Skip to content

Release cascade: wait for nuget.org indexing before dispatching to siblings #236

Closed

Description

@thomas-stegemann

Problem

When the Bowire Release workflow finishes a new version, it does these steps back-to-back:

  1. dotnet nuget push to nuget.org (release.yml step 27)
  2. repository_dispatch to every sibling repo tagged bowire-cascade (step 28)

Each sibling's bowire-released.yml listener bumps Kuestenlogik.Bowire to the new version, opens a PR, auto-merges if CI passes, then tags + runs its own Release.

The race: dotnet nuget push succeeds when the upload finishes, but nuget.org's package index doesn't surface the new version through /v3-flatcontainer/... for typically 5–15 minutes after the upload. Step 28 fires immediately after step 27, so every sibling's CI / Release tries dotnet restore against a nuget.org that doesn't know v2.0.0 exists yet, and fails with:

error NU1102: Unable to find package Kuestenlogik.Bowire with version (>= 2.0.0)
  - Found 33 version(s) in nuget.org [ Nearest version: 1.9.0 ]

Hit this on every v2.0.0 cascade sibling (Dis / Kafka / Surgewave / Udp / Akka). Each needed a manual rerun ~15 min later once nuget.org had indexed.

Fix

Insert a Wait for nuget.org indexing step between steps 27 and 28 in .github/workflows/release.yml. The step polls https://api.nuget.org/v3-flatcontainer/kuestenlogik.bowire/index.json until the just-pushed version appears in the versions[] array, with a generous deadline (15 minutes; if nuget.org is slower than that we'd rather fail loud than dispatch broken cascades).

- name: Wait for nuget.org indexing
  shell: bash
  env:
    VERSION: ${{ steps.version.outputs.VERSION }}
  run: |
    deadline=$(( $(date +%s) + 900 ))
    while [ "$(date +%s)" -lt "$deadline" ]; do
      if curl -sf "https://api.nuget.org/v3-flatcontainer/kuestenlogik.bowire/index.json" \
           | grep -q "\"$VERSION\""; then
        echo "::notice::nuget.org has indexed Kuestenlogik.Bowire $VERSION"
        exit 0
      fi
      echo "still waiting for $VERSION on nuget.org ..."
      sleep 30
    done
    echo "::error::nuget.org never indexed $VERSION within 15 min — cascade would dispatch broken bumps."
    exit 1

Acceptance

  • Release workflow's NuGet push → Cascade dispatch sequence interleaves an explicit wait step
  • Step polls api.nuget.org/v3-flatcontainer/kuestenlogik.bowire/index.json (no auth required)
  • Polls every 30 s; deadline 15 min; fails loud if missed
  • No-op when the version is already indexed (idempotent)
  • Same step optionally added to the sibling-template bowire-released.yml for when a sibling pushes its OWN cascade-bumped package — same race exists there

Out of scope

  • Mirroring the wait into GitHub Packages publish (faster index, never observed this race there)
  • Rate-limit handling — nuget.org's flatcontainer is CDN-fronted, 30 s polling × 30 attempts is well below any limit
  • Notification when wait > 5 min — log noise is enough for now

Where this came from

Observed live during the v2.0.0 publish on 2026-06-21. The full failure chain was: v2.0.0 published → cascade dispatched at 21:38 → 5 of 7 listener-equipped siblings hit NU1102 between 21:42 and 21:48 → manual gh run rerun per sibling at 22:00 worked because nuget.org had indexed by then (verified via curl …flatcontainer/kuestenlogik.bowire/index.json at 21:59Z).

Activity

  1. changed the issue type fromtoon Jun 24, 2026
  2. thomas-stegemann commented on Aug 17, 2026

    @thomas-stegemann
    MemberAuthor

    Erledigt — Audit-Sweep 2026-08-18.

    The release pipeline waits for nuget.org to index the just-pushed release before dispatching the sibling cascade:

    • .github/workflows/release.yml:400-420 — "Wait for nuget.org to index the release": polls v3-flatcontainer/kuestenlogik.bowire/index.json until the new version appears.
    • Ordering holds: nuget.org push (:384) → indexing wait → cascade dispatch (:441+), so siblings never bump against an unindexed version.

    Minor tuning drift from the body's snippet (architecture is the same): the poll runs every 20s with a 40-min ceiling (not 30s/15min), and on timeout it logs a ::warning:: and dispatches anyway rather than failing hard — a deliberate choice so a slow-but-eventually-successful index doesn't block the whole cascade. Core acceptance (wait-before-dispatch) is met.

  3. moved this from Backlog to Done in Bowireon Aug 17, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    Projects

    • Status
      Done

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions