Zero-code tracing only counts if it does no harm. The wedge gate now proves it: CI keeps a deliberately fragile workload — tight CPU limit, aggressive liveness probe — Ready under the sensor, and a new opt-in mode lets cautious fleets adopt uprobes label by label.
A new Alerting module closes the loop on service health: when a group, service, or tier crosses into a bad state and stays there, avuru-obs fires a webhook — into Slack, PagerDuty, Opsgenie, Alertmanager, or any endpoint that accepts one. It's derived from the health you already compute; no probes, no new signal.
The service map now shows connection-level health on its edges: per-edge TCP RTT and failed/reset connections, measured in the kernel by OBI — no traces, no SDKs, no app changes. A link that's slow or refusing connections is now visible on the same topology you already watch.
A new Service Health board rolls each service's RED health up into consolidated group status — with criticality tiers and a rule that a service can't read green while a critical dependency is red. It is derived from the traces you already send: no health checks, no probes, no new storage.
A new Error tracking module turns the exceptions already flowing through avuru-obs — span exception events, error spans, and ERROR/FATAL logs — into first-class issues: deduplicated by fingerprint, each with a stack trace, an occurrence timeline, a link to a representative trace, and a triage lifecycle. No code change required.
Installing no longer means cloning the repo. The Helm chart and every image are
now published to GHCR — cosign-signed and built for amd64 + arm64 — so a
single command brings up the whole stack.
avuru obs v0.1.0 is tagged: install the Helm chart on a fresh Kubernetes cluster and watch a live service map appear in under five minutes, zero app changes — a promise enforced in CI, not a demo.
The Profiling screen is live: cluster-wide CPU flame graphs, sampled by
eBPF with zero app changes and stored in the same ClickHouse as your traces,
logs and metrics. Experimental and opt-in for v0.1
(sensor.profiler.enabled=true) — the upstream alpha loader does not support
every kernel yet.
The Nodes screen is live: CPU, memory and network per node, with the busiest pods underneath — collected by the sensor's node agent (kubeletstats), stored in the same ClickHouse as everything else.
The Metrics screen is live: request rate, error percentage and latency percentiles over time for the busiest services of the window — derived from the same spans the trace explorer uses, so there is nothing extra to instrument or operate.