-null Boolean filter cases across the existing selectivities.
Adds prepared RowConverter cases for rotating 8,192-key ranges backed by unique string dictionaries with 1M and 4M values.
- Measures the full-value encoding cost from RowConverter encodes every dictionary value for each batch, not just the values its rows reference #11901 without including RowConverter construction.
Adds concat cases for 12 distinct dictionaries sliced to 682 keys, with 12, 16, and 24 values per selected key.
- Straddles Perf: Improve dictionary slice concatenation #12008's 16-values-per-selected-key cutoff so we can assess the threshold.

Are these changes tested?

Built and ran each filtered benchmark locally.

Benchmark bot runs:

Are there any significant user-facing changes?

No. This only adds benchmark coverage.

Rich-T-kid · 2026-11-03T15:41:01Z

run benchmark filter_kernels
env:
BENCH_FILTER: "filter boolean"

Rich-T-kid · 2026-11-03T15:41:02Z

run benchmark row_format
env:
BENCH_FILTER: "convert_columns_prepared 8192 string_dictionary"

adriangbot · 2026-11-03T15:41:08Z

🤖 Benchmark starting (GKE) | trigger

Target: row_format

Sharding factor: 1 (1 workers).

Comparing rich-T-kid/dictionary-concat-ratio-bench (37da834) to b231cf1 (merge-base) diff

Run configuration

run benchmark row_format
env:
  BENCH_FILTER: convert_columns_prepared 8192 string_dictionary
shards: 1

Results will be posted when all workers finish.

File an issue against this benchmark runner

adriangbot · 2026-11-03T15:41:09Z

🤖 Benchmark starting (GKE) | trigger

Target: filter_kernels

Sharding factor: 1 (1 workers).

Comparing rich-T-kid/dictionary-concat-ratio-bench (37da834) to b231cf1 (merge-base) diff

Run configuration

run benchmark filter_kernels
env:
  BENCH_FILTER: filter boolean
shards: 1

Results will be posted when all workers finish.

File an issue against this benchmark runner

Rich-T-kid · 2026-11-03T15:41:10Z

run benchmark concatenate_kernel
env:
BENCH_FILTER: "concat str_dict_slice"

adriangbot · 2026-11-03T15:41:14Z

🤖 Benchmark starting (GKE) | trigger

Target: concatenate_kernel

Sharding factor: 1 (1 workers).

Comparing rich-T-kid/dictionary-concat-ratio-bench (37da834) to b231cf1 (merge-base) diff

Run configuration

run benchmark concatenate_kernel
env:
  BENCH_FILTER: concat str_dict_slice
shards: 1

Results will be posted when all workers finish.

File an issue against this benchmark runner

adriangbot · 2026-11-03T15:43:47Z

🤖 Benchmark completed (GKE) | trigger

Target: row_format

Comparing rich-T-kid/dictionary-concat-ratio-bench (37da834) to b231cf1 (merge-base) diff

Run configuration

run benchmark row_format
env:
  BENCH_FILTER: convert_columns_prepared 8192 string_dictionary
shards: 1

Details

group                                                              changed
-----                                                              -------
convert_columns_prepared 8192 string_dictionary(1000000 values)    1.00      6.8±0.05ms        ? ?/sec
convert_columns_prepared 8192 string_dictionary(4000000 values)    1.00     67.9±0.43ms        ? ?/sec

Per-runner information

row_format — shard 1/1

Node: gk3-benchmark-cluster-nap-1q2slmnr-b33373f7-5pzx

Instance: c4a-highmem-16 (12 vCPU / 65 GiB)

uname:

Linux bench-c6608167940-4309-zqbz7 6.12.105+ #1 SMP Sat Sep 19 07:55:25 UTC 2026 aarch64 GNU/Linux

BENCH_COMMAND:

cargo bench --features=arrow,async,test_common,experimental,object_store --bench row_format

CPU Details (lscpu)

Architecture:                            aarch64
CPU op-mode(s):                          64-bit
Byte Order:                              Little Endian
CPU(s):                                  16
On-line CPU(s) list:                     0-15
Vendor ID:                               ARM
Model name:                              Neoverse-V2
Model:                                   1
Thread(s) per core:                      1
Core(s) per cluster:                     16
Socket(s):                               -
Cluster(s):                              1
Stepping:                                r0p1
BogoMIPS:                                2000.00
Flags:                                   fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 sm3 sm4 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm sb paca pacg dcpodp sve2 sveaes svepmull svebitperm svesha3 svesm4 flagm2 frint svei8mm svebf16 i8mm bf16 dgh rng bti
L1d cache:                               1 MiB (16 instances)
L1i cache:                               1 MiB (16 instances)
L2 cache:                                32 MiB (16 instances)
L3 cache:                                80 MiB (1 instance)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-15
Vulnerability Gather data sampling:      Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Not affected
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; CSV2, BHB
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Not affected

