Optimize take_fixed_size_binary For Predefined Value Lengths

[tobixdev]

Which issue does this PR close?

• Related to Performance improvements for take #279

Rationale for this change

The take kernel is very important for many operations (e.g., HashJoin in DataFusion IIRC). Currently, there is a gap between the performance of the take kernel for primitive arrays (e.g., DataType::UInt32) and fixed size binary arrays of the same length (e.g., FixedSizeBinary<4>).

In our case this lead to a performance reduction when moving from an integer-based id column to a fixed-size-binary-based id column. This PR aims to address parts of this gap.

The 16-bytes case would especially benefit operations on UUID columns.

What changes are included in this PR?

• Add take_fixed_size that can be called for set of predefined fsb-lengths that we want to support. This is a "flat buffer" version of the take_native kernel.

Are these changes tested?

I've added another test that still exercises the non-optimized code path.

Are there any user-facing changes?

No

[tobixdev]

run benchmark take_kernels

[alamb-ghbot]

🤖 Hi @tobixdev, thanks for the request (#9535 (comment)). scrape_comments.py only responds to whitelisted users. Allowed users: Dandandan, Jefffrey, Omega359, adriangb, alamb, comphead, etseidl, gabotechs, geoffreyclaude, klion26, rluvaton, xudong963, zhuqi-lucas.

[tobixdev]

I assumed that it would not work but it was worth a shot. Would appreciate someone running the benchmark. I saw approximately -80% on my machine.

[alamb]

run benchmark take_kernels

[alamb-ghbot]

🤖 ./gh_compare_arrow.sh gh_compare_arrow.sh Running
Linux aal-dev 6.14.0-1018-gcp #19~24.04.1-Ubuntu SMP Wed Sep 24 23:23:09 UTC 2025 x86_64 x86_64 x86_64 GNU/Linux
Comparing improve-take-fsb (7aee3d9) to 33aed33 diff
BENCH_NAME=take_kernels
BENCH_COMMAND=cargo bench --features=arrow,async,test_common,experimental,object_store --bench take_kernels
BENCH_FILTER=
BENCH_BRANCH_NAME=improve-take-fsb
Results will be posted here when complete

[alamb-ghbot]

