Sort filenames when reading parquet to ensure consistent schema

[thomas-k-cameron]

Which issue does this PR close?

Closes #5779.

Rationale for this change

Sorting is passed to ensure that files returned is in the same order when the OS is different.

What changes are included in this PR?

This PR adds a lines of code to FileFormat::infer_schema to ParquetFormat; It sorts parquet file by it's location.

On test_util::store_parquet is changed to sort temp file by it's path and and write batches in the sorted order.

This change was introduced as it made the some tests, such as datasource::file_format::parquet::tests::read_merged_batches to fail.

It appears that these tests assumed that infer_schema processed the files in the same order that store_parquet's returning value files.

Are these changes tested?

Yes. This PR introduces new test case async fn issue_5779().

Are there any user-facing changes?

The order of Fields may change.

[alamb]

Thanks @thomas-k-cameron -- given my reading of #5779 it seems like perhaps a more direct solution might be to sort the list of files so they are always processed in the same order

Perhaps somehow figure out how to sort the objects by the location field, for example?

        let schemas: Vec<_> = futures::stream::iter(objects)
            .map(|object| fetch_schema(store.as_ref(), object, self.metadata_size_hint))
            .boxed() // Workaround https://github.com/rust-lang/rust/issues/64552
            .buffered(SCHEMA_INFERENCE_CONCURRENCY)
            .try_collect()
            .await?;

[thomas-k-cameron]

Thank you for your advice.

I couldn't figure out what I was getting it wrong.
Let me try it out.

[thomas-k-cameron]

Let me give you an update.

I tried to sort by location, but it hasn't been working.
I end up getting error from tests for fetching statistics.

It seems like this is coming from how parquet files for tests are created with random names.

I have few other ideas. I will see if it works out.

[thomas-k-cameron]

@alamb
It's been a while but I believe it is ready for review.

I updated the PR comment to reflect my changes.

[alamb]

Thank you @thomas-k-cameron -- I think this looks great. The only potential downside of this change I can think of is that this will change behavior for any object store that lists files sorted deterministicially in a different order than sort()

However, I think most object stores sort their locations lexographically so this PR is likely to simply make local filesystems consistent.

Thanks again

[thomas-k-cameron]

@alamb

Thank you very much for your feedback and approval.

Change that I just pushed simply applies a cargo fmt to the codebase; I think all tests should pass now.

[alamb]

Thanks @thomas-k-cameron -- once the check pass I'll plan to merge this PR.

[alamb]

🚀