GH-41560: [C++] ChunkResolver: Implement ResolveMany and add unit tests #41561

felipecrv · 2024-05-06T22:20:19Z

Rationale for this change

I want ResolveMany to support me in the implementation of Take that doesn't Concatenate all the chunks from a ChunkedArray values parameter.

What changes are included in this PR?

Implementation of ChunkResolver::ResolveMany()
Addition of missing unit tests for ChunkResolver

Are these changes tested?

Yes. By new unit tests.

Are there any user-facing changes?

No. ChunkResolver is an internal API at the moment (see #34535 for future plans).

GitHub Issue: [C++] Add ChunkResolver::ResolveMany() API #41560

github-actions · 2024-05-06T22:20:43Z

⚠️ GitHub issue #41560 has been automatically assigned in GitHub to PR creator.

pitrou

There seems to be some (unwarranted? gratuitous?) complexity here. Also, it's not obvious why this would be better than calling Resolve multiple times in a row.

pitrou · 2024-05-13T13:35:33Z

cpp/src/arrow/chunk_resolver.h

  int64_t index_in_chunk = 0;
+
+  /// \brief Create a ChunkLocation without asserting any preconditions.
+  static ChunkLocation Forge(int64_t chunk_index, int64_t index_in_chunk) {


Can we avoid inventing terminology that's currently not used in the project?

What's your actionable suggestion? I'm trying to make this API less error-prone. MakeUnsafe?

Is there an actual need for a checked constructor?

It adds a layer of safety and by poisoning the index_in_chunk value when chunk_index is invalid, I (and other users of the library) are more likely to detect unguarded use of chunk_index this way.

I don't understand how such unguarded uses could occur. The primary use case for ChunkLocation is as a return value for ChunkResolver. Unless the user has passed an invalid logical index, ChunkResolver should always return valid ChunkLocations.

cpp/src/arrow/chunk_resolver.h

cpp/src/arrow/chunk_resolver.cc

cpp/src/arrow/compute/kernels/vector_sort.cc

felipecrv · 2024-05-13T23:37:07Z

There seems to be some (unwarranted? gratuitous?) complexity here.

You can comment on what you think is "unwarranted" and "gratuitous" complexity. This kind of feedback is demotivating and not actionable. I try to simplify the code before opening a PR instead of pushing the first thing that works.

Also, it's not obvious why this would be better than calling Resolve multiple times in a row.

I tried that in the Take kernels, but things got messy and ugly because:

Resolve and ChunkLocation use int64_t for everything while Take kernels use unsigned integer of all widths. The loops would be full of unsafe casts with confusing safety justifications.
Take kernels have complex loops handling different cases (the 4 combinations of nullability of the values and indices), so Resolve would be called (inlined) in 4 branches per kernel loop specialization.
Ensuring Resolve gets called only once per index is tricky because in the Take loops you need to resolve chunks for reading values and validity bits from multiple functions if you modularize the loop. By having a buffer with all the chunks resolved at once, code that performs these loads can be more cleanly separated and accept an index array position as parameter instead of a ChunkLocation
That makes the code that deals with ChunkedArray look very similar to the code that handles flat arrays — the ChunkedArray version needs a preparation pass to resolve the chunks
By having this chunk-resolution pass, I can make Take loops for ChunkedArray of nested types more easily

In addition to that, ResolveMany can be further optimized with different search strategies (e.g. branchless and interleaved binary searches), the step that calculates index_in_chunk can be vectorized. All of that without contributing to binary size of all the numerous Take specializations that will rely on it because ResolveManyImpl is not inlined.

pitrou · 2024-05-14T12:25:10Z

You can comment on what you think is "unwarranted" and "gratuitous" complexity.

I did post specific comments :-)

I should have done this renaming in a previous PR. Correcting it now.