🤖: Benchmark completed

Details

group                                                                     improve-take-fsb                       main
-----                                                                     ----------------                       ----
take bool 1024                                                            1.00   1328.7±5.30ns        ? ?/sec    1.00   1328.3±2.79ns        ? ?/sec
take bool 512                                                             1.00    728.2±8.58ns        ? ?/sec    1.00    727.1±2.58ns        ? ?/sec
take bool null indices 1024                                               1.00  1113.3±13.46ns        ? ?/sec    1.11  1232.8±24.14ns        ? ?/sec
take bool null values 1024                                                1.00      2.6±0.02µs        ? ?/sec    1.00      2.6±0.01µs        ? ?/sec
take bool null values null indices 1024                                   1.00      2.1±0.07µs        ? ?/sec    1.42      2.9±0.04µs        ? ?/sec
take check bounds i32 1024                                                1.10   930.9±68.94ns        ? ?/sec    1.00    844.0±4.14ns        ? ?/sec
take check bounds i32 512                                                 1.00    515.4±4.77ns        ? ?/sec    1.14    588.5±3.92ns        ? ?/sec
take i32 1024                                                             1.00    712.3±2.60ns        ? ?/sec    1.01    716.9±6.95ns        ? ?/sec
take i32 512                                                              1.00    441.8±1.44ns        ? ?/sec    1.00    442.8±1.33ns        ? ?/sec
take i32 null indices 1024                                                1.00    995.6±2.39ns        ? ?/sec    1.02  1015.2±126.82ns        ? ?/sec
take i32 null values 1024                                                 1.01      2.0±0.02µs        ? ?/sec    1.00      2.0±0.03µs        ? ?/sec
take i32 null values null indices 1024                                    1.00      2.1±0.02µs        ? ?/sec    1.05      2.2±0.08µs        ? ?/sec
take primitive fsb value len: 12, indices: 1024                           1.00  1057.2±10.66ns        ? ?/sec    3.33      3.5±0.17µs        ? ?/sec
take primitive fsb value len: 12, null values, indices: 1024              1.00      2.4±0.03µs        ? ?/sec    2.01      4.9±0.08µs        ? ?/sec
take primitive run logical len: 1024, physical len: 512, indices: 1024    1.00     20.9±0.21µs        ? ?/sec    1.00     20.8±0.09µs        ? ?/sec
take str 1024                                                             1.00     11.2±0.06µs        ? ?/sec    1.01     11.3±0.16µs        ? ?/sec
take str 512                                                              1.00      5.4±0.03µs        ? ?/sec    1.02      5.5±0.06µs        ? ?/sec
take str null indices 1024                                                1.00      7.9±0.04µs        ? ?/sec    1.00      7.9±0.05µs        ? ?/sec
take str null indices 512                                                 1.00      3.8±0.01µs        ? ?/sec    1.00      3.8±0.03µs        ? ?/sec
take str null values 1024                                                 1.00      8.7±0.03µs        ? ?/sec    1.00      8.6±0.07µs        ? ?/sec
take str null values null indices 1024                                    1.00      7.0±0.03µs        ? ?/sec    1.00      7.0±0.08µs        ? ?/sec
take stringview 1024                                                      1.00    809.3±8.35ns        ? ?/sec    1.10    893.3±7.40ns        ? ?/sec
take stringview 512                                                       1.00    579.9±4.53ns        ? ?/sec    1.01    587.5±8.20ns        ? ?/sec
take stringview null indices 1024                                         1.00  1440.7±21.77ns        ? ?/sec    1.01   1451.7±4.98ns        ? ?/sec
take stringview null indices 512                                          1.00   718.1±19.57ns        ? ?/sec    1.11    800.5±0.89ns        ? ?/sec
take stringview null values 1024                                          1.00      2.1±0.00µs        ? ?/sec    1.00      2.1±0.00µs        ? ?/sec
take stringview null values null indices 1024                             1.00      2.3±0.01µs        ? ?/sec    1.06      2.4±0.04µs        ? ?/sec

[tobixdev]

take primitive fsb value len: 12, indices: 1024                           1.00  1057.2±10.66ns        ? ?/sec    3.33      3.5±0.17µs        ? ?/sec
take primitive fsb value len: 12, null values, indices: 1024              1.00      2.4±0.03µs        ? ?/sec    2.01      4.9±0.08µs        ? ?/sec

It's not quite 80% but still significant. Maybe this difference is due to Aarch64 vs. x86 (my PC).

take i32 1024                                                             1.00    712.3±2.60ns        ? ?/sec    1.01    716.9±6.95ns        ? ?/sec

We're also close to the primitive kernel! Parts of the gap could be explained by the entry size (4 vs. 12 bytes).

[alamb]

run benchmark take_kernels

[alamb]

Will run once more to ensure we can reproduce the results. They look good

[alamb-ghbot]

🤖 ./gh_compare_arrow.sh gh_compare_arrow.sh Running
Linux aal-dev 6.14.0-1018-gcp #19~24.04.1-Ubuntu SMP Wed Sep 24 23:23:09 UTC 2025 x86_64 x86_64 x86_64 GNU/Linux
Comparing improve-take-fsb (7aee3d9) to 33aed33 diff
BENCH_NAME=take_kernels
BENCH_COMMAND=cargo bench --features=arrow,async,test_common,experimental,object_store --bench take_kernels
BENCH_FILTER=
BENCH_BRANCH_NAME=improve-take-fsb
Results will be posted here when complete

[alamb]

as I read this it results in 7 copies of the code which is probably ok here but we do have to be careful in general to avoid too much bloat

[tobixdev]

Yes that's definetely a drawback (for compile time and binary size).

