perf : Optimize count distinct by coderfender · Pull Request #21456 · apache/datafusion

coderfender · 2026-04-08T08:25:40Z

Which issue does this PR close?

Remove hashset based accumulators for smaller int data types and use bitmaps. Follow up of : #21453

Closes Use bitmap for count_distinct expression for u8/16 and i8/16 [perf] #21488

Rationale for this change

What changes are included in this PR?

Are these changes tested?

Are there any user-facing changes?

coderfender · 2026-04-08T08:26:42Z

benchmark results :

count_distinct i16 bitmap                      1.00      3.3±0.43µs        ? ?/sec    23.87    78.4±0.84µs        ? ?/sec
count_distinct i8 bitmap                       1.00      2.3±0.49µs        ? ?/sec    7.13     16.7±0.55µs        ? ?/sec
count_distinct u16 bitmap                      1.00      3.1±0.18µs        ? ?/sec    25.45    78.8±3.92µs        ? ?/sec
count_distinct u8 bitmap                       1.00      2.3±0.34µs        ? ?/sec    7.37     16.9±0.14µs        ? ?/sec

It seems like we are 25x faster for u16 bitmap based accumulators (or I am sleepy :) )

Dandandan · 2026-04-08T09:11:51Z

I think we can do the same for 16 bit types, it is just 65_536 bytes 8192 if we use a bitmap.

Dandandan · 2026-04-08T09:12:46Z

Oh wait, you're already doing that :)

coderfender · 2026-04-08T21:38:57Z

Query 0 in clickbench_extended dataset (which uses count distinct on u8 is now ~ 11 % faster :

┃ Query     ┃    main_cb ┃ bitmap_cb_2 ┃        Change ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ QQuery 0  │  529.46 ms │   478.99 ms │ +1.11x faster │

(Other queries are faster but I believe that is more around variance )

┏━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Query     ┃    main_cb ┃ bitmap_cb_2 ┃        Change ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ QQuery 0  │  529.46 ms │   478.99 ms │ +1.11x faster │
│ QQuery 1  │  107.43 ms │   102.59 ms │     no change │
│ QQuery 2  │  250.89 ms │   240.76 ms │     no change │
│ QQuery 3  │  207.67 ms │   207.49 ms │     no change │
│ QQuery 4  │  391.43 ms │   353.05 ms │ +1.11x faster │
│ QQuery 5  │ 4144.11 ms │  4084.08 ms │     no change │
│ QQuery 6  │  676.03 ms │   622.21 ms │ +1.09x faster │
│ QQuery 7  │  719.78 ms │   599.06 ms │ +1.20x faster │
│ QQuery 8  │  238.30 ms │   207.27 ms │ +1.15x faster │
│ QQuery 9  │ 1531.52 ms │  1406.34 ms │ +1.09x faster │
│ QQuery 10 │  435.27 ms │   403.28 ms │ +1.08x faster │
│ QQuery 11 │ 1043.68 ms │   955.22 ms │ +1.09x faster │
│ QQuery 12 │  113.16 ms │   106.31 ms │ +1.06x faster │
└───────────┴────────────┴─────────────┴───────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
┃ Benchmark Summary          ┃            ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
│ Total Time (main_cb)       │ 10388.73ms │
│ Total Time (bitmap_cb_2)   │  9766.66ms │
│ Average Time (main_cb)     │   799.13ms │
│ Average Time (bitmap_cb_2) │   751.28ms │
│ Queries Faster             │          9 │
│ Queries Slower             │          0 │
│ Queries with No Change     │          4 │
│ Queries with Failure       │          0 │
└────────────────────────────┴────────────┘

coderfender · 2026-04-08T21:41:04Z

cc : @neilconway , @alamb , @martin-g . Please take a look whenever you get a chance

alamb

This looks like a great idea. Thank you @coderfender

alamb · 2026-04-09T12:40:38Z

datafusion/functions-aggregate/Cargo.toml

 harness = false
+
+[[bench]]
+name = "count_distinct"


can you please add this benchmark as a separate PR (so we can use our standard benchmark runner to confirm the results)?

Thats sounds like a good idea . Let me put up a PR to add benchmarks to count_distinct , approx_distinct on a separate PR

Raised : #21521

alamb · 2026-04-09T12:41:56Z

datafusion/functions-aggregate-common/src/aggregate/count_distinct/native.rs

+/// Uses 256 bytes to track all possible u8 values.
+#[derive(Debug)]
+pub struct BoolArray256DistinctCountAccumulator {
+    seen: Box<[bool; 256]>,


I think you can probably use a BooleanBuffer from Arrow to make this signifcantly faster (I think [bool uses a byte for each booelan) 🤔

Great idea ! Let me try and experiment with that

For small arrays this might not be necessarily faster (as it fits in L1 cache) - perhaps only if you can use instructions like popcount etc.?

