[Variant] Avoid collecting offset iterator #7934

codephage2020 · 2025-07-16T05:38:02Z

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.

Closes [Variant] Avoid collecting offset iterator #7901 .

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.

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.

Are these changes tested?

We typically require tests for all PRs in order to:

Prevent the code from being accidentally broken by subsequent changes
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)?

Are there any user-facing changes?

If there are user-facing changes then we may require documentation to be updated before approving the PR.

If there are any breaking changes to public APIs, please call them out.

Signed-off-by: codephage2020 <[email protected]>

codephage2020 · 2025-07-16T06:16:44Z

CC @friendlymatthew @alamb. Would it be possible for you to review this PR at your convenience? Thank you in advance.

alamb

Thank you @codephage2020 -- I went through this PR carefully and it all makes sense to me and I think it is good to merge

cc @scovich @Samyak2 and @friendlymatthew

alamb · 2025-07-16T20:03:39Z

parquet-variant/src/variant/object.rs

-                    return Err(ArrowError::InvalidArgumentError(
-                        "field names not sorted".to_string(),
-                    ));
+                let mut current_field_name = match field_ids_iter.next() {


It wasn't added in this PR, but this check for the field names being sorted doesn't seem right to me -- I thought the only requirement on an object's fields were that the field_ids were sorted (so lookup by field_id can be fast) but the corresponding names of the fields don't have to be sorted

Maybe @friendlymatthew can help

Yes, you are right. VariantEncoding.md is described like this.

Hi, this is how I understood the spec

The field ids and field offsets must be in lexicographical order of the corresponding field names in the metadata dictionary

If we have an object with a sorted dictionary, the field ids are already ordered by the lexicographical order of field names.

If we don't have a sorted dictionary, we iterate through the field ids and probe the object for the corresponding field name. If the field names are lexicographically ordered, we can also verify that the field ids are in lexicographical order.

Got it -- I re-read the spec and I was confused and agree this check is doign the right thing

The field ids and field offsets must be in lexicographical order of the corresponding field names in the metadata dictionary. However, the actual value entries do not need to be in any particular order. This implies that the field_offset values may not be monotonically increasing. For example, for the following object:

alamb · 2025-07-16T20:05:34Z

parquet-variant/src/variant/object.rs

-                // we also know all field ids are smaller than the dictionary size and in-bounds.
-                if let Some(&last_field_id) = field_ids.last() {
-                    if last_field_id >= self.metadata.dictionary_size() {
+                    if current_id >= dictionary_size {


It took me a few times to figure out why this (redundant) check was still needed

I am not sure if there is some way to refactor the loop to avoid this (perhaps by keeping previous_id: Option<u32> as you did in the loop above 🤔

No changes needed, I just figured I would point it out

Sorry, I didn't expect this to be confusing. This loop has considered many ways of writing, and in the end, I referred to https://github.com/apache/arrow-rs/blob/main/parquet-variant/src/variant/list.rs#L225-L232. I think this is an elegant implementation. Perhaps I need to add comments or use the following writing style to make the code more readable.

let mut previous_id: Option<u32> = None; for field_id in field_ids_iter { if field_id >= dictionary_size { return Err(ArrowError::InvalidArgumentError( "field id is not valid".to_string(), )); } if let Some(prev_id) = previous_id { if field_id <= prev_id { return Err(ArrowError::InvalidArgumentError( "field names not sorted".to_string(), )); } } previous_id = Some(field_id); }

alamb · 2025-07-16T20:07:04Z

🤖 ./gh_compare_arrow.sh Benchmark Script Running
Linux aal-dev 6.11.0-1016-gcp #16~24.04.1-Ubuntu SMP Wed May 28 02:40:52 UTC 2025 x86_64 x86_64 x86_64 GNU/Linux
Comparing refactor/Avoid-collecting-offset-iterator (5cbf8cc) to c40830e diff
BENCH_NAME=variant_validation
BENCH_COMMAND=cargo bench --features=arrow,async,test_common,experimental --bench variant_validation
BENCH_FILTER=
BENCH_BRANCH_NAME=refactor_Avoid-collecting-offset-iterator
Results will be posted here when complete