Resource Usage

base (merge-base)

Metric	Value
Wall time	5.0s
Peak memory	5.8 MiB
Avg memory	2.9 MiB
CPU user	0.0s
CPU sys	0.0s
Peak spill	0 B

branch

Metric	Value
Wall time	25.0s
Peak memory	221.7 MiB
Avg memory	92.1 MiB
CPU user	13.5s
CPU sys	6.5s
Peak spill	0 B

File an issue against this benchmark runner

adriangbot · 2026-11-03T15:45:14Z

🤖 Benchmark completed (GKE) | trigger

Target: concatenate_kernel

Comparing rich-T-kid/dictionary-concat-ratio-bench (37da834) to b231cf1 (merge-base) diff

Run configuration

run benchmark concatenate_kernel
env:
  BENCH_FILTER: concat str_dict_slice
shards: 1

Details

group                                                        changed
-----                                                        -------
concat str_dict_slice rows=682 values_per_key=12 count=12    1.00    203.3±0.83µs        ? ?/sec
concat str_dict_slice rows=682 values_per_key=16 count=12    1.00    209.2±0.91µs        ? ?/sec
concat str_dict_slice rows=682 values_per_key=24 count=12    1.00    222.6±1.17µs        ? ?/sec

Per-runner information

concatenate_kernel — shard 1/1

Node: gk3-benchmark-cluster-nap-35k6403m-e6737b58-4fjg

Instance: c4a-highmem-16 (12 vCPU / 65 GiB)

uname:

Linux bench-c6608170975-4311-mb2tg 6.12.105+ #1 SMP Sat Sep 19 07:55:25 UTC 2026 aarch64 GNU/Linux

BENCH_COMMAND:

cargo bench --features=arrow,async,test_common,experimental,object_store --bench concatenate_kernel

CPU Details (lscpu)

Architecture:                            aarch64
CPU op-mode(s):                          64-bit
Byte Order:                              Little Endian
CPU(s):                                  16
On-line CPU(s) list:                     0-15
Vendor ID:                               ARM
Model name:                              Neoverse-V2
Model:                                   1
Thread(s) per core:                      1
Core(s) per cluster:                     16
Socket(s):                               -
Cluster(s):                              1
Stepping:                                r0p1
BogoMIPS:                                2000.00
Flags:                                   fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 sm3 sm4 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm sb paca pacg dcpodp sve2 sveaes svepmull svebitperm svesha3 svesm4 flagm2 frint svei8mm svebf16 i8mm bf16 dgh rng bti
L1d cache:                               1 MiB (16 instances)
L1i cache:                               1 MiB (16 instances)
L2 cache:                                32 MiB (16 instances)
L3 cache:                                80 MiB (1 instance)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-15
Vulnerability Gather data sampling:      Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Not affected
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; CSV2, BHB
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Not affected

Resource Usage

base (merge-base)

Metric	Value
Wall time	5.0s
Peak memory	129.6 MiB
Avg memory	38.9 MiB
CPU user	1.6s
CPU sys	0.1s
Peak spill	0 B

branch

Metric	Value
Wall time	30.0s
Peak memory	130.7 MiB
Avg memory	94.3 MiB
CPU user	28.0s
CPU sys	0.1s
Peak spill	0 B

File an issue against this benchmark runner

adriangbot · 2026-11-03T15:46:19Z

🤖 Benchmark completed (GKE) | trigger

Target: filter_kernels

Comparing rich-T-kid/dictionary-concat-ratio-bench (37da834) to b231cf1 (merge-base) diff

Run configuration

run benchmark filter_kernels
env:
  BENCH_FILTER: filter boolean
shards: 1

Details

group                                                       changed
-----                                                       -------
filter boolean (kept 1/2)                                   1.00      8.6±0.00µs        ? ?/sec
filter boolean high selectivity (kept 1023/1024)            1.00      2.5±0.00µs        ? ?/sec
filter boolean low selectivity (kept 1/1024)                1.00    683.2±1.19ns        ? ?/sec
filter boolean w NULLs (kept 1/2)                           1.00     17.1±0.01µs        ? ?/sec
filter boolean w NULLs high selectivity (kept 1023/1024)    1.00      4.8±0.01µs        ? ?/sec
filter boolean w NULLs low selectivity (kept 1/1024)        1.00   1187.8±1.19ns        ? ?/sec

Per-runner information

filter_kernels — shard 1/1

Node: gk3-benchmark-cluster-nap-1q2slmnr-f9d718d9-5nfz

Instance: c4a-highmem-16 (12 vCPU / 65 GiB)

uname:

Linux bench-c6608167474-4310-lchtl 6.12.105+ #1 SMP Sat Sep 19 07:55:25 UTC 2026 aarch64 GNU/Linux

