Skip to content

Draft Poc for Unified select (Enum for bitmap and range) #7454

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Draft
wants to merge 6 commits into
base: main
Choose a base branch
from
Draft

Draft Poc for Unified select (Enum for bitmap and range) #7454

wants to merge 6 commits into from

Conversation

zhuqi-lucas
Copy link
Contributor

@zhuqi-lucas zhuqi-lucas commented Apr 29, 2025 •
edited by alamb
Loading

Which issue does this PR close?

Rationale for this change

What changes are included in this PR?

Are there any user-facing changes?

@zhuqi-lucas zhuqi-lucas marked this pull request as draft April 29, 2025 04:39
@github-actions github-actions bot added the parquet Changes to the parquet crate label Apr 29, 2025
@zhuqi-lucas zhuqi-lucas mentioned this pull request Apr 29, 2025
Open
5 tasks
@alamb alamb mentioned this pull request Apr 29, 2025
Closed
26 tasks
@alamb
Copy link
Contributor

alamb commented Apr 29, 2025

🤖 ./gh_compare_arrow.sh Benchmark Script Running
Linux aal-dev 6.11.0-1013-gcp #13~24.04.1-Ubuntu SMP Wed Apr 2 16:34:16 UTC 2025 x86_64 x86_64 x86_64 GNU/Linux
Comparing unified_select (e5aad7c) to 5f0aed6 diff
BENCH_NAME=arrow_reader_row_filter
BENCH_COMMAND=cargo bench --all-features --bench arrow_reader_row_filter
BENCH_FILTER=
BENCH_BRANCH_NAME=unified_select
Results will be posted here when complete

