Summary

Admin metrics rollup tables stay empty during the benchmark window (load tests < 1 hour), so admin metrics queries fall back to raw scans only. This increases DB load and admin metrics latency.

Evidence

Rollup tables during benchmark:

• tool_metrics_hourly: 0
• resource_metrics_hourly: 0
• prompt_metrics_hourly: 0
• server_metrics_hourly: 0
• a2a_agent_metrics_hourly: 0

After manual rollup (, hours_back=24):

• tool_metrics_hourly: 5
• resource_metrics_hourly: 4
• prompt_metrics_hourly: 3
• server_metrics_hourly: 0
• a2a_agent_metrics_hourly: 0

Rollup summary:

• total_records_aggregated=1,569,843
• rollups_updated=12 (created=0)
• duration_seconds=13.12

Raw data window observed:

• tool_metrics min_ts=2026-01-06 17:19:56
• tool_metrics max_ts=2026-01-06 18:06:12

Note: current load test duration <1 hour, so rollups do not cover the window and combined raw+rollup queries still scan raw metrics.

Impact

• Admin metrics pages trigger expensive raw scans.
• Increased DB CPU and higher /admin* latency under load.

Potential fixes

1. Ensure a rollup job runs on hour boundaries (or more frequently) during tests.
2. Run a manual rollup before benchmarking if the test window is <1 hour.
3. Increase metrics cache TTLs so admin UI does not recompute heavy metrics every request.

Validation

• Rollup tables contain data for the benchmark window after an hour boundary.
• Admin metrics queries use rollups and show reduced total time in pg_stat_statements.
• /admin metrics endpoints show lower p95 latency under load.