Observability

Reference for the telemetry Superset exposes over its observability relations. For setup instructions, see Observe Superset.

Metrics endpoint

Every unit exposes Prometheus metrics on port 9102. What produces them depends on the application’s charm-function:

Application	Exporter	Content
UI (app-gunicorn, app)	StatsD exporter	Superset application metrics, emitted by the workload over UDP on port 9125 and translated to Prometheus format
Beat (beat)	StatsD exporter	Scheduler activity
Worker (worker)	Celery exporter	Celery worker and task metrics, read from the Redis broker

Metric namespaces

Prefix	Source	Examples
superset_*	Superset StatsD logger	superset_welcome, superset_log, superset_DashboardRestApi_get_success, superset_ChartDataRestApi_data_time_count, superset_sqllab_query_time_executing_query_sum
celery_*	Celery exporter on worker units	celery_worker_up, task counters and timings

Superset’s REST API metrics follow the pattern superset_<ApiClass>_<endpoint>_<outcome> for counters and superset_<ApiClass>_<endpoint>_time_<sum\|count> for timings, so per-endpoint latency is the ratio of the two.

Grafana dashboard

The grafana-dashboard relation ships the Superset Metrics dashboard with these panels:

Group	Panels
Metrics	Is Superset up?, rate of all Superset events, dashboard loads, chart data loads
SQL Lab	SQL Lab query execution time, time to fetch SQL Lab results
Timings	Time to load charts, time to get dashboard dataset
Workers	Worker count and availability, derived from celery_worker_up
Logs	Separate panels for UI, worker, and beat logs

The dashboard filters on Juju topology, so it works per model, application, and unit.

Alert rules

Three Prometheus alert rules load automatically over the metrics-endpoint relation:

Alert	Severity	Condition
SupersetDown	critical	No Superset instance has served the welcome endpoint for 2 minutes.
SupersetHalfOrMoreDown	high	Half or more of the Superset instances have stopped serving for 2 minutes.
WorkersDown	critical	No Celery worker has reported as up for 2 minutes.

Note

The availability rules are traffic-derived: they measure requests served rather than a health endpoint, so a deployment with no user traffic at all can trigger SupersetDown. Take that into account when tuning thresholds for a deployment that is idle for long periods.

Logs

The logging relation forwards the workload logs of every related application to Loki, labeled with Juju topology. The workload also writes a rotated application log to /var/log/superset.log inside the container.