@alamb
Copy link
Contributor

alamb commented Apr 29, 2025

I kicked off running the benchmarks on this PR

@alamb
Copy link
Contributor

alamb commented Apr 29, 2025

🤖: Benchmark completed

Details 

group                                                                                 main                                   unified_select
-----                                                                                 ----                                   --------------
arrow_reader_row_filter/Composite/all_columns/async                                   1.00      2.9±0.01ms        ? ?/sec    1.00      2.9±0.03ms        ? ?/sec
arrow_reader_row_filter/Composite/all_columns/sync                                    1.00      3.2±0.02ms        ? ?/sec    1.00      3.2±0.03ms        ? ?/sec
arrow_reader_row_filter/Composite/exclude_filter_column/async                         1.00      2.5±0.01ms        ? ?/sec    1.00      2.5±0.01ms        ? ?/sec
arrow_reader_row_filter/Composite/exclude_filter_column/sync                          1.00      2.8±0.02ms        ? ?/sec    1.00      2.8±0.02ms        ? ?/sec
arrow_reader_row_filter/ModeratelySelectiveClustered/all_columns/async                1.00      3.0±0.01ms        ? ?/sec    1.00      3.0±0.01ms        ? ?/sec
arrow_reader_row_filter/ModeratelySelectiveClustered/all_columns/sync                 1.00      3.2±0.02ms        ? ?/sec    1.00      3.2±0.02ms        ? ?/sec
arrow_reader_row_filter/ModeratelySelectiveClustered/exclude_filter_column/async      1.00      2.8±0.01ms        ? ?/sec    1.00      2.8±0.02ms        ? ?/sec
arrow_reader_row_filter/ModeratelySelectiveClustered/exclude_filter_column/sync       1.01      3.1±0.01ms        ? ?/sec    1.00      3.1±0.01ms        ? ?/sec
arrow_reader_row_filter/ModeratelySelectiveUnclustered/all_columns/async              1.00      6.5±0.03ms        ? ?/sec    1.40      9.1±0.12ms        ? ?/sec
arrow_reader_row_filter/ModeratelySelectiveUnclustered/all_columns/sync               1.00      6.6±0.03ms        ? ?/sec    1.00      6.6±0.04ms        ? ?/sec
arrow_reader_row_filter/ModeratelySelectiveUnclustered/exclude_filter_column/async    1.00      5.6±0.03ms        ? ?/sec    1.58      8.9±0.07ms        ? ?/sec
arrow_reader_row_filter/ModeratelySelectiveUnclustered/exclude_filter_column/sync     1.00      5.7±0.03ms        ? ?/sec    1.00      5.7±0.03ms        ? ?/sec
arrow_reader_row_filter/PointLookup/all_columns/async                                 1.00      2.0±0.01ms        ? ?/sec    1.00      2.0±0.01ms        ? ?/sec
arrow_reader_row_filter/PointLookup/all_columns/sync                                  1.00      2.2±0.01ms        ? ?/sec    1.01      2.2±0.01ms        ? ?/sec
arrow_reader_row_filter/PointLookup/exclude_filter_column/async                       1.00  1984.8±12.54µs        ? ?/sec    1.01      2.0±0.01ms        ? ?/sec
arrow_reader_row_filter/PointLookup/exclude_filter_column/sync                        1.00      2.2±0.02ms        ? ?/sec    1.01      2.2±0.01ms        ? ?/sec
arrow_reader_row_filter/SelectiveUnclustered/all_columns/async                        1.00      3.1±0.02ms        ? ?/sec    2.80      8.7±0.07ms        ? ?/sec
arrow_reader_row_filter/SelectiveUnclustered/all_columns/sync                         1.00      3.4±0.02ms        ? ?/sec    1.00      3.4±0.02ms        ? ?/sec
arrow_reader_row_filter/SelectiveUnclustered/exclude_filter_column/async              1.00      2.7±0.01ms        ? ?/sec    2.98      8.2±0.05ms        ? ?/sec
arrow_reader_row_filter/SelectiveUnclustered/exclude_filter_column/sync               1.00      3.0±0.01ms        ? ?/sec    1.01      3.0±0.03ms        ? ?/sec
arrow_reader_row_filter/UnselectiveClustered/all_columns/async                        1.03      7.9±0.03ms        ? ?/sec    1.00      7.7±0.04ms        ? ?/sec
arrow_reader_row_filter/UnselectiveClustered/all_columns/sync                         1.03      8.2±0.04ms        ? ?/sec    1.00      8.0±0.05ms        ? ?/sec
arrow_reader_row_filter/UnselectiveClustered/exclude_filter_column/async              1.03      7.7±0.05ms        ? ?/sec    1.00      7.4±0.05ms        ? ?/sec
arrow_reader_row_filter/UnselectiveClustered/exclude_filter_column/sync               1.03      7.9±0.03ms        ? ?/sec    1.00      7.6±0.03ms        ? ?/sec
arrow_reader_row_filter/UnselectiveUnclustered/all_columns/async                      1.00      3.1±0.02ms        ? ?/sec    2.80      8.7±0.09ms        ? ?/sec
arrow_reader_row_filter/UnselectiveUnclustered/all_columns/sync                       1.00      3.4±0.02ms        ? ?/sec    1.00      3.4±0.02ms        ? ?/sec
arrow_reader_row_filter/UnselectiveUnclustered/exclude_filter_column/async            1.00      2.7±0.01ms        ? ?/sec    2.98      8.2±0.06ms        ? ?/sec
arrow_reader_row_filter/UnselectiveUnclustered/exclude_filter_column/sync             1.00      3.0±0.02ms        ? ?/sec    1.00      3.0±0.02ms        ? ?/sec
arrow_reader_row_filter/Utf8ViewNonEmpty/all_columns/async                            1.13     23.7±0.09ms        ? ?/sec    1.00     20.9±0.26ms        ? ?/sec
arrow_reader_row_filter/Utf8ViewNonEmpty/all_columns/sync                             1.02     23.9±0.12ms        ? ?/sec    1.00     23.5±0.10ms        ? ?/sec
arrow_reader_row_filter/Utf8ViewNonEmpty/exclude_filter_column/async                  1.57     15.0±0.05ms        ? ?/sec    1.00      9.6±0.06ms        ? ?/sec
arrow_reader_row_filter/Utf8ViewNonEmpty/exclude_filter_column/sync                   1.02     15.1±0.04ms        ? ?/sec    1.00     14.9±0.07ms        ? ?/sec

@zhuqi-lucas
Copy link
Contributor Author

Thank you @alamb , from the benchmark it seems this PR only improve the Utf8ViewNonEmpty cases, and regression for some cases.

