Faster strpos() string function for ASCII-only case #12401

goldmedal · 2024-09-09T18:27:22Z

Which issue does this PR close?

Rationale for this change

For ASCII-only cases, I tried to use the slice window to find the position instead of the string pattern matching.
We can find that the ASCII-only StringViewArray case performs 50% faster than before. ~~However, I'm not really sure why the normal string array can't be improved. 🤔~~

What changes are included in this PR?

If both arguments are ASCII-only, we can perform this faster method.

The benchmark result:

group                                        after                                   before
-----                                        -----                                   ------
strpos_StringArray_ascii_str_len_128         1.00  1585.2±49.78µs        ? ?/sec     1.74      2.8±0.07ms        ? ?/sec
strpos_StringArray_ascii_str_len_32          1.00   558.2±24.13µs        ? ?/sec     2.22  1240.1±51.67µs        ? ?/sec
strpos_StringArray_ascii_str_len_4096        1.00     47.3±2.79ms        ? ?/sec     1.88     89.0±1.68ms        ? ?/sec
strpos_StringArray_ascii_str_len_8           1.00    199.8±6.46µs        ? ?/sec     2.42   483.6±16.36µs        ? ?/sec
strpos_StringArray_utf8_str_len_128          1.00      4.9±0.14ms        ? ?/sec     1.00      4.9±0.12ms        ? ?/sec
strpos_StringArray_utf8_str_len_32           1.00  1852.6±47.05µs        ? ?/sec     1.01  1875.0±66.89µs        ? ?/sec
strpos_StringArray_utf8_str_len_4096         1.00    136.1±2.14ms        ? ?/sec     1.01    136.8±2.63ms        ? ?/sec
strpos_StringArray_utf8_str_len_8            1.00   831.5±43.31µs        ? ?/sec     1.00   830.9±35.23µs        ? ?/sec
strpos_StringViewArray_ascii_str_len_128     1.00  1700.8±63.55µs        ? ?/sec     1.66      2.8±0.12ms        ? ?/sec
strpos_StringViewArray_ascii_str_len_32      1.00   626.3±27.34µs        ? ?/sec     2.02  1264.6±88.59µs        ? ?/sec
strpos_StringViewArray_ascii_str_len_4096    1.00     47.4±1.11ms        ? ?/sec     1.88     88.9±1.96ms        ? ?/sec
strpos_StringViewArray_ascii_str_len_8       1.00   275.8±12.18µs        ? ?/sec     1.69   467.0±17.50µs        ? ?/sec
strpos_StringViewArray_utf8_str_len_128      1.00      4.9±0.14ms        ? ?/sec     1.00      4.9±0.21ms        ? ?/sec
strpos_StringViewArray_utf8_str_len_32       1.00  1898.4±121.22µs        ? ?/sec    1.00  1889.9±51.64µs        ? ?/sec
strpos_StringViewArray_utf8_str_len_4096     1.00    136.5±2.14ms        ? ?/sec     1.00    136.4±1.78ms        ? ?/sec
strpos_StringViewArray_utf8_str_len_8        1.00   848.3±35.50µs        ? ?/sec     1.02  861.6±172.97µs        ? ?/sec

Are these changes tested?

yes

Are there any user-facing changes?

2010YOUY01

Great 🚀 The implementation looks good to me

However, I'm not really sure why the normal string array can't be improved. 🤔

I think it's not executed and triggered some error, see benchmark StringArray cases only take several ns, and StringViewArray cases are at least several us
I guess it's because in micro bench, arguments are of different physical string type and break the function implementation (strpos(string_col, string_view_col))

~~With #12415 this kind of bug can be more throughly tested~~
Update: It's not a bug for SQL API, and this can only be triggered if we use raw expression API as in the benchmark

2010YOUY01 · 2024-09-11T10:47:05Z

datafusion/functions/src/unicode/strpos.rs

 {
-    let string_iter = ArrayIter::new(string_array);
-    let substring_iter = ArrayIter::new(substring_array);
+    let ascii_only = string_array.is_ascii() && substring_array.is_ascii();


Suggested change

let ascii_only = string_array.is_ascii() && substring_array.is_ascii();

let ascii_only = substring_array.is_ascii() && string_array.is_ascii() ;

I think we can check substring first, since it's cheaper in common cases and we can then shortcircuit checking string_array in some case

Nice suggestion! I did the benchmark again. The performance is improved!

