13803: Fix behavior change on Oj 3.x

[arvados.git] / services / crunch-dispatch-slurm / squeue.go

diff --git a/services/crunch-dispatch-slurm/squeue.go b/services/crunch-dispatch-slurm/squeue.go
index 365b4e85706ccd46f8a77ad30af2700cfaf1ac60..742943f197580e186e7fd1f7b8084a1357f3661d 100644 (file)
--- a/services/crunch-dispatch-slurm/squeue.go
+++ b/services/crunch-dispatch-slurm/squeue.go
@@ -91,7 +91,15 @@ func (sqc *SqueueChecker) reniceAll() {
        }
 
        sort.Slice(jobs, func(i, j int) bool {
-               return jobs[i].wantPriority > jobs[j].wantPriority
+               if jobs[i].wantPriority != jobs[j].wantPriority {
+                       return jobs[i].wantPriority > jobs[j].wantPriority
+               } else {
+                       // break ties with container uuid --
+                       // otherwise, the ordering would change from
+                       // one interval to the next, and we'd do many
+                       // pointless slurm queue rearrangements.
+                       return jobs[i].uuid > jobs[j].uuid
+               }
        })
        renice := wantNice(jobs, sqc.PrioritySpread)
        for i, job := range jobs {
@@ -149,7 +157,7 @@ func (sqc *SqueueChecker) check() {
                replacing.nice = n
                newq[uuid] = replacing
 
-               if state == "PENDING" && reason == "BadConstraints" && p == 0 && replacing.wantPriority > 0 {
+               if state == "PENDING" && ((reason == "BadConstraints" && p == 0) || reason == "launch failed requeued held") && replacing.wantPriority > 0 {
                        // When using SLURM 14.x or 15.x, our queued
                        // jobs land in this state when "scontrol
                        // reconfigure" invalidates their feature
@@ -163,7 +171,14 @@ func (sqc *SqueueChecker) check() {
                        // reappeared, so rather than second-guessing
                        // whether SLURM is ready, we just keep trying
                        // this until it works.
+                       //
+                       // "launch failed requeued held" seems to be
+                       // another manifestation of this problem,
+                       // resolved the same way.
+                       log.Printf("releasing held job %q", uuid)
                        sqc.Slurm.Release(uuid)
+               } else if p < 1<<20 && replacing.wantPriority > 0 {
+                       log.Printf("warning: job %q has low priority %d, nice %d, state %q, reason %q", uuid, p, n, state, reason)
                }
        }
        sqc.queue = newq