@zhuqi-lucas
Copy link
Contributor Author

But from clickbench result, it seems performance better than the original default push down:

./bench.sh compare  older_push_down test_default_parquet_push_down
Comparing older_push_down and test_default_parquet_push_down
--------------------
Benchmark clickbench_1.json
--------------------
┏━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Query        ┃ older_push_down ┃ test_default_parquet_push_down ┃        Change ┃
┡━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ QQuery 0     │          0.29ms │                         0.40ms │  1.39x slower │
│ QQuery 1     │         50.39ms │                        46.59ms │ +1.08x faster │
│ QQuery 2     │         77.56ms │                        75.08ms │     no change │
│ QQuery 3     │         77.95ms │                        79.83ms │     no change │
│ QQuery 4     │        579.45ms │                       504.25ms │ +1.15x faster │
│ QQuery 5     │        560.83ms │                       541.37ms │     no change │
│ QQuery 6     │          0.33ms │                         0.33ms │     no change │
│ QQuery 7     │         63.95ms │                        56.08ms │ +1.14x faster │
│ QQuery 8     │        718.15ms │                       663.54ms │ +1.08x faster │
│ QQuery 9     │        786.44ms │                       751.61ms │     no change │
│ QQuery 10    │        206.47ms │                       214.36ms │     no change │
│ QQuery 11    │        213.81ms │                       239.89ms │  1.12x slower │
│ QQuery 12    │        679.51ms │                       705.69ms │     no change │
│ QQuery 13    │       1002.85ms │                       868.43ms │ +1.15x faster │
│ QQuery 14    │        763.19ms │                       689.47ms │ +1.11x faster │
│ QQuery 15    │        649.33ms │                       612.41ms │ +1.06x faster │
│ QQuery 16    │       1390.12ms │                      1339.99ms │     no change │
│ QQuery 17    │       1219.49ms │                      1134.36ms │ +1.08x faster │
│ QQuery 18    │       2932.03ms │                      2566.26ms │ +1.14x faster │
│ QQuery 19    │         63.44ms │                        58.59ms │ +1.08x faster │
│ QQuery 20    │        752.05ms │                       679.10ms │ +1.11x faster │
│ QQuery 21    │        927.06ms │                       811.46ms │ +1.14x faster │
│ QQuery 22    │       1558.98ms │                      1461.95ms │ +1.07x faster │
│ QQuery 23    │       3402.25ms │                      2604.51ms │ +1.31x faster │
│ QQuery 24    │        471.04ms │                       445.48ms │ +1.06x faster │
│ QQuery 25    │        420.43ms │                       423.96ms │     no change │
│ QQuery 26    │        522.57ms │                       482.58ms │ +1.08x faster │
│ QQuery 27    │       1375.58ms │                      1304.33ms │ +1.05x faster │
│ QQuery 28    │       8767.85ms │                      8484.28ms │     no change │
│ QQuery 29    │        458.33ms │                       446.29ms │     no change │
│ QQuery 30    │        689.41ms │                       450.94ms │ +1.53x faster │
│ QQuery 31    │        758.38ms │                       574.79ms │ +1.32x faster │
│ QQuery 32    │       2981.93ms │                      2339.84ms │ +1.27x faster │
│ QQuery 33    │       2886.54ms │                      2587.16ms │ +1.12x faster │
│ QQuery 34    │       3298.54ms │                      2760.24ms │ +1.20x faster │
│ QQuery 35    │        834.81ms │                       893.25ms │  1.07x slower │
│ QQuery 36    │         40.41ms │                        36.52ms │ +1.11x faster │
│ QQuery 37    │         35.59ms │                        33.71ms │ +1.06x faster │
│ QQuery 38    │         39.18ms │                        36.37ms │ +1.08x faster │
│ QQuery 39    │         35.07ms │                        36.88ms │  1.05x slower │
│ QQuery 40    │         36.55ms │                        34.98ms │     no change │
│ QQuery 41    │         36.18ms │                        36.72ms │     no change │
│ QQuery 42    │         37.00ms │                        34.64ms │ +1.07x faster │
└──────────────┴─────────────────┴────────────────────────────────┴───────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
┃ Benchmark Summary                             ┃            ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
│ Total Time (older_push_down)                  │ 42401.30ms │
│ Total Time (test_default_parquet_push_down)   │ 38148.50ms │
│ Average Time (older_push_down)                │   986.08ms │
│ Average Time (test_default_parquet_push_down) │   887.17ms │
│ Queries Faster                                │         26 │
│ Queries Slower                                │          4 │
