Observability

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

Metrics endpoints

The GMS and frontend rocks embed the JMX Prometheus exporter, enabled by the charm. Both services share the pod network namespace, so they use distinct ports:

Target	Port	Content
datahub-gms	4318	JVM metrics and DataHub GMS application metrics
datahub-frontend	4319	JVM metrics
datahub-actions	-	No metrics endpoint (logs only)

Prometheus distinguishes the two targets by scrape job name; the instance label carries Juju topology, not the host and port.

Metric namespaces

Prefix	Source	Examples
java_lang_*	JVM MBeans	java_lang_Memory_HeapMemoryUsage_used, java_lang_Threading_ThreadCount
jvm_*	JMX exporter built-in collectors	jvm_gc_collection_seconds_sum
metrics_com_linkedin_metadata_*	DataHub application (Dropwizard)	Kafka consumer lag, aspect ingest latency, OpenSearch bulk-processor rates

Grafana dashboard

The grafana-dashboard relation ships the DataHub Monitoring dashboard with four rows:

Row	Panels
DataHub availability	Availability stat per metrics target
GMS Operations	Kafka consumer lag, database ingest latency, Elasticsearch search latency, Elasticsearch bulk writes
JVM	Heap usage ratio, heap used, GC time rate, thread count, process CPU load
Logs	Per-container log panels: GMS, frontend, and actions

The dashboard uses the standard COS datasource and Juju topology variables, so it filters per model, application, and unit.

Alert rules

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

Alert	Severity	Condition
DatahubServiceDown	critical	A DataHub metrics target has been unreachable (up == 0) for more than 5 minutes.
DatahubHeapStarvation	warning	A JVM has been above 90% heap usage for 10 minutes.
DatahubExcessiveGC	warning	A JVM spent more than 50% of wall-clock time in garbage collection over the last 5 minutes, sustained for 10 minutes.
DatahubKafkaConsumerLag	warning	End-to-end lag of the embedded metadata consumers (MCL/MCP) has been above 5 minutes for 15 minutes; search and graph indices are falling behind ingestion.

Logs

The logging relation forwards Pebble-captured logs from all three containers (datahub-gms, datahub-frontend, datahub-actions) to Loki, labeled with Juju topology and the pebble_service label per container.