perf: Cast entire Date32 array to Date64 on 1st failure by huymq1710 · Pull Request #21948 · apache/datafusion

huymq1710 · 2026-04-30T03:21:23Z

Which issue does this PR close?

Closes Optimize to_char for array conversions #17152

Rationale for this change

1. Vec<Option<String>> extra allocations
- Resolved in perf: Optimize to_char to allocate less, fix NULL handling #20635 by using StringBuilder
2. Per-row cast on fallback
- Will be resolved in this PR

// When `Date32` formatting fails (due to time specifiers),
// the code called cast individually for each row instead of converting the entire array
let date64_value = cast(&data_array.slice(idx, 1), &Date64)?;

What changes are included in this PR?

Cast the entire Date32 array to Date64 once on first failure, instead of per-row

cast(data_array.as_ref(), &Date64

Are these changes tested?

Yes

Add 2 benchmarks:
- Date32 + datetime patterns (all rows trigger the fallback)
- Date32 + mixed patterns (roughly half do)

  ┌─────────────────────────────────────┬───────────────────┬───────────┐
  │              Benchmark              │      Format       │  Change   │
  ├─────────────────────────────────────┼───────────────────┼───────────┤
  │ array_date_only_patterns_1000       │ array, date-only  │ No change │
  ├─────────────────────────────────────┼───────────────────┼───────────┤
  │ array_datetime_patterns_1000        │ array, datetime   │ -3.21%    │
  ├─────────────────────────────────────┼───────────────────┼───────────┤
  │ array_mixed_patterns_1000           │ array, mixed      │ +1.41%    │
  ├─────────────────────────────────────┼───────────────────┼───────────┤
  │ scalar_date_only_pattern_1000       │ scalar, date-only │ No change │
  ├─────────────────────────────────────┼───────────────────┼───────────┤
  │ scalar_datetime_pattern_1000        │ scalar, datetime  │ No change │
  ├─────────────────────────────────────┼───────────────────┼───────────┤
  │ array_date32_datetime_patterns_1000 │ array, datetime   │ -20.24%   │
  ├─────────────────────────────────────┼───────────────────┼───────────┤
  │ array_date32_mixed_patterns_1000    │ array, mixed      │ -15.74%   │
  └─────────────────────────────────────┴───────────────────┴───────────┘

Detail

cargo bench --bench to_char          
   Compiling datafusion-functions v53.1.0 (/Users/qmac/Documents/GitHub/datafusion/datafusion/functions)
    Finished `bench` profile [optimized] target(s) in 1m 10s
     Running benches/to_char.rs (target/release/deps/to_char-d2acea4b7a3e2fba)
Gnuplot not found, using plotters backend
Benchmarking to_char_array_date_only_patterns_1000: Warming up for 3.00Benchmarking to_char_array_date_only_patterns_1000: Collecting 100 sampto_char_array_date_only_patterns_1000
                        time:   [104.01 µs 104.37 µs 104.74 µs]
                        change: [−0.5692% +0.3244% +1.0788%] (p = 0.47 > 0.05)
                        No change in performance detected.
Found 3 outliers among 100 measurements (3.00%)
  1 (1.00%) low mild
  2 (2.00%) high mild

Benchmarking to_char_array_datetime_patterns_1000: Warming up for 3.000Benchmarking to_char_array_datetime_patterns_1000: Collecting 100 samplto_char_array_datetime_patterns_1000
                        time:   [165.96 µs 166.43 µs 166.94 µs]
                        change: [−4.1660% −3.2055% −2.3626%] (p = 0.00 < 0.05)
                        Performance has improved.
Found 6 outliers among 100 measurements (6.00%)
  2 (2.00%) low mild
  3 (3.00%) high mild
  1 (1.00%) high severe

Benchmarking to_char_array_mixed_patterns_1000: Collecting 100 samples to_char_array_mixed_patterns_1000
                        time:   [139.58 µs 140.02 µs 140.49 µs]
                        change: [+1.0116% +1.4122% +1.8178%] (p = 0.00 < 0.05)
                        Performance has regressed.
Found 3 outliers among 100 measurements (3.00%)
  1 (1.00%) low mild
  2 (2.00%) high mild

Benchmarking to_char_scalar_date_only_pattern_1000: Warming up for 3.00Benchmarking to_char_scalar_date_only_pattern_1000: Collecting 100 sampto_char_scalar_date_only_pattern_1000
                        time:   [60.073 µs 63.184 µs 66.309 µs]
                        change: [−4.0228% +0.6839% +6.1581%] (p = 0.79 > 0.05)
                        No change in performance detected.

Benchmarking to_char_scalar_datetime_pattern_1000: Warming up for 3.000Benchmarking to_char_scalar_datetime_pattern_1000: Collecting 100 samplto_char_scalar_datetime_pattern_1000
                        time:   [117.65 µs 124.76 µs 131.84 µs]
                        change: [−3.3009% +2.6889% +8.6148%] (p = 0.37 > 0.05)
                        No change in performance detected.
Found 30 outliers among 100 measurements (30.00%)
  16 (16.00%) low mild
  14 (14.00%) high mild

Benchmarking to_char_array_date32_datetime_patterns_1000: Warming up foBenchmarking to_char_array_date32_datetime_patterns_1000: Collecting 10to_char_array_date32_datetime_patterns_1000
                        time:   [289.67 µs 290.77 µs 291.88 µs]
                        change: [−20.661% −20.240% −19.863%] (p = 0.00 < 0.05)
                        Performance has improved.

Benchmarking to_char_array_date32_mixed_patterns_1000: Warming up for 3Benchmarking to_char_array_date32_mixed_patterns_1000: Collecting 100 sto_char_array_date32_mixed_patterns_1000
                        time:   [194.47 µs 195.28 µs 196.11 µs]
                        change: [−16.230% −15.738% −15.285%] (p = 0.00 < 0.05)
                        Performance has improved.

Are there any user-facing changes?

No

huymq1710 · 2026-04-30T11:00:12Z

Thank you. I updated

Omega359 · 2026-05-03T02:04:59Z

Previous work on this might show some insight - #15361 (comment)

alamb

Thanks @huymq1710 and @kumarUjjawal and @Omega359

THis looks good to me at a high level -- I would like to get the benchmark in place so we can verify performance

alamb · 2026-05-03T13:12:27Z

+    // handle time specifiers and falls back to a Date64 cast.
+
+    // Covers full fallback (every row triggers the cast)
+    c.bench_function("to_char_array_date32_datetime_patterns_1000", |b| {


Thank you @huymq1710

Can you please put this new benchmark into a separate PR so that we can use our automated benchmark runner to compare the performance of main and your proposal?

Got it, I temporary removed benchmark changes, I will create other PR to add back after this PR is merged

@alamb One clarification, do you want this merged first or the pr with benchmark?

Typically I like to merge the PR with the benchmark first, then merge up the PR with performance improvements, and then validate the benchmark results using the run benchmarks command.

Okay, thanks!

…10/datafusion into cast-entire-date32-array

kumarUjjawal · 2026-05-04T04:19:31Z

@huymq1710 can you open the pr with the benchmark

huymq1710 · 2026-05-04T04:28:07Z

I see, created

## Which issue does this PR close?  - Closes #. ## Rationale for this change  ## What changes are included in this PR?  Benchmark for apache#21948 ## Are these changes tested?  ## Are there any user-facing changes?

kumarUjjawal · 2026-05-04T06:34:35Z

run benchmark

adriangbot · 2026-05-04T06:34:38Z

Hi @kumarUjjawal, run benchmark requires benchmark names (#21948 (comment)).

Supported benchmarks:

Standard: clickbench_1, clickbench_extended, clickbench_partitioned, clickbench_pushdown, external_aggr, smj, sort_pushdown, sort_pushdown_inexact, sort_pushdown_inexact_overlap, sort_pushdown_inexact_unsorted, sort_pushdown_sorted, topk_tpch, tpcds, tpch, tpch10, tpch_mem, tpch_mem10
Criterion: (any)

Usage:

run benchmark <name>           # run specific benchmark(s)
run benchmarks                 # run default suite
run benchmarks <name1> <name2> # run specific benchmarks

Per-side configuration (run benchmark tpch followed by):

env:
SHARED_SETTING: enabled
baseline:
ref: v45.0.0
env:
DATAFUSION_RUNTIME_MEMORY_LIMIT: 1G
changed:
ref: v46.0.0
env:
DATAFUSION_RUNTIME_MEMORY_LIMIT: 2G

File an issue against this benchmark runner

kumarUjjawal · 2026-05-04T06:40:18Z

run benchmark to_char

adriangbot · 2026-05-04T06:44:24Z

🤖 Criterion benchmark running (GKE) | trigger
Instance: c4a-highmem-16 (12 vCPU / 65 GiB) | Linux bench-c4368798241-2000-xvvnf 6.12.68+ #1 SMP Wed Apr 1 02:23:28 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 cast-entire-date32-array (205f116) to b2fd2d3 (merge-base) diff
BENCH_NAME=to_char
BENCH_COMMAND=cargo bench --features=parquet --bench to_char
BENCH_FILTER=
Results will be posted here when complete

File an issue against this benchmark runner

adriangbot · 2026-05-04T06:49:42Z

🤖 Criterion benchmark completed (GKE) | trigger

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

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

Details

group                                          cast-entire-date32-array               main
-----                                          ------------------------               ----
to_char_array_date32_datetime_patterns_1000    1.00    621.4±4.11µs        ? ?/sec    1.32    821.3±3.57µs        ? ?/sec
to_char_array_date32_mixed_patterns_1000       1.00    402.7±9.22µs        ? ?/sec    1.25    504.8±8.98µs        ? ?/sec
to_char_array_date_only_patterns_1000          1.01    175.3±1.79µs        ? ?/sec    1.00    173.9±2.11µs        ? ?/sec
to_char_array_datetime_patterns_1000           1.00    372.1±4.27µs        ? ?/sec    1.01    376.7±3.66µs        ? ?/sec
to_char_array_mixed_patterns_1000              1.00    282.6±4.08µs        ? ?/sec    1.01    285.5±3.33µs        ? ?/sec
to_char_scalar_date_only_pattern_1000          1.04   149.9±21.47µs        ? ?/sec    1.00   144.8±21.11µs        ? ?/sec
to_char_scalar_datetime_pattern_1000           1.00   293.6±59.07µs        ? ?/sec    1.03   302.4±47.62µs        ? ?/sec

Resource Usage

base (merge-base)

Metric	Value
Wall time	80.0s
Peak memory	4.3 GiB
Avg memory	4.2 GiB
CPU user	92.7s
CPU sys	0.8s
Peak spill	0 B

branch

Metric	Value
Wall time	75.0s
Peak memory	4.2 GiB
Avg memory	4.2 GiB
CPU user	91.7s
CPU sys	0.3s
Peak spill	0 B

File an issue against this benchmark runner

kumarUjjawal · 2026-05-04T07:17:16Z

Benchmark looks good!

huymq1710 added 2 commits April 30, 2026 11:54

perf: Cast entire Date32 array to Date64 on 1st failure

2862ee4

bench: add to_char_array_date32

e0eb0fd

github-actions Bot added the functions Changes to functions implementation label Apr 30, 2026

huymq1710 changed the title ~~Cast entire date32 array~~ perf: Cast entire Date32 array to Date64 on 1st failure Apr 30, 2026

huymq1710 mentioned this pull request Apr 30, 2026

Optimize to_char for array conversions #17152

Open

kumarUjjawal reviewed Apr 30, 2026

View reviewed changes

Comment thread datafusion/functions/benches/to_char.rs Outdated

Comment thread datafusion/functions/src/datetime/to_char.rs

Comment thread datafusion/functions/src/datetime/to_char.rs

chore: update test

a5631c8

Merge branch 'main' into cast-entire-date32-array

8900de5

kumarUjjawal requested a review from alamb May 2, 2026 04:40

alamb approved these changes May 3, 2026

View reviewed changes

huymq1710 and others added 4 commits May 3, 2026 22:22

perf: Cast entire Date32 array to Date64 on 1st failure

b194da4

Merge branch 'cast-entire-date32-array' of https://github.com/huymq17…

c566146

…10/datafusion into cast-entire-date32-array

bench: move to other PR to run benchmark

13f9504

Merge branch 'main' into cast-entire-date32-array

0fe0be2

huymq1710 mentioned this pull request May 4, 2026

bench: add to_char_array_date32 #22007

Merged

Merge branch 'main' into cast-entire-date32-array

205f116

Conversation

huymq1710 commented Apr 30, 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!

Uh oh!

Uh oh!

Uh oh!

huymq1710 commented Apr 30, 2026

Uh oh!

Omega359 commented May 3, 2026

Uh oh!

alamb left a comment

Choose a reason for hiding this comment

Uh oh!

alamb May 3, 2026

Choose a reason for hiding this comment

Uh oh!

huymq1710 May 3, 2026

Choose a reason for hiding this comment

Uh oh!

kumarUjjawal May 3, 2026

Choose a reason for hiding this comment

Uh oh!

alamb May 3, 2026

Choose a reason for hiding this comment

Uh oh!

kumarUjjawal May 4, 2026

Choose a reason for hiding this comment

Uh oh!

kumarUjjawal commented May 4, 2026

Uh oh!

huymq1710 commented May 4, 2026

Uh oh!

kumarUjjawal commented May 4, 2026

Uh oh!

adriangbot commented May 4, 2026

Uh oh!

kumarUjjawal commented May 4, 2026

Uh oh!

adriangbot commented May 4, 2026

Uh oh!

adriangbot commented May 4, 2026

Uh oh!

kumarUjjawal commented May 4, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

huymq1710 commented Apr 30, 2026 •

edited

Loading