[Data] decrease parquet metadata storage usage

[sword865]

Why are these changes needed?

When working with search and recommendation systems, datasets often contain numerous columns, resulting in large metadata overhead in Parquet files (sometimes a few MBs or more for each file). Currently, the driver fetches first all metadata, then simplifies and merges them to reduce memory usage. However, this process can cause memory peaks proportional to the number of fragments multiplied by their metadata size, potentially leading to OOM issues.

This PR addresses the problem by simplifying and merging the dataset metadata within each _fetch_metadata task before sending it back to the driver. This change helps lower memory consumption and reduces the risk of OOM errors.

Test script:

from pyarrow._fs import FileSystem
from ray.data.datasource.file_meta_provider import _get_file_infos
import os
import ray
import psutil

hdfs_path = "hdfs://path/to/dataset"

def list_files(remote_path):
    filesystem, remote_path = FileSystem.from_uri(remote_path)
    from ray.data.datasource.file_meta_provider import _get_file_infos
    files = _get_file_infos(remote_path, filesystem)
    return filesystem, files

filesystem, files = list_files(hdfs_path)
files = [f"{fs[0]}" for fs in files if fs[0].endswith(".parquet")]

process = psutil.Process()

print(f"total file_num: {len(files)}")
start_mem = process.memory_info().rss / 1024**2
dataset = ray.data.read_parquet(paths=files[:500], filesystem=filesystem)
end_mem = process.memory_info().rss / 1024**2
print(f"datasize: {dataset.count()}, col number: {len(dataset.columns())}")
print(f"mem diff {end_mem - start_mem:.3f}MiB [start: {start_mem:.3f}MiB, end: {end_mem:.3f}MiB]")

Output before this PR:

total file_num: 2358
Metadata Fetch Progress 0: 100%|███████████████████████|500/500 [02:13<00:00, 3.75 task/s]
Parquet Files Sample 0: 100%|███████████████████████| 5.00/5.00 [00:13<00:00, 2.62s/ file]
datasize: 22630452, col number: 1200
mem diff 13727.605MiB [start: 570.617MiB, end: 14298.223MiB]

Output after this PR:

total file_num: 2358
Metadata Fetch Progress 0: 100%|███████████████████████| 500/500 [01:55<00:00, 4.35 task/s]
Parquet Files Sample 0: 100%|███████████████████████| 5.00/5.00 [00:03<00:00, 1.56 file/s]
datasize: 22630452, col number: 1200
mem diff 69.113MiB [start: 575.820MiB, end: 644.934MiB]

We can see the memory usage reduce from 13GBs to 69MBs.

Note: This approach is most effective for large-scale datasets. If len(fragments) < PARALLELIZE_META_FETCH_THRESHOLD, there will be no performance improvements.

Checks

• I've signed off every commit(by using the -s flag, i.e., git commit -s) in this PR.
• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
  • I've added any new APIs to the API Reference. For example, if I added a
    method in Tune, I've added it in doc/source/tune/api/ under the
    corresponding .rst file.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[gemini-code-assist]

Summary of Changes

Hello @sword865, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request significantly reduces memory consumption when reading Parquet datasets, especially those with a high number of columns, by optimizing the handling of Parquet file metadata. The metadata simplification and deduplication now occur directly within the remote tasks, preventing large memory peaks and potential Out-Of-Memory (OOM) errors on the driver.

Highlights

• Memory Optimization for Parquet Metadata: The core change shifts the metadata deduplication process from the driver to the remote worker tasks that fetch the metadata. This prevents large memory spikes on the driver when dealing with numerous Parquet fragments, particularly those with complex schemas, as demonstrated by a reduction in memory usage from 13GB to 65MB in the provided test case.
• Early Metadata Deduplication: The _fetch_metadata function now performs schema deduplication using _dedupe_metadata before returning the metadata to the driver, ensuring that only simplified and merged metadata is transferred.
• New Deduplication Function: A new helper function, _dedupe_fragment_metadata, has been introduced to specifically handle deduplication of _ParquetFileFragmentMetaData objects, ensuring that identical schemas across fragments reference the same in-memory object.
• Type Hint Refinement: The return type hints for _fetch_metadata_serialization_wrapper and _fetch_metadata have been updated to reflect that they now return lists of _ParquetFileFragmentMetaData objects, indicating the metadata is already in a simplified, deduplicated form.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in issue comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments or fill out our survey to provide feedback.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

The pull request reduces memory usage when reading Parquet metadata by moving deduplication to worker tasks. The changes are well-motivated with impressive performance gains. I've provided feedback on a redundant function call and a simplification opportunity.