Increase Default Value of TerminationDelay#199
Merged
unmarshall merged 2 commits intoSep 23, 2025
Merged
Conversation
…in for most cases Signed-off-by: Rohan Varma <rohanv@nvidia.com>
Signed-off-by: Rohan Varma <rohanv@nvidia.com>
531199e to
8c5a51a
Compare
unmarshall
approved these changes
Sep 23, 2025
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
PR to increase default value of terminationDelay from 30 seconds to 4 hours.
Rationale:
terminationDelay is tied to gang termination semantics, which are meant to prevent large amounts of resources from staying reserved when a system is non-functional (e.g., when minAvailable is breached). The delay provides a grace period for the system to heal before triggering termination.
The current default of 30 seconds is too short and often disruptive. Kubernetes users are accustomed to workloads continually trying to self-heal, so shutting things down after 30 seconds is jarring. In practice, we’ve also seen inference workloads terminate before they had time to fully initialize—especially when models are large. This usually stems from user or application error: worker pods mark themselves as Ready before the model has actually finished loading, either because startup probes are not configured or because the inference framework itself signals readiness too early, before completing all initialization.
Nevertheless, we want terminationDelay and gang termination semantics to stay out of the user’s way by default. The goal is for users to be deliberate about opting in and explicitly deciding how long a workload should be given to self-heal before being considered irreparably broken. By raising the default to 4 hours, we provide enough time for workloads with heavy initialization—such as loading very large models like DeepSeek-R1—to complete without being prematurely torn down. At the same time, the system will still eventually terminate if it remains in a broken state for an extended period, preserving the original intent of gang termination.