[Iceberg] Add StatisticsFileCache

[ZacBlanco]

Description

This change adds a connector-wide cache for StatisticsFiles. This prevents loading the same statisticsFile multiple times into the connector. It improves query planning times and reduces memory usage by having only one set of statistics available from each file.

The cache is designed to be connector-wide to reduce memory consumption. Since separate queries can reference the same tables, we prevent memory utilization by sharing the cache across queries. The files are immutable and unique to a table and snapshot (hashed and equated on properties of file path, footer size, blob metadata, etc), so caching between queries/users/sessions shouldn't be an issue. Cache entries should never become stale

Motivation and Context

I noticed after experimenting with histograms in the Iceberg connector that planning was quite slow for two reasons

1. We were always hitting the FS for StatisticsFiles
2. Every time we load the blobs from a StatisticFile, even if using the same file as a previous query, we would allocate a new buffer to represent the statistics. These statistics are not evacuated from memory until the query metadata for an execution expires from query.max-age. For smaller statistics, this is not really a problem. However, histograms can be large (KBs to MBs in size). Having the cache can prevent making copies of redundant data, leading to lower memory pressure and smaller chance of choking up the coordinator.

The cache inside of the Iceberg connector does not fully alleviate the issue. If a user has thousands of tables/columns, memory pressure can still occur when there is large amount of historical query metadata in the heap. The only way to prevent the pressure is to eventually store some query metadata (or maybe just statistics?) in a cache which is freed once the coordinator reaches high memory pressure.

Impact

• Addition to the ConnectorHistogram SPI interface to add a getMemoryUtilization method. This will be useful in the future when implementing real histograms in [Iceberg] Add Histogram Statistic Support #22365 .

Test Plan

• Added an E2E test to verify cache is hit when table statistics are retrieved.

Contributor checklist

• Please make sure your submission complies with our development, formatting, commit message, and attribution guidelines.
• PR description addresses the issue accurately and concisely. If the change is non-trivial, a GitHub Issue is referenced.
• Documented new properties (with its default value), SQL syntax, functions, or other functionality.
• If release notes are required, they follow the release notes guidelines.
• Adequate tests were added if applicable.
• CI passed.

Release Notes

== RELEASE NOTES ==

General Changes
* Add a configurable-sized cache for Iceberg table puffin files to improve query planning time controlled by the  `iceberg.max-statistics-file-cache-size` configuration property. :pr:`23177`

[steveburnett]

LGTM! (docs)

Pull branch, local docs build, reviewed compiled page in local build. Looks good, thanks!

[vivek-bharathan]

lgtm overall

[hantangwangd]

Overall looks good to me. Bring up one question for discussing, as the Guava cache is a light weight LRU implementation, it may not adapt well to some scenarios. So do you think it's useful to provide a way to configure not to use statistics file cache in some situations, or have a way to extend and configure to use other enhanced cache implementations, for example the ones based on Caffeine or Redis?

[hantangwangd]

Thanks for the fix, lgtm.

[aaneja]

LGTM % one question

[ZacBlanco]

@tdcmeehan could you take a look and give a stamp?

[tdcmeehan]

Just some nits. Main thing is it would be nice to have a test that breaks if the memory estimates change for some reason.

[hantangwangd]

One little thing, otherwise lgtm.