[vaclav-dvorak]

Summary

• push no longer downloads the repo index when --skip-reindex is set. Bumps plugin.yaml to 0.1.1.

Details

FetchRaw sat outside the if !act.skipReindex guard, and the bytes it returns are read only inside that guard:

b, err := storage.FetchRaw(ctx, repoEntry.IndexURL())   // always ran
if err != nil { ... }
if !act.skipReindex {
    idx := helmutil.NewIndex()
    if err := idx.UnmarshalBinary(b); err != nil { ... } // only reader of b
    ...
} else {
    act.printer.Printf("[DEBUG] Skipping reindex ...")
}

So with --skip-reindex every push did a full s3manager download of index.yaml and threw it away.

Measured on the prod chart repo (exa-helm-charts-prod-c42d): index.yaml is 22.3 MB (41,987 chart versions across 130 charts). v2-helm-charts-s3-prod pushes 129 charts sequentially with --skip-reindex --force, so one build spent ~2.8 GB on GETs it never read. From that build's own log:

step	time
helm package ×129	12s
helm s3 push ×129	440s
helm s3 reindex + index re-stamp	22s

Per-push gaps were tightly clustered (median 4.0s, min 2s, max 5s) — the signature of a fixed-size download, not of the chart uploads, which vary a lot in size. It also explains a steady slowdown with no step change: median build time went from ~240–330s in late August to ~460–490s now, tracking index growth (16.6 MB → 22.3 MB in four days) while chart count only went 111 → 130.

The comment above the fetch — "The gap between index fetching and uploading should be as small as possible to make the best effort to avoid race conditions" — is about the index read-modify-write, which does not happen on the skip-reindex path. It moves with the code it describes.

Unit tests and go vet are unchanged from the base commit (internal/awsutil's TestDynamicBucketRegion fails identically before and after — it reads the ambient AWS region). The minio-backed e2e suite could not run here: the minio/minio image pull is denied without a Docker Hub login.