alamb · 2025-07-16T20:08:12Z

🤖: Benchmark completed

Details

group                               main                                   refactor_Avoid-collecting-offset-iterator
-----                               ----                                   -----------------------------------------
bench_validate_complex_object       1.07    241.0±0.51µs        ? ?/sec    1.00    224.3±0.58µs        ? ?/sec
bench_validate_large_nested_list    1.00     18.6±0.04ms        ? ?/sec    1.21     22.5±0.03ms        ? ?/sec
bench_validate_large_object         1.00     55.0±0.09ms        ? ?/sec    1.21     66.3±0.08ms        ? ?/sec

alamb · 2025-07-16T20:09:22Z

🤖 ./gh_compare_arrow.sh Benchmark Script Running
Linux aal-dev 6.11.0-1016-gcp #16~24.04.1-Ubuntu SMP Wed May 28 02:40:52 UTC 2025 x86_64 x86_64 x86_64 GNU/Linux
Comparing refactor/Avoid-collecting-offset-iterator (5cbf8cc) to c40830e diff
BENCH_NAME=variant_validation
BENCH_COMMAND=cargo bench --features=arrow,async,test_common,experimental --bench variant_validation
BENCH_FILTER=
BENCH_BRANCH_NAME=refactor_Avoid-collecting-offset-iterator
Results will be posted here when complete

alamb · 2025-07-16T20:09:44Z

🤔 interesting that the benchmark seems to imply this approach is slower than collecting in a Vec

#7934 (comment)

bench_validate_large_object         1.00     55.0±0.09ms        ? ?/sec    1.21     66.3±0.08ms        ? ?/sec

I reran to see if we can see a difference

alamb · 2025-07-16T20:10:28Z

🤖: Benchmark completed

Details

group                               main                                   refactor_Avoid-collecting-offset-iterator
-----                               ----                                   -----------------------------------------
bench_validate_complex_object       1.07    242.5±0.83µs        ? ?/sec    1.00    227.0±0.48µs        ? ?/sec
bench_validate_large_nested_list    1.01     19.0±0.03ms        ? ?/sec    1.00     18.8±0.04ms        ? ?/sec
bench_validate_large_object         1.01     54.5±0.08ms        ? ?/sec    1.00     53.8±0.09ms        ? ?/sec

alamb · 2025-07-17T13:54:20Z

Next benchmark run looks good to me

alamb · 2025-07-17T13:56:31Z

🚀

alamb · 2025-07-17T13:56:36Z

Thanks agian @codephage2020 and @friendlymatthew

scovich · 2025-07-17T19:07:24Z

parquet-variant/src/variant/metadata.rs

-                let are_offsets_monotonic = offsets.is_sorted_by(|a, b| a < b);
-                if !are_offsets_monotonic {
-                    return Err(ArrowError::InvalidArgumentError(
-                        "offsets not monotonically increasing".to_string(),
-                    ));
+                for next_offset in offsets_iter {
+                    if next_offset <= current_offset {
+                        return Err(ArrowError::InvalidArgumentError(
+                            "offsets not monotonically increasing".to_string(),
+                        ));
+                    }
+                    current_offset = next_offset;


Belated comment but -- AFAIK this specific snippet of original code was perfectly fine? This PR would have just changed it from https://doc.rust-lang.org/std/primitive.slice.html#method.is_sorted_by to Iterator::is_sorted_by?

Ah, but this PR unconditionally consumes the first offset outside the if/else; that part would need to move inside the if so this else can keep doing what it always did.

Maybe worth a follow on PR

Oh! There's also a bug: Neither variant nor JSON forbids the empty string as a field name. So we have to allow empty ranges. In case the dictionary is sorted, "" compares less-than every other string, and would be the first string; the sortedness check would naturally catch any later empty strings as breaking the sort order.

I'll throw up a quick PR for this.

