Reduce size of `Expr` struct by alamb · Pull Request #14366 · apache/datafusion

alamb · 2025-01-29T21:15:26Z

Which issue does this PR close?

Related to Build time regression #14256
ClosesReduce size of Expr struct #16199

Rationale for this change

@waynexia 's comment on #14256 (comment) got me thinking maybe the build time regression had something to do with the size of Expr

So I poked around for ways to reduce the size, and I found that currently Expr is 272 bytes

What changes are included in this PR?

Add a test for the size of Expr
Change Expr::WindowFunction(WindowFunction) --> Expr::WindowFunction(Box<WindowFunction>) -- which drops the size of Exprfrom 272 to112` bytes

Are these changes tested?

functionally by CI

TODO:

test impact on build time
test impact on planning performance benchmarks

Are there any user-facing changes?

findepi · 2025-01-29T21:29:40Z

+ return Some(idx + input_len);
+ } else {
+ None


inconsistent return vs value

Co-authored-by: Felipe Oliveira Carvalho <felipekde@gmail.com>

alamb · 2025-02-12T01:06:10Z

I ran some build benchmarks on a GPC machine and I conclude this change does not improve the build timings

Building main....
+ rm -rf target
+ cargo build --release --timings --lib --quiet
real 4m28.125s
user 36m43.568s
sys 1m19.993s
+ rm -rf target
+ cargo build --release --timings --lib --quiet
real 4m32.743s
user 36m47.328s
sys 1m19.786s
+ rm -rf target
+ cargo build --release --timings --lib --quiet
real 4m33.336s
user 36m50.926s
sys 1m19.246s
+ git reset --hard
HEAD is now at cb5d42e135 Disable extended tests (#14604)
+ git checkout alamb/make_expr_smaller
Switched to branch 'alamb/make_expr_smaller'
Your branch is up to date with 'alamb/alamb/make_expr_smaller'.
+ echo 'Building make_expr_smaller....'
Building make_expr_smaller....
+ rm -rf target
+ cargo build --release --timings --lib --quiet
real 4m30.234s
user 37m15.027s
sys 1m19.688s
+ rm -rf target
+ cargo build --release --timings --lib --quiet
real 4m30.580s
user 37m6.178s
sys 1m19.885s
+ rm -rf target
+ cargo build --release --timings --lib --quiet
real 4m32.293s
user 36m54.063s
sys 1m19.557s
alamb@aal-dev:~/arrow-datafusion$

crepererum · 2025-02-17T09:50:55Z

TBH I would be surprised if the build performance is affected by a few memcpy calls. I think the question here is rather if our runtime performance changes, potentially due to less stress on the memory allocator.

alamb · 2025-02-19T14:15:25Z

I'll try and find time to run some sql planning benchmarks

github-actions · 2025-04-22T02:07:06Z

Thank you for your contribution. Unfortunately, this pull request is stale because it has been open 60 days with no activity. Please remove the stale label or comment or this will be closed in 7 days.

alamb · 2025-05-16T19:14:12Z

🤖 ./gh_compare_branch_bench.shBenchmark 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 alamb/make_expr_smaller (279f074) to cb5d42ediff
BENCH_NAME=sql_planner
BENCH_COMMAND=cargo bench --bench sql_planner
BENCH_FILTER=
BENCH_BRANCH_NAME=alamb_make_expr_smaller
Results will be posted here when complete