group before after ----- ----- ------- strpos_StringArray_ascii_str_len_128 1.06 370.5±9.46ns ? ?/sec 1.00 348.6±11.38ns ? ?/sec strpos_StringArray_ascii_str_len_32 1.07 372.5±10.23ns ? ?/sec 1.00 346.9±9.45ns ? ?/sec strpos_StringArray_ascii_str_len_4096 1.08 378.4±12.05ns ? ?/sec 1.00 349.4±16.83ns ? ?/sec strpos_StringArray_ascii_str_len_8 1.07 371.4±14.53ns ? ?/sec 1.00 346.1±28.44ns ? ?/sec strpos_StringArray_utf8_str_len_128 1.06 377.2±18.04ns ? ?/sec 1.00 356.4±21.92ns ? ?/sec strpos_StringArray_utf8_str_len_32 1.08 374.9±34.32ns ? ?/sec 1.00 345.8±11.05ns ? ?/sec strpos_StringArray_utf8_str_len_4096 1.09 381.6±16.68ns ? ?/sec 1.00 351.4±23.14ns ? ?/sec strpos_StringArray_utf8_str_len_8 1.09 372.9±20.83ns ? ?/sec 1.00 343.3±11.78ns ? ?/sec strpos_StringViewArray_ascii_str_len_128 1.79 3.2±0.15ms ? ?/sec 1.00 1763.8±44.16µs ? ?/sec strpos_StringViewArray_ascii_str_len_32 1.03 648.7±18.27µs ? ?/sec 1.00 628.1±21.69µs ? ?/sec strpos_StringViewArray_ascii_str_len_4096 1.24 62.8±7.27ms ? ?/sec 1.00 50.5±2.05ms ? ?/sec strpos_StringViewArray_ascii_str_len_8 1.00 280.2±10.44µs ? ?/sec 1.05 294.8±42.47µs ? ?/sec strpos_StringViewArray_utf8_str_len_128 1.03 5.3±0.13ms ? ?/sec 1.00 5.1±0.12ms ? ?/sec strpos_StringViewArray_utf8_str_len_32 1.01 1961.9±57.34µs ? ?/sec 1.00 1944.1±73.19µs ? ?/sec strpos_StringViewArray_utf8_str_len_4096 1.03 147.4±9.24ms ? ?/sec 1.00 142.5±3.61ms ? ?/sec strpos_StringViewArray_utf8_str_len_8 1.01 874.5±26.02µs ? ?/sec 1.00 863.4±40.68µs ? ?/sec

goldmedal · 2024-09-11T16:04:35Z

I think it's not executed and triggered some error, see benchmark StringArray cases only take several ns, and StringViewArray cases are at least several us I guess it's because in micro bench, arguments are of different physical string type and break the function implementation (strpos(string_col, string_view_col))

You're right. There are some errors here. I ran the benchmark in the strpos test case and got the following error:

Error: Execution("Unsupported data type combination (Utf8, Utf8View) for function strpos")

It seems that strpos doesn't allow the second argument to be a StringView if the first one is a StringArray. I'm not sure if this case makes sense. Maybe we can file another issue for it if it's reasonable.

Then, I realized it was my mistake 😢. When I generated the substring for the StringArray, I mistakenly set its type to StringView. After fixing it, the benchmark for StringArray works fine.

goldmedal · 2024-09-11T16:32:00Z

datafusion/functions/benches/strpos.rs

+        ]
+    } else {
+        let string_array: StringArray = output_string_vec.clone().into_iter().collect();
+        let sub_string_array: StringArray =


I rebase to make all benchmark-related changes in the first commit. It was

let sub_string_array: StringViewArray = output_string_vec.clone().into_iter().collect();

After changing to StringArray, the benchmark works fine.

goldmedal · 2024-09-11T16:37:02Z

@2010YOUY01 Thank you for the review. I updated the benchmark result. After fixing the benchmark, we can find the StringArray case also be improved.

alamb

Thank you @goldmedal -- this PR looks really nice 🙏

🚀

goldmedal · 2024-09-18T01:54:50Z

Thanks @alamb @2010YOUY01

github-actions bot added the functions label Sep 9, 2024

goldmedal marked this pull request as ready for review September 10, 2024 01:52

2010YOUY01 approved these changes Sep 11, 2024

View reviewed changes

alamb mentioned this pull request Sep 11, 2024

DataFusion weekly project plan (Andrew Lamb) - Sep 9, 2024 #12391

Closed

5 tasks

goldmedal added 4 commits September 12, 2024 00:07

add strpos benchmark

9c60c63

add faster path for strpos in ascii-only case

4aebfe4

clippy

a854dd4

compare substring first

62f2307

goldmedal force-pushed the feature/12366-strpos-ascii branch from 5aae0fb to 62f2307 Compare September 11, 2024 16:28

goldmedal commented Sep 11, 2024

View reviewed changes

cargo fmt

81c3fa1

alamb mentioned this pull request Sep 16, 2024

DataFusion weekly project plan (Andrew Lamb) - Sep 16, 2024 #12494

Closed

8 tasks

alamb approved these changes Sep 16, 2024

View reviewed changes

alamb merged commit 269a473 into apache:main Sep 17, 2024
26 checks passed

goldmedal deleted the feature/12366-strpos-ascii branch September 18, 2024 01:54

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Faster strpos() string function for ASCII-only case #12401

Faster strpos() string function for ASCII-only case #12401

goldmedal commented Sep 9, 2024 •

edited

Loading

2010YOUY01 left a comment •

edited

Loading

2010YOUY01 Sep 11, 2024

goldmedal Sep 11, 2024

goldmedal commented Sep 11, 2024

goldmedal Sep 11, 2024

goldmedal commented Sep 11, 2024 •

edited

Loading

alamb left a comment

goldmedal commented Sep 18, 2024

	let ascii_only = string_array.is_ascii() && substring_array.is_ascii();
	let ascii_only = substring_array.is_ascii() && string_array.is_ascii() ;

Faster strpos() string function for ASCII-only case #12401

Faster strpos() string function for ASCII-only case #12401

Conversation

goldmedal commented Sep 9, 2024 • edited Loading

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?

2010YOUY01 left a comment • edited Loading

Choose a reason for hiding this comment

2010YOUY01 Sep 11, 2024

Choose a reason for hiding this comment

goldmedal Sep 11, 2024

Choose a reason for hiding this comment

goldmedal commented Sep 11, 2024

goldmedal Sep 11, 2024

Choose a reason for hiding this comment

goldmedal commented Sep 11, 2024 • edited Loading

alamb left a comment

Choose a reason for hiding this comment

goldmedal commented Sep 18, 2024

goldmedal commented Sep 9, 2024 •

edited

Loading

2010YOUY01 left a comment •

edited

Loading

goldmedal commented Sep 11, 2024 •

edited

Loading