Uh oh!
There was an error while loading. Please reload this page.
Call take arrays once per repartitioned input batch - #22159
Merged
Dandandan merged 2 commits intoMay 16, 2026
Merged
Conversation
rluvaton
commented
May 13, 2026
Member
run benchmarks |
adriangbot
commented
May 13, 2026
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing gene.bordegaray/2026/05/repartition-grouped-hash-take (a0a727c) to 937dfda (merge-base) diff using: tpch File an issue against this benchmark runner |
adriangbot
commented
May 13, 2026
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing gene.bordegaray/2026/05/repartition-grouped-hash-take (a0a727c) to 937dfda (merge-base) diff using: clickbench_partitioned File an issue against this benchmark runner |
adriangbot
commented
May 13, 2026
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing gene.bordegaray/2026/05/repartition-grouped-hash-take (a0a727c) to 937dfda (merge-base) diff using: tpcds File an issue against this benchmark runner |
adriangbot
commented
May 13, 2026
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)DetailsResource Usagetpch — base (merge-base)
tpch — branch
File an issue against this benchmark runner |
adriangbot
commented
May 13, 2026
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)DetailsResource Usagetpcds — base (merge-base)
tpcds — branch
File an issue against this benchmark runner |
adriangbot
commented
May 13, 2026
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)DetailsResource Usageclickbench_partitioned — base (merge-base)
clickbench_partitioned — branch
File an issue against this benchmark runner |
ContributorAuthor
This is inended for fanout on larger scale factor. The benchmarks in my description are run with Can this be run with |
gene-bordegaray
commented
May 13, 2026
ContributorAuthor
cc: @gabotechs |
gene-bordegaray
marked this pull request as ready for review
May 14, 2026 00:38
gabotechs
commented
May 14, 2026
Contributor
run benchmarks |
adriangbot
commented
May 14, 2026
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing gene.bordegaray/2026/05/repartition-grouped-hash-take (a0a727c) to 937dfda (merge-base) diff using: tpcds File an issue against this benchmark runner |
adriangbot
commented
May 14, 2026
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing gene.bordegaray/2026/05/repartition-grouped-hash-take (a0a727c) to 937dfda (merge-base) diff using: clickbench_partitioned File an issue against this benchmark runner |
adriangbot
commented
May 14, 2026
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing gene.bordegaray/2026/05/repartition-grouped-hash-take (a0a727c) to 937dfda (merge-base) diff using: tpch File an issue against this benchmark runner |
adriangbot
commented
May 14, 2026
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)DetailsResource Usagetpch — base (merge-base)
tpch — branch
File an issue against this benchmark runner |
adriangbot
commented
May 14, 2026
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)DetailsResource Usageclickbench_partitioned — base (merge-base)
clickbench_partitioned — branch
File an issue against this benchmark runner |
adriangbot
commented
May 14, 2026
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)DetailsResource Usagetpcds — base (merge-base)
tpcds — branch
File an issue against this benchmark runner |
Dandandan
commented
May 14, 2026
Contributor
This is very smart ❤️ |
