A deep dive into the performance characteristics of pgx's CollectRows and AppendRows convenience APIs for Go. The author benchmarks these APIs against a hand-written loop, finds they are 1.8× slower and allocate 10× more memory, then explains why: repeated reflection-based struct field lookups, heap escapes of temporary values, and per-row allocation of scan target slices. The post covers patches submitted to pgx (landed in v5.6) that added caching, and introduces pgx-collect, a drop-in compatible library that redesigns the core abstraction from a per-row RowToFunc to a Scanner interface with separate Initialize and ScanRowInto methods. This allows per-query setup to run once, scan targets to be reused, and data to be written directly into the output slice — reducing overhead from 180% to near-baseline. Key lessons: benchmark convenience APIs, choose abstractions that match the concrete work, cache reflection metadata by type, and minimize heap allocations.