alamb · 2025-05-16T20:11:18Z

🤖: Benchmark completed

Details

group alamb_make_expr_smaller main
----- ----------------------- ----
logical_aggregate_with_join 1.00 1486.0±16.67µs ? ?/sec 1.02 1519.2±14.28µs ? ?/sec
logical_select_all_from_1000 1.00 4.5±0.02ms ? ?/sec 1.01 4.6±0.03ms ? ?/sec
logical_select_one_from_700 1.00 1182.9±12.15µs ? ?/sec 1.03 1213.7±19.22µs ? ?/sec
logical_trivial_join_high_numbered_columns 1.00 1148.5±16.11µs ? ?/sec 1.02 1171.8±12.08µs ? ?/sec
logical_trivial_join_low_numbered_columns 1.00 1140.3±19.70µs ? ?/sec 1.02 1164.6±17.11µs ? ?/sec
physical_intersection 1.00 2.4±0.02ms ? ?/sec 1.02 2.5±0.02ms ? ?/sec
physical_join_consider_sort 1.00 3.2±0.02ms ? ?/sec 1.02 3.3±0.02ms ? ?/sec
physical_join_distinct 1.00 1148.5±15.20µs ? ?/sec 1.00 1150.0±10.96µs ? ?/sec
physical_many_self_joins 1.01 17.5±0.12ms ? ?/sec 1.00 17.3±0.18ms ? ?/sec
physical_plan_clickbench_all 1.00 285.4±4.98ms ? ?/sec 1.14 324.2±6.52ms ? ?/sec
physical_plan_clickbench_q1 1.00 4.6±0.12ms ? ?/sec 1.10 5.0±0.13ms ? ?/sec
physical_plan_clickbench_q10 1.00 5.7±0.14ms ? ?/sec 1.16 6.5±0.42ms ? ?/sec
physical_plan_clickbench_q11 1.00 5.8±0.22ms ? ?/sec 1.13 6.6±0.21ms ? ?/sec
physical_plan_clickbench_q12 1.00 6.0±0.17ms ? ?/sec 1.12 6.7±0.24ms ? ?/sec
physical_plan_clickbench_q13 1.00 5.5±0.11ms ? ?/sec 1.12 6.1±0.17ms ? ?/sec
physical_plan_clickbench_q14 1.00 5.8±0.17ms ? ?/sec 1.12 6.5±0.15ms ? ?/sec
physical_plan_clickbench_q15 1.00 5.6±0.14ms ? ?/sec 1.13 6.3±0.18ms ? ?/sec
physical_plan_clickbench_q16 1.00 5.1±0.18ms ? ?/sec 1.09 5.5±0.19ms ? ?/sec
physical_plan_clickbench_q17 1.00 5.0±0.13ms ? ?/sec 1.14 5.8±0.15ms ? ?/sec
physical_plan_clickbench_q18 1.00 4.9±0.13ms ? ?/sec 1.12 5.4±0.15ms ? ?/sec
physical_plan_clickbench_q19 1.00 5.7±0.16ms ? ?/sec 1.12 6.4±0.14ms ? ?/sec
physical_plan_clickbench_q2 1.00 4.9±0.13ms ? ?/sec 1.10 5.4±0.12ms ? ?/sec
physical_plan_clickbench_q20 1.00 4.5±0.10ms ? ?/sec 1.11 5.0±0.14ms ? ?/sec
physical_plan_clickbench_q21 1.00 4.9±0.15ms ? ?/sec 1.13 5.5±0.20ms ? ?/sec
physical_plan_clickbench_q22 1.00 5.7±0.17ms ? ?/sec 1.15 6.6±0.17ms ? ?/sec
physical_plan_clickbench_q23 1.00 6.2±0.16ms ? ?/sec 1.15 7.1±0.17ms ? ?/sec
physical_plan_clickbench_q24 1.00 6.9±0.18ms ? ?/sec 1.12 7.8±0.18ms ? ?/sec
physical_plan_clickbench_q25 1.00 5.1±0.13ms ? ?/sec 1.14 5.8±0.12ms ? ?/sec
