Skip to content

ci(release): switch PyPI watch to JSON API instead of pip index - #70

Merged
lukeocodes merged 1 commit into
mainfrom
fix/pypi-watch-json-api
May 8, 2026
Merged

lukeocodes merged 1 commit into
mainfrom
fix/pypi-watch-json-api

Conversation

@lukeocodes

Copy link
Copy Markdown
Contributor

Problem

The "Wait for new deepctl version on PyPI" step in .github/workflows/release.yml polls pip index versions deepctl. That endpoint queries PyPI's simple index, which sits behind a CDN cache that can lag 5-15 minutes after a release is actually published.

Two consecutive releases hit this:

In both cases PyPI confirmed the release was installable (any user could pip install deepctl==X.Y.Z), but the cached index hadn't picked it up before the 10-minute window (30 × 20s) ran out. The bump-brew-formula job failed → we had to regenerate the formula manually.

Fix

GET https://pypi.org/pypi/deepctl/<version>/json — the PyPI JSON API endpoint flips to HTTP 200 the instant a release is published, no CDN cache between us and the source-of-truth catalog.

URL="https://pypi.org/pypi/deepctl/${VERSION}/json"
if [ "$(curl -fsS -o /dev/null -w '%{http_code}' "${URL}")" = "200" ]; then
  echo "deepctl==${VERSION} is live on PyPI"
  exit 0
fi

Same 30 × 20s loop shape, total worst-case budget unchanged at 10 minutes. Typical wait should drop from 5-15 minutes to a single iteration.

Verification

While diagnosing v0.2.22 today:

$ curl -fsS -o /dev/null -w "HTTP %{http_code}\n" https://pypi.org/pypi/deepctl/0.2.22/json
HTTP 200

$ pip index versions deepctl
deepctl (0.2.21)        # ← still didn't see 0.2.22

JSON API: live. Simple index: lagging. The fix is to probe the source of truth.

What it doesn't change

  • The 30 × 20s budget shape is preserved so we don't accidentally let bad releases through faster than the existing flow.
  • bump-brew-formula's downstream pipeline (homebrew-pypi-poet resource generation, etc.) is unchanged. That step's own pip install deepctl==<version> may still hit the simple-index lag for ~30-60s after this watch step succeeds, but that's a much smaller window and the install retry is implicit in the script's own behaviour.

The 'Wait for new deepctl version on PyPI' step in the release workflow
polled `pip index versions deepctl`, which queries PyPI's simple index
behind a CDN cache that can lag 5-15 minutes after a successful publish.
v0.2.21 (PR #3 on homebrew-tap) and v0.2.22 (PR #4 on homebrew-tap) both
hit this — PyPI confirmed the release was installable, but the cached
simple index hadn't picked it up by the time the 10-minute window
(30 × 20s) ran out, so bump-brew-formula failed and we had to re-run
the formula bump manually.

Switch to GET https://pypi.org/pypi/deepctl/<version>/json. That endpoint
flips to HTTP 200 the instant a release is published. Same 30 × 20s loop
shape, just a different probe — total runtime budget unchanged at
10 minutes worst case but the typical wait should drop to a single
iteration.

Verified locally that the JSON endpoint already returned 200 for 0.2.22
while `pip index versions` still missed it. No other changes needed —
release-please outputs and downstream jobs are untouched.
@lukeocodes
lukeocodes merged commit 63903ea into main May 8, 2026
39 checks passed
@lukeocodes
lukeocodes deleted the fix/pypi-watch-json-api branch May 8, 2026 19:34
dg-coreylweathers added a commit that referenced this pull request Nov 12, 2026
## Summary

After the release workflow opens the "Bump deepctl to X" PR on
deepgram/homebrew-tap, it now merges that PR in the same step, using the
same `CLI_TAP_SYNC_PAT` that pushed the branch. The step is also made
safe to re-run and safe against two releases racing on the tap.

## Why

The tap has no branch protection, no checks and no reviewers, so an open
bump PR only delays Homebrew users. The 0.3.1, 0.3.2 and 0.3.3 bumps
(homebrew-tap #9, #10, #11) all sat unmerged while the tap's formula
stayed at 0.2.26. #11 was merged by hand; this keeps that from happening
again.

The formula is generated from PyPI metadata by
`scripts/bump_brew_formula.py`, so there is nothing for a reviewer to
judge in the tap PR itself. The merged PR stays as the audit trail, and
a broken formula is fixed by editing the template, as the generated PR
body already says.

## What changed (`bump-brew-formula` job only)

- **Serialized.** The job gets `concurrency: {group: homebrew-tap-bump,
cancel-in-progress: false}`, so two root releases in quick succession
bump the tap one after the other instead of both reading the same
starting formula.
- **Version guard.** Before touching git, the step reads the version the
tap currently serves from `git show HEAD:Formula/deepgram.rb` (the
regenerate step has already overwritten the working tree, so HEAD is the
tap's state) and the regenerated version from the working tree.
- It fails the job if the tag, the current formula or the regenerated
formula is not a plain `X.Y.Z` version, or if the regenerated formula
does not point at the tagged version.
- It exits 0 with a warning unless the new version is strictly greater
than what the tap serves. Re-running an older release can no longer
downgrade the tap, and re-running a release that already landed is a
no-op.
- **Formula syntax check.** `ruby -c Formula/deepgram.rb` runs before
the commit. This is a Ruby syntax check only: it catches a broken
template or a mangled resource block, not a wrong sha256 or a failing
`brew install`.
- **Re-runnable push and PR.** The branch is pushed with `--force`, so a
re-run replaces a branch left behind by an earlier failed attempt. If an
open PR for that branch already exists, it is reused instead of `gh pr
create` failing on the duplicate.
- The existing-PR lookup only accepts same-repo PRs (`isCrossRepository
== false`) authored by the PAT's own login. The tap is public, so anyone
can open a fork PR whose head branch is named `bump-deepctl-X`; an
unfiltered lookup could have merged it with the PAT.
- The resulting PR URL must match
`https://github.com/deepgram/homebrew-tap/pull/N` or the job fails.
- **Merge with retry.** GitHub computes mergeability after a PR is
created and gh does not wait for it, so `gh pr merge <url> --merge
--delete-branch` is tried up to five times with a short backoff. If a
merge attempt fails but the PR's state is `MERGED` (merged by hand
mid-loop, or merged and then the branch delete failed), that counts as
success. After five failed attempts the job fails with an error
annotation.
- **Stale bump cleanup.** After a successful merge, open
`bump-deepctl-*` PRs that are same-repo, authored by the PAT owner, and
for a strictly older plain `X.Y.Z` version are closed with their branch
deleted and a comment linking the superseding PR. This is best effort:
listing or closing failures emit a warning and never fail the job.

## Checked

- The workflow YAML parses, and `bash -n` passes on the extracted step
script.
- actionlint 1.7.12 with shellcheck 0.11.0 reports 0 findings on
`.github/workflows/release.yml`. The one intentional single-quoted jq
`$login` carries a `shellcheck disable=SC2016` with a comment.
- No job other than `bump-brew-formula` is modified.
- Typed `ci:` so release-please adds no changelog line and bumps no
version for CI plumbing (precedent: #66, #70).

🤖 Generated with [Claude Code](https://claude.com/claude-code)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant