System resources page

performance/system-resources/ — CPU, memory and per-tier metrics from SystemResourceSnapshot. A service selector switches between the web, celery_worker, redis and postgres tiers. See Resource metrics for how CPU normalisation and peaks work.

Which tiers have data

The web tier snapshots itself every RESOURCE_SNAPSHOT_INTERVAL_MINUTES (default 15) via the built-in timer, so it populates without Celery. The Celery worker, Redis and Postgres tiers are populated by Celery tasks — if those tiles are empty, schedule the snapshot tasks with Beat. See Celery setup.

CPU trend

Per-snapshot cpu_percent for the web process over the range. CPU is normalised by logical core count, so it reads 0–100% of total host capacity, not per-core (a process on two of four cores reads 50%, not 200%).

Memory trend

ram_percent over time, with ram_used_mb / ram_total_mb behind it so you can see both the percentage and the absolute footprint.

File descriptors and threads

open_file_descriptors and active_threads for the web process. A steadily climbing FD count is a classic leak signature (unclosed sockets/files); thread count helps spot runaway thread pools.

CPU heatmap (day × hour)

Average web CPU by weekday and hour of day, over the last 30 days (independent of the page range). Seven rows (Mon–Sun) × 24 columns, each cell the mean CPU for that slot. It reveals load patterns — nightly batch jobs, weekday peaks, weekend lulls — that a linear trend flattens out.

Disk usage

The latest snapshot's disk_used_gb / disk_total_gb / disk_percent. Shown as the most recent reading rather than a trend, since disk moves slowly.

Celery tier

For the celery_worker tier: active, reserved and queued task counts over time, plus a peak queued-tasks figure. A queue depth that climbs and does not drain means workers are saturated. These come from Celery's inspect API during snapshot collection, so they need the Celery snapshot task scheduled.

Redis tier

redis_used_memory_mb, redis_connected_clients and redis_blocked_clients over time, with a peak blocked-clients figure. Blocked clients climbing indicates contention on blocking operations.

Postgres tier

postgres_active_connections and postgres_idle_connections over time, a database-size trend (postgres_db_size_mb), and a connection-pressure figure — the peak share of active connections out of active + idle. High active pressure signals the database is a bottleneck for concurrency.

Refresh endpoint

performance/api/resources/ returns the page context as JSON, staff-only. The service selector is passed through as a query parameter.