│ Queries with No Change                        │         13 │
└───────────────────────────────────────────────┴────────────┘
--------------------
Benchmark clickbench_partitioned.json
--------------------
┏━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Query        ┃ older_push_down ┃ test_default_parquet_push_down ┃        Change ┃
┡━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ QQuery 0     │          1.12ms │                         1.10ms │     no change │
│ QQuery 1     │         25.44ms │                        25.93ms │     no change │
│ QQuery 2     │         54.65ms │                        55.48ms │     no change │
│ QQuery 3     │         56.82ms │                        61.02ms │  1.07x slower │
│ QQuery 4     │        484.52ms │                       470.33ms │     no change │
│ QQuery 5     │        524.38ms │                       546.59ms │     no change │
│ QQuery 6     │          1.24ms │                         1.16ms │ +1.07x faster │
│ QQuery 7     │         43.00ms │                        38.75ms │ +1.11x faster │
│ QQuery 8     │        632.15ms │                       593.78ms │ +1.06x faster │
│ QQuery 9     │        658.82ms │                       680.13ms │     no change │
│ QQuery 10    │        154.14ms │                       184.81ms │  1.20x slower │
│ QQuery 11    │        183.86ms │                       198.64ms │  1.08x slower │
│ QQuery 12    │        664.23ms │                       678.02ms │     no change │
│ QQuery 13    │        916.31ms │                       771.84ms │ +1.19x faster │
│ QQuery 14    │        717.15ms │                       637.99ms │ +1.12x faster │
│ QQuery 15    │        561.85ms │                       562.13ms │     no change │
│ QQuery 16    │       1430.77ms │                      1346.40ms │ +1.06x faster │
│ QQuery 17    │       1249.19ms │                      1117.41ms │ +1.12x faster │
│ QQuery 18    │       2739.07ms │                      2427.62ms │ +1.13x faster │
│ QQuery 19    │         41.63ms │                        45.73ms │  1.10x slower │
│ QQuery 20    │        724.72ms │                       694.00ms │     no change │
│ QQuery 21    │        819.39ms │                       790.52ms │     no change │
│ QQuery 22    │       1551.09ms │                      1517.14ms │     no change │
│ QQuery 23    │       3565.46ms │                      2681.59ms │ +1.33x faster │
│ QQuery 24    │        457.95ms │                       386.72ms │ +1.18x faster │
│ QQuery 25    │        381.44ms │                       370.83ms │     no change │
│ QQuery 26    │        526.80ms │                       435.73ms │ +1.21x faster │
│ QQuery 27    │       1415.36ms │                      1354.63ms │     no change │
│ QQuery 28    │       8528.35ms │                      9813.62ms │  1.15x slower │
│ QQuery 29    │        401.68ms │                       396.41ms │     no change │
│ QQuery 30    │        728.25ms │                       431.94ms │ +1.69x faster │
│ QQuery 31    │        744.53ms │                       528.87ms │ +1.41x faster │
│ QQuery 32    │       3257.23ms │                      2323.61ms │ +1.40x faster │
│ QQuery 33    │       3187.75ms │                      2517.60ms │ +1.27x faster │
│ QQuery 34    │       3735.45ms │                      2840.70ms │ +1.31x faster │
│ QQuery 35    │        877.66ms │                       728.25ms │ +1.21x faster │
│ QQuery 36    │         26.23ms │                        22.64ms │ +1.16x faster │
│ QQuery 37    │         23.04ms │                        21.68ms │ +1.06x faster │
│ QQuery 38    │         21.55ms │                        21.89ms │     no change │
│ QQuery 39    │         22.02ms │                        22.69ms │     no change │
│ QQuery 40    │         21.54ms │                        21.98ms │     no change │
│ QQuery 41    │         21.64ms │                        21.95ms │     no change │
│ QQuery 42    │         23.77ms │                        22.51ms │ +1.06x faster │
└──────────────┴─────────────────┴────────────────────────────────┴───────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
┃ Benchmark Summary                             ┃            ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
│ Total Time (older_push_down)                  │ 42203.20ms │
│ Total Time (test_default_parquet_push_down)   │ 38412.36ms │
│ Average Time (older_push_down)                │   981.47ms │
│ Average Time (test_default_parquet_push_down) │   893.31ms │
│ Queries Faster                                │         20 │
│ Queries Slower                                │          5 │
│ Queries with No Change                        │         18 │
└───────────────────────────────────────────────┴────────────┘