[alamb-ghbot]

🤖: Benchmark completed

Details

group                                                                     improve-take-fsb                       main
-----                                                                     ----------------                       ----
take bool 1024                                                            1.00   1327.8±4.95ns        ? ?/sec    1.00  1329.2±12.95ns        ? ?/sec
take bool 512                                                             1.00    726.9±3.22ns        ? ?/sec    1.00   727.1±14.49ns        ? ?/sec
take bool null indices 1024                                               1.00   1097.8±5.99ns        ? ?/sec    1.11  1215.4±31.46ns        ? ?/sec
take bool null values 1024                                                1.00      2.6±0.02µs        ? ?/sec    1.00      2.6±0.01µs        ? ?/sec
take bool null values null indices 1024                                   1.00  1977.1±21.64ns        ? ?/sec    1.48      2.9±0.05µs        ? ?/sec
take check bounds i32 1024                                                1.08    917.4±7.58ns        ? ?/sec    1.00   846.9±12.66ns        ? ?/sec
take check bounds i32 512                                                 1.00   516.9±21.67ns        ? ?/sec    1.14    587.1±1.14ns        ? ?/sec
take i32 1024                                                             1.00    712.5±1.52ns        ? ?/sec    1.01    717.9±2.41ns        ? ?/sec
take i32 512                                                              1.00    441.9±0.76ns        ? ?/sec    1.01    444.9±9.14ns        ? ?/sec
take i32 null indices 1024                                                1.00    993.5±2.18ns        ? ?/sec    1.00    993.2±2.50ns        ? ?/sec
take i32 null values 1024                                                 1.01      2.0±0.00µs        ? ?/sec    1.00      2.0±0.00µs        ? ?/sec
take i32 null values null indices 1024                                    1.00      2.1±0.02µs        ? ?/sec    1.05      2.2±0.02µs        ? ?/sec
take primitive fsb value len: 12, indices: 1024                           1.00   1056.9±6.85ns        ? ?/sec    3.28      3.5±0.02µs        ? ?/sec
take primitive fsb value len: 12, null values, indices: 1024              1.00      2.4±0.04µs        ? ?/sec    1.99      4.9±0.01µs        ? ?/sec
take primitive run logical len: 1024, physical len: 512, indices: 1024    1.00     21.0±0.24µs        ? ?/sec    1.00     21.0±0.10µs        ? ?/sec
take str 1024                                                             1.01     11.3±0.09µs        ? ?/sec    1.00     11.2±0.11µs        ? ?/sec
take str 512                                                              1.00      5.4±0.02µs        ? ?/sec    1.02      5.5±0.09µs        ? ?/sec
take str null indices 1024                                                1.00      7.9±0.20µs        ? ?/sec    1.00      7.8±0.02µs        ? ?/sec
take str null indices 512                                                 1.00      3.8±0.02µs        ? ?/sec    1.00      3.8±0.01µs        ? ?/sec
take str null values 1024                                                 1.00      8.6±0.13µs        ? ?/sec    1.00      8.6±0.03µs        ? ?/sec
take str null values null indices 1024                                    1.00      6.9±0.06µs        ? ?/sec    1.02      7.0±0.09µs        ? ?/sec
take stringview 1024                                                      1.00    811.9±1.46ns        ? ?/sec    1.09   887.2±12.82ns        ? ?/sec
take stringview 512                                                       1.00    582.0±7.36ns        ? ?/sec    1.01    588.3±2.38ns        ? ?/sec
take stringview null indices 1024                                         1.00  1436.1±22.29ns        ? ?/sec    1.01  1449.0±24.56ns        ? ?/sec
take stringview null indices 512                                          1.00    727.5±2.41ns        ? ?/sec    1.10    801.0±4.78ns        ? ?/sec
take stringview null values 1024                                          1.00      2.1±0.05µs        ? ?/sec    1.00      2.1±0.02µs        ? ?/sec
take stringview null values null indices 1024                             1.00      2.3±0.02µs        ? ?/sec    1.05      2.4±0.03µs        ? ?/sec

[alamb]

take primitive fsb value len: 12, indices: 1024                           1.00   1056.9±6.85ns        ? ?/sec    3.28      3.5±0.02µs        ? ?/sec
take primitive fsb value len: 12, null values, indices: 1024              1.00      2.4±0.04µs        ? ?/sec    1.99      4.9±0.01µs        ? ?/sec

🚀