[Variant] VariantMetadata is allowed to contain the empty string #7956

friendlymatthew · 2025-07-18T11:47:07Z

parquet-variant/src/variant/object.rs

+                    let next_field_name = self.metadata.get(field_id)?;
+
+                    if let Some(current_name) = current_field_name {
+                        if next_field_name <= current_name {


Hi, I caught this post-merge but in this branch we can't assume the field names in the metadata dictionary are unique. So when we perform the probing mentioned in https://github.com/apache/arrow-rs/pull/7934/files#r2213356637, we should only check if next_field_name < current_name.

fixed in #7961

# Which issue does this PR close? - Follow-up to #7901 # Rationale for this change - #7934 Introduced a minor regression, in (accidentally?) forbidding the empty string as a dictionary key. Fix the bug and simplify the code a bit further while we're at it. # What changes are included in this PR? Revert the unsorted dictionary check back to what it had been (it just uses `Iterator::is_sorted_by` now, instead of `primitive.slice::is_sorted_by`). Remove the redundant offset monotonicity check from the ordered dictionary path, relying on the fact that string slice extraction will anyway fail if the offsets are not monotonic. Improve the error message now that it does double duty. # Are these changes tested? New unit tests for dictionaries containing the empty string. As a side effect, we now have at least a little coverage for sorted dictionaries -- somehow, I couldn't find any existing unit test that creates a sorted dictionary?? # Are there any user-facing changes? No --------- Co-authored-by: Andrew Lamb <[email protected]>

# Rationale for this change If a variant has an unsorted dictionary, you can't assume fields are unique nor ordered by name. This PR updates the logical equality check among `VariantMetadata` to properly handle this case. - Closes #7952 It also fixes a bug in #7934 where we do a uniqueness check when probing an unsorted dictionary --------- Co-authored-by: Andrew Lamb <[email protected]>

perf(parquet-variant): optimize dictionary and field id validation

8d3e802

Signed-off-by: codephage2020 <[email protected]>

github-actions bot added the parquet Changes to the parquet crate label Jul 16, 2025

codephage2020 added 2 commits July 16, 2025 13:40

fmt&clippy

1c8025f

Signed-off-by: codephage2020 <[email protected]>

(parquet-variant)pass all tests

5cbf8cc

Signed-off-by: codephage2020 <[email protected]>

codephage2020 marked this pull request as ready for review July 16, 2025 06:08

alamb approved these changes Jul 16, 2025

View reviewed changes

alamb merged commit 7089786 into apache:main Jul 17, 2025
12 checks passed

codephage2020 deleted the refactor/Avoid-collecting-offset-iterator branch July 17, 2025 14:02

scovich reviewed Jul 17, 2025

View reviewed changes

scovich mentioned this pull request Jul 17, 2025

[Variant] VariantMetadata is allowed to contain the empty string #7956

Merged

friendlymatthew mentioned this pull request Jul 18, 2025

[Variant] Revisit VariantMetadata and Object equality #7961

Merged

friendlymatthew reviewed Jul 18, 2025

View reviewed changes

[Variant] Avoid collecting offset iterator #7934

[Variant] Avoid collecting offset iterator #7934

Conversation

codephage2020 commented Jul 16, 2025

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!

codephage2020 commented Jul 16, 2025

Uh oh!

alamb left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

alamb Jul 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

codephage2020 Jul 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

alamb commented Jul 16, 2025

Uh oh!

alamb commented Jul 16, 2025

Uh oh!

alamb commented Jul 16, 2025

Uh oh!

alamb commented Jul 16, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

alamb commented Jul 16, 2025

Uh oh!

alamb commented Jul 17, 2025

Uh oh!

alamb commented Jul 17, 2025

Uh oh!

alamb commented Jul 17, 2025

Uh oh!

Uh oh!

scovich Jul 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

scovich Jul 17, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

alamb Jul 17, 2025 •

edited

Loading

codephage2020 Jul 17, 2025 •

edited

Loading

alamb commented Jul 16, 2025 •

edited

Loading

scovich Jul 17, 2025 •

edited

Loading

scovich Jul 17, 2025 •

edited

Loading