Repository navigation
docs: add removal of a deployment's retained Docker resources - #12929
Open
bowenzhu21 wants to merge 10 commits into
Open
bowenzhu21 wants to merge 10 commits into
bowenzhu21 wants to merge 10 commits into
Conversation
After its tests pass, cargo ci live-docker fails if OpenShell left a sandbox object in one of its gateways' namespaces, or if a container, volume, or network labelled with one of its run UUIDs survives removal. It removes those objects, containers first, whether or not the tests pass, and removes the images it built last. Signed-off-by: Bowen Zhu <bowenzhu66@gmail.com>
Add the removal procedure to the destroy guide, the managed Kubernetes gateway and the Docker object names to the retention reference, and links or limits on the pages that mention retained data. Verified on Linux ARM64 with Docker 29.5.2 in a Colima VM at 9e2f241, using the SDK's managed gateway recovery test, which applies and destroys only the runtime stage, with no sandboxes or services. After destroy, the initializer, gateway volume, and bridge remained, labelled with the UID and named after its workspace. The section's commands, run as written, found no sandbox objects, removed the three objects, and left the lists empty. A rerun changed nothing, applying with the old state directory failed with the recreation guard, and a new state directory applied again. Signed-off-by: Bowen Zhu <bowenzhu66@gmail.com>
Follow the existing procedures: prose prerequisites, one lead-in per command block, deployment-prefixed variables, the documented UID-derived workspace placeholder instead of a hash command, and the model runtime guide's instruction to run on an SSH-placed host. The section shrinks from four command blocks to two. Run as written on a destroyed managed Docker gateway (Linux ARM64, Docker 29.5.2 in a Colima VM): the listing showed only the initializer, gateway volume, and bridge, named after the workspace; the sandbox commands printed nothing; removal and a rerun succeeded; and every list command then printed nothing. Signed-off-by: Bowen Zhu <bowenzhu66@gmail.com>
List and remove a deployment's objects by its UID label alone, without the workspace placeholder or the sandbox namespace commands, which the live Docker suite still checks. Name every engine a deployment can select, including a service's SSH placement.engine, and point at Docker's in-use error, which names the container, when a volume or network stays. Run as written on a destroyed managed Docker gateway (Linux ARM64, Docker 29.5.2 in a Colima VM): the listing showed only the initializer, gateway volume, and bridge, removal and a rerun succeeded, and every list command then printed nothing. Signed-off-by: Bowen Zhu <bowenzhu66@gmail.com>
Writing the fake Docker CLI and executing it from parallel tests raced their forks: on Linux, 33 of 150 runs failed with Text file busy, which the listing reported as a cleanup failure. The fake now runs through sh, and 300 runs passed. It also needs --all for stopped containers, --force for running ones, and the matching kind on rm, so dropping any of them fails a test, and the namespace test compares against the SDK's workspace instead of a literal. A run whose tests fail now prints any cleanup failure instead of discarding it, listing failures are reported once as cannot list, and Docker's removal errors are no longer hidden. Signed-off-by: Bowen Zhu <bowenzhu66@gmail.com>
Key the Docker object table by the entries plan --destroy prints under Retained resources, and keep the state directory until every entry's objects are gone, so a missed engine no longer looks finished. Say why the UID must be unique and why an old state directory cannot be reused, link the live suite without overclaiming, and drop the either/or from bundle removal. Signed-off-by: Bowen Zhu <bowenzhu66@gmail.com>
bowenzhu21
marked this pull request as ready for review
October 9, 2026 20:41
Bundle removal keeps state for anything the removal procedure does not cover, so link the issue that tracks it, as the other pages that state a limit do. Signed-off-by: Bowen Zhu <bowenzhu66@gmail.com>
Bind the sandbox and owner filters by name instead of indexing an array, build the full workspace name as the SDK does, and name the test after what it checks. Signed-off-by: Bowen Zhu <bowenzhu66@gmail.com>
A full managed deployment also retains its OpenShell workspace, which lives in the gateway storage volume, so the table maps that entry and the procedure can finish. Split the closing checks into separate steps, set the engine directly, give each limit its own row, and restore the model diagnostics sentence. Signed-off-by: Bowen Zhu <bowenzhu66@gmail.com>
…removal Signed-off-by: Bowen Zhu <bowenzhu66@gmail.com>
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Part of #12640.
Failure
Destroy keeps gateway storage, model downloads, and service credentials by design, but the guide had no way to remove them and pointed at #12640 instead.
cargo ci live-dockerdiscarded every cleanup error, so anything its tests left behind went unseen.Decision
Document removal on Docker engines. Remove Retained Resources in
usage.mdhas the reader confirm destroy finished, list the deployment's objects by its UID label on each engine, match them to the entriesplan --destroyprints underRetained resources:using a new table instate.md, and remove them, containers first. The state directory stays until every entry's objects are gone, so a missed engine does not look finished.state.mdalso gains the missing managed Kubernetes row, and pages that mention retained data link the procedure or state what stays open.cargo ci live-dockernow fails, naming the objects, when OpenShell leaves a sandbox object or when a labelled object cannot be listed or survives removal.A purge command, Podman, Kubernetes, an external gateway's workspace, and verified removal of service volumes and of objects on SSH engines stay under #12640.
Validation
OpenShell workspaceandgateway storageafter destroy, the listbrev.rsasserts for a full deployment; its agent did not become ready in this environment. On Linux ARM64 with Docker 29.5.2 in a Colima VM, the section's commands, run as written, listed exactly the initializer, volume, and bridge, removed them, and a rerun changed nothing.gateway storagealone and was removed the same way. Applying with its old state directory then failed with the recreation guard, and a fresh UID and state directory deployed again.cargo ci live-docker, OpenShell labelled the sandbox containers and volumes it created withopenshell.ai/sandbox-namespaceand their gateway's name, which is what the check selects.--alland--forceand refuses to remove a volume or network while any container remains. Each of eleven deliberate breaks, such as removing volumes before containers or dropping--all, fails at least one, and 400 repeated runs passed.v1at 29e4506 merged,cargo cipassed on Linux ARM64 (1,334 tests, 129 lifecycle tests), andcargo ci live-dockerpassed all 10 tests and left no labelled or sandbox object. A fork run of CI / Native passed on Linux ARM64 and AMD64 with both live suites, macOS, and Windows, and CI / Images passed on both architectures.Signed-off-by: Bowen Zhu bowenzhu66@gmail.com