Projection pushdown is a query optimization technique that reads only the columns required by a query. With Parquet files, the storage layer can leave unneeded columns untouched, reducing data transfer and improving selective-query performance. section 5.4
Projection pushdown is especially useful when:
- Querying large Parquet datasets directly without populating a database. section 10.4
- Analyzing files stored remotely, where avoiding unnecessary network transfer lowers latency and potential egress costs. section 10.4
- Combining it with predicate pushdown so queries restrict both the columns and data segments they read. section 5.4
Query only the columns needed
A DuckDB query against Parquet can request a narrow result instead of reading every stored field: SELECT ranking_date, rank, points FROM 'atp_rankings.parquet' ORDER BY ranking_date DESC LIMIT 5; The same technique supports one-off analysis of very large New York taxi datasets stored in cloud buckets, transferring only data needed for the query. section 5.4 section 10.4