gabotechs
commented
May 14, 2026
Contributor
Yeap, this is pretty smart indeed, nice work @gene-bordegaray! |
gene-bordegaray
commented
May 14, 2026
ContributorAuthor
if you guys are interested, this was stacked with the "internal" metrics level to repartition: #21152 along with a microbenchmark for repartition cases that will follow up the metrics PR 👍 |
gabotechs
approved these changes
May 14, 2026
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Contributor
run benchmark tpch10 |
adriangbot
commented
May 14, 2026
🤖 Benchmark running (GKE) | trigger CPU Details (lscpu)Comparing gene.bordegaray/2026/05/repartition-grouped-hash-take (96b9172) to 937dfda (merge-base) diff using: tpch10 File an issue against this benchmark runner |
adriangbot
commented
May 14, 2026
🤖 Benchmark completed (GKE) | trigger Instance: CPU Details (lscpu)DetailsResource Usagetpch10 — base (merge-base)
tpch10 — branch
File an issue against this benchmark runner |
ContributorAuthor
I think that dyn filter CI test that is fialing is flaky I can address or this could be separate PR |
gabotechs
commented
May 14, 2026
Contributor
Just re-run the job and succeeded. That test seems to be flaky. For me this is good to go! I'll give some time for @Dandandan (or anyone else) to give some feedback. |
Dandandan
approved these changes
May 15, 2026
Dandandan
commented
May 15, 2026
Contributor
Thank you @gene-bordegaray 🚀 |
github-merge-queueBot
removed this pull request from the merge queue due to no response for status checks
May 15, 2026
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
gabotechs pushed a commit
to DataDog/datafusion
that referenced
this pull request
May 18, 2026
## Which issue does this PR close? <!-- We generally require a GitHub issue to be filed for all bug fixes and enhancements and this helps us generate change logs for our releases. You can link an issue to this PR using the GitHub syntax. For example `Closes#123` indicates that this PR will close issue #123. --> - Does not close an issue. - This is targeted toward high fanout repartitions we see in shuffles in distributed-datafusion (datafusion-contrib/datafusion-distributed#385, datafusion-contrib/datafusion-distributed#353). - Related ongoing effort to improve metrics and repartition performance: apache#21148, apache#22155. ## Rationale for this change <!-- Why are you proposing this change? If this is already explained clearly in the issue then this section is not needed. Explaining clearly why changes are proposed helps reviewers understand your changes and offer better suggestions for fixes. --> Hash repartition currently builds one output batch per non-empty target partition by calling `take_arrays` separately for each partition. At high fanout this means an input batch can issue many take kernels, which shows in repartition-heavy queries. This changes hash repartition to concatenate the per-partition row indices, call `take_arrays` once for the input batch, and then slice the reordered batch back into per-partition output batches. This is complementary to apache#22010: that PR reduces channel/gate traffic from many small batches, while this PR reduces the Arrow take-kernel work required to create the repartitioned batches. ## What changes are included in this PR? <!