Spur nightly (20260622 · 38bffdb)

Automated nightly build from main at 2026-06-22 07:57 UTC.

Commit: 38bffdb

1 commit(s) since previous nightly.

Install

curl -fsSL https://raw.githubusercontent.com/ROCm/spur/main/install.sh | bash -s -- nightly

Docker

docker pull rocm/spur:nightly

Spur v0.3.0

What's Changed

• CLAUDE.md: add don'ts for writing unit tests by @shiv-tyagi in #112
• Fix test t50_106: test real RawModeGuard, not simulation (#111) by @powderluv in #114
• GPU over-scheduling fix, cancel deallocation, K8s pending failure det… by @codetap-developer in #115
• Add secretEnv field to SpurJob CRD (#117) by @powderluv in #118
• Add auto-update: startup check + CLI self-update by @powderluv in #119
• fix(spurd): deadlock in exec_in_job due to double mutex acquisition by @shiv-tyagi in #120
• fix(spurctld): propagate Raft propose errors to RPC callers by @shiv-tyagi in #121
• fix(spurctld): refactor submit_array_job to avoid deadlock by @shiv-tyagi in #122
• chore: sync Cargo.lock after spur-update crate addition by @shiv-tyagi in #123
• fix(spurctld): return error from snapshot_state instead of swallowing serde failures by @shiv-tyagi in #124
• fix(spur-cli): group sinfo rows by node state within partition by @shiv-tyagi in #127
• Wire spurctld accounting notifications to spurdbd by @shiv-tyagi in #129
• fix(spurd): drop privilege inside unshare wrapper, not before exec by @powderluv in #130
• chore: bump workspace version to 0.3.0 by @powderluv in #131

New Contributors

• @codetap-developer made their first contribution in #115

Full Changelog: v0.2.2...v0.3.0

Spur v0.2.2

Changes since v0.2.1

• #107/#108: Set supplementary groups (video, render) for non-root job processes — fixes GPU device access when running as non-root
• #109: Replace GetNode heartbeat hack with proper Heartbeat RPC — cleaner separation of read/write paths, real telemetry in heartbeats
• #110: Sync CLAUDE.md and README.md with current codebase

Spur v0.2.1

Hotfix: fix namespace wrapper mount ordering (PR #106). Removes private /tmp mount that broke job output file visibility.

Spur v0.2.0

Spur v0.2.0

Major release with 99 commits since v0.1.0. Highlights:

Raft-based HA (PR #84, #93)

• Replaced K8s Lease leader election with OpenRaft consensus
• Works on both bare-metal and Kubernetes — unified HA
• Config: controller.peers = ["node1:6817", "node2:6817"]

Topology-aware scheduling (PR #80)

• topology/tree — fat-tree fabric-aware scheduling, minimize switch hops
• topology/block — rack-level job co-location
• Config: [topology] section with switch hierarchy

Bare-metal job isolation (PRs #100, #102, #103, #104, #105)

• UID/GID enforcement via setuid (jobs run as submitting user)
• PID + mount namespace isolation (private /tmp, /dev/shm, process view)
• GPU device restriction via selective bind-mount
• seccomp-BPF syscall whitelist (~150 syscalls, blocks ptrace/mount/bpf)
• Landlock LSM filesystem access control (kernel 5.13+)
• [isolation] config section for operator control
• Fork bomb protection (pids.max) and OOM isolation (memory.oom.group)

K8s operator improvements (PRs #85, #93, #98)

• hostNetwork, privileged, hostIPC, shmSize fields now applied to pods
• Extra device plugin resources (RDMA, InfiniBand)
• Cross-namespace SpurJob support
• Node state machine with auto-recovery from heartbeat timeout

Scheduler fixes (PR #94)

• Fixed misleading PendingReason::Priority on new jobs
• Dispatch failures now requeue instead of marking Failed
• salloc timeout + pending reason display

Binaries

• spur-v0.2.0-linux-x86_64.tar.gz — all binaries (dynamically linked, glibc 2.39+)
• spurd-static — statically linked agent (musl, works on Ubuntu 22.04+)

Spur v0.1.0

What's Changed

• Users/powderluv/k8s operator and bare metal deploy by @powderluv in #1
• Multi-node Slurm feature tests + nodelist/exclude/time-limit fixes by @powderluv in #2
• Add release workflow and curl-pipe-bash installer by @powderluv in #3

New Contributors

• @powderluv made their first contribution in #1

Full Changelog: https://github.com/powderluv/spur/commits/v0.1.0