Although it probably wouldn't be very different anyway.

It seems like BooleanBuffer is immutable. I didnt have luck using [u8:8] while trying to optimize approx_distinct :
#21453 (comment) and resorted the approach which yielded good speed up (~2x) vs regression using[u8;8]

Results using MutableBuffer : (slower than directly using bitmap)

group bitmap_count_distinct bitmap_count_distinct_mutable_buffer main ----- --------------------- ------------------------------------ ---- 6.2±0.22µs ? ?/sec count_distinct i16 bitmap 1.03 3.1±0.14µs ? ?/sec 1.00 3.0±0.02µs ? ?/sec 26.16 77.9±3.43µs ? ?/sec count_distinct i64 80% distinct 1.03 48.2±0.44µs ? ?/sec 1.00 47.0±0.40µs ? ?/sec 1.02 48.0±1.49µs ? ?/sec count_distinct i64 99% distinct 1.02 48.3±0.46µs ? ?/sec 1.00 47.2±0.92µs ? ?/sec 1.02 48.3±2.14µs ? ?/sec count_distinct i8 bitmap 1.00 2.9±0.43µs ? ?/sec 1.37 3.9±0.06µs ? ?/sec 5.75 16.4±0.26µs ? ?/sec count_distinct u16 bitmap 1.06 3.2±0.30µs ? ?/sec 1.00 3.0±0.16µs ? ?/sec 25.58 77.7±1.82µs ? ?/sec count_distinct u8 bitmap 1.00 2.1±0.03µs ? ?/sec 1.90 3.9±0.08µs ? ?/sec 8.14

alamb · 2026-04-09T20:25:40Z

run benchmark count_distinct

alamb · 2026-04-09T20:26:12Z

run benchmark clickbench

adriangbot · 2026-04-09T20:28:17Z

🤖 Criterion benchmark running (GKE) | trigger
Instance: c4a-highmem-16 (12 vCPU / 65 GiB) | Linux bench-c4217267905-1036-wmjln 6.12.55+ #1 SMP Sun Feb 1 08:59:41 UTC 2026 aarch64 GNU/Linux

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

Comparing optimize_count_distinct (a13bcaa) to fbdf770 (merge-base) diff
BENCH_NAME=count_distinct
BENCH_COMMAND=cargo bench --features=parquet --bench count_distinct
BENCH_FILTER=
Results will be posted here when complete

File an issue against this benchmark runner

adriangbot · 2026-04-09T20:28:18Z

Benchmark for this request failed.

Last 20 lines of output:

Click to expand

rustc 1.94.1 (e408947bf 2026-03-25)
a13bcaad5b82e69257e82b741cf72619da838990
fbdf7703a96408b4eba27801431be8bf468734d8
error: failed to load manifest for workspace member `/workspace/datafusion-branch/datafusion/catalog`
referenced by workspace at `/workspace/datafusion-branch/Cargo.toml`

Caused by:
  failed to load manifest for dependency `datafusion-datasource`

Caused by:
  failed to load manifest for dependency `datafusion-physical-plan`

Caused by:
  failed to load manifest for dependency `datafusion-functions-aggregate`

Caused by:
  failed to parse manifest at `/workspace/datafusion-branch/datafusion/functions-aggregate/Cargo.toml`

Caused by:
  found duplicate bench name count_distinct, but all bench targets must have a unique name

File an issue against this benchmark runner

adriangbot · 2026-04-09T20:28:35Z

🤖 Criterion benchmark running (GKE) | trigger
Instance: c4a-highmem-16 (12 vCPU / 65 GiB) | Linux bench-c4217270691-1037-8wdsm 6.12.55+ #1 SMP Sun Feb 1 08:59:41 UTC 2026 aarch64 GNU/Linux

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

Comparing optimize_count_distinct (a13bcaa) to fbdf770 (merge-base) diff
BENCH_NAME=clickbench
BENCH_COMMAND=cargo bench --features=parquet --bench clickbench
BENCH_FILTER=
Results will be posted here when complete

File an issue against this benchmark runner

adriangbot · 2026-04-09T20:28:36Z

Benchmark for this request failed.