BENCH_COMMAND:

cargo bench --features=arrow,async,test_common,experimental,object_store --bench filter_kernels

CPU Details (lscpu)

Architecture:                            aarch64
CPU op-mode(s):                          64-bit
Byte Order:                              Little Endian
CPU(s):                                  16
On-line CPU(s) list:                     0-15
Vendor ID:                               ARM
Model name:                              Neoverse-V2
Model:                                   1
Thread(s) per core:                      1
Core(s) per cluster:                     16
Socket(s):                               -
Cluster(s):                              1
Stepping:                                r0p1
BogoMIPS:                                2000.00
Flags:                                   fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 sm3 sm4 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm sb paca pacg dcpodp sve2 sveaes svepmull svebitperm svesha3 svesm4 flagm2 frint svei8mm svebf16 i8mm bf16 dgh rng bti
L1d cache:                               1 MiB (16 instances)
L1i cache:                               1 MiB (16 instances)
L2 cache:                                32 MiB (16 instances)
L3 cache:                                80 MiB (1 instance)
NUMA node(s):                            1
NUMA node0 CPU(s):                       0-15
Vulnerability Gather data sampling:      Not affected
Vulnerability Indirect target selection: Not affected
Vulnerability Itlb multihit:             Not affected
Vulnerability L1tf:                      Not affected
Vulnerability Mds:                       Not affected
Vulnerability Meltdown:                  Not affected
Vulnerability Mmio stale data:           Not affected
Vulnerability Reg file data sampling:    Not affected
Vulnerability Retbleed:                  Not affected
Vulnerability Spec rstack overflow:      Not affected
Vulnerability Spec store bypass:         Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1:                Mitigation; __user pointer sanitization
Vulnerability Spectre v2:                Mitigation; CSV2, BHB
Vulnerability Srbds:                     Not affected
Vulnerability Tsa:                       Not affected
Vulnerability Tsx async abort:           Not affected
Vulnerability Vmscape:                   Not affected

Resource Usage

base (merge-base)

Metric	Value
Wall time	5.0s
Peak memory	4.9 MiB
Avg memory	829.3 KiB
CPU user	0.0s
CPU sys	0.0s
Peak spill	0 B

branch

Metric	Value
Wall time	65.0s
Peak memory	27.7 MiB
Avg memory	25.0 MiB
CPU user	60.1s
CPU sys	0.0s
Peak spill	0 B

File an issue against this benchmark runner

Rich-T-kid · 2026-11-04T00:18:07Z

@etseidl could you take a look, just benchmarks and I double checked that they run

etseidl · 2026-11-04T00:20:09Z

Sure, but it will be later this evening.

etseidl

Nvm, it was smaller than I thought. 😅

Rich-T-kid · 2026-11-04T01:09:17Z

thanks @etseidl

…2043) ## Why Closes apache#11901. The benchmark coverage for this change is in the consolidated benchmark PR apache#12054. `RowConverter` currently encodes every value in a dictionary before converting a batch. For an 8,192-row slice backed by a large dictionary, the work therefore scales with the retained values array rather than the rows being converted. ```text DictionaryArray keys: 8,192 rows --------+ +--> RowConverter --> 8,192 encoded rows values: 1,000,000 strings ---+ before: encode all 1,000,000 values --> look up 8,192 keys after: gather 8,192 row values --> encode the row-aligned values ``` This is a smaller follow-up to apache#11904. It avoids the hash map and per-row remap used there, and uses the benchmark crossover requested in [review](apache#11904 (comment)) to keep small and dense dictionaries on the existing path. ## What changed When a dictionary has at least 32K values and its values array is at least four times longer than the input, gather one value per row and encode that row-aligned array. All other dictionaries keep the existing dictionary-key lookup path. The benchmark sweep put the crossover between 20K and 24.6K values. The 32K / 4x guard is deliberately conservative. ## Results The [hosted exact-ref run](apache#12043 (comment)) is flat at 10K, 1.3x faster at 32K, 3.45x at 100K, and about 33x at 1M. Row decoding is unchanged, and peak memory drops from 56.8 MiB to 29.5 MiB. ## Test coverage One test checks row equality and round-trip decoding for a sliced dictionary with repeated high keys and nulls.

