portfolio/platform-projects/hpc-slurm-ai-cluster/playbooks/upgrade/rolling-upgrade-slurm-workers.yml

---
- name: Rolling upgrade Slurm worker nodes
  hosts: slurm_compute:slurm_gpu
  become: true
  gather_facts: true
  serial: 1

  vars:
    skip_canary_node: "{{ canary_node | default('slurm-c02') }}"
    do_skip_canary: "{{ skip_canary | default(true) | bool }}"

  pre_tasks:
    - name: Skip canary node if requested
      ansible.builtin.meta: end_host
      when:
        - do_skip_canary
        - inventory_hostname == skip_canary_node

    - name: Drain node before OS upgrade
      ansible.builtin.command:
        cmd: scontrol update NodeName={{ inventory_hostname }} State=DRAIN Reason="rolling OS upgrade"
      delegate_to: "{{ groups['slurm_controller'][0] }}"
      changed_when: true

    - name: Wait until no jobs are running on this node
      ansible.builtin.shell: |
        set -euo pipefail
        squeue -h -w {{ inventory_hostname }} || true
      args:
        executable: /bin/bash
      delegate_to: "{{ groups['slurm_controller'][0] }}"
      register: jobs_on_node
      retries: 120
      delay: 10
      until: jobs_on_node.stdout | trim == ""
      changed_when: false

  tasks:
    - name: Update apt cache
      ansible.builtin.apt:
        update_cache: true
        cache_valid_time: 1800

    - name: Full upgrade packages
      ansible.builtin.apt:
        upgrade: full
        autoremove: true
        autoclean: true
      register: apt_upgrade_result

    - name: Check if reboot is required
      ansible.builtin.stat:
        path: /var/run/reboot-required
      register: reboot_required

    - name: Show upgrade status
      ansible.builtin.debug:
        msg:
          - "Node: {{ inventory_hostname }}"
          - "Apt changed: {{ apt_upgrade_result.changed }}"
          - "Reboot required: {{ reboot_required.stat.exists }}"

    - name: Reboot node if required
      ansible.builtin.reboot:
        msg: "Reboot after rolling OS upgrade"
        reboot_timeout: 900
        connect_timeout: 20
        pre_reboot_delay: 5
        post_reboot_delay: 20
      when: reboot_required.stat.exists

    - name: Restart munge
      ansible.builtin.systemd:
        name: munge
        state: restarted
        enabled: true

    - name: Restart slurmd
      ansible.builtin.systemd:
        name: slurmd
        state: restarted
        enabled: true

    - name: Validate local slurm services
      ansible.builtin.shell: |
        set -euo pipefail
        systemctl is-active munge
        systemctl is-active slurmd
        munge -n | unmunge >/dev/null
        scontrol ping
      args:
        executable: /bin/bash
      changed_when: false

  post_tasks:
    - name: Restart controller to refresh state after node upgrade
      ansible.builtin.systemd:
        name: slurmctld
        state: restarted
      delegate_to: "{{ groups['slurm_controller'][0] }}"
      run_once: false

    - name: Wait for controller after restart
      ansible.builtin.command:
        cmd: scontrol ping
      delegate_to: "{{ groups['slurm_controller'][0] }}"
      register: slurmctld_ping
      retries: 15
      delay: 2
      until: slurmctld_ping.rc == 0
      changed_when: false

    - name: Clear upgraded node maintenance state
      ansible.builtin.shell: |
        set -euo pipefail

        scontrol update NodeName={{ inventory_hostname }} State=RESUME 2>/dev/null || true
        scontrol update NodeName={{ inventory_hostname }} State=UNDRAIN 2>/dev/null || true
        scontrol update NodeName={{ inventory_hostname }} State=IDLE 2>/dev/null || true

        sleep 3
        sinfo -N -n {{ inventory_hostname }}
        scontrol show node {{ inventory_hostname }}
      args:
        executable: /bin/bash
      delegate_to: "{{ groups['slurm_controller'][0] }}"
      register: resume_node
      changed_when: true

    - name: Wait until node is healthy
      ansible.builtin.shell: |
        set -euo pipefail
        sinfo -N -n {{ inventory_hostname }}
        scontrol show node {{ inventory_hostname }}
      args:
        executable: /bin/bash
      delegate_to: "{{ groups['slurm_controller'][0] }}"
      register: upgraded_node_state
      retries: 30
      delay: 5
      until:
        - upgraded_node_state.rc == 0
        - "'not_responding' not in upgraded_node_state.stdout.lower()"
        - "'down' not in upgraded_node_state.stdout.lower()"
        - "'drain' not in upgraded_node_state.stdout.lower()"
        - "'idle*' not in upgraded_node_state.stdout.lower()"
      changed_when: false

    - name: Submit node-local post-upgrade test job
      ansible.builtin.shell: |
        set -euo pipefail

        job_id="$(
          sudo -iu slurmuser sbatch --parsable <<SBATCH
        #!/bin/bash
        #SBATCH --job-name=rolling-upgrade-test
        #SBATCH --partition=all
        #SBATCH --nodelist={{ inventory_hostname }}
        #SBATCH --cpus-per-task=1
        #SBATCH --mem=256M
        #SBATCH --time=00:02:00
        #SBATCH --output=/shared/rolling-upgrade-test-%j.out

        echo "HOST=\$(hostname)"
        echo "SLURM_JOB_ID=\$SLURM_JOB_ID"
        echo "SLURM_JOB_NODELIST=\$SLURM_JOB_NODELIST"
        echo "CPUS_ALLOWED=\$(grep Cpus_allowed_list /proc/self/status)"
        echo "KERNEL=\$(uname -r)"
        date
        SBATCH
        )"

        echo "JOB_ID=$job_id"

        for i in $(seq 1 90); do
          if squeue -h -j "$job_id" | grep -q .; then
            squeue -j "$job_id"
            sleep 1
          else
            break
          fi
        done

        echo "### sacct"
        sacct -j "$job_id" --format=JobID,JobName,User,Partition,State,ExitCode,Elapsed,AllocCPUS,ReqMem,NodeList

        echo "### output"
        cat "/shared/rolling-upgrade-test-${job_id}.out"
      args:
        executable: /bin/bash
      delegate_to: "{{ groups['slurm_controller'][0] }}"
      register: node_test_job
      changed_when: true

    - name: Show node post-upgrade test result
      ansible.builtin.debug:
        var: node_test_job.stdout_lines