Ollie

Getting started

Mon, 01 Jan 0001 00:00:00 +0000

This walks from a fresh clone to per-pod HTTP metrics on a real nginx Deployment, scraped from the agent’s :9090 endpoint with k8s.pod.name / k8s.namespace.name / k8s.deployment.name labels attached by OBI.

Kind is the reference cluster — the same flow works on any Linux 6.x cluster with eBPF support (GKE 1.33+ on cos-125+, EKS with bottlerocket, k3s with --snapshotter=overlayfs, etc.); only the image-loading step differs.

Prerequisites

A clone of github.com/gke-labs/in-cluster-observability.
docker, kubectl, kind, go 1.26+.
A kernel with BTF + CO-RE (any 5.10+ kernel built with CONFIG_DEBUG_INFO_BTF=y; Kind’s default node image qualifies).

1. Spin up a Kind cluster

kind create cluster --name ollie-v03

The default single-node cluster is enough. Multi-node works too — the agent is a DaemonSet, you’ll get one agent per node.

What works today

Mon, 01 Jan 0001 00:00:00 +0000

Ollie is at v0.3 dev preview. The capture path works end-to-end on real clusters, but several pieces from the original requirements (CRD-driven onboarding, in-cluster store + PromQL, HPA custom-metrics API, AI-agent streaming, TLS uprobes, peer.k8s.* on the L7 path) are still ahead. This page is the inventory.

What’s captured

L4 TCP — every flow on the node

OBI’s network mode attaches an in-kernel socket filter and observes every TCP byte flowing through the node, regardless of which process owns it. The metric you’ll see is:

Architecture

Mon, 01 Jan 0001 00:00:00 +0000

Ollie is built on top of OpenTelemetry eBPF Instrumentation (OBI). OBI does the eBPF capture and K8s metadata enrichment; Ollie adds the production deployment scaffolding plus an OTLP receiver carve-out that future milestones use as the hook point for CRD-driven onboarding (v0.4) and the in-cluster store (v0.5).

This page describes the data path that’s actually shipping in v0.3. For the full design log and the reasoning behind each choice, see the ADRs in the repo — especially ADR-0018 (sibling-container model) and ADR-0021 (lean v0.3 pivot to OBI native enrichment).

Roadmap

Mon, 01 Jan 0001 00:00:00 +0000

The project is on a milestone cadence — v0.1 → v0.2 → v0.3 → … → v1.0. Each milestone branch in upstream is the integration point for that milestone’s work, and milestone PRs stack (each one PRs to the previous) so the bottom of the stack collapses into main once a milestone is ready to ship.

Issue numbers below link to upstream gke-labs/in-cluster-observability.

Shipped

v0.1 Foundation — single Go module, public-package skeletons

Issues #64–#69. Established the single AP root, the pkg/{capture,store,query,sink,topology,controller,schema,obsapi} public API skeletons with stability tags, the OBI adapter shell, the distroless agent image, and the minimal DaemonSet manifest. Nothing observable yet — the binary stays alive on signals; that’s it.

Contributing

Mon, 01 Jan 0001 00:00:00 +0000

The canonical contributor reference is AGENTS.md in the repo root — it covers conventions, build commands, layout, and the OBI integration boundary. This page is a short pointer.

Filing issues

All issues and milestones live in upstream gke-labs/in-cluster-observability. Even if you’re working on a fork, file issues against upstream — the fork’s issue tracker is intentionally empty. gh repo set-default is configured in the upstream clone so gh issue list / gh issue create target upstream by default.