fix(world-postgres): create runs and initial events atomically

[komly]

Description

Fixes #3656.

Create the run row, its slot marker, and the initial run_created event in one transaction at both creation sites. A concurrent writer can no longer observe a new run before its slot identity exists, and a failed initial-event insert rolls back the run and marker. Existing creator-conflict handling is preserved.

How did you test your changes?

• Seven integration tests against an isolated local PostgreSQL 18.2 database, using the package migrations. Four lock barriers cover visibility before the marker and before the first event for both creation paths; two check rollback on a real constraint violation; one checks duplicate creation. Unmodified main fails six tests and passes the duplicate test; the fix passes all seven.
• Existing package source tests: 27 passed.
• Package build passed with Node 24.21.0 and pnpm 11.24.0.
• Biome: no errors (existing complexity and non-null assertion warnings remain). Changeset status and git diff --check passed.

The new test file uses Testcontainers by default and accepts an explicit loopback WORKFLOW_POSTGRES_URL for local development. It creates and drops only its isolated database. No mocks are added.

PR Checklist - Required to merge

• 📦 Patch changeset included; pnpm changeset status --since=main passes.
• 🔒 DCO sign-off included.
• 📝 Ping @vercel/workflow in a comment once the PR is ready, and the above checklist is complete.

Diff size

Docs — 2 files · +14 / -0

Describe the atomic creation guarantee, local test command, and patch release.

Implementation — 1 file · +81 / -58

Keep both existing creation paths in a transaction through their first event; most of the diff is moving those inserts into the transaction.

Tests — 1 file · +169 / -0

Use real PostgreSQL locks and constraint failures to verify visibility, rollback, and duplicate creation.

[changeset-bot]

🦋 Changeset detected

Latest commit: 561ffa1

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 1 package

Name	Type
@workflow/world-postgres	Patch

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[vercel]

@komly is attempting to deploy a commit to the Vercel Labs Team on Vercel.

A member of the Team first needs to authorize it.

[komly]

@vercel/workflow Ready for review. This fixes #3656 with seven real-PostgreSQL regression cases (six fail on the upstream base), a patch changeset, and a verified, DCO-signed commit.

[VaguelySerious]

Remove all the changes in this file

[VaguelySerious]

(AI) Reverted in cef47e2. Since that removed the documented WORKFLOW_POSTGRES_URL path, I also rewrote run-creation.test.ts to follow the setup of the other world-postgres integration tests (Testcontainers plus pnpm db:push, in-describe win32 skip). Verified the rewritten file still fails 6/7 on main and passes 7/7 with the fix; the full world-postgres suite (219 tests) passes on the branch.

[VaguelySerious]

Suggested change

	Create a workflow run, its slot marker, and its first event in one transaction so concurrent execution and interrupted creation cannot observe partial runs.
	Fix run creation and slot allocation not happening transactionally

[VaguelySerious]

(AI) Applied in cef47e2.

[VaguelySerious]

(AI) Pushed cef47e2 addressing the review comments (changeset wording, README reverted, test aligned with the other world-postgres integration tests). Full world-postgres suite passes on the branch (219 tests).

Additional validation, not committed: a live creation storm (60 run ids, each hit by 3 concurrent run_created and 3 resilient run_started from separate connections) produced no errors other than EntityConflictError on this branch, no duplicate run_created, and no legacy wevt_ ids. It also did not reproduce the window on unmodified main, so the lock-barrier tests in this PR are the right way to pin the behavior.

One note, not blocking: run creation now costs BEGIN/COMMIT round trips around the three inserts (5 statements instead of 3 autocommits). Creation-only and the same shape as the existing step_started / step_created transactions.

[github-actions]

Backport PR opened against stable: #4132. Merge conflicts were resolved by AI — please review carefully. (backport job run)