@zhuqi-lucas
Copy link
Contributor Author

And here is the result compared with main(No push down), still some regression:

./bench.sh compare main test_default_parquet_push_down
Comparing main and test_default_parquet_push_down
--------------------
Benchmark clickbench_1.json
--------------------
┏━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Query        ┃      main ┃ test_default_parquet_push_down ┃        Change ┃
┡━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ QQuery 0     │    0.32ms │                         0.40ms │  1.24x slower │
│ QQuery 1     │   46.96ms │                        46.59ms │     no change │
│ QQuery 2     │   75.59ms │                        75.08ms │     no change │
│ QQuery 3     │   74.12ms │                        79.83ms │  1.08x slower │
│ QQuery 4     │  556.03ms │                       504.25ms │ +1.10x faster │
│ QQuery 5     │  563.52ms │                       541.37ms │     no change │
│ QQuery 6     │    0.31ms │                         0.33ms │  1.06x slower │
│ QQuery 7     │   52.23ms │                        56.08ms │  1.07x slower │
│ QQuery 8     │  720.15ms │                       663.54ms │ +1.09x faster │
│ QQuery 9     │  741.10ms │                       751.61ms │     no change │
│ QQuery 10    │  171.95ms │                       214.36ms │  1.25x slower │
│ QQuery 11    │  187.66ms │                       239.89ms │  1.28x slower │
│ QQuery 12    │  597.16ms │                       705.69ms │  1.18x slower │
│ QQuery 13    │  877.71ms │                       868.43ms │     no change │
│ QQuery 14    │  605.11ms │                       689.47ms │  1.14x slower │
│ QQuery 15    │  630.66ms │                       612.41ms │     no change │
│ QQuery 16    │ 1422.47ms │                      1339.99ms │ +1.06x faster │
│ QQuery 17    │ 1221.90ms │                      1134.36ms │ +1.08x faster │
│ QQuery 18    │ 2773.23ms │                      2566.26ms │ +1.08x faster │
│ QQuery 19    │   66.30ms │                        58.59ms │ +1.13x faster │
│ QQuery 20    │  682.62ms │                       679.10ms │     no change │
│ QQuery 21    │  800.86ms │                       811.46ms │     no change │
│ QQuery 22    │ 1521.09ms │                      1461.95ms │     no change │
│ QQuery 23    │ 4223.95ms │                      2604.51ms │ +1.62x faster │
│ QQuery 24    │  286.83ms │                       445.48ms │  1.55x slower │
│ QQuery 25    │  274.47ms │                       423.96ms │  1.54x slower │
│ QQuery 26    │  320.45ms │                       482.58ms │  1.51x slower │
│ QQuery 27    │  945.72ms │                      1304.33ms │  1.38x slower │
│ QQuery 28    │ 8206.32ms │                      8484.28ms │     no change │
│ QQuery 29    │  459.59ms │                       446.29ms │     no change │
│ QQuery 30    │  493.35ms │                       450.94ms │ +1.09x faster │
│ QQuery 31    │  585.82ms │                       574.79ms │     no change │
│ QQuery 32    │ 2436.43ms │                      2339.84ms │     no change │
│ QQuery 33    │ 2916.52ms │                      2587.16ms │ +1.13x faster │
│ QQuery 34    │ 2975.16ms │                      2760.24ms │ +1.08x faster │
│ QQuery 35    │  866.75ms │                       893.25ms │     no change │
│ QQuery 36    │  104.22ms │                        36.52ms │ +2.85x faster │
│ QQuery 37    │   62.50ms │                        33.71ms │ +1.85x faster │
│ QQuery 38    │  107.57ms │                        36.37ms │ +2.96x faster │
│ QQuery 39    │  167.64ms │                        36.88ms │ +4.55x faster │
│ QQuery 40    │   46.49ms │                        34.98ms │ +1.33x faster │
│ QQuery 41    │   45.49ms │                        36.72ms │ +1.24x faster │
│ QQuery 42    │   42.51ms │                        34.64ms │ +1.23x faster │
└──────────────┴───────────┴────────────────────────────────┴───────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
┃ Benchmark Summary                             ┃            ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
│ Total Time (main)                             │ 39956.84ms │
│ Total Time (test_default_parquet_push_down)   │ 38148.50ms │
│ Average Time (main)                           │   929.23ms │
│ Average Time (test_default_parquet_push_down) │   887.17ms │
│ Queries Faster                                │         17 │
│ Queries Slower                                │         12 │
│ Queries with No Change                        │         14 │