Last 20 lines of output:

Click to expand

rustc 1.94.1 (e408947bf 2026-03-25)
a13bcaad5b82e69257e82b741cf72619da838990
fbdf7703a96408b4eba27801431be8bf468734d8
error: failed to load manifest for workspace member `/workspace/datafusion-branch/datafusion/catalog`
referenced by workspace at `/workspace/datafusion-branch/Cargo.toml`

Caused by:
  failed to load manifest for dependency `datafusion-datasource`

Caused by:
  failed to load manifest for dependency `datafusion-physical-plan`

Caused by:
  failed to load manifest for dependency `datafusion-functions-aggregate`

Caused by:
  failed to parse manifest at `/workspace/datafusion-branch/datafusion/functions-aggregate/Cargo.toml`

Caused by:
  found duplicate bench name count_distinct, but all bench targets must have a unique name

File an issue against this benchmark runner

coderfender · 2026-04-09T21:34:04Z

Rebased with main

coderfender · 2026-04-09T21:57:53Z

Update benchmarks after rebase with main

Command :

cargo bench -p datafusion-functions-aggregate --bench count_distinct

group                                          bitmap_count_distinct                  main
-----                                          ---------------------                  ----
count_distinct i16 bitmap                      1.00      3.1±0.14µs        ? ?/sec    25.35    77.9±3.43µs        ? ?/sec
count_distinct i64 80% distinct                1.00     48.2±0.44µs        ? ?/sec    1.00     48.0±1.49µs        ? ?/sec
count_distinct i64 99% distinct                1.00     48.3±0.46µs        ? ?/sec    1.00     48.3±2.14µs        ? ?/sec
count_distinct i8 bitmap                       1.00      2.9±0.43µs        ? ?/sec    5.75     16.4±0.26µs        ? ?/sec
count_distinct u16 bitmap                      1.00      3.2±0.30µs        ? ?/sec    24.08    77.7±1.82µs        ? ?/sec
count_distinct u8 bitmap                       1.00      2.1±0.03µs        ? ?/sec    8.14     16.8±0.05µs        ? ?/sec

github-actions bot added the functions Changes to functions implementation label Apr 8, 2026

This comment has been minimized.

Sign in to view

alamb added the performance Make DataFusion faster label Apr 9, 2026

alamb reviewed Apr 9, 2026

View reviewed changes

This comment has been minimized.

Sign in to view

alamb reviewed Apr 9, 2026

View reviewed changes

This comment has been minimized.

Sign in to view

coderfender mentioned this pull request Apr 9, 2026

chore: create benches small ints for count_distinct #21521

Merged

coderfender force-pushed the optimize_count_distinct branch from 93acd98 to 7e67e2e Compare April 9, 2026 18:53

coderfender added 7 commits April 9, 2026 14:33

bitmap_smaller_datatypes

8d49dfe

bitmap_smaller_datatypes

0b179ff

bitmap_instead_of_hll_smaller_datatypes

c6095ab

bitmap_instead_of_hll_smaller_datatypes

9d06408

bitmap_instead_of_hll_smaller_datatypes

f185fdc

bitmap_instead_smaller_datatypes

f7c487a

bitmap_instead_smaller_datatypes

3f091d9

coderfender added 2 commits April 9, 2026 14:33

bitmap_instead_smaller_datatypes

e09f68a

cleanup_pr

f8c01a1

coderfender force-pushed the optimize_count_distinct branch from a13bcaa to f8c01a1 Compare April 9, 2026 21:33

rebase_main

df90ef5

coderfender force-pushed the optimize_count_distinct branch from 3db92e3 to df90ef5 Compare April 9, 2026 21:38

Merge branch 'main' into optimize_count_distinct

bf0f95c

Conversation

coderfender commented Apr 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Which issue does this PR close?

Rationale for this change

What changes are included in this PR?

Are these changes tested?

Are there any user-facing changes?

Uh oh!

coderfender commented Apr 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Dandandan commented Apr 8, 2026

Uh oh!

Dandandan commented Apr 8, 2026

Uh oh!

coderfender commented Apr 8, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

coderfender commented Apr 8, 2026

Uh oh!

This comment has been minimized.

alamb left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

This comment has been minimized.

This comment has been minimized.

This comment has been minimized.

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

coderfender Apr 9, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

This comment has been minimized.

This comment has been minimized.

alamb commented Apr 9, 2026

Uh oh!

alamb commented Apr 9, 2026

Uh oh!

adriangbot commented Apr 9, 2026

Uh oh!

adriangbot commented Apr 9, 2026

Uh oh!

adriangbot commented Apr 9, 2026

Uh oh!

adriangbot commented Apr 9, 2026

Uh oh!

coderfender commented Apr 9, 2026

Uh oh!

coderfender commented Apr 9, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

coderfender commented Apr 8, 2026 •

edited

Loading

coderfender commented Apr 8, 2026 •

edited

Loading

coderfender commented Apr 8, 2026 •

edited

Loading

coderfender Apr 9, 2026 •

edited

Loading