Updating Sourcegraph with Kubernetes

A new version of Sourcegraph is released every month (with patch releases in between, released as needed). Check the Sourcegraph blog for release announcements.

Steps

These steps assume that you have created a release branch following the instructions in the configuration guide.

Merge the new version of Sourcegraph into your release branch.

cd $DEPLOY_SOURCEGRAPH_FORK
# get updates
git fetch upstream
# to merge the upstream release tag into your release branch.
git checkout release
# Choose which version you want to deploy from https://github.com/sourcegraph/deploy-sourcegraph/releases
git merge $NEW_VERSION

Deploy the updated version of Sourcegraph to your Kubernetes cluster:
```
./kubectl-apply-all.sh
```
Monitor the status of the deployment.
```
kubectl get pods -o wide --watch
```

Rollback

You can rollback by resetting your release branch to the old state and proceeding with step 2 above.

If an update includes a database migration, rollback will require some manual DB modifications. We plan to eliminate these in the near future, but for now, email mailto:[email protected] if you have concerns before updating to a new release.

Improving update reliability and latency with node selectors

Some of the services that comprise Sourcegraph require more resources than others, especially if the default CPU or memory allocations have been overridden. During an update when many services restart, you may observe that the more resource-hungry pods (e.g., gitserver, indexed-search) fail to restart, because no single node has enough available CPU or memory to accommodate them. This may be especially true if the cluster is heterogeneous (i.e., not all nodes have the same amount of CPU/memory).

If this happens, do the following:

Use kubectl drain $NODE to drain a node of existing pods, so it has enough allocation for the larger service.
Run watch kubectl get pods -o wide and wait until the node has been drained. Run kubectl get pods to check that all pods except for the resource-hungry one(s) have been assigned to a node.
Run kubectl uncordon $NODE to enable the larger pod(s) to be scheduled on the drained node.

Note that the need to run the above steps can be prevented altogether with node selectors, which tell Kubernetes to assign certain pods to specific nodes. See the docs on enabling node selectors for Sourcegraph on Kubernetes.

High-availability updates

Sourcegraph is designed to be a high-availability (HA) service, but upgrades by default require a 10m downtime window. If you need zero-downtime upgrades, please contact us. Services employ health checks to test the health of newly updated components before switching live traffic over to them by default. HA-enabling features include the following:

Replication: nearly all of the critical services within Sourