OpenTelemetry
The telemetry standard and instrumentation layer used by AI services.
Component category
Telemetry and instrumentation
Component description
OpenTelemetry is the standard instrumentation and transport layer used by BullSequana AI services to emit traces, metrics, and logs in a consistent way.
Why it is used
It connects application activity to the shared Foundation observability stack. The BSQAI API and MCP Gateway export application telemetry to the mTLS-authenticated Alloy gateway. Tempo also sends its own operational telemetry to Alloy.
Alloy remote-writes metrics to Prometheus, which forwards them to Grafana Mimir for long-term storage per tenant, and routes application logs to Loki and traces to Tempo using server-authorized tenant context. Log records carry tenant_id; trace spans carry tenant.id. Missing or malformed OTLP tenant context is quarantined, not assigned to the platform tenant. Infrastructure stdout/stderr follows a separate platform-scoped collection path. See Observability and audit for exporter and query prerequisites. KServe metrics are collected separately through Prometheus scraping; KServe workloads are not configured as general OTLP trace or log emitters.
GenAI-specific workflow spans also use an isolated OpenTelemetry provider to reach MLflow without replacing the platform-wide exporter.
Learn more
Content in Portal chat traces
The BSQAI API traces Portal chat sessions of signed-in users without conversation content. Model, tool, run, and history-compaction spans omit messages, tool definitions, call IDs, run metadata, exception messages, and status descriptions. This applies to the platform traces and to MLflow GenAI tracing. The spans keep token counts, timing, tool names, and failure status. Automatic HTTP spans are not recorded inside these runs. The model and the saved conversation still receive the full content.
Skills telemetry
The BSQAI API and its Skills publication worker export these metrics through the platform OpenTelemetry pipeline:
| Metric | Measures | Labels |
|---|---|---|
skill_operations_total | Skills operation attempts, including failures and cancellations | tenant_id, operation, outcome |
skill_operation_duration_seconds | Latency of each operation attempt | tenant_id, operation, outcome |
skill_failures_total | Failed operation attempts | tenant_id, operation, outcome, error_code |
skill_publications_total | Publication outcomes, ready or failed | tenant_id, kind, outcome |
skill_publication_duration_seconds | Time from submission to its publication outcome | tenant_id, kind, outcome |
tenant_id is the organization ID; the platform-wide staging cleanup uses platform. Operation outcomes are success, failure, or cancelled. Skills operations also produce skills.<operation> spans that record the tenant, operation, outcome, and error code. Skills error logs record error types or codes, not package contents. These metrics are operational signals, not an audit or billing record. Usage and BCU consumption come from the BSQAI API's usage records, not from OpenTelemetry; see Usage analytics.
Interacts with
BSQAI APIandMCP Gatewayas instrumented application servicesMLflowas the separate destination for GenAI observability traces- Foundation observability services such as
Grafana Alloy,Tempo,Loki,Prometheus, andGrafana Mimir