Skip to main content
Version: 3.6.0

Telemetry Export

The platform can send copies of four telemetry streams to an external monitoring backend. You switch each stream on separately. None of them is anonymized: records leave the cluster as they were produced, so decide per stream what you are willing to send.

StreamPersonal data by designPersonal data possibleRisk
MetricsNoResource namesLow
AlertsNoResource namesLow
TracesPseudonymous user, API key and tenant IDsURL paths and query stringsMedium
LogsPseudonymous user, API key and tenant IDsAnything a workload writes to its outputHigh

Pseudonymous IDs are internal identifiers, not names or email addresses. Anyone with access to your identity provider can map them back to a person, so treat them as personal data.

Metrics​

Numeric time series about the cluster and the platform: resource usage, request rates, latencies, GPU use, and the health of databases and the service mesh. Every series the cluster's Prometheus collects is sent.

Labels identify Kubernetes objects, such as namespaces, pods, nodes and model deployments. The platform's own metrics carry no user or tenant identifiers. Names you choose for your resources appear as label values, so a deployment named after a person sends that name.

Alerts​

Alerts that fire in the cluster, with their labels and descriptions. They contain the same kind of data as metrics.

Traces​

One record per step of a request as it moves between platform services: which service handled it, how long it took, and whether it failed. Prompt and answer content is never part of a trace.

Each authenticated request carries the caller's user ID, API key ID and tenant ID. Records also contain request URLs, including query strings, and the text of database queries without their parameter values. Metric scraping and telemetry traffic are left out.

Logs​

The output of every container in the cluster, except the monitoring, kube-system and argocd namespaces and the service mesh proxies. This includes workloads your tenants deploy and third-party components such as the identity provider and the databases.

Three kinds of records are never sent: audit events, captured prompt and answer content, and per-request mesh access logs. Authorization headers are removed. The rest of each record is sent unchanged.

The platform cannot guarantee what other software writes to its logs. Expect pseudonymous IDs on request logs, names of uploaded files, and usernames, email addresses and client IP addresses from sign-in events. A model server with request logging enabled also logs prompts. If logs must not leave the cluster with personal data in them, keep this stream off.