└───────────────────────────────────────────────┴────────────┘
--------------------
Benchmark clickbench_partitioned.json
--------------------
┏━━━━━━━━━━━━━━┳━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━┓
┃ Query        ┃      main ┃ test_default_parquet_push_down ┃        Change ┃
┡━━━━━━━━━━━━━━╇━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━┩
│ QQuery 0     │    1.10ms │                         1.10ms │     no change │
│ QQuery 1     │   24.27ms │                        25.93ms │  1.07x slower │
│ QQuery 2     │   58.03ms │                        55.48ms │     no change │
│ QQuery 3     │   58.91ms │                        61.02ms │     no change │
│ QQuery 4     │  478.56ms │                       470.33ms │     no change │
│ QQuery 5     │  549.38ms │                       546.59ms │     no change │
│ QQuery 6     │    1.18ms │                         1.16ms │     no change │
│ QQuery 7     │   40.05ms │                        38.75ms │     no change │
│ QQuery 8     │  645.79ms │                       593.78ms │ +1.09x faster │
│ QQuery 9     │  671.10ms │                       680.13ms │     no change │
│ QQuery 10    │  133.89ms │                       184.81ms │  1.38x slower │
│ QQuery 11    │  159.81ms │                       198.64ms │  1.24x slower │
│ QQuery 12    │  561.74ms │                       678.02ms │  1.21x slower │
│ QQuery 13    │  750.14ms │                       771.84ms │     no change │
│ QQuery 14    │  525.03ms │                       637.99ms │  1.22x slower │
│ QQuery 15    │  553.88ms │                       562.13ms │     no change │
│ QQuery 16    │ 1417.77ms │                      1346.40ms │ +1.05x faster │
│ QQuery 17    │ 1104.70ms │                      1117.41ms │     no change │
│ QQuery 18    │ 3037.46ms │                      2427.62ms │ +1.25x faster │
│ QQuery 19    │   45.81ms │                        45.73ms │     no change │
│ QQuery 20    │  733.71ms │                       694.00ms │ +1.06x faster │
│ QQuery 21    │  789.70ms │                       790.52ms │     no change │
│ QQuery 22    │ 1299.84ms │                      1517.14ms │  1.17x slower │
│ QQuery 23    │ 3952.24ms │                      2681.59ms │ +1.47x faster │
│ QQuery 24    │  273.40ms │                       386.72ms │  1.41x slower │
│ QQuery 25    │  274.14ms │                       370.83ms │  1.35x slower │
│ QQuery 26    │  320.12ms │                       435.73ms │  1.36x slower │
│ QQuery 27    │  900.06ms │                      1354.63ms │  1.51x slower │
│ QQuery 28    │ 7812.82ms │                      9813.62ms │  1.26x slower │
│ QQuery 29    │  390.07ms │                       396.41ms │     no change │
│ QQuery 30    │  420.68ms │                       431.94ms │     no change │
│ QQuery 31    │  571.58ms │                       528.87ms │ +1.08x faster │
│ QQuery 32    │ 2585.80ms │                      2323.61ms │ +1.11x faster │
│ QQuery 33    │ 2621.59ms │                      2517.60ms │     no change │
│ QQuery 34    │ 3144.83ms │                      2840.70ms │ +1.11x faster │
│ QQuery 35    │  855.71ms │                       728.25ms │ +1.18x faster │
│ QQuery 36    │   80.10ms │                        22.64ms │ +3.54x faster │
│ QQuery 37    │   35.71ms │                        21.68ms │ +1.65x faster │
│ QQuery 38    │   79.06ms │                        21.89ms │ +3.61x faster │
│ QQuery 39    │  123.78ms │                        22.69ms │ +5.46x faster │
│ QQuery 40    │   29.03ms │                        21.98ms │ +1.32x faster │
│ QQuery 41    │   27.67ms │                        21.95ms │ +1.26x faster │
│ QQuery 42    │   25.98ms │                        22.51ms │ +1.15x faster │
└──────────────┴───────────┴────────────────────────────────┴───────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┓
┃ Benchmark Summary                             ┃            ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━┩
│ Total Time (main)                             │ 38166.21ms │
│ Total Time (test_default_parquet_push_down)   │ 38412.36ms │
│ Average Time (main)                           │   887.59ms │
│ Average Time (test_default_parquet_push_down) │   893.31ms │
│ Queries Faster                                │         16 │
│ Queries Slower                                │         11 │
│ Queries with No Change                        │         16 │
└───────────────────────────────────────────────┴────────────┘

