Issues

• Dispatch failure reason drops the node names, so clients cannot tell which node refused a job

  ROCm/spur#1118 · Filter by author zoroyihan7 opened on Oct 23, 2026

• A prolog failure drains the node and holds the job at priority 0 with no automatic release

  ROCm/spur#1117 · Filter by author zoroyihan7 opened on Oct 23, 2026

• Restoring state on controller restart re-stamps every job's submit_time and start_time to the restart time (and accounting keeps the re-stamped values)

  ROCm/spur#1116 · Filter by author zoroyihan7 opened on Oct 23, 2026

• 0.13.0: a QOS with GrpTRES node=3 is running six 1-node jobs on six nodes

  ROCm/spur#1096 · Filter by author zoroyihan7 opened on Oct 20, 2026
  2

  

• spurd refuses every dispatch while the controller reports the node idle: a job held with JobHoldMaxRequeue is still running on the node

  ROCm/spur#1088 · Filter by author zoroyihan7 opened on Oct 18, 2026

• spurctld: a write to a leader that lost its quorum blocks with no timeout

  area: controllerarea: fault-tolerancetype: bug
  ROCm/spur#1086 · Filter by author pre opened on Oct 17, 2026

• spurd: container rootfs overlay mount always fails, so each container job extracts the full image

  area: job-lifecycletype: bug
  ROCm/spur#1085 · Filter by author pre opened on Oct 17, 2026

• A job's second run keeps its node's GPUs for a day after the controller records the job ended; every dispatch there is refused

  ROCm/spur#1083 · Filter by author zoroyihan7 opened on Oct 16, 2026

• private_data has no category for users, accounts or reservations, and "jobs" does not cover sacct`

  ROCm/spur#1074 · Filter by author amd-vpenumal opened on Oct 14, 2026

• v0.14.0: QOS cannot be set per account\

  ROCm/spur#1072 · Filter by author amd-vpenumal opened on Oct 14, 2026

• v0.14.0: No partition scoping for users or accounts — associations.partition_name cannot be set

  ROCm/spur#1071 · Filter by author amd-vpenumal opened on Oct 14, 2026

• v0.14.0: sacctmgr show qos omits preempt, preempt_exempt_time and four other stored fields

  ROCm/spur#1070 · Filter by author amd-vpenumal opened on Oct 14, 2026

• v0.14.0: qos.preemptmode accepts any value — no validation

  ROCm/spur#1068 · Filter by author amd-vpenumal opened on Oct 14, 2026

• 0.14.0: Preemption does not occur with preempt_type = "qos_priority"

  ROCm/spur#1067 · Filter by author amd-vpenumal opened on Oct 14, 2026

• 0.13.0: SubmitJob intermittently fails to decode JobSpec.script as non-UTF-8 although the client sent an ASCII script; one accepted job ran a changed script

  ROCm/spur#1066 · Filter by author zoroyihan7 opened on Oct 14, 2026

• 0.13.0: with a follower down, the leader's RaftCore stops on "Read Snapshot(None): snapshot not found"; the controller keeps claiming leadership and every write to it fails

  ROCm/spur#1059 · Filter by author zoroyihan7 opened on Oct 14, 2026
  1

• Job updates (TimeLimit, Partition, Account, QOS, Comment) are applied only in the leader's memory, so a leader change reverts them

  ROCm/spur#1056 · Filter by author zoroyihan7 opened on Oct 14, 2026
  1

• srun exits 0 when tasks fail on nodes running more than one task

  ROCm/spur#1054 · Filter by author nikhilsk opened on Oct 13, 2026

• 0.13.0: with one controller down, the leader ends jobs on healthy nodes in per-minute batches (-1:0 / NODE_FAIL)

  ROCm/spur#1040 · Filter by author zoroyihan7 opened on Oct 13, 2026
  7

• A job held after a prolog failure no longer says it's held

  ROCm/spur#1038 · Filter by author ammarhd opened on Oct 12, 2026

• Host-dockerd containers are not tied to the job lifecycle; one non-compliant job can keep using GPUs on nodes later allocated to others

  ROCm/spur#1030 · Filter by author zengleixin-amd opened on Oct 9, 2026
  1

• Docs say Spur-managed k0s GPU nodes advertise amd.com/gpu, but Spur installs no device plugin

  managed-k8stype: docs
  ROCm/spur#1028 · Filter by author nikhilsk opened on Oct 8, 2026
  1

• spur k8s status reports active for a worker that never joined the Kubernetes cluster

  area: controllermanaged-k8stype: bug
  ROCm/spur#1027 · Filter by author nikhilsk opened on Oct 8, 2026

• spur k8s up is accepted and the cluster stays in provisioning forever when the controllers' spur.conf has no [cluster] section

  area: controllermanaged-k8stype: bug
  ROCm/spur#1026 · Filter by author nikhilsk opened on Oct 8, 2026
  

• spurctld: k0s reconcile loop can act on a half-replayed Raft log after a controller restart

  ROCm/spur#1024 · Filter by author nikhilsk opened on Oct 7, 2026