physical_plan_clickbench_q26 1.00 4.8±0.13ms ? ?/sec 1.12 5.4±0.13ms ? ?/sec
physical_plan_clickbench_q27 1.00 5.2±0.16ms ? ?/sec 1.11 5.8±0.19ms ? ?/sec
physical_plan_clickbench_q28 1.00 5.9±0.18ms ? ?/sec 1.14 6.7±0.20ms ? ?/sec
physical_plan_clickbench_q29 1.00 6.9±0.15ms ? ?/sec 1.17 8.1±0.27ms ? ?/sec
physical_plan_clickbench_q3 1.00 4.9±0.11ms ? ?/sec 1.10 5.3±0.12ms ? ?/sec
physical_plan_clickbench_q30 1.00 19.5±0.32ms ? ?/sec 1.15 22.3±0.25ms ? ?/sec
physical_plan_clickbench_q31 1.00 6.1±0.17ms ? ?/sec 1.15 7.0±0.20ms ? ?/sec
physical_plan_clickbench_q32 1.00 6.0±0.14ms ? ?/sec 1.15 7.0±0.18ms ? ?/sec
physical_plan_clickbench_q33 1.00 5.6±0.17ms ? ?/sec 1.13 6.3±0.17ms ? ?/sec
physical_plan_clickbench_q34 1.00 5.1±0.15ms ? ?/sec 1.15 5.9±0.16ms ? ?/sec
physical_plan_clickbench_q35 1.00 5.2±0.15ms ? ?/sec 1.14 6.0±0.16ms ? ?/sec
physical_plan_clickbench_q36 1.00 6.4±0.15ms ? ?/sec 1.15 7.3±0.16ms ? ?/sec
physical_plan_clickbench_q37 1.00 6.4±0.16ms ? ?/sec 1.17 7.5±0.15ms ? ?/sec
physical_plan_clickbench_q38 1.00 6.3±0.15ms ? ?/sec 1.16 7.3±0.16ms ? ?/sec
physical_plan_clickbench_q39 1.00 6.0±0.17ms ? ?/sec 1.15 6.8±0.14ms ? ?/sec
physical_plan_clickbench_q4 1.00 4.6±0.12ms ? ?/sec 1.09 5.0±0.13ms ? ?/sec
physical_plan_clickbench_q40 1.00 6.6±0.22ms ? ?/sec 1.16 7.6±0.19ms ? ?/sec
physical_plan_clickbench_q41 1.00 6.3±0.18ms ? ?/sec 1.16 7.3±0.15ms ? ?/sec
physical_plan_clickbench_q42 1.00 6.2±0.17ms ? ?/sec 1.14 7.0±0.21ms ? ?/sec
physical_plan_clickbench_q43 1.00 6.2±0.19ms ? ?/sec 1.18 7.3±0.22ms ? ?/sec
physical_plan_clickbench_q44 1.00 4.7±0.13ms ? ?/sec 1.13 5.3±0.17ms ? ?/sec
physical_plan_clickbench_q45 1.00 4.7±0.14ms ? ?/sec 1.10 5.2±0.11ms ? ?/sec
physical_plan_clickbench_q46 1.00 5.4±0.22ms ? ?/sec 1.10 5.9±0.16ms ? ?/sec
physical_plan_clickbench_q47 1.00 6.1±0.18ms ? ?/sec 1.09 6.6±0.19ms ? ?/sec
physical_plan_clickbench_q48 1.00 6.6±0.19ms ? ?/sec 1.12 7.4±0.20ms ? ?/sec
physical_plan_clickbench_q49 1.00 6.8±0.17ms ? ?/sec 1.14 7.8±0.23ms ? ?/sec
physical_plan_clickbench_q5 1.00 4.8±0.12ms ? ?/sec 1.11 5.4±0.12ms ? ?/sec
physical_plan_clickbench_q6 1.00 4.9±0.14ms ? ?/sec 1.08 5.3±0.14ms ? ?/sec
physical_plan_clickbench_q7 1.00 5.3±0.11ms ? ?/sec 1.13 6.0±0.19ms ? ?/sec
physical_plan_clickbench_q8 1.00 5.1±0.14ms ? ?/sec 1.11 5.7±0.18ms ? ?/sec
