Repository navigation
Conversation
🦋 Changeset detectedLatest commit: 0aac38f The changes in this PR will be included in the next version bump. This PR includes changesets to release 18 packages
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 |
|
You must have Developer access to commit code to Vercel Labs on Vercel. If you contact an administrator and receive Developer access, commit again to see your changes. Learn more: https://vercel.com/docs/accounts/team-members-and-roles/access-roles#team-level-roles |
🧪 E2E Test Results✅ All tests passed 🛠 Infra Events (absorbed by the harness)Platform anomalies the e2e harness detected and worked around (e.g. a run the queue never picked up, replaced by a fresh run). Clustered timestamps indicate a backend blip; a steady drip indicates a platform issue worth escalating.
E2E Test SummarySummary
Details by Category✅ 💻 Local Development
✅ 📦 Local Production
✅ 🐘 Local Postgres
✅ 🪟 Windows
✅ 🌐 Cross-language Conformance
✅ dynamic-runs
✅ local-postgres-spec9
✅ local-spec9
✅ vercel-http-transport
✅ vercel-spec9
✅ vercel-ws-transport
|
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Sim WorldSimulated world deterministic testing for races. Traces 🟢 world-sim scenario book — 0 fail of 45 total
Full trace: |
fde70e1 to
c76ceea
Compare
c76ceea to
055aa7f
Compare
055aa7f to
60ece23
Compare
📊 Workflow Benchmarks❌ The benchmark run for No benchmark results were produced. ℹ️ Metric definitions & methodologyMetrics — All timestamps are deployment-side; runs are triggered in-deployment, so the CI runner and api.vercel.com sit outside every measured window. TTFS = Cold starts stay in the numbers (real bursty-workload latency, inflates P75+); Best is the warm floor. |
About these numbersSizes are gzip; parentheses show the change against
|
60ece23 to
7c64607
Compare
7c64607 to
affdd51
Compare
Signed-off-by: benpankow <ben.pankow@vercel.com>
The divergence handling landed with the rebase onto the drain fix: a parked VM that diverges on its first pass is replayed from cold without spending the divergence budget. Its test runs with that budget at zero. The stale-resume test now also asserts that no cold replay ran. Signed-off-by: benpankow <ben.pankow@vercel.com>
…s crossed Signed-off-by: benpankow <ben.pankow@vercel.com>
…-of-order park, and a crossed divergence Signed-off-by: benpankow <ben.pankow@vercel.com>
affdd51 to
0aac38f
Compare
Description
Stacked on #5467. There, a parked VM is dropped when another orchestrator wrote in-band since it parked (the read's
seqInBandmoved past the parked count), and the delivery loads and replays the whole log from cold.That case is common once one run's messages land on different instances. I hit it with eve on Vercel running deliveries inline (#5365): each message runs in whichever server instance received it. In one 8-follow-up run, 2 of the 4 parked VMs found were stale and fell back.
The events that orchestrator wrote are in the incremental read, after the parked cursor. This PR resumes the parked VM on those events too, and falls back only when the read could not continue from the cursor (
reloadedFromStart). The fence is unchanged: the delivery adopts the read's snapshot, so its next in-band write still has to match the World's count, and an orchestrator that overlaps after the read is refused there, as before.Replaying another orchestrator's events isn't a new kind of trust. They go through the same events consumer a cold replay uses, and a parked VM that diverges on them falls back to a cold replay of the log, which decides; only that replay's divergence counts against
WORKFLOW_REPLAY_DIVERGENCE_MAX_RETRIES(workflow.parked_session_diverged, withcrossedWriteson the warning). The unchecked part is a warm VM writing new events from state a cold replay didn't build, and #5467 already has that whenever it resumes on a newhook_received.Worth flagging, though not new here: a stale VM can resume while the orchestrator that wrote past it is still lingering with a newer VM. The fence lets one of them write and refuses the other, but with run-ahead on, a retriable step's body can start before its start commits, so the refused racer has run it too. That's the overlap any redelivery makes (see "The in-band fence" in the spec 9 README), and a
maxRetries: 0step never runs ahead. eve doesn't run ahead today; it would matter if it did.How did you test your changes?
parked-session.runtime.test.tsreplaces "loads and replays the whole log when another orchestrator wrote in-band meanwhile", which only bumped the World's count, with a real overlap: the test takes the VMaparked aside, sobis handled from cold, then puts the stale VM back forc. The run finishes with each step run once and no further full load. New tests:maxRetries: 0step, and a retriable one with run-ahead off);newer_parked);bwith a different step, so the old VM diverges on its events and falls back to one cold replay, without spending the divergence budget.srcin core passes locally.On Vercel, with eve on #5365 + #5467 + this, 8 follow-ups 35s apart on a 20-turn session resumed 7 parked VMs (4 of them stale) with no errors. The server's
post.start→ turn step was 402ms median, against 568ms with queued delivery and no parking. Client first output was ~760ms against ~860ms on #5365 + #5467 without this change.PR Checklist - Required to merge
pnpm changesetwas run to create a changelog for this PRpatchfor bug fixes,minorfor new features,majorfor breaking changes.pnpm changeset --emptyif you are changing documentation or workbench appsgit commit --signoffon your commits)@vercel/workflowin a comment once the PR is ready, and the above checklist is complete🤖