Statement statistics

Statement statistics help you find queries that consume the most time overall, even when each execution is too fast for the slow query log. CamusDB groups similar statements and keeps their execution counts, durations, rows read, failures, spills, and plan changes. This serves a similar purpose to PostgreSQL's pg_stat_statements.

Enable collection

Collection is off by default. Configure the node:

statement_stats_enabled: true
statement_stats_max_shapes: 2000

You can also enable it at runtime:

SET CLUSTER SETTING statement_stats_enabled = true;
SHOW STATEMENT STATS;
SHOW STATEMENT STATS LIKE '%orders%';

statement_stats_enabled is node-scoped and affects the next statement. Turning it off preserves existing totals. statement_stats_max_shapes requires a restart. Sample SQL length is bounded by slow_query_log_max_sql_length. The inspection statement requires a superuser.

How statements are grouped

A shape identifies a statement with its literal values and placeholders removed. WHERE id = 1, WHERE id = 2, and WHERE id = @id share one shape. Constant-only IN lists and multi-row VALUES lists share a shape regardless of their length. Identifiers are compared without case.

Each row is keyed by database and a 16-character shape_id, stable across nodes and restarts for the same build. Totals are recorded for SELECT, INSERT, UPDATE, and DELETE, including RETURNING and INSERT ... SELECT. DDL, transaction control, and SHOW statements are excluded.

An execution is recorded when it finishes, fails, or its cursor is abandoned. An open cursor has not yet contributed its final counters.

Read the result

Columns	Meaning
node, database, shape_id, kind	Node, database, statement fingerprint, and kind (select, insert, update, or delete).
sample_sql	First statement text seen for the shape, with password literals redacted and text truncated to the configured limit. Other literals can remain.
calls, errors	Executions that ended, and executions that failed.
total_ms, sum_sq_ms	Sum of durations and sum of squared durations.
min_ms, max_ms	Shortest and longest execution since the shape was first seen.
mean_ms, stddev_ms, cv	Mean, standard deviation, and coefficient of variation (standard deviation divided by mean).
rows_read, rows_returned	Storage rows read and rows returned; for writes, returned means affected rows. Uses the slow query log's counting rules.
full_scans, spills	Executions with a full relation scan or a blocking operator that spilled to disk.
retries	Failed executions with CADB0502, CADB0504, or CADB0505; counts retryable failures, not successful client retries.
plan_changes, last_plan	Changes from the previous execution's plan and the latest plan hash. last_plan is null when no plan was built.
first_seen_at, last_seen_at	UTC timestamps for the shape's first and latest observations.

Results are ordered by cv, highest first. High variation can point to plan changes or different cache states; inspect the plan and workload to establish the cause. Plan hashes include node kinds, tables, access paths, and indexes, without literal values, bounds, or estimates.

For an initial investigation:

• High total_ms identifies expensive shapes across the whole observation period.
• High rows_read relative to rows_returned suggests checking filters and indexes.
• full_scans, spills, and plan_changes direct you to specific plan behavior.

Use EXPLAIN for the plan of one execution, and index usage to check which indexes serve the workload.

Measure a time window

Totals are cumulative. Take two readings and subtract additive columns to measure a window. For example, window mean duration is Δtotal_ms / Δcalls; window variance is Δsum_sq_ms / Δcalls − (Δtotal_ms / Δcalls)². The reported minimum, maximum, and derived statistics describe the whole life of the shape, so they cannot be subtracted to obtain window values.

The table never evicts existing shapes. When capacity is full, new shapes are not admitted. Inspect the camusdb metrics:

SHOW ENGINE STATS LIKE 'statement_stats.%';

statement_stats.shapes reports the current count and statement_stats.dropped counts refused admissions. A growing dropped count means the report is incomplete; increase the capacity and restart if needed.

Counters live in memory and reset at restart. They describe one node only; collect from every node for a cluster view and retain readings externally for history. The slow query log complements these totals with individual executions over a threshold.