spi-fleet

The fleet is the forks: list in osdu-spi-stack ops/environments/shared.yaml. .github/fleet/rollout.json sets lanes, the number of deploy lanes allowed at once, and the few merge-ordering exceptions. fleet.py reads both, so a new fork needs no edit here.

Never

• Start more than lanes validations at once. Every fork PR and every main merge borrows the shared backing environment, and the environment has no queue across forks. Dispatch gh workflow run sync-template.yml -R <repo> for lanes forks, then wait until watch shows no pending runs before dispatching more.
• Rerun a failed run without the person's go. Rerun an environment failure once, after the other lanes drain: gh run rerun <id> --failed -R Azure/osdu-spi-<service>.
• Patch a fork by hand to get it green. Stop the line and fix forward in the template.
• Approve a pull request, or merge without the person's go.

Watch

python3 .github/skills/spi-fleet/fleet.py watch --wait            # template-sync PRs, poll until nothing is pending
python3 .github/skills/spi-fleet/fleet.py watch --prefix release-please--   # any head-branch prefix
python3 .github/skills/spi-fleet/fleet.py watch --main            # the latest push run on each main

Each fork is judged by the Validation run for its PR's head SHA. A failure is an environment failure when Validation raised it with the annotation title Environment: the environment stayed undeployable, the pin stayed refused, or the digest never went live. Anything else is a failure of the change, and so is an environment failure that has already been rerun once. Exit codes: 0 green, 1 real failure, 2 pending, 3 environment failure that waits for a rerun.

Merge

python3 .github/skills/spi-fleet/fleet.py merge          # print the groups
python3 .github/skills/spi-fleet/fleet.py merge --go     # merge them, after the person says go

merge re-reads every PR first. It merges green, approved PRs in fleet order, and ordering exceptions follow rollout.json. Before each group it counts the Validation runs already queued or running across the whole fleet, PR and main alike, and merges only into the lanes left free. It waits for each merge SHA's push run on main and stops the line on any failure, or on any PR that did not merge. A merge whose push changed only paths Validation ignores shows no run; after --grace seconds (default 600) it is reported unvalidated and the line carries on.

Each merge is written to fleet-merge-<date>.jsonl as it happens, with the PR, merge SHA, image and the main-snapshot digest it replaced, and a second record follows with the main run's result. When a forward fix is slow, that digest is the last-good pin the person can choose.

Change the rollout

Raise lanes in rollout.json once a rollout at the current value runs clean. An ordering entry may name only forks in the fleet list; fleet.py refuses to run otherwise. A new environment failure is classified by giving its ::error in validate.yml the title Environment, not by adding a rule here. Run bash .github/local-actions/spi-fleet-tests/run-tests.sh.