trips 15 files · 6.0 MB

• trip_id int64
• pickup_hour int32
• passengers int32
• distance double
• price double
• tip double
• payment string
• region part
• date part

Partition columns live in the directory path, not inside the files — filtering on one costs zero I/O.

1 Full scan (baseline) Every column, no WHERE clause — nothing can be pruned, so this reads 100% of the dataset. Run it first: it is the number every other query is measured against. 2 Partition pruning region is a partition column — it lives in the directory path, not inside the files. The engine drops every file that cannot match without opening a single one. 3 Row-group skipping A highly selective filter on trip_id. Every Parquet footer carries min/max statistics per row-group, so most row-groups can be proven irrelevant and skipped before a single byte is decompressed. 4 Column projection Only distance is referenced, so only distance comes off disk — and region is free because it comes from the path, not the data. Parquet is columnar, so the unreferenced columns are never touched. 5 All three at once A query someone would actually write, with all three optimizations firing. It reads more than the single-optimization demos above because it needs a real measure column rather than just a row count — but still a small fraction of the baseline.

Each one isolates a single mechanism, so the percentages do not decrease monotonically down the list.