After learning that GCC can't compile the previous code. Well... this is simpler now.

pitrou

Sorry for the number of back-and-forths. Here are a number of followup questions and comments, but feel free to merge in any case.

pitrou · 2024-05-15T14:40:16Z

cpp/src/arrow/chunk_resolver.h

+
+  /// \brief Resolve `n` logical indices to chunk indices.
+  ///
+  /// \pre 0 <= logical_index_vec[i] < n (for well-defined and valid chunk index results)


You mean logical_array_length() rather than n?

pitrou · 2024-05-15T14:40:37Z

cpp/src/arrow/chunk_resolver.h

+  ///
+  /// \pre 0 <= logical_index_vec[i] < n (for well-defined and valid chunk index results)
+  /// \pre out_chunk_index_vec has space for `n_indices`
+  /// \post chunk_hint in [0, chunks.size()]


Isn't it rather a pre-condition?

pitrou · 2024-05-15T14:42:15Z

cpp/src/arrow/chunk_resolver.h

+  /// \param n_indices The number of logical indices to resolve
+  /// \param logical_index_vec The logical indices to resolve
+  /// \param out_chunk_index_vec The output array where the chunk indices will be written
+  /// \param chunk_hint 0 or the last chunk_index produced by ResolveMany


Meaning the caller is supposed to pass the value of out_chunk_index_vec[n_indices - 1] from the previous call to ResolveMany?

Yes. That's the plan. Or 0 if n_indices == 0. That's why I didn't want to write the formula.

pitrou · 2024-05-15T14:44:56Z

cpp/src/arrow/chunk_resolver.h

+  /// \param chunk_hint 0 or the last chunk_index produced by ResolveMany
+  /// \param out_index_in_chunk_vec If not NULLPTR, the output array where the
+  ///                               within-chunk indices will be written
+  /// \return false iff chunks.size() > std::numeric_limits<IndexType>::max()


I'm assuming ResolveMany will be invoked in batches? This condition doesn't need to be checked in each ResolveMany call, only once for each call to the larger operation (such as Take). That's not necessarily a problem.

I'm assuming ResolveMany will be invoked in batches?

It could be. I'm not doing it like that for Take at the moment. It's a very predictable branch though and I'm allowing it to be inlined at the caller.

pitrou · 2024-05-15T14:48:10Z

cpp/src/arrow/chunk_resolver.h

+      //
+      // Negative logical indices can become large values when cast to unsigned, but
+      // they are gracefully handled by ResolveManyImpl. Although both the chunk index
+      // and the index in chunk values will be undefined in these cases.


That doesn't seem to be the case: if a logical index is negative, its unsigned counterpart will be out of bounds and the corresponding out_chunk_index_vec value will therefore be equal to chunks_.size().

~~You're right. The overflow check guarantees it's impossible for negative logical indices to become valid indices.~~

Not really. INT8_MIN becomes 128 when cast to uint8_t, -1 becomes 255, so depending on the chunks, they won't be an out-of-bounds logical indices.

I'm tweaking the comment here and improving the tests.

conbench-apache-arrow · 2024-05-16T04:15:51Z

After merging your PR, Conbench analyzed the 7 benchmarking runs that have been run so far on merge-commit e04f5b4.

There were 7 benchmark results indicating a performance regression:

Commit Run on ursa-i9-9960x at 2024-05-16 00:13:14Z
- dataset-selectivity (Python) with dataset=nyctaxi_multi_ipc_s3, selectivity=10%
- dataset-selectivity (Python) with dataset=nyctaxi_multi_ipc_s3, selectivity=100%
and 5 more (see the report linked below)

The full Conbench report has more details. It also includes information about 16 possible false positives for unstable benchmarks that are known to sometimes produce them.

…it tests (apache#41561) ### Rationale for this change I want `ResolveMany` to support me in the implementation of `Take` that doesn't `Concatenate` all the chunks from a `ChunkedArray` `values` parameter. ### What changes are included in this PR? - Implementation of `ChunkResolver::ResolveMany()` - Addition of missing unit tests for `ChunkResolver` ### Are these changes tested? Yes. By new unit tests. ### Are there any user-facing changes? No. `ChunkResolver` is an internal API at the moment (see apache#34535 for future plans). * GitHub Issue: apache#41560 Authored-by: Felipe Oliveira Carvalho <felipekde@gmail.com> Signed-off-by: Felipe Oliveira Carvalho <felipekde@gmail.com>

github-actions bot added Component: C++ awaiting committer review Awaiting committer review labels May 6, 2024

zeroshade requested a review from pitrou May 8, 2024 14:32

pitrou reviewed May 13, 2024

View reviewed changes

github-actions bot added awaiting changes Awaiting changes awaiting review Awaiting review and removed awaiting committer review Awaiting committer review labels May 13, 2024

github-actions bot added awaiting changes Awaiting changes awaiting change review Awaiting change review and removed awaiting review Awaiting review awaiting changes Awaiting changes labels May 14, 2024

github-actions bot added awaiting changes Awaiting changes and removed awaiting change review Awaiting change review labels May 14, 2024

felipecrv added 9 commits May 14, 2024 20:49

ChunkResolver: Rename ResolveWith{ChunkIndex->}Hint()

3e2ff8f

I should have done this renaming in a previous PR. Correcting it now.

