parquet: support predicate pushdown for partial reads / exclusions

[derekperkins]

Feature request

Is your feature request related to a problem? Please describe.

When accessing a remote parquet file using FILES, the entire file is fetched across the network before executing. This can result in waiting for hundreds of megabytes to download, then seeing an error like #37169 where the encoding isn't supported

• Error when reading parquet file using DELTA_BINARY_PACKED encoding (IOError: Not yet implemented: Unsupported encoding.) #37169

As of StarRocks v3.3.0-rc02, an unsupported encoding in a parquet file, even if it isn't referenced by the query, makes the entire file unqueryable. Only the specific columns in the SELECT should be fetched, which saves both network, and should make it so that StarRocks can read columns even if others aren't supported.

This is listed in the 2024 roadmap, but I couldn't find a tracking issue for it

• StarRocks Roadmap 2024 #39686

Describe the solution you'd like

Support parquet predicate pushdown, so that only specific metadata and/or columns are read.

1. By looking at the metadata, unsupported encodings could throw an error without reading the entire file
2. By utilizing object store range reads, only fetch the column data requested by the query, rather than the whole file

Describe alternatives you've considered

DuckDB, Clickhouse, etc

Additional context

• https://db-blog.web.cern.ch/node/194
• https://clickhouse.com/blog/apache-parquet-clickhouse-local-querying-writing-internals-row-groups
• https://duckdb.org/2021/06/25/querying-parquet.html