mattrobinsonsre/terrapod v0.33.1 on GitHub

v0.33.1 is a fix release covering two mirror-endpoint issues, both surfaced in production against v0.33.0.

Bug Fixes

Mirror OOM-killed the API when serving large providers — the v0.33.0 lazy h1 backfill loaded entire provider archives into RAM (twice: once via storage.get(key), then again inside the hash function via fh.read() per zip entry). hashicorp/aws at ~500 MB per platform blew past the 2 GiB pod limit, exit-137-killed both API pods in a tight crashloop, and surfaced as HTTP 500 from /v1/providers/.../version.json to the runner. Fixed by chunked-per-entry h1 compute (~1 MB peak) and a new _compute_h1_from_zip_path that operates on an on-disk tempfile.
Mirror returned empty archives for self-hosted providers — the platform terrapod provider (and any operator-published provider) lives in the registry tables, not the cache tables that the mirror consulted. Result: the runner's lock-extender logged no h1 from mirror for provider and fell back to tofu providers lock for every self-served lookup. Added a "Tier 0" registry lookup at the top of get_or_fetch_platforms for requests whose hostname matches the configured external_url; eager h1 compute at upload time avoids lazy-backfill on first read.

API tempfile policy — h1-backfill streams now write to the CSP-attached ephemeral PVC (settings.vcs.tmpdir, default /var/lib/terrapod/tmp) instead of /tmp (which is an emptyDir{} tmpfs on the API pod and would only move the OOM, not fix it). Audited the rest of services/; policy_vcs_poller._download_archive had the same pattern and got the same one-line fix. The CSP-disk rule is now Architecture Principle #14.
New migration adds registry_provider_platforms.h1_hash (nullable string, lazy-backfilled for existing rows, eagerly computed for new uploads).

Stable — same code shape as v0.33.0 plus the mirror fixes; no API contract changes.

Full Changelog: v0.33.0...v0.33.1