-- There is no need to duplicate the description in the issue here but it is sometimes worth providing a summary of the individual changes in this PR. --> - Replaces per-partition hash repartition `take_arrays` calls with one grouped `take_arrays` call per input batch. - Tracks partition ranges into the grouped reordered batch and returns zero-copy `RecordBatch::slice` outputs for each non-empty partition. How the grouped take works: ```text input rows: 0 1 2 3 4 5 6 partition 0: [2, 5] partition 1: [] partition 2: [0, 3, 4] partition 3: [1, 6] grouped indices: [2, 5, 0, 3, 4, 1, 6] partition ranges: [(0, start=0, len=2), (2, start=2, len=3), (3, start=5, len=2)] take once: rows [2, 5, 0, 3, 4, 1, 6] slice outputs: partition 0 = slice(0, 2) partition 2 = slice(2, 3) partition 3 = slice(5, 2) ``` ## Are these changes tested? <!-- We typically require tests for all PRs in order to: 1. Prevent the code from being accidentally broken by subsequent changes 2. Serve as another way to document the expected behavior of the code If tests are not included in your PR, please explain why (for example, are they covered by existing tests)? --> - `cargo test -p datafusion-physical-plan repartition --lib` **Benchmarks:** Default TPCH SF10 summary, with no `--batch-size` override: | Partitions | main total ms | grouped total ms | change | wins | losses | biggest win | biggest loss |---:|---:|---:|---:|---:|---:|---:|---:| | 8 | 6234.28 | 6149.98 | -1.35% | 10 | 12 | Q3 -22.12% | Q21 5.47% | | 16 | 5602.63 | 5427.40 | -3.13% | 18 | 4 | Q21 -10.67% | Q10 4.47% | | 32 | 6097.10 | 5738.12 | -5.89% | 20 | 2 | Q8 -10.45% | Q6 0.47% | | 64 | 7194.70 | 6693.30 | -6.97% | 15 | 7 | Q21 -15.92% | Q1 5.02% | | 300 | 26276.60 | 23701.32 | -9.80% | 19 | 3 | Q21 -24.24% | Q1 9.45% | <details> <summary>TPCH SF10 default batch size, 8 partitions, all queries</summary> | Query | main ms | grouped ms | change | speedup | |---:|---:|---:|---:|---:| | Q1 | 411.40 | 356.46 | -13.35% | 1.154x | | Q2 | 117.97 | 95.68 | -18.89% | 1.233x | | Q3 | 278.88 | 217.20 | -22.12% | 1.284x | | Q4 | 104.92 | 98.47 | -6.15% | 1.065x | | Q5 | 311.36 | 302.12 | -2.97% | 1.031x | | Q6 | 133.82 | 135.13 | 0.98% | 0.990x | | Q7 | 366.32 | 360.50 | -1.59% | 1.016x | | Q8 | 408.86 | 391.14 | -4.33% | 1.045x | | Q9 | 509.62 | 515.74 | 1.20% | 0.988x | | Q10 | 278.97 | 273.80 | -1.85% | 1.019x | | Q11 | 76.55 | 75.34 | -1.58% | 1.016x | | Q12 | 175.56 | 180.87 | 3.03% | 0.971x | | Q13 | 211.03 | 221.03 | 4.74% | 0.955x | | Q14 | 183.05 | 189.95 | 3.77% | 0.964x | | Q15 | 318.46 | 328.85 | 3.26% | 0.968x | | Q16 | 57.66 | 57.70 | 0.08% | 0.999x | | Q17 | 505.05 | 503.77 | -0.25% | 1.003x | | Q18 | 596.49 | 602.77 | 1.05% | 0.990x | | Q19 | 273.60 | 283.41 | 3.58% | 0.965x | | Q20 | 255.73 | 267.30 | 4.52% | 0.957x | | Q21 | 592.44 | 624.83 | 5.47% | 0.948x | | Q22 | 66.54 | 67.91 | 2.05% | 0.980x | </details> <details> <summary>TPCH SF10 default batch size, 16 partitions, all queries</summary> | Query | main ms | grouped ms | change | speedup | |---:|---:|---:|---:|---:| | Q1 | 278.95 | 276.85 | -0.76% | 1.008x | | Q2 | 95.84 | 94.86 | -1.02% | 1.010x | | Q3 | 205.24 | 195.05 | -4.97% | 1.052x | | Q4 | 93.60 | 90.58 | -3.23% | 1.033x | | Q5 | 292.34 | 289.93 | -0.82% | 1.008x | | Q6 | 103.59 | 106.02 | 2.35% | 0.977x | | Q7 | 370.86 | 365.40 | -1.47% | 1.015x | | Q8 | 392.01 | 364.63 | -6.98% | 1.075x | | Q9 | 512.66 | 483.45 | -5.70% | 1.060x | | Q10 | 246.01 | 257.01 | 4.47% | 0.957x | | Q11 | 73.33 | 70.21 | -4.26% | 1.045x | | Q12 | 143.58 | 143.53 | -0.04% | 1.000x | | Q13 | 193.30 | 193.57 | 0.14% | 0.999x | | Q14 | 157.01 | 146.14 | -6.92% | 1.074x | | Q15 | 255.58 | 258.73 | 1.23% | 0.988x | | Q16 | 58.90 | 57.14 | -2.97% | 1.031x | | Q17 | 515.22 | 491.16 | -4.67% | 1.049x | | Q18 | 518.50 | 516.90 | -0.31% | 1.003x | | Q19 | 212.28 | 211.63 | -0.30% | 1.003x | | Q20 | 229.34 | 224.54 | -2.09% | 1.021x | | Q21 | 594.29 | 530.87 | -10.67% | 1.119x | | Q22 | 60.20 | 59.18 | -1.70% | 1.017x | </details> <details> <summary>TPCH SF10 default batch size, 32 partitions, all queries</summary> | Query | main ms | grouped ms | change | speedup | |---:|---:|---:|---:|---:| | Q1 | 294.06 | 286.80 | -2.47% | 1.025x | | Q2 | 110.59 | 103.55 | -6.37% | 1.068x | | Q3 | 227.86 | 215.08 | -5.61% | 1.059x | | Q4 | 110.99 | 102.77 | -7.41% | 1.080x | | Q5 | 347.34 | 319.49 | -8.02% | 1.087x | | Q6 | 109.39 | 109.91 | 0.47% | 0.995x | | Q7 | 423.98 | 385.42 | -9.09% | 1.100x | | Q8 | 428.83 | 384.00 | -10.45% | 1.117x | | Q9 | 559.81 | 510.72 | -8.77% | 1.096x | | Q10 | 269.17 | 265.02 | -1.54% | 1.016x | | Q11 | 85.99 | 80.21 | -6.72% | 1.072x | | Q12 | 153.67 | 150.36 | -2.15% | 1.022x | | Q13 | 197.90 | 188.72 | -4.64% | 1.049x | | Q14 | 161.28 | 153.93 | -4.55% | 1.048x | | Q15 | 259.92 | 261.01 | 0.42% | 0.996x | | Q16 | 64.90 | 63.70 | -1.85% | 1.019x | | Q17 | 574.18 | 531.76 | -7.39% | 1.080x | | Q18 | 572.23 | 538.68 | -5.86% | 1.062x | | Q19 | 227.95 | 220.19 | -3.41% | 1.035x | | Q20 | 232.25 | 225.80 | -2.78% | 1.029x | | Q21 | 622.54 | 579.94 | -6.84% | 1.073x | | Q22 | 62.29 | 61.06 | -1.98% | 1.020x | </details> <details> <summary>TPCH SF10 default batch size, 64 partitions, all queries</summary> | Query | main ms | grouped ms | change | speedup | |---:|---:|---:|---:|---:| | Q1 | 285.10 | 299.41 | 5.02% | 0.952x | | Q2 | 161.42 | 153.51 | -4.90% | 1.052x | | Q3 | 297.85 | 272.09 | -8.65% | 1.095x | | Q4 | 147.28 | 140.69 | -4.47% | 1.047x | | Q5 | 428.29 | 381.28 | -10.98% | 1.123x | | Q6 | 106.27 | 108.41 | 2.02% | 0.980x | | Q7 | 494.50 | 443.89 | -10.23% | 1.114x | | Q8 | 507.01 | 446.69 | -11.90% | 1.135x | | Q9 | 667.11 | 624.78 | -6.34% | 1.068x | | Q10 | 294.91 | 299.05 | 1.40% | 0.986x | | Q11 | 112.17 | 104.79 | -6.57% | 1.070x | | Q12 | 168.23 | 166.87 | -0.81% | 1.008x | | Q13 | 198.74 | 196.30 | -1.23% | 1.012x | | Q14 | 175.31 | 177.74 | 1.39% | 0.986x | | Q15 | 265.68 | 265.71 | 0.01% | 1.000x | | Q16 | 85.69 | 82.93 | -3.22% | 1.033x | | Q17 | 691.32 | 629.09 | -9.00% | 1.099x | | Q18 | 697.85 | 617.88 | -11.46% | 1.129x | | Q19 | 237.00 | 243.78 | 2.86% | 0.972x | | Q20 | 272.55 | 278.57 | 2.21% | 0.978x | | Q21 | 813.29 | 683.82 | -15.92% | 1.189x | | Q22 | 87.14 | 76.04 | -12.73% | 1.146x | </details> <details> <summary>TPCH SF10 default batch size, 300 partitions, all queries</summary> | Query | main ms | grouped ms | change | speedup | |---:|---:|---:|---:|---:| | Q1 | 277.84 | 304.08 | 9.45% | 0.914x | | Q2 | 1303.10 | 1268.65 | -2.64% | 1.027x | | Q3 | 1496.14 | 1393.15 | -6.88% | 1.074x | | Q4 | 681.20 | 652.85 | -4.16% | 1.043x | | Q5 | 1680.43 | 1469.91 | -12.53% | 1.143x | | Q6 | 100.65 | 105.75 | 5.07% | 0.952x | | Q7 | 1880.26 | 1652.83 | -12.10% | 1.138x | | Q8 | 1956.81 | 1760.72 | -10.02% | 1.111x | | Q9 | 1787.75 | 1454.84 | -18.62% | 1.229x | | Q10 | 1334.62 | 1296.02 | -2.89% | 1.030x | | Q11 | 1018.99 | 994.13 | -2.44% | 1.025x | | Q12 | 