Service map
The live dependency map Avuru Obs draws itself — services, the databases and brokers behind them, and the connections nobody instrumented — with no code changes.
Traces
Investigate distributed requests with waterfalls, trace comparison and correlated logs from eBPF and OpenTelemetry spans in Avuru Obs.
Logs
Collect logs, preserve trace and span IDs, and investigate requests with correlated OpenTelemetry logs in Avuru Obs.
Metrics
Read service rate, errors and duration alongside traces, and ingest OpenTelemetry infrastructure metrics in Avuru Obs.
Cluster X-Ray
Inspect Kubernetes Nodes, Pods and trace-backed connections in an interactive isometric view.
Profiling
Continuous, eBPF-based CPU profiling with per-service flame graphs.
Errors
Exceptions in your telemetry become deduplicated, triageable issues — derived in-database or ingested from Sentry SDKs.
Service health
One board that answers "what's broken and how bad" — group status with criticality tiers and dependency propagation, derived from the traces you already send.
Alerting
A webhook when a service goes down — and another when it recovers. Rules on service-health status, config-driven, hot-reloaded, guarded outbound.
Network health
Per-edge TCP RTT, failed connections and retransmits on the service map, measured in the kernel by OBI — connection health with no traces, SDKs, or app changes.
Green
Per-service energy (Wh) and carbon (gCO2e) from CNCF Kepler, correlated with the service map you already run — carbon budgets, per-request intensity, and a CSRD-ready export. Zero code changes, no egress.
Service mesh
The fabric your services run on — every proxy by role with what it carries and what it counted, whether the control plane is still programming the mesh, every workload the cluster runs with what was declared for it beside what its proxy observed, and what your mesh configuration gets wrong without emitting a single span.
Cost
Compare Kubernetes CPU and memory requests with observed peaks to investigate overprovisioning, with optional rates and no automatic workload changes.
AI
What your applications asked models to do — per model and per caller, with tokens, latency, failures, truncation and an optional cost. Read from spans you are already sending, and with prompts dropped at the gateway by default.