@zhuqi-lucas
Copy link
Contributor Author

zhuqi-lucas commented Apr 30, 2025 •
edited
Loading

It helps part of the regression about the read record/skip record too dense, which is the original regression:

Here is the result for page cache without this PR:
#7363 (comment)
The regression will from Q24->28 and Q30 -> Q31.

Q30 / Q31 no regression now for current PR:

│ QQuery 30    │  420.68ms │                       431.94ms │     no change │
│ QQuery 31    │  571.58ms │                       528.87ms │ +1.08x faster │

But Q24 -> Q 28 still have regression, same with original result:

│ QQuery 24    │  273.40ms │                       386.72ms │  1.41x slower │
│ QQuery 25    │  274.14ms │                       370.83ms │  1.35x slower │
│ QQuery 26    │  320.12ms │                       435.73ms │  1.36x slower │
│ QQuery 27    │  900.06ms │                      1354.63ms │  1.51x slower │
│ QQuery 28    │ 7812.82ms │                      9813.62ms │  1.26x slower │

@zhuqi-lucas
Copy link
Contributor Author

zhuqi-lucas commented Apr 30, 2025 •
edited
Loading

I take Q27 for example, it looks like the regression still comes from decompress, even the page cache case, it still use more decompress time, which cause the regression. We need to investigate more about the regression.

  1. The no pushdown case:
    flamegraphFast

  2. The page cache case:

flamegraphPageCache

  1. Current PR case:
    flamegraphSlow

@alamb
Copy link
Contributor

alamb commented Apr 30, 2025

BTW there is a benchmark for RowSelection: https://github.com/apache/arrow-rs/blob/a2cc42639b4ad5579d052e6f2317a413e2407e0f/parquet/benches/row_selector.rs#L1-L0

@alamb
Copy link
Contributor

alamb commented Apr 30, 2025 •
edited
Loading

It helps part of the regression about the read record/skip record too dense, which is the original regression:

Here is the result for page cache without this PR: #7363 (comment) The regression will from Q24->28 and Q30 -> Q31.

Q30 / Q31 no regression now for current PR:

│ QQuery 30    │  420.68ms │                       431.94ms │     no change │
│ QQuery 31    │  571.58ms │                       528.87ms │ +1.08x faster │

https://github.com/apache/datafusion/blob/7b370e26fea75fcd17121272eec1bd9447b2cb8f/benchmarks/queries/clickbench/queries.sql#L31-L32

These queries have a predicate like

WHERE "SearchPhrase" <> ''

But SearchPhrase is not used except for filtering (aka it is not in the projection)

For example

SELECT "SearchEngineID", "ClientIP", COUNT(*) AS c, SUM("IsRefresh"), AVG("ResolutionWidth") FROM hits WHERE "SearchPhrase" <> '' GROUP BY "SearchEngineID", "ClientIP" ORDER BY c DESC LIMIT 10;

But Q24 -> Q 28 still have regression, same with original result:

│ QQuery 24    │  273.40ms │                       386.72ms │  1.41x slower │
│ QQuery 25    │  274.14ms │                       370.83ms │  1.35x slower │
│ QQuery 26    │  320.12ms │                       435.73ms │  1.36x slower │
│ QQuery 27    │  900.06ms │                      1354.63ms │  1.51x slower │
│ QQuery 28    │ 7812.82ms │                      9813.62ms │  1.26x slower │

https://github.com/apache/datafusion/blob/7b370e26fea75fcd17121272eec1bd9447b2cb8f/benchmarks/queries/clickbench/queries.sql#L25-L29

These queries have the same predicate

WHERE "SearchPhrase" <> ''

But in this case SearchPhrase is also used in the rest of the query (and thus the projection)

For example

SELECT "SearchPhrase" FROM hits WHERE "SearchPhrase" <> '' ORDER BY "EventTime" LIMIT 10;

@zhuqi-lucas
Copy link
Contributor Author

It helps part of the regression about the read record/skip record too dense, which is the original regression:
Here is the result for page cache without this PR: #7363 (comment) The regression will from Q24->28 and Q30 -> Q31.
Q30 / Q31 no regression now for current PR:

│ QQuery 30    │  420.68ms │                       431.94ms │     no change │
│ QQuery 31    │  571.58ms │                       528.87ms │ +1.08x faster │

https://github.com/apache/datafusion/blob/7b370e26fea75fcd17121272eec1bd9447b2cb8f/benchmarks/queries/clickbench/queries.sql#L31-L32

These queries have a predicate like

WHERE "SearchPhrase" <> ''

But SearchPhrase is not used except for filtering (aka it is not in the projection)

For example

SELECT "SearchEngineID", "ClientIP", COUNT(*) AS c, SUM("IsRefresh"), AVG("ResolutionWidth") FROM hits WHERE "SearchPhrase" <> '' GROUP BY "SearchEngineID", "ClientIP" ORDER BY c DESC LIMIT 10;

But Q24 -> Q 28 still have regression, same with original result:

│ QQuery 24    │  273.40ms │                       386.72ms │  1.41x slower │
│ QQuery 25    │  274.14ms │                       370.83ms │  1.35x slower │
│ QQuery 26    │  320.12ms │                       435.73ms │  1.36x slower │
│ QQuery 27    │  900.06ms │                      1354.63ms │  1.51x slower │
│ QQuery 28    │ 7812.82ms │                      9813.62ms │  1.26x slower │

https://github.com/apache/datafusion/blob/7b370e26fea75fcd17121272eec1bd9447b2cb8f/benchmarks/queries/clickbench/queries.sql#L25-L29

These queries have the same predicate

WHERE "SearchPhrase" <> ''

But in this case SearchPhrase is also used in the rest of the query (and thus the projection)

For example

SELECT "SearchPhrase" FROM hits WHERE "SearchPhrase" <> '' ORDER BY "EventTime" LIMIT 10;

Thank you @alamb , good finding, so in theory we can combine the unified select(this PR) and also the page cache, in theory we can get the best performance until now. I will try to do a poc.

Dandandan
Dandandan reviewed May 2, 2025
View reviewed changes
parquet/src/arrow/arrow_reader/boolean_selector.rs
/// Unlike intersection, the `other` [`BooleanRowSelection`] must have exactly as many set bits as `self`.
/// This method will keep only the bits in `self` that are also set in `other`
/// at the positions corresponding to `self`'s set bits.
pub fn and_then(&self, other: &Self) -> Self {
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This should be able to use bitwise and instead?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I will refactor all this new file to the enum file, and we can remove this separate boolean selector.

@zhuqi-lucas
Copy link
Contributor Author

I take Q27 for example, it looks like the regression still comes from decompress, even the page cache case, it still use more decompress time, which cause the regression. We need to investigate more about the regression.

  1. The no pushdown case:
    flamegraphFast
  2. The page cache case:

flamegraphPageCache

  1. Current PR case:
    flamegraphSlow

Updated, after investigation, i found the root cause for page cache PR use more time to decode pages, i will try to update the polish_page_cache PR to address the changes, i hope we can solve all the regression for the page cache.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@Dandandan Dandandan Dandandan left review comments

Assignees
No one assigned
Labels
parquet Changes to the parquet crate
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.
3 participants
@zhuqi-lucas @alamb @Dandandan