768.97 | 780.88 | 1.55% | 0.985x | | Q13 | 671.88 | 638.51 | -4.97% | 1.052x | | Q14 | 603.10 | 586.19 | -2.80% | 1.029x | | Q15 | 302.19 | 295.28 | -2.29% | 1.023x | | Q16 | 597.64 | 585.39 | -2.05% | 1.021x | | Q17 | 1963.57 | 1712.46 | -12.79% | 1.147x | | Q18 | 1942.46 | 1634.96 | -15.83% | 1.188x | | Q19 | 818.85 | 808.19 | -1.30% | 1.013x | | Q20 | 1499.39 | 1468.55 | -2.06% | 1.021x | | Q21 | 3006.35 | 2277.49 | -24.24% | 1.320x | | Q22 | 584.39 | 560.47 | -4.09% | 1.043x | </details> **Stress cases:** - These runs use `--batch-size 1024` to stress the repartition path. They are included to show the mechanism under smaller input batches and higher output fanout, not as the primary end-to-end performance claim. <details> <summary>TPCH SF10, 8 partitions, all queries</summary> | Query | main ms | grouped ms | change | speedup | |---:|---:|---:|---:|---:| | Q1 | 640.78 | 451.82 | -29.49% | 1.420x | | Q2 | 315.81 | 150.07 | -52.48% | 2.100x | | Q3 | 899.21 | 375.88 | -58.20% | 2.390x | | Q4 | 469.31 | 217.07 | -53.75% | 2.160x | | Q5 | 1131.37 | 446.36 | -60.55% | 2.530x | | Q6 | 376.40 | 163.66 | -56.52% | 2.300x | | Q7 | 1388.40 | 484.36 | -65.11% | 2.870x | | Q8 | 1369.67 | 571.62 | -58.27% | 2.400x | | Q9 | 1834.81 | 739.88 | -59.68% | 2.480x | | Q10 | 813.73 | 361.94 | -55.52% | 2.250x | | Q11 | 267.06 | 114.84 | -57.00% | 2.330x | | Q12 | 526.41 | 250.39 | -52.43% | 2.100x | | Q13 | 760.54 | 324.78 | -57.30% | 2.340x | | Q14 | 446.91 | 221.04 | -50.54% | 2.020x | | Q15 | 764.64 | 375.67 | -50.87% | 2.040x | | Q16 | 167.74 | 80.36 | -52.09% | 2.090x | | Q17 | 1801.72 | 763.58 | -57.62% | 2.360x | | Q18 | 3303.89 | 1649.87 | -50.06% | 2.000x | | Q19 | 694.16 | 354.97 | -48.86% | 1.960x | | Q20 | 693.91 | 323.17 | -53.43% | 2.150x | | Q21 | 3112.83 | 1065.36 | -65.78% | 2.920x | | Q22 | 205.88 | 95.97 | -53.38% | 2.150x | </details> <details> <summary>TPCH SF10, 16 partitions, all queries</summary> | Query | main ms | grouped ms | change | speedup | |---:|---:|---:|---:|---:| | Q1 | 518.84 | 328.66 | -36.66% | 1.580x | | Q2 | 350.47 | 148.64 | -57.59% | 2.360x | | Q3 | 1003.55 | 371.04 | -63.03% | 2.700x | | Q4 | 589.70 | 258.05 | -56.24% | 2.290x | | Q5 | 1343.64 | 506.01 | -62.34% | 2.660x | | Q6 | 322.21 | 130.93 | -59.37% | 2.460x | | Q7 | 1527.85 | 550.88 | -63.94% | 2.770x | | Q8 | 1476.46 | 578.77 | -60.80% | 2.550x | | Q9 | 2091.16 | 785.54 | -62.44% | 2.660x | | Q10 | 817.98 | 331.02 | -59.53% | 2.470x | | Q11 | 341.46 | 123.31 | -63.89% | 2.770x | | Q12 | 493.51 | 221.61 | -55.10% | 2.230x | | Q13 | 690.52 | 290.54 | -57.92% | 2.380x | | Q14 | 410.54 | 171.45 | -58.24% | 2.390x | | Q15 | 733.96 | 290.56 | -60.41% | 2.530x | | Q16 | 197.35 | 86.09 | -56.37% | 2.290x | | Q17 | 2089.12 | 828.96 | -60.32% | 2.520x | | Q18 | 2712.00 | 1097.77 | -59.52% | 2.470x | | Q19 | 602.77 | 260.74 | -56.74% | 2.310x | | Q20 | 661.20 | 288.58 | -56.35% | 2.290x | | Q21 | 5490.50 | 1151.50 | -79.03% | 4.770x | | Q22 | 198.38 | 103.13 | -48.01% | 1.920x | </details> <details> <summary>TPCH SF10, 32 partitions, all queries</summary> | Query | main ms | grouped ms | change | speedup | |---:|---:|---:|---:|---:| | Q1 | 533.86 | 338.54 | -36.59% | 1.580x | | Q2 | 439.59 | 199.50 | -54.62% | 2.200x | | Q3 | 1242.19 | 510.11 | -58.93% | 2.440x | | Q4 | 743.92 | 363.33 | -51.16% | 2.050x | | Q5 | 1711.97 | 666.50 | -61.07% | 2.570x | | Q6 | 325.39 | 134.07 | -58.80% | 2.430x | | Q7 | 1947.59 | 722.22 | -62.92% | 2.700x | | Q8 | 1914.31 | 775.62 | -59.48% | 2.470x | | Q9 | 2662.07 | 976.47 | -63.32% | 2.730x | | Q10 | 902.80 | 362.71 | -59.82% | 2.490x | | Q11 | 400.93 | 170.81 | -57.40% | 2.350x | | Q12 | 572.19 | 265.06 | -53.68% | 2.160x | | Q13 | 736.31 | 296.82 | -59.69% | 2.480x | | Q14 | 430.11 | 180.93 | -57.93% | 2.380x | | Q15 | 732.36 | 327.12 | -55.33% | 2.240x | | Q16 | 245.97 | 116.24 | -52.74% | 2.120x | | Q17 | 2711.18 | 1100.17 | -59.42% | 2.460x | | Q18 | 2946.70 | 1176.02 | -60.09% | 2.510x | | Q19 | 600.47 | 258.58 | -56.94% | 2.320x | | Q20 | 765.20 | 337.01 | -55.96% | 2.270x | | Q21 | 10062.70 | 1534.95 | -84.75% | 6.560x | | Q22 | 250.50 | 128.27 | -48.79% | 1.950x | </details> <details> <summary>TPCH SF10, 64 partitions, all queries</summary> | Query | main ms | grouped ms | change | speedup | |---:|---:|---:|---:|---:| | Q1 | 595.70 | 324.74 | -45.49% | 1.830x | | Q2 | 663.08 | 305.30 | -53.96% | 2.170x | | Q3 | 1744.90 | 727.81 | -58.29% | 2.400x | | Q4 | 1070.72 | 566.20 | -47.12% | 1.890x | | Q5 | 2447.07 | 938.91 | -61.63% | 2.610x | | Q6 | 315.47 | 132.73 | -57.93% | 2.380x | | Q7 | 2807.33 | 1004.83 | -64.21% | 2.790x | | Q8 | 2674.51 | 1069.64 | -60.01% | 2.500x | | Q9 | 3777.94 | 1424.08 | -62.31% | 2.650x | | Q10 | 1086.91 | 469.38 | -56.82% | 2.320x | | Q11 | 575.59 | 264.02 | -54.13% | 2.180x | | Q12 | 841.83 | 387.25 | -54.00% | 2.170x | | Q13 | 867.57 | 379.90 | -56.21% | 2.280x | | Q14 | 470.87 | 214.58 | -54.43% | 2.190x | | Q15 | 762.07 | 340.55 | -55.31% | 2.240x | | Q16 | 337.20 | 179.25 | -46.84% | 1.880x | | Q17 | 3953.82 | 1701.46 | -56.97% | 2.320x | | Q18 | 3763.51 | 1606.90 | -57.30% | 2.340x | | Q19 | 644.43 | 314.27 | -51.23% | 2.050x | | Q20 | 973.56 | 453.24 | -53.45% | 2.150x | | Q21 | 19356.91 | 2396.96 | -87.62% | 8.080x | | Q22 | 366.20 | 195.40 | -46.64% | 1.870x | </details> <details> <summary>TPCH SF10, 300 partitions, targeted high-fanout queries</summary> Yes this is a real use case for fanout in distributed-datafusion | Query | main ms | grouped ms | change | speedup | |---:|---:|---:|---:|---:| | Q3 | 2543.94 | 2250.91 | -11.52% | 1.130x | | Q9 | 6495.22 | 4755.78 | -26.78% | 1.370x | | Q10 | 1869.05 | 1709.18 | -8.55% | 1.090x | | Q13 | 1238.63 | 1157.47 | -6.55% | 1.070x | | Q15 | 461.51 | 446.25 | -3.31% | 1.030x | | Q21 | 37810.29 | 5594.01 | -85.21% | 6.760x | | Q22 | 1084.95 | 1058.74 | -2.42% | 1.020x | </details> <details> <summary>TPCH SF10, 300 partitions, peak RSS stress</summary> Measured with `/usr/bin/time -l`, one iteration, `--batch-size 1024`, `--partitions 300`, and no DataFusion memory limit. RSS is process peak resident set size from the OS. | Query | main ms | grouped ms | time change | main peak RSS | grouped peak RSS | RSS change | |---:|---:|---:|---:|---:|---:|---:| | Q7 | 5171.45 | 4151.15 | -19.73% | 3.69 GiB | 3.75 GiB | 1.61% | | Q9 | 6055.57 | 4758.10 | -21.43% | 4.01 GiB | 4.01 GiB | 0.04% | | Q21 | 36300.80 | 5810.14 | -83.99% | 2.96 GiB | 2.05 GiB | -30.79% | </details> #### Memory concern and follow-up work This PR changes output batches from materializing per-partition batches to slices of one reordered batch. This means sibling slices can share the same buffers. Potential concern: ```text one reordered batch allocation -> slice for partition 0 -> slice for partition 1 -> slice for partition 2 ``` A slow output partition can keep the shared reordered batch buffers alive until its slice is dropped. Also, `RecordBatch::get_array_memory_size()` may count shared slice buffers repeatedly when repartition reserves memory per output batch. The peak RSS stress above did not show a process-memory regression in the measured queries. Follow-up work should add buffer-aware accounting. ## Are there any user-facing changes? <!