profiling-statement-fingerprints

Query crdb_internal.statement_statistics filtered by a time window and ordered by the runLat mean extracted from the statistics JSONB column. This returns the slowest statement fingerprints with execution counts, full scan flags, and index recommendations.

A statement fingerprint is a normalized SQL pattern with parameterized constants, so queries like SELECT with different literal values aggregate into one fingerprint. Statistics are collected per fingerprint in hourly aggregated_ts buckets.

You need the VIEWACTIVITY or VIEWACTIVITYREDACTED cluster privilege for cluster-wide visibility. VIEWACTIVITYREDACTED hides query constants, which is recommended for multi-tenant production environments.

CPU, contention, memory, and admission metrics live in execution_statistics, which is probabilistically sampled rather than collected for every execution. Filter with WHERE (statistics->'execution_statistics'->>'cnt') IS NOT NULL to exclude unsampled buckets.

Retention is bounded by row count via the sql.stats.persisted_rows.max cluster setting (default 1,000,000 rows), not by a fixed TTL. The effective time window depends on workload diversity, so always filter by aggregated_ts.

Use live triage of currently running queries for immediate incident response, such as canceling runaway queries. Statement fingerprint analysis is for historical patterns, trends, and optimization opportunities over hours or days.