Skip to content
Runner-Only Image Binding Without Deploying Pending API Changes
joseph0426
OP

10 days ago

Hello Railway Team,

We are using a Limited Trial workspace and need guidance on safely configuring and deploying an existing one-shot Docker-image service.

Our staging environment contains 22 pending changes for a separate API service. We need to preserve these changes without committing or deploying them.

Could you please confirm:

  1. Can an explicit environmentPatchCommit with skipDeploys: true configure only an existing service's digest-pinned public GHCR image, service-local variables, and sealed policy without consuming unrelated staged changes or triggering deployment?

  2. What exact supported mutation and arguments should be used for this operation?

  3. How can we independently verify that the image digest, variable configuration, and disabled image auto-updates are committed while the unrelated staged changes remain intact?

  4. Which supported operation will launch only the selected service once, using the committed image digest, without applying the environment's pending changes?

  5. If binding partially succeeds or the launch request times out before returning a deployment ID, what read-only query can establish the actual state? How can we restore only that service's prior configuration without deploying anything?

  6. Is this procedure available on Limited Trial without upgrading to Pro or adding a payment method?

  7. How can we verify or exclude automatic deployment triggers, retries, schedules, image auto-updates, and other sources of additional execution?

  8. Does Railway provide a per-invocation cost ceiling? Please distinguish it from a workspace-wide usage limit, and clarify whether image pulling, initialization, startup, execution, and shutdown are included. We want to use trial credit only, with no paid charges.

We are requesting procedural confirmation only. Please do not modify any project, service, deployment, or billing configuration.

Thank you.

$20 Bounty

1 Reply

Status changed to Awaiting Railway Response Railway • 10 days ago


Railway
BOT

10 days ago

This thread has been opened as a bounty so the community can help solve it.

Status changed to Open Railway • 10 days ago


skyelabz210

10 days ago

I'd separate this into a runner-only bind, verification, and one selected-service launch. The attached Railway Image-Runner Toolkit has adaptable GraphQL templates, recovery checks, and sources. Here's the workflow for keeping your API's 22-change patch pending

Live introspection on 10 November confirmed environmentPatchCommit accepts skipDeploys; all 12 schema checks passed. That establishes the argument contract. EnvironmentConfig and deployment meta are opaque scalars, so nested patch fields need implementation/read-back checks. Isolation of this complete image sequence under concurrent edits or partial failure remains a behavior to validate

1- Capture what must stay unchanged.

Resolve exact IDs and arrange a quiet writer window, with no runner-local staged changes or active/queued runner deployment. Save committed configuration and the staged patch's ID, status, updatedAt, and full content. Baseline both services' deployment histories across all pages/statuses, including deleted records. Treat reads as separate observations and repeat until consistent. A count of 22 or unchanged configEtag alone isn't enough.

Check the image source, effective entrypoint/start command, config-file overrides, pre-deploy hooks, healthcheck, and serverless behavior. Require one replica total, restart NEVER, disabled image auto-updates, and inactive cron/repository/CI triggers. Confirm the GHCR package is anonymously pullable.

2- Bind only intended runner fields.

Use environmentPatchCommit(environmentId, patch: RUNNER_ONLY_PATCH, skipDeploys: true). Pin the image by digest. Use an existing ordinary service-local RUN_ID key with its prior value retained; leave sealed/shared/reference entries and sealing policy untouched. Redacted snapshots can mask ordinary values too, so compare approved ordinary keys privately. Sealed bytes cannot be read back, and an unknown original cannot be restored from a mask.

Keep Deploy, Alt+Deploy, and environmentPatchCommitStaged reserved for the API release: they apply the pending batch. Retain the bind's workflow ID, poll to Complete, then require read-back showing intended runner changes, unchanged API configuration and staged patch, and no unexpected deployments. If the runner is already correctly configured, skip the bind.

3- Launch separately and verify the actual run.

Use serviceInstanceDeployV2(serviceId, environmentId), omit commitSha for an image service, and save the deployment ID. Verify meta.imageDigest, allowing for a multi-platform index's selected child manifest, then recheck configuration and staging. Require a durable RUN_ID result and instance-stop evidence, not just deployment status.

Railway's image-specific employee explanation is here:

https://station.railway.com/questions/reusing-a-private-image-and-verifying-it-710bfc12

4- Reconcile uncertainty before another write.

Transport/GraphQL errors, partial/null data, or missing expected IDs require reconciliation. Retain available IDs. For a lost launch response, scan every page from after: null, including all statuses/deleted records, and compare new IDs, times, digest evidence, and durable receipts with the baseline. Missing/repeated cursors or empty results leave the attempt unresolved. Use bounded backoff; don't blindly retry or guess which deployment is yours.

For partial binding, first establish a settled attempt and quiet writers. Restore only changed fields from retained values through the explicit skipDeploys patch, and only while current fields match recorded post-attempt values. Drift means reconcile. This isn't server-side CAS; new-key deletion and sealed rollback need separately verified recovery procedures.

5- Protect effects and budget before launch.

Singleton deployments can overlap, and platform-initiated redeployments can recur. Use atomic duplicate protection before effects, including hooks: commit database changes and the unique run/step receipt together. For external effects, use a stable RUN_ID:STEP_ID key, retain provider references, and respect concurrency, retention, and same-payload rules. A local file or a claim written before an external effect is insufficient.

Limited Trial permits deployment with resource/network restrictions. Verify actual plan, credit, and reachable dependencies. Use modest resources, bounded hook/runtime deadlines, and a credit reserve; account for external charges separately. Hard spending controls cover the workspace billing cycle, including the API. A strict all-phases per-run dollar ceiling remains an unmet acceptance criterion until a supported mechanism is established.

Caution: A timed-out launch may appear later. Check all deployment history before retrying; an empty recent list isn’t proof it failed

https://docs.railway.com/pricing/cost-control

Human-led, AI-augmented research drawing on GraphQL, distributed systems, and deployment-safety practices, informed by Railway's API, implementation, and community resources.


Welcome!

Sign in to your Railway account to join the conversation.

Loading...