-- If there are user-facing changes then we may require documentation to be updated before approving the PR. --> No. <!-- If there are any breaking changes to public APIs, please add the `api change` label. --> (cherry picked from commit 1ab146a)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for freeto join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Which issue does this PR close?
Rationale for this change
Hash repartition currently builds one output batch per non-empty target partition by calling
take_arraysseparately for each partition. At high fanout this means an input batch can issue many take kernels, which shows in repartition-heavy queries.This changes hash repartition to concatenate the per-partition row indices, call
take_arraysonce for the input batch, and then slice the reordered batch back into per-partition output batches.This is complementary to #22010: that PR reduces channel/gate traffic from many small batches, while this PR reduces the Arrow take-kernel work required to create the repartitioned batches.
What changes are included in this PR?
take_arrayscalls with one groupedtake_arrayscall per input batch.RecordBatch::sliceoutputs for each non-empty partition.How the grouped take works:
Are these changes tested?
cargo test -p datafusion-physical-plan repartition --libBenchmarks:
Default TPCH SF10 summary, with no
--batch-sizeoverride:TPCH SF10 default batch size, 8 partitions, all queries
TPCH SF10 default batch size, 16 partitions, all queries
TPCH SF10 default batch size, 32 partitions, all queries
TPCH SF10 default batch size, 64 partitions, all queries
TPCH SF10 default batch size, 300 partitions, all queries
Stress cases:
--batch-size 1024to stress the repartition path. They are included to show the mechanism under smaller input batches and higher output fanout, not as the primary end-to-end performance claim.TPCH SF10, 8 partitions, all queries
TPCH SF10, 16 partitions, all queries
TPCH SF10, 32 partitions, all queries
TPCH SF10, 64 partitions, all queries
TPCH SF10, 300 partitions, targeted high-fanout queries
Yes this is a real use case for fanout in distributed-datafusion
TPCH SF10, 300 partitions, peak RSS stress
Measured with
/usr/bin/time -l, one iteration,--batch-size 1024,--partitions 300, and no DataFusion memory limit. RSS is process peak resident set size from the OS.Memory concern and follow-up work
This PR changes output batches from materializing per-partition batches to slices of one reordered batch. This means sibling slices can share the same buffers.
Potential concern:
A slow output partition can keep the shared reordered batch buffers alive until its slice is dropped. Also,
RecordBatch::get_array_memory_size()may count shared slice buffers repeatedly when repartition reserves memory per output batch.The peak RSS stress above did not show a process-memory regression in the measured queries. Follow-up work should add buffer-aware accounting.
Are there any user-facing changes?
No.