A small Kubernetes pod-log collector. Watches a namespace (or a whole cluster), tails logs from every container in every running pod, and writes them to disk in a per-run folder along with a manifest describing what was captured.

cargo build --release

The binary lands at target/release/driftwood.

# Tail every container in the current kube context's default namespace.
driftwood

# Scope to a namespace and a label selector.
driftwood -n nativelink --label-selector app=nativelink

# Cross-namespace, on a non-default kube context.
driftwood --context staging --all-namespaces

# Use a specific kubeconfig file and pin a run name (useful for retries).
driftwood --kubeconfig ~/.kube/foo.yaml --run-name repro-issue-42

Stream output to stdout while logs accumulate on disk. Ctrl-C (or SIGTERM) finalizes the run manifest before exit.

Each invocation creates a run folder:

logs/<run-id>/
├── run.yaml                                # manifest for this run
└── <namespace>/
    └── <pod>/
        ├── pod.yaml                        # if --dump-config
        ├── <container>.log                 # live tail
        └── <container>.previous.log        # if the container had restarted

<run-id> defaults to a UTC timestamp (20260506T142301Z) and can be overridden with --run-name. The root (logs/) can be moved with --logs-dir.

Written at start, updated as pods are picked up, finalized on shutdown. Captures everything you need to reproduce or interpret a run:

run_name: repro-issue-42
run_dir: logs/repro-issue-42
start_time: 2026-05-06T14:23:01Z
end_time: 2026-05-06T14:51:18Z
context: staging
cluster_url: https://10.0.0.1
namespace: nativelink
all_namespaces: false
label_selector: app=nativelink
field_selector: null
argv: [driftwood, --context, staging, -n, nativelink, --label-selector, app=nativelink]
captured:
  - pod: scheduler-0
    namespace: nativelink
    uid: 8c1f...
    started_at: 2026-05-06T14:23:04Z
    pod_yaml_path: null
    containers:
      - name: scheduler
        log_path: logs/repro-issue-42/nativelink/scheduler-0/scheduler.log
        previous_log_path: null

• Pods are picked up when their phase first becomes Running. A pod's UID is remembered so the same pod isn't logged twice in one run.
• If a container has a non-zero restart_count at the moment driftwood attaches, driftwood fetches that container's previous-instance logs (previous: true) into <container>.previous.log. Live restarts that happen during a run are not handled yet.
• Watching is in-memory only; if driftwood is restarted, it starts a fresh run folder and reattaches to whatever pods are running at that moment.

make cleanup-logs

Removes everything tracked under logs/ via git clean -dfX.