physical_plan_clickbench_q9 1.00 5.4±0.14ms ? ?/sec 1.13 6.1±0.23ms ? ?/sec
physical_plan_tpcds_all 1.00 1333.8±11.64ms ? ?/sec 1.10 1466.0±19.74ms ? ?/sec
physical_plan_tpch_all 1.00 85.9±0.70ms ? ?/sec 1.11 95.5±0.88ms ? ?/sec
physical_plan_tpch_q1 1.00 3.1±0.03ms ? ?/sec 1.10 3.4±0.02ms ? ?/sec
physical_plan_tpch_q10 1.00 4.4±0.05ms ? ?/sec 1.06 4.7±0.04ms ? ?/sec
physical_plan_tpch_q11 1.00 3.9±0.06ms ? ?/sec 1.09 4.2±0.05ms ? ?/sec
physical_plan_tpch_q12 1.00 2.9±0.03ms ? ?/sec 1.11 3.3±0.03ms ? ?/sec
physical_plan_tpch_q13 1.00 2.4±0.03ms ? ?/sec 1.08 2.6±0.04ms ? ?/sec
physical_plan_tpch_q14 1.00 2.7±0.03ms ? ?/sec 1.09 3.0±0.03ms ? ?/sec
physical_plan_tpch_q16 1.00 3.6±0.06ms ? ?/sec 1.11 4.0±0.03ms ? ?/sec
physical_plan_tpch_q17 1.00 3.5±0.03ms ? ?/sec 1.10 3.9±0.04ms ? ?/sec
physical_plan_tpch_q18 1.00 4.0±0.03ms ? ?/sec 1.09 4.3±0.04ms ? ?/sec
physical_plan_tpch_q19 1.00 5.2±0.04ms ? ?/sec 1.17 6.0±0.04ms ? ?/sec
physical_plan_tpch_q2 1.00 7.3±0.10ms ? ?/sec 1.08 7.9±0.08ms ? ?/sec
physical_plan_tpch_q20 1.00 4.5±0.05ms ? ?/sec 1.10 5.0±0.04ms ? ?/sec
physical_plan_tpch_q21 1.00 5.7±0.05ms ? ?/sec 1.10 6.3±0.05ms ? ?/sec
physical_plan_tpch_q22 1.00 3.4±0.02ms ? ?/sec 1.12 3.8±0.03ms ? ?/sec
physical_plan_tpch_q3 1.00 3.2±0.03ms ? ?/sec 1.08 3.4±0.03ms ? ?/sec
physical_plan_tpch_q4 1.00 2.5±0.02ms ? ?/sec 1.08 2.7±0.03ms ? ?/sec
physical_plan_tpch_q5 1.00 4.4±0.05ms ? ?/sec 1.06 4.6±0.03ms ? ?/sec
physical_plan_tpch_q6 1.00 1799.7±24.60µs ? ?/sec 1.07 1931.1±14.84µs ? ?/sec
physical_plan_tpch_q7 1.00 5.5±0.05ms ? ?/sec 1.09 6.0±0.07ms ? ?/sec
physical_plan_tpch_q8 1.00 6.6±0.06ms ? ?/sec 1.09 7.2±0.09ms ? ?/sec
physical_plan_tpch_q9 1.00 5.2±0.05ms ? ?/sec 1.07 5.6±0.04ms ? ?/sec
physical_select_aggregates_from_200 1.00 30.7±0.19ms ? ?/sec 1.08 33.1±0.21ms ? ?/sec
physical_select_all_from_1000 1.00 41.5±0.17ms ? ?/sec 1.02 42.5±0.20ms ? ?/sec
physical_select_one_from_700 1.00 3.3±0.02ms ? ?/sec 1.01 3.4±0.02ms ? ?/sec
physical_sorted_union_orderby 1.00 113.2±0.52ms ? ?/sec 1.05 119.2±0.60ms ? ?/sec
physical_theta_join_consider_sort 1.00 3.7±0.06ms ? ?/sec 1.01 3.7±0.04ms ? ?/sec
physical_unnest_to_join 1.00 3.3±0.03ms ? ?/sec 1.01 3.4±0.02ms ? ?/sec
with_param_values_many_columns 1.00 123.3±0.94µs ? ?/sec 1.28 157.8±1.29µs ? ?/sec

