TODO

Refactor: host-namespace paths everywhere in the file source

What

Move the file source so that the exporter reads, caches, and labels every on-disk artefact by its host-namespace path (the path the operator reads kubectl exec-ing into a node, e.g. /var/lib/kubelet/pki/...), and applies PathMapping.From → To translation uniformly at every filesystem operation — open, Stat, Lstat, ReadDir, Readlink.

Today (post-symlink-mapping change) the chart writes in-pod paths into the configmap (/mnt/watch/file-<sha1>/var/lib/kubelet/pki/...), and PathMapping is consulted only when an absolute symlink target is read back — i.e. only inside pkg/fileglob.handleSymlink. The proposal is to flip the polarity: have the chart write raw host paths, and let the runtime translate transparently for every FS op.

Why we'd want to

Drop trimPathComponents. chart/templates/configmap.yaml currently sets metrics.trimPathComponents: 3 for hostPath sources so that /mnt/watch/file-<sha1>/var/lib/kubelet/pki/foo.pem displays as var/lib/kubelet/pki/foo.pem in filepath Prometheus labels. That's a workaround for a label-cosmetics problem that goes away if the runtime knows the host path natively.
One translation point, not two. Right now pkg/fileglob.handleSymlink has bespoke translation + containment logic. With a universal translator, the same logic covers every code path that touches the FS, so any new FS-using feature is covered for free.
Operator UX. Labels match what humans type. No mental mapping from /mnt/watch/file-9dff…/var/lib/kubelet/pki/... back to the real file.

Why this is not done as part of the symlink-mapping change

Blast radius. It touches every FS op in fileglob, not just symlink resolution. The cache key in pkg/source/file/file.go becomes the host path (today it's the walker's Path = in-pod), affecting cache invariants and the SkipUnchanged machinery.
Public API ripple. pkg/cert.SourceRef.Location and the registry's filepath label semantics change in user-visible ways. Needs a deprecation cycle for downstream consumers reading those labels (Grafana dashboards, alert rules).
Test churn. Every fileglob/source/file/cmd test that asserts on paths needs to be revisited. Today's symlink-mapping change is surgical; this would be a multi-package rewrite.
trimPathComponents removal is a chained breaking change. The values key would have to be deprecated separately.

When to revisit

After the symlink-mapping change has had at least one release in the wild and we have feedback from real DaemonSet operators about the shape of filepath labels they'd prefer.
When we next plan a pkg/cert API revision (anything that already forces consumers to update their dashboards is a natural moment to bundle this in).
If a third concrete need for "exporter knows the host path" surfaces (e.g. exposing the host path in error messages, in /healthz diagnostics, in a future audit trail). Two needs is a coincidence, three is a pattern that justifies the refactor.

Sketch of the change

pkg/fileglob/walkfs.go (new): a small WalkFS decorator that takes []PathMapping and a base WalkFS, applies the longest-prefix From → To rewrite to every method's name argument, and forwards. The walker becomes oblivious to the translation — it sees host paths everywhere.
pkg/source/file/file.go: drop the special-case if e.LinkTo != "" { readPath = e.LinkTo } (no longer needed — Reader would also be wrapped to apply the translation).
chart/templates/configmap.yaml: write host paths in paths: and pathMappings:; remove metrics.trimPathComponents:.
chart/values.yaml: deprecate metrics.trimPathComponents (keep for backwards compat for one release, then remove).
Tests: every test that hardcodes /mnt/watch/file-<sha1>/... flips to host paths.

3.9 KiB Raw Permalink Blame History