Repository navigation
fleet coord: fall back to on-demand on a failed spot launch; watch the host groups until they are up - #1263
Merged
Conversation
…e host groups until they are up The on-demand fallback (#1215) waited --spot-wait (120 s on staging) and was decided only while the fleet was waking. The EC2 half makes the fleet awake about 50 s into a wake (53 s on 2026-10-25), so with no spot capacity the spot group stayed short and the on-demand group was never raised: on 2026-10-24 21:44 the spot group failed 12 launches over 13 minutes (UnfulfillableCapacity) and the on-demand group shows no activity. - The groups are watched (polled, instances reported, the fallback decided) from the wake, or the coordinator's start, until they hold --awake-capacity instances in service, also after the fleet is awake. - A short spot group falls back at once when Auto Scaling has recorded a failed launch of it since the wake raised it (DescribeScalingActivities, a Failed activity; Auto Scaling records one about 10 s after the raise and keeps retrying spot on its own). The spot wait stays as the backstop, and is the only rule when the activities cannot be read (logged once). - Which rule fell back is logged ("spot launch failed; falling back to on-demand now" with the activity's message, or "spot group still short after the spot wait"), and when the groups are up, "wake: host capacity in service" with path=spot, on-demand, or spot and on-demand. - The poll while the groups come up is 2 s (was 5 s). Tests: - a_failed_spot_launch_falls_back_to_on_demand_without_waiting: with the failed-launch rule removed it fails ("desired {fleet-hosts: 1} after 5s"); a failure from before the wake does not count. - a_spot_group_still_short_once_the_fleet_is_awake_falls_back: with the groups polled only while waking it fails the same way. - failed_spot_launches_parse_from_the_scaling_activities (an answer shaped like staging's), a_failed_spot_launch_falls_back_without_the_wait, the_path_names_the_groups_the_hosts_came_from. - a_spot_group_short_after_the_wait_falls_back_to_on_demand_and_sleeps_to_zero gives its "awake at start" phase a spot host in service: an awake fleet whose spot group has nothing now falls back, which that phase did not expect. The coordinator's role needs autoscaling:DescribeScalingActivities for the fast rule (ejc3/aws); without it the coordinator logs the refusal once and keeps the spot wait. make test-unit (test_fleet_coord, fcvm fleet::coord lib tests): 68 passed. make lint: ok. Refs #1257
Owner
Author
|
@codex review |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
Warning Review limit reachedYou've used all free OSS reviews for now. Wait for the free limit to reset to keep reviewing this public repository. Next included review available in 44 minutes. View limit details
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
|
Codex Review: Didn't find any major issues. Bravo. Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
This was referenced Oct 25, 2026
ejc3
added a commit
to ejc3/aws
that referenced
this pull request
Oct 25, 2026
…ies (#670) ejc3/fcvm#1263 moves a wake to the on-demand host group as soon as Auto Scaling records a failed spot launch (UnfulfillableCapacity or MaxSpotInstanceCountExceeded, about 10 s after the raise), instead of after --spot-wait (120 s). It reads that from DescribeScalingActivities, which the coordinator's task role lacks; without it the coordinator logs the refusal once and keeps the spot wait. DescribeScalingActivities takes no resource-level permission, so it joins DescribeAutoScalingGroups on Resource "*". scripts/test-fcvm-fleet.py pins the action list. Tested: python3 scripts/test-fcvm-fleet.py (33 tests OK), terraform fmt -check. Co-authored-by: ejc3 <ejc3@users.noreply.github.com>
Owner
Author
|
NOT-A-DEFECT: the CodeRabbit comment is a review-limit notice, not a finding. |
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.
Part of #1257 (GPU wall wake latency): a spot shortage reaches the on-demand group in seconds, and it reaches it at all once the fleet is awake.
The defect
The fallback (#1215) decided only while the power state was
waking, after--spot-wait(120 s on staging). The EC2 half makes the fleetawakeabout 50 s into a wake, before the spot wait ends, so a spot group with no capacity was never moved to on demand. On staging, 2026-10-24 21:44 UTC, with a coordinator task definition that passed--on-demand-group(registered 14:07 that day), the spot group failed 12 launches over 13 minutes (UnfulfillableCapacity) and the on-demand group has no scaling activity at all.What changed (
src/fleet/coord/scale.rs)--awake-capacityinstances in service, also after the fleet is awake.DescribeScalingActivities, aFailedactivity). Auto Scaling records one about 10 s after the raise (UnfulfillableCapacityat +7 s andMaxSpotInstanceCountExceededon staging) and then keeps retrying spot without telling anyone.--spot-waitstays as the backstop and is the only rule when the activities cannot be read (logged once).spot launch failed; falling back to on-demand now(with the activity's message) orspot group still short after the spot wait, thenwake: host capacity in servicewithpath=spot,on-demandorspot and on-demand, andsince_wake_ms.Expected effect with no spot capacity: on-demand raised about 12 s into the wake instead of never (or 120 s if the fleet were still waking). The on-demand group then takes Auto Scaling's own reaction (11 s on 2026-10-25) and a cold host boot.
Evidence
a_failed_spot_launch_falls_back_to_on_demand_without_waiting(spot wait 60 s)desired {"fleet-hosts": 1} after 5s, calls [("fleet-hosts", 1)]a_spot_group_still_short_once_the_fleet_is_awake_falls_backfailed_spot_launches_parse_from_the_scaling_activitiesa_failed_spot_launch_falls_back_without_the_wait,the_path_names_the_groups_the_hosts_came_froma_spot_group_short_after_the_wait_falls_back_to_on_demand_and_sleeps_to_zeronow gives its "awake at start" phase a spot host in service: an awake fleet whose spot group has nothing now falls back, which that phase did not expect.Deploy
The coordinator's task role needs
autoscaling:DescribeScalingActivities(ejc3/aws#670). Without it the coordinator logs the refusal once and falls back after the spot wait, which this PR still applies after the fleet is awake.