Deploying the Upgrade

These are the activities that are performed when the time of deployment comes around. You will need access to a number of resources (Summit, TTS) at the sites so be sure that you have the credentials to do so. A number of scripts are available for easing the bare metal docker-compose deployment handling. The scripts can be retrieved from this repo: https://github.com/lsst-ts/docker-compose-admin. Use the scripts from the appropriate site directory. If changes are necessary to these scripts from work described in the previous section, use the appropriate site Jira ticket.

Note

The deployment is only concerned with CSCs and systems (Summit, TTS) in the production domain (domainId = 0). All other domains are left alone.

Attention

As the person running the deployment, you have absolute control over the system to complete this process. No one should do anything without your express consent.

  1. Log into the machines for bare metal deployments and Kubernetes access (Summit, TTS)

  2. Use LOVE or Nublado to send all CSC systems to OFFLINE state.

    • WARNING: Not all CSCs report OFFLINE; these will instead report STANDBY as the last state seen. Check for heartbeats to be sure. (Summit)
    • Preference is to use LOVE, but Nublado is a good fall back in case LOVE isn’t working.
    • An overall status view is available from LOVE (Summit, TTS).
    • You can also consult these dashboards on Chronograf. The names are the same across sites.
      • AT Summary State Monitor
      • MT Summary State Monitor
      • EAS Summary State Monitor
      • Observatory Systems Summary State Monitor
    • The Watcher and the ScriptQueues should come down last if running from LOVE.

  3. With all the systems OFFLINE, you can log out of your Nubaldo instance as we will clean them up soon.

  4. Once all systems are in OFFLINE, still running CSCs/systems and OSPL daemons need to be cleaned up.
    1. Get number of currently running daemons/CSCs still in OFFLINE from main OSPL daemon: (TTS)
      • docker exec ospl-daemon grep “federations” durability.log
      • You must give the daemon some time (30 seconds to 2 minutes) before getting worried that the number isn’t going down once you start shutting down daemons.
      • You can check this after every shutdown or just periodically.
    2. Cleanup CSCs and Daemons on DM and Camera machines (Summit, TTS).
      • One can work with the system principles to shutdown the services.
    3. Shutdown LOVE and associated daemon (Summit, TTS).
      • This step will become unnecessary when it moves to Kubernetes.
    4. Shutdown and Cleanup Bare Metal Deployments (Summit, TTS).
    5. Cleanup Kubernetes Deployment.
      • Below uses scripts in this repo: https://github.com/lsst-ts/k8s-admin.
      • Execute the following (Summit, TTS): ./cleanup_all -d
        • The -d is important as that cleans up the OSPL daemons too.
      • If the script fails, you can use Argo CD to delete the job/deployment/daemonset associated with each application you wish to stop. Be sure to delete the job/deployment/daemonset box not the application itself. Note that auxtel and the other “app of apps” meta applications have no jobs; you have to deal with each application individually.
      • Execute the following (TTS): ./cleanup_love
    6. Cleanup Nublado namespaces (Summit, TTS).
    7. Check to ensure all daemons have disconnected.
      • If the reported number is not 0, you will need to investigate further to find the source of the rogue process.
    8. Shutdown and Cleanup Main Daemon (Summit, TTS).
  5. With everything shutdown, the configurations need to be updated before deployment starts.
    • Ensure SQuaRE has approved the cachemachine PR and then merge the PR.
    • Rebase Argo CD branch to single commit, create a PR and merge PR.
    • All other configuration repositories should have the necessary commits already on branches and pushed to GitHub.
    • Update configuration repositories on bare metal machine deployments (Summit, TTS).
      • Unlike shutdown, only the T&S systems are handled here. DM and Camera are handled by the system principles.
      • Also, only certain T&S systems are handled here, the rest need to be coordinated with system principles.
  6. Once all configuration in place, deployment of the new system can now begin.
    • Be patient with container pulling (goes for everything containerized here).
    1. Update Nublado
      • From the site specific Argo CD UI, find the cachemachine app.
      • It should indicate OutOfSync (yellow) status, so click the Sync button to begin the process.
      • Once it syncs, a new pod will start from the cachemachine deployment.
      • Delete that running pod so a new one starts up.
      • The Nublado pull will be completed when the child processes from the new pod all complete and no downstream APIs are shown in the UI.
    2. Startup Main OSPL daemon (Summit, TTS) and verify that it has started.
      • Verify that each daemon has actually started by running: docker logs ospl-daemon and checking for a line that says “daemon ready”.
      • To monitor the number of daemons (“federations”) as you bring up daemons and single-process CSCs, run the following: docker exec ospl-daemon grep “federations” durability.log (TTS)
    3. Startup Minimal Kubernetes Services (TTS)
      • This uses the sync_apps.py script found in https://github.com/lsst-ts/argocd-csc/bin.
      • The script is run in the same place that Kubernetes (kubectl) interactions are run.
      • Log into the argocd pod by doing the following:
        • python argocd_login.py /path/to/auth_file
        • The auth_file must contain the appropriate site username/password on separate lines.
      • python sync_apps.py -p –no-sync=love
      • csc-cluster-config, ospl-config and ospl-daemon apps will be synced automatically.
      • Once the ospl-daemon app is synced, the script will pause.
      • Check the logs on Argo CD UI to see if daemons are ready.
      • Type go and enter to move onto syncing the kafka-producers app.
      • Script will again pause once the kafka-producers are synced.
      • Check that all the logs say “Running” at the bottom. M1M3 now has an indicator saying: “Partial producers are all running”.
      • Once all the kafka-producers are verified to be running, stop here but leave the script alive.
    4. Startup Bare Metal LOVE (Summit, TTS).
    5. Continue Kubernetes Deployment
      • Go back to where your running sync_apps.py script is and type go and enter to move onto syncing the obssys apps.
      • Use the Argo CD UI to verify that the containers are pulling and running.
      • The script will pause again here. Do not proceed further with the script, but leave it alive.
    6. The above now represents a minimal system that other system principles can be allowed to start their daemons/CSCs.
      • Use the site specific Slack channel (Summit, TTS) to inform the system principles.
    7. Startup Rest of Kubernetes Services.
      • Go back to where your running sync_apps.py script is and type go and enter to proceed with syncing the rest of the apps.
      • The rest of the apps will be synced automatically so no further intervention is necessary.
    8. Startup Services on Bare Metal Deployments (Summit, TTS).
  7. Once the deployment steps have been executed, the system should be monitored to see if all CSCs come up into STANDBY/OFFLINE. Daemons can also be monitored for connection using the methods listed above.
    • Use the site specific resources to help ascertain this condition.
    • Report any issues directly to the system principles (DMs are OK) to get issue corrected.
    • This step is completed when either all CSCs are in STANDBY/OFFLINE or CSCs with issues cannot be fixed in a reasonable (~30 minutes) amount of time.
    • If leaving this step with CSCs in non-working order, make sure to report that on the site specific Slack channel.

  8. Certain sites require that some CSCs be put into ENABLED state (Summit, TTS).

  9. If not carrying on with integration testing, folks can be told they can use Nublado again via the site specific Slack channel.

Site Specific Variations