# Performance patterns

Use pools and release checked-out clients in `finally`, especially for cursors, streams, and COPY.
Route ordinary reads to a replica when its lag is acceptable; route read-after-write, locking, and
transaction-consistent reads to the writer. Do not hold a client across unrelated application work.

For large reads, use keyset pagination, cursors, or streams with cancellation and bounded batches.
For writes, prefer set-based batches such as UNNEST or COPY when they preserve validation and error
handling. Measure query plans with representative cardinality before changing an index or query.

Check query predicates, joins, ordering, and selected columns against index shape. Use EXPLAIN
evidence to confirm planner behavior. Add extended statistics only when observed estimates show a
correlation problem; verify the statistics are collected and used. Partitioning can reduce scanned
data, but it does not replace suitable local indexes or predicates that permit pruning.