ChunkResolver: Implement ResolveMany and add unit tests

99a2a4d

ChunkResolver: Use if-constexpr instead of template specialization

18c84ad

After learning that GCC can't compile the previous code. Well... this is simpler now.

ChunkResolver: Fix linting error

7d26515

ChunkResolver: Remove redundant public: from ChunkLocation

848eaa3

ChunkResolver: Make readability improvements

eb7a1bd

ChunkResolver: Remove asserts on ChunkLocation construction

d0db90f

ChunkResolver: Remove ARROW_RESTRICT annotations

dbfc796

ChunkResolver: Handle vectors of signed logical indices

ce7840b

felipecrv force-pushed the resolve_many branch from 9c16963 to ce7840b Compare May 15, 2024 00:02

github-actions bot added awaiting change review Awaiting change review and removed awaiting changes Awaiting changes labels May 15, 2024

felipecrv requested a review from pitrou May 15, 2024 01:40

pitrou approved these changes May 15, 2024

View reviewed changes

github-actions bot added awaiting changes Awaiting changes and removed awaiting change review Awaiting change review labels May 15, 2024

felipecrv added 2 commits May 15, 2024 16:48

ChunkResolver: Fix the handling of negative indices

f5a7d1d

ChunkResolver: Fix comments

5745c7d

github-actions bot added awaiting change review Awaiting change review and removed awaiting changes Awaiting changes labels May 15, 2024

felipecrv merged commit e04f5b4 into apache:main May 15, 2024
36 of 37 checks passed

felipecrv removed the awaiting change review Awaiting change review label May 15, 2024

felipecrv mentioned this pull request May 15, 2024

[C++] Add ChunkResolver::ResolveMany() API #41560

Closed

felipecrv deleted the resolve_many branch May 16, 2024 00:00

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

GH-41560: [C++] ChunkResolver: Implement ResolveMany and add unit tests #41561

GH-41560: [C++] ChunkResolver: Implement ResolveMany and add unit tests #41561

felipecrv commented May 6, 2024 •

edited by github-actions bot

Loading

github-actions bot commented May 6, 2024

pitrou left a comment

pitrou May 13, 2024

felipecrv May 13, 2024

pitrou May 14, 2024

felipecrv May 14, 2024 •

edited

Loading

pitrou May 14, 2024

felipecrv commented May 13, 2024

pitrou commented May 14, 2024

pitrou left a comment

pitrou May 15, 2024

pitrou May 15, 2024

pitrou May 15, 2024

felipecrv May 15, 2024

pitrou May 15, 2024

felipecrv May 15, 2024

pitrou May 15, 2024

felipecrv May 15, 2024 •

edited

Loading

conbench-apache-arrow bot commented May 16, 2024

GH-41560: [C++] ChunkResolver: Implement ResolveMany and add unit tests #41561

GH-41560: [C++] ChunkResolver: Implement ResolveMany and add unit tests #41561

Conversation

felipecrv commented May 6, 2024 • edited by github-actions bot Loading

Rationale for this change

What changes are included in this PR?

Are these changes tested?

Are there any user-facing changes?

github-actions bot commented May 6, 2024

pitrou left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

felipecrv May 14, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

felipecrv commented May 13, 2024

pitrou commented May 14, 2024

pitrou left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

felipecrv May 15, 2024 • edited Loading

Choose a reason for hiding this comment

conbench-apache-arrow bot commented May 16, 2024

felipecrv commented May 6, 2024 •

edited by github-actions bot

Loading

felipecrv May 14, 2024 •

edited

Loading

felipecrv May 15, 2024 •

edited

Loading