3 hours ago
Advice only: please do not stop, restart, redeploy, or modify any service.
A multithreaded Python application runs as namespace PID 1 and currently stores data under /tmp without a volume. An encrypted off-container backup and isolated restore have been tested. A final snapshot requires all writers to stop; some write paths do not use the application advisory lock.
The maintenance session is in the same PID namespace. cgroup.freeze is not writable and CAP_SYS_ADMIN/CAP_SYS_PTRACE are unavailable. We do not intend to bypass these restrictions.
Is there a supported platform-side pause that retains the container and ephemeral filesystem while a separate maintenance session completes the encrypted snapshot and transfer, with a confirmed paused state and reliable resume on cancellation?
If unsupported, what supported migration procedure to a persistent volume preserves every write? Please distinguish pause/resume from stop/redeploy and serverless sleep. No service action is authorized by this question.
3 Replies
3 hours ago
We can't confirm a platform-side pause that freezes the running container while keeping its process and ephemeral filesystem intact, with a confirmed paused state and resume. Serverless is not that: it stops a service when it has been inactive, and the first request afterwards wakes it with a cold boot, so it is a stop, not a freeze of the live process.
Stop and redeploy both end the current container, and anything written outside a volume (including /tmp) is discarded when a new deployment replaces it. For services with a volume attached, we only allow one deployment to be active and mounted at a time, so there is a short downtime on each redeploy.
Because of that, write quiescing has to happen inside your application before the final snapshot, and the migration path is the one you've already tested: take the final off-container backup once all writers are stopped, attach a volume, then restore that backup into the volume path and point the app at it. Since some write paths bypass your advisory lock, writes made after the snapshot and before the new deployment would not be captured, so the cutover should only proceed once the app is no longer accepting writes. We'd recommend rehearsing the full sequence in an isolated test environment before running it against this service.
Status changed to Awaiting User Response Railway • about 3 hours ago
Railway
We can't confirm a platform-side pause that freezes the running container while keeping its process and ephemeral filesystem intact, with a confirmed paused state and resume. Serverless is not that: it stops a service when it has been inactive, and the first request afterwards wakes it with a cold boot, so it is a stop, not a freeze of the live process. Stop and redeploy both end the current container, and anything written outside a volume (including /tmp) is discarded when a new deployment replaces it. For services with a volume attached, we only allow one deployment to be active and mounted at a time, so there is a short downtime on each redeploy. Because of that, write quiescing has to happen inside your application before the final snapshot, and the migration path is the one you've already tested: take the final off-container backup once all writers are stopped, attach a volume, then restore that backup into the volume path and point the app at it. Since some write paths bypass your advisory lock, writes made after the snapshot and before the new deployment would not be captured, so the cutover should only proceed once the app is no longer accepting writes. We'd recommend rehearsing the full sequence in an isolated test environment before running it against this service.
3 hours ago
Thank you. This remains unresolved: the existing running application does not yet have a complete write-quiescing mechanism, and deploying one would itself replace the container holding the ephemeral data. Could a human confirm whether any supported procedure addresses this initial migration without losing intervening writes? If none exists, please explicitly confirm that limitation. We will not treat serverless sleep, stop, or redeploy as a pause. Advice only; please do not change or stop any service.
Status changed to Awaiting Railway Response Railway • about 3 hours ago
2 hours ago
We can't point you to a platform-side pause that freezes the running container, keeps its process and ephemeral filesystem, reports a confirmed paused state and resumes it. The controls we can describe all end the current container: stopping, redeploying, attaching a volume (which takes effect through a new deployment), and serverless sleep, which stops an idle service and cold-boots it on the next request.
Because of that, we can't confirm any procedure that guarantees every write is preserved while the running application has no way to stop all of its own writers. Anything written to /tmp after your last snapshot is lost when the container is replaced. That includes redeploys you don't start, since we sometimes redeploy services to move them between hosts for maintenance.
The data in /tmp is also at risk until the app is moved onto a volume, so whatever gap you accept during a planned cutover is the same exposure the service already carries today.
Status changed to Awaiting User Response Railway • about 2 hours ago