Data stores

OpenRUM keeps control data and telemetry in different stores on purpose. Control data is small, relational and must never be lost. Telemetry is large, append-only and queried in aggregate. Each store below has one job, and only two of them are a source of truth.

Store	Holds	Source of truth
PostgreSQL	Accounts, organizations, Projects, settings and workflow state	Yes, for control data
ClickHouse	Events and the aggregates built from them	Yes, for telemetry
Kafka	Accepted Envelopes waiting for the Consumer, plus a dead-letter topic	No, a buffer
Redis	Rate-limit counters, caches and connection status	No, rebuildable
Object storage	Source Map Artifacts, optional	For the artifact files

PostgreSQL

PostgreSQL is the control plane. The tables fall into a few groups:

Group	Examples
Identity and access	users, sessions, auth_providers, auth_identities, organizations, organization_members, instance_members
Projects	projects, project_keys, project_environments, project_upload_tokens
Delivery and issues	releases, sourcemap_artifacts, issue_states
Alerts	alert_rules, notification_channels, alert_evaluations, alert_deliveries
Instance operations	instance_settings, instance_secrets, maintenance_jobs, emergency_cleanup_jobs, project_deletions, project_data_purges
Audit	audit_logs, instance_audit_logs

Losing PostgreSQL means nobody can sign in or change configuration, and Ingest loses the source of its Project keys and settings. Events already in ClickHouse are untouched. The PostgreSQL runbook covers operations and recovery.

ClickHouse

ClickHouse holds every stored Event in one raw table, rum_events, and a set of aggregate tables kept up to date by materialized views as Events are inserted. Dashboards and the Console read the aggregates; investigation reads the raw Events.

Table	Used for
rum_events	Every Event, with its ingest_flags array and Fingerprint
project_metrics_1m	Per-minute Project metrics
behavior_metrics_1m	Page views, visitors and custom Event aggregates
measurement_metrics_1m	Web Vitals and other measurements
api_metrics_1m	API request volume, latency and failures
issue_metrics_5m	Issue occurrence counts
usage_records, usage_metrics_1h	Per-Project volume shown under Usage
event_stack_mappings	Mapped stack frames for errors

Events expire by the Project’s retention policy, and aggregates by their own, so raw data can be shorter-lived than the charts built from it. See Data lifecycle and ClickHouse runbook.

Kafka

Ingest writes accepted Envelopes to the event topic, rum-events-v1 by default, and the Consumer reads them as a consumer group. Failed items go to a second topic with the same name plus .dlq. The default Compose setup creates both with 12 partitions.

Kafka is the one store where loss means losing accepted Events, because Ingest has already answered 202. It is also the buffer that lets the Consumer fall behind safely, so its retention must outlast your longest expected ClickHouse outage. Capacity planning starts there, and the Kafka runbook covers lag and recovery.

Redis

Redis holds short-lived state that is safe to lose: the fixed-window counters behind rate limits, query caches, sign-in throttling and each Project’s last-seen SDK status shown during onboarding. When Redis is unavailable, each Ingest process enforces half the configured limit locally, so rate limiting degrades rather than failing open. Nothing in Redis needs a backup. See the Redis runbook.

Object storage

Object storage is optional. It holds the Source Map files uploaded for each Release; PostgreSQL keeps their metadata. Without it, every other feature works, but new Source Maps cannot be uploaded and minified stack frames stay unmapped. Use an OSS or S3-compatible bucket and give it its own lifecycle rules, because artifacts do not follow Event retention. See the object storage runbook.

Which stores a backup must cover

Back up PostgreSQL, ClickHouse and the artifact bucket as one recovery set. Kafka and Redis are not restored from a backup; Kafka only matters for the events that were accepted but not yet stored. Backup and restore gives the procedure.