# Which issue does this PR close? Closes apache#11885 # Why is this change needed? Slicing a dictionary array only slices its keys; it still retains the full values array. When concat or interleave merges short slices from large dictionaries, the current code builds a mask and key mapping for every dictionary value, including values that no selected row references. The work therefore scales with the full dictionary instead of the number of output rows. ```text Example: one 682-row slice backed by 1,000,000 dictionary values slice keys: [900_000, 17, 900_000, ...] 682 entries | | v v retained values: [v0, ... v17, ... v900_000, ...] 1,000,000 entries old new | | allocate and scan by values.len() scan the selected keys | | 1,000,000-slot referenced mask sort + deduplicate 1,000,000-slot key mapping {17, 900_000, ...} | | +---------------+----------------+ v same merged dictionary for concat / interleave ``` # What changes are included in this PR? When a dictionary has more than 256 values per key lookup, collect only the referenced keys and store the old-to-new key mappings sparsely. The gate uses the number of remaps performed by concat or interleave, rather than the number of distinct selected rows, so mostly-null and repeatedly selected inputs stay on the dense path when hashing would be slower. Let `r` be the number of selected rows, `l` the number of key lookups, `u` the number of unique referenced keys, and `v` the number of retained dictionary values. The sparse path collects keys in `O(r)`, sorts them in `O(r log r)`, deduplicates them in `O(r)`, builds the hash mapping in `O(u)` expected time, and remaps the output in `O(l)` expected time. This replaces the dense path's `O(v)` mask and mapping work. Sorting also makes repeated keys adjacent and preserves the ascending dictionary-key order produced by the dense mask path, so both paths build the merged dictionary in the same order. Both concat and interleave use this shared dictionary merge path. BatchCoalescer has a separate dictionary merge path and is intentionally left for a follow-up. # Are these changes tested? Tests cover concat and interleave of short slices backed by large dictionaries, the key-lookup gate, sparse primitive dictionaries, null dictionary values, null keys, `UInt8` key overflow, and the existing dense path. The benchmark now covers ratios of 12, 16, 24, 128, 256, 400, and 1024 values per key with both contiguous and scattered keys. [Review measurements](apache#12008 (comment)) placed the scattered-key crossover near 256:1: at 400:1 the measured head/base ratios were 0.57-0.84, and at 1024:1 they were 0.35-0.56. The earlier [hosted run](apache#12008 (comment)) improved the 12:1, 16:1, and 24:1 dense-path cases by 10.0%, 8.5%, and 8.7%. The benchmark harness was introduced in apache#12054. # Are there any significant user-facing changes? No. This only changes how dictionary values are merged internally.

Add dictionary slice concat ratio benchmarks

17cc21f

Rich-T-kid mentioned this pull request Nov 3, 2026

Perf: Improve dictionary slice concatenation #12008

Merged

Rich-T-kid added 4 commits November 3, 2026 10:28

Narrow dictionary concat ratio sweep

73d9a65

Focus dictionary ratio sweep on cutoff

efbc5b1

Add Boolean filter kernel benchmarks

6d334f0

bench(row): cover oversized dictionaries

37da834

Rich-T-kid changed the title ~~Bench: Add dictionary concat ratio sweep~~ Bench: Cover Boolean filtering and oversized dictionaries Nov 3, 2026

github-actions Bot added the arrow-core Core crates related to representing arrow data label Nov 3, 2026

This was referenced Nov 3, 2026

benchmark: Add Boolean filter kernel benchmarks #11980

Closed

Add row dictionary slice benchmarks #12041

Closed

Rich-T-kid added the development-process Internal changes; PRs with this label are excluded from changelog label Nov 3, 2026

Rich-T-kid mentioned this pull request Nov 3, 2026

perf(arrow-row): avoid encoding oversized dictionary values #12043

Merged

Merge upstream main

1882a8a

github-actions Bot removed the development-process Internal changes; PRs with this label are excluded from changelog label Nov 3, 2026

etseidl self-requested a review November 4, 2026 00:19

etseidl approved these changes Nov 4, 2026

View reviewed changes

Rich-T-kid added this pull request to the merge queue Nov 4, 2026

Merged via the queue into apache:main with commit 258b0d2 Nov 4, 2026
21 checks passed

Jefffrey added the development-process Internal changes; PRs with this label are excluded from changelog label Nov 4, 2026

Conversation

Rich-T-kid commented Nov 3, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Which issue does this PR close?

Why is this change needed?

What changes are included in this PR?

Are these changes tested?

Are there any significant user-facing changes?

Uh oh!

Rich-T-kid commented Nov 3, 2026

Uh oh!

Rich-T-kid commented Nov 3, 2026

Uh oh!

adriangbot commented Nov 3, 2026

Uh oh!

adriangbot commented Nov 3, 2026

Uh oh!

Rich-T-kid commented Nov 3, 2026

Uh oh!

adriangbot commented Nov 3, 2026

Uh oh!

adriangbot commented Nov 3, 2026

Uh oh!

adriangbot commented Nov 3, 2026

Uh oh!

adriangbot commented Nov 3, 2026

Uh oh!

Rich-T-kid commented Nov 4, 2026

Uh oh!

etseidl commented Nov 4, 2026

Uh oh!

etseidl left a comment

Choose a reason for hiding this comment

Uh oh!

Rich-T-kid commented Nov 4, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Rich-T-kid commented Nov 3, 2026 •

edited

Loading