GH-45750: [C++][Python][Parquet] Implement Content-Defined Chunking for the Parquet writer

[kszucs]

Rationale for this change

I have been working on to improve Parquet's deduplication efficiency for content-addressable storages. These system generally use some kind of a CDC algorithm which are better suited for uncompressed row-major formats. Although thanks to Parquet's unique features I was able to reach good deduplication results by consistently chunking data pages by maintaining a gearhash based chunker for each column.

Deduplication efficiency

The feature enables efficient data deduplication for compressed parquet files on content addressable storage (CAS) systems such as Hugging Face Hub. There is a purpose built evaluation tool is available at https://github.com/kszucs/de used during development to continuously check the improvements and to visually inspect the results. Please take a look at the repository's readme to see how different changes made to parquet files affect the deduplication ratio when they are stored in CAS systems.

Some results calculated on all revisions of datasets.parquet

❯ de stats /tmp/datasets Writing CDC Parquet files with ZSTD compression 100%|███████████████████████████████████████████████████████████████████████████████████| 194/194 [00:12<00:00, 15.73it/s]
Writing CDC Parquet files with Snappy compression 100%|███████████████████████████████████████████████████████████████████████████████████| 194/194 [00:10<00:00, 17.95it/s]
Estimating deduplication for Parquet Estimating deduplication for CDC ZSTD Estimating deduplication for CDC Snappy ┏━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┓
┃ ┃ ┃ ┃ Compressed Chunk ┃ ┃ Compressed Dedup ┃ Transmitted XTool ┃
┃ Title ┃ Total Size ┃ Chunk Size ┃ Size ┃ Dedup Ratio ┃ Ratio ┃ Bytes ┃
┡━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━┩
│ Parquet │ 16.2 GiB │ 15.0 GiB │ 13.4 GiB │ 93% │ 83% │ 13.5 GiB │
│ CDC ZSTD │ 8.8 GiB │ 5.6 GiB │ 5.6 GiB │ 64% │ 64% │ 6.0 GiB │
│ CDC Snappy │ 16.2 GiB │ 8.6 GiB │ 8.1 GiB │ 53% │ 50% │ 9.4 GiB │
└────────────┴────────────┴────────────┴──────────────────────┴─────────────┴─────────────────────┴──────────────────────┘

Some results calculated on all revisions of food.parquet

❯ de stats /tmp/food --max-processes 4 Writing CDC Parquet files with ZSTD compression 100%|█████████████████████████████████████████████████████████████████████████████████████| 32/32 [10:28<00:00, 19.64s/it]
Writing CDC Parquet files with Snappy compression 100%|█████████████████████████████████████████████████████████████████████████████████████| 32/32 [08:11<00:00, 15.37s/it]
Estimating deduplication for Parquet Estimating deduplication for CDC ZSTD Estimating deduplication for CDC Snappy ┏━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┓
┃ ┃ ┃ ┃ Compressed Chunk ┃ ┃ Compressed Dedup ┃ Transmitted XTool ┃
┃ Title ┃ Total Size ┃ Chunk Size ┃ Size ┃ Dedup Ratio ┃ Ratio ┃ Bytes ┃
┡━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━┩
│ Parquet │ 182.6 GiB │ 148.0 GiB │ 140.5 GiB │ 81% │ 77% │ 146.4 GiB │
│ CDC ZSTD │ 107.1 GiB │ 58.0 GiB │ 57.9 GiB │ 54% │ 54% │ 66.2 GiB │
│ CDC Snappy │ 176.7 GiB │ 79.6 GiB │ 77.2 GiB │ 45% │ 44% │ 101.0 GiB │
└────────────┴────────────┴────────────┴──────────────────────┴─────────────┴─────────────────────┴──────────────────────┘

Chunk size shows the actual storage required to store the CDC chunked parquet files in a simple CAS implementation.

What changes are included in this PR?

A new column chunker implementation based on CDC algorithm, see more details in the docstrings. The implementation is added to the C++ Parquet writer and exposed in PyArrow as well.

Are these changes tested?

Yes. Tests have been added to the C++ implementation as well as the exposed PyArrow API.

Are there any user-facing changes?

There are two new parquet writer properties on the C++ side:

• enable_content_defined_chunking() to enable the feature
• content_defined_chunking_options(min_chunk_size, max_chunk_size, norm_factor) to provide additional options

There is a new pq.write_table(..., use_content_defined_chunking=) keyword argument to expose the feature on the Python side.

I marked all user-facing changes as EXPERIMENTAL.

• GitHub Issue: [C++][Python][Parquet] Support Content-Defined Chunking of Parquet files #45750

[mapleFU]

Is cdc a part of the parquet spec? Or is it a poc?

[kszucs]

Is cdc a part of the parquet spec? Or is it a poc?

It is not. You can think of it as an implementation specific feature similar to the existing options to specify how record batches and pages are being split.

[rok]

Thanks for doing this @kszucs ! I like how this doesn't need any changes to readers.

Questions:

• As it stands in this PR, CDC is either on or off for all columns. How about enabling it per column? In general case some columns might not be worthy candidates for it.
• Use case described in HF blogpost describes cases where rows are added or removed but not much else is changed. Wouldn't it then make sense to first try a shortcut deduplication where if we identify a duplication in the first column we first check for the same duplication at the same indices in all other columns before running a full hashing pass?

[kszucs]

@github-actions crossbow submit test-conda-cpp-valgrind

[github-actions]

Revision: 1cc2e4b

Submitted crossbow builds: ursacomputing/crossbow @ actions-6a43d39b56

Task	Status
test-conda-cpp-valgrind

[kszucs]

I collected the possible follow-ups, once the PR is merged I will create the corresponding tickets:

• [C++][Parquet] Support CDC in WriteBatch and WriteBatchSpacedGH-45750: [C++][Python][Parquet] Implement Content-Defined Chunking for the Parquet writer #45360 (comment)
• [C++][Parquet] Center the chunk size around avg_chunk_size if norm_level is not 0GH-45750: [C++][Python][Parquet] Implement Content-Defined Chunking for the Parquet writer #45360 (comment)
• [C++][Parquet] Consider enforcing little-endian intepretation during chunking GH-45750: [C++][Python][Parquet] Implement Content-Defined Chunking for the Parquet writer #45360 (comment)
• [C++][Parquet] Avoid unconditional array slice in WriteArrow, WriteArrowDense and WriteArrowDictionary if value_offset == 0GH-45750: [C++][Python][Parquet] Implement Content-Defined Chunking for the Parquet writer #45360 (comment)
• [C++][Python][Parquet] Store the properties used to write the parquet file as arbitrary metadata GH-45750: [C++][Python][Parquet] Implement Content-Defined Chunking for the Parquet writer #45360 (comment)
• [Python][Parquet] Implement a hierarchy of Parquet exceptions rather than raising OSErrorGH-45750: [C++][Python][Parquet] Implement Content-Defined Chunking for the Parquet writer #45360 (comment)

[pitrou]

@github-actions crossbow submit preview-docs

[pitrou]

Congratulations @kszucs :)

[github-actions]

Revision: 1cc2e4b

Submitted crossbow builds: ursacomputing/crossbow @ actions-81dc2a98bf

Task	Status
preview-docs

[wgtmac]

+1 (for the C++ part)

[kszucs]

Thanks @pitrou@wgtmac@kou@mapleFU for the reviews!

[conbench-apache-arrow]

After merging your PR, Conbench analyzed the 4 benchmarking runs that have been run so far on merge-commit dd94c90.

There were no benchmark performance regressions. 🎉

The full Conbench report has more details. It also includes information about 8 possible false positives for unstable benchmarks that are known to sometimes produce them.