crepererum · 2025-05-26T09:16:12Z

If these benchmarks are reliable, then this would be a pretty neat improvement.

alamb · 2025-05-26T10:05:50Z

If these benchmarks are reliable, then this would be a pretty neat improvement.

Oh, man I missed the benchmark results (I forgot to check them).

I'll ressurect this PR over the next week or so and rerun benchmarks

Thank you very much for the shout @crepererum

alamb · 2025-05-27T18:10:52Z

Filed #16199 to track

Add a test for Expr size
2797cf7

github-actionsBot added the logical-expr Logical plan and expressions label Jan 29, 2025

findepi reviewed Jan 29, 2025
View reviewed changes

Comment threaddatafusion/expr/src/expr.rs

findepi reviewed Jan 29, 2025
View reviewed changes

Reduce size of Expr
f5df767

alamb force-pushed the alamb/make_expr_smaller branch from b25ba57 to f5df767Compare January 29, 2025 21:42

github-actionsBot added sql SQL Planner optimizer Optimizer rules core Core DataFusion crate substrait Changes to the substrait crate proto Related to proto crate labels Jan 29, 2025

alamb commented Jan 29, 2025
View reviewed changes

Comment threaddatafusion/expr/src/expr.rs Outdated

alamb mentioned this pull request Jan 29, 2025
Build time regression #14256
Closed

alamband others added 2 commits January 30, 2025 09:08

Update datafusion/expr/src/expr.rs
ec8b6ad
Co-authored-by: Felipe Oliveira Carvalho <felipekde@gmail.com>

Merge remote-tracking branch 'apache/main' into alamb/make_expr_smaller
279f074

github-actionsBot added the Stale PR has not had any activity for some time label Apr 22, 2025

github-actionsBot closed this May 1, 2025

alamb mentioned this pull request May 16, 2025
chore(CI) Upgrade toolchain to Rust-1.87 #16068
Merged

Dandandan reopened this May 26, 2025

github-actionsBot removed the Stale PR has not had any activity for some time label May 27, 2025

This was referenced May 27, 2025
Weekly Plan: Andrew Lamb 2025-05-26 #16198
Closed
Reduce size of Expr struct #16199
Closed

alamb closed this May 27, 2025

Navigation Menu

Uh oh!

Reduce size of `Expr` struct - #14366

Reduce size of `Expr` struct#14366
alamb wants to merge 4 commits into
apache:mainfrom
alamb:alamb/make_expr_smaller

alamb commented Jan 29, 2025•
edited
Loading

Uh oh!

Uh oh!

findepiJan 29, 2025

Uh oh!

Uh oh!

alamb commented Feb 12, 2025

Uh oh!

crepererum commented Feb 17, 2025

Uh oh!

alamb commented Feb 19, 2025

Uh oh!

github-actionsBot commented Apr 22, 2025

Uh oh!

alamb commented May 16, 2025

Uh oh!

alamb commented May 16, 2025

Uh oh!

crepererum commented May 26, 2025

Uh oh!

alamb commented May 26, 2025

Uh oh!

alamb commented May 27, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Uh oh!

Conversation

alamb commented Jan 29, 2025• 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!

findepiJan 29, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

alamb commented Feb 12, 2025

Uh oh!

crepererum commented Feb 17, 2025

Uh oh!

alamb commented Feb 19, 2025

Uh oh!

github-actionsBot commented Apr 22, 2025

Uh oh!

alamb commented May 16, 2025

Uh oh!

alamb commented May 16, 2025

Uh oh!

crepererum commented May 26, 2025

Uh oh!

alamb commented May 26, 2025

Uh oh!

alamb commented May 27, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

alamb commented Jan 29, 2025•
edited
Loading