Repository navigation
Conversation
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
shiju-nv
force-pushed
the
fix/startup-conflict-recovery
branch
from
October 4, 2026 15:29
b4a4566 to
b2675ba
Compare
YasirBinMudassir
approved these changes
Oct 5, 2026
YasirBinMudassir
approved these changes
Oct 5, 2026
shiju-nv
force-pushed
the
fix/startup-conflict-recovery
branch
from
October 13, 2026 14:06
b7664dc to
20879b9
Compare
shiju-nv
marked this pull request as ready for review
October 13, 2026 14:08
|
This pull request has had no activity for 14 days and is now marked stale. |
|
Label |
shiju-nv
force-pushed
the
fix/startup-conflict-recovery
branch
from
November 2, 2026 14:46
3664c2e to
7c820f3
Compare
A startup image-policy upload can read an empty sandbox baseline, then wait behind an operator's repair. When that upload commits, it can replace the repair. This preserves the management baseline and makes startup read fresh configuration after a write conflict. This covers a backfill selected before the competing repair. General fencing of delayed startup snapshots remains separate work. Signed-off-by: Shiju <shiju@nvidia.com>
shiju-nv
force-pushed
the
fix/startup-conflict-recovery
branch
from
November 3, 2026 11:46
7c820f3 to
e723e90
Compare
github-actions
Bot
requested review from
FrostGod,
TaylorMutch,
alangou,
drew,
johntmyers,
krishicks,
pimlock and
purp
November 15, 2026 02:39
Keep the latest sandbox timeout guidance when resolving the docs conflict. Explain the queued image upload race, name the saved policy directly, and describe the remaining delayed-upload limitation. Signed-off-by: Shiju <shiju@nvidia.com>
Name replacement permission directly and share its authorization check. Use the existing error-chain iterator for startup conflict detection. Fold transport retry coverage into the repair regression and remove duplicate projector coverage. Signed-off-by: Shiju <shiju@nvidia.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
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.
Summary
When a sandbox starts without a policy saved in the gateway, its supervisor uploads the policy found in the image. That upload can wait for the write lock while an operator saves a repair in the gateway's database. Without this fix, the waiting upload can overwrite the repair. The gateway now rejects the upload with
ABORTEDif it would replace the operator's repaired policy. The supervisor then fetches the latest configuration, including the repaired policy, from the gateway before trying again.Related Issue
Related to #3784 and #3785, which fixed startup after refused policy writes. This PR fixes a different case: the gateway accepts a queued upload and overwrites a repair. A separate issue covering this race still needs to be filed and linked before landing.
Changes
ABORTEDfrom either startup policy write. Keep transport retries and the existing retry limit.Startup uploads still don't send an expected configuration version. If an old upload reaches the handler after a repair is already saved, it can still overwrite newer network rules when its filesystem, process, and Landlock settings pass validation. Closing that race requires checking the version startup originally read; this PR only protects uploads already queued to fill a missing policy.
Release Notes
A policy repair saved in the gateway during first startup now survives an older image policy upload that was already queued to fill a missing policy.
Testing
All seven focused tests pass: the three new regressions and four existing storage and activation tests. The startup repair test covers a transport retry followed by a conflict, so it checks both retrying the upload and fetching current configuration. Formatting, Markdown, license checks, scoped Clippy, and independent source review pass. The gateway race regression previously failed on main at
c92c2c798, which accepted the stale upload as revision 2. Final-head CI and E2E results are pending.Checklist