Known limitations

​

Deployment

• No WAL replication. Intake nodes do not replicate the write-ahead log. Losing an intake volume loses the records that were not flushed to object storage yet.
• Jobs that run on one node. The primary alert node runs cross-cluster event push and gossip, service graph recompute, APM rollup, status page maintenance, synthetic monitoring automation, slow query analysis, and incident root cause analysis. These jobs do not scale with the number of alert nodes. They resume on another node after the primary leaves the membership. See Role-based deployment.
• PostgreSQL only. PostgreSQL is the only metadata database.
• Restart to apply configuration. A running process does not apply changes to the configuration file. See Configuration.
• Forward-only migrations. A database that a newer build migrated cannot be opened by an older build, and builds before 1.0 can change the baseline schema. See Upgrade.

​

Intake

• Request size. A request body larger than 2 MiB is rejected with 413.
• OTLP over HTTP compression. Compressed request bodies are rejected with 400. Disable compression in the sender.
• Histogram buckets. OTLP histograms store count, sum, min, and max. Bucket boundaries, bucket counts, and summary quantiles are not stored.
• Duplicate spans. A span that is sent again with the same trace_id and span_id is dropped.
• Loki push. The endpoint decodes the JSON request shape only. Senders that use snappy-compressed protobuf, such as Promtail and Grafana Alloy, cannot push to it.
• Prometheus remote_write. Native histograms and metric metadata are skipped.
• OTLP gRPC without TLS. The receiver does not terminate TLS. Put a TLS-terminating proxy in front of it for traffic that leaves a trusted network.
• Rate limiting. Only the router role limits requests. The limit counts the requests of the authenticated organization, and every router replica counts separately.
• Reserved stream name. _molesignal is reserved for the service’s own telemetry and rejects public intake.

​

Query

• Long ranges run asynchronously. A query whose estimated size reaches [querier].auto_async_threshold_rows returns 202 with a job to follow. With the defaults, that is a range of about 13.9 hours. Send Prefer: respond-sync to run it in the request. See Query.
• PromQL over the query endpoint. POST /api/v1/query evaluates PromQL as an instant query at time_range.end.
• PromQL subset. The engine implements a subset of PromQL. See PromQL support matrix.
• Rows per query. A query that would read more than [querier].max_scan_rows rows (100 million by default) is rejected with 400. Narrow the time range or add filters. PromQL applies the limit to each selector.
• Freshness in a role-split cluster. A query reads the files that the catalog lists, so a record that an intake node still buffers becomes visible after the next flush, within [intake].flush_interval_secs. A standalone node also reads its own buffer.

​

Alerting

• Real-time rules count per intake node. A real_time rule counts the records that one intake node receives, and the counts live in memory. In a role-split cluster every intake node counts on its own, and a restart starts the count at zero. The statement is limited to SELECT count(*) FROM <stream> [WHERE <condition>]. See Real-time rules.

​

Security and identity

• Internal gRPC is unauthenticated. Port 5082 carries node-to-node traffic without authentication. Keep it on a private network.
• No multi-factor authentication. The web UI has no multi-factor authentication or passkeys. Use SSO with an identity provider that enforces it.
• Password reset needs email. Resetting a password sends a link by email, so [notify.smtp] and [http].external_url must be configured.
• Sessions. The account settings page manages only the session of the current browser.

​

Reports

• PDF and PNG rendering needs headless Chrome. The release container image includes Chromium. A binary installation needs a Chrome installation, and [scheduled_reports.renderer].base_url must reach the web UI.