perf: Use file_metadata_cache in geoparquet

[petern48]

Use file metadata cache for geoparquet

closes #250

[petern48]

Well that was suspiciously simple...

[petern48]

with_file_metadata_cache() is called for each iteration of the loop (.map()), we need a clone for each separate iteration. get_file_metadata_cache() returns a cloned Arc, already, so no need to call another .clone().

https://github.com/apache/datafusion/blob/28755b1d7eb5222a8f5fb5417134dd6865ac1311/datafusion/execution/src/cache/cache_manager.rs#L174-L176

[petern48]

It would be nice to see how effective this is, with some benchmarks involving geoparquet reading. That doesn't seem to exist yet, does it?

[paleolimbot]

Thank you!

Can you also the with_file_metadata_cache() line here?

sedona-db/rust/sedona-geoparquet/src/file_opener.rs

Lines 116 to 120 in f49016d

	let parquet_metadata =
	DFParquetMetadata::new(&self_clone.object_store, &file_meta.object_meta)
	.with_metadata_size_hint(self_clone.metadata_size_hint)
	.fetch_metadata()
	.await?;

This seems a bit slower than on main at the moment and 0.1 and I don't see an impact on multiple calls to read_parquet(). Querying a big table with lots of Parquet files is probably a good way to check (but it might be better to list files individually for this particular benchmark, since there's also a cost to querying S3 to list the files).

import os
os.environ["AWS_SKIP_SIGNATURE"] = "true"
os.environ["AWS_DEFAULT_REGION"] = "us-west-2"
import sedona.db

sd = sedona.db.connect()

# 16s on main, 22s on this PR
sd.read_parquet(
    "s3://overturemaps-us-west-2/release/2025-10-22.0/theme=buildings/type=building/"
).to_view("buildings")

# Second time: 16s on main, 19s on this PR
sd.read_parquet(
    "s3://overturemaps-us-west-2/release/2025-10-22.0/theme=buildings/type=building/"
).to_view("buildings", overwrite=True)

From https://datafusion.apache.org/blog/2025/08/15/external-parquet-indexes/ , I wonder if there's something in the ListingTable we need to port over to SedonaContext::read_parquet() as well (or something we need to propagate in the FileConfig).

[petern48]

Seems like you're right. Here's what I got in terms of numbers (4 trials each). (Before adding the extra changes)

# On branch:
20.807693004608154 seconds. 19.300298929214478 seconds. 30.610652208328247. 20.58704400062561

# On main:
19.879024744033813 seconds. 17.96437692642212 seconds. 21.73752498626709. 17.080259084701538.

Need more time to investigate.

[petern48]

Found this neat datafusion feature for checking the metadata cache. I'll give it a try once I find the time to circle back to this.
apache/datafusion#17126

[petern48]

Can you also the with_file_metadata_cache() line here?

Tried this before, and ran into some complications when trying to pass the state variable I need there. Datafusion's opener doesn't seem to use the metadata cache state either. I haven't yet looked deeper into what they're doing, but it's probably something.

(As a note to myself) I need to next look into what's happening here with ArrowReaderMetadata and async_file_reader.

[paleolimbot]

I just noticed that DataFusion 51 updated the default metadata cache hint. It may be worth trying some of this with

sd.sql("SET datafusion.execution.parquet.metadata_size_hint = 524288")