What are best practices for writing Parquet files in production?

Producing high-quality Parquet files that perform well at query time requires attention at write time:

1. Target 128 MB–512 MB row groups — too small wastes footer reads; too large makes predicate skipping coarse.
2. Sort data before writing on filter columns — tight min/max ranges per row group dramatically improve skipping.
3. Choose the right compression codec — ZSTD for I/O-bound workloads; Snappy for CPU-bound (fast decompression).
4. Enable dictionary encoding on low-cardinality columns — automatic in most frameworks but verify it is not being disabled.
5. Partition on moderate-cardinality columns (e.g., date, country) — never on user IDs or UUIDs.
6. Avoid tiny files — compact regularly if streaming or incremental writes produce many small files.
7. Enable Bloom filters on high-cardinality equality columns (UUIDs, hashed IDs).
8. Embed correct schema types — use TIMESTAMP_MICROS not INT96 (deprecated); use DECIMAL not DOUBLE for monetary values.
9. Test with query benchmarks after schema changes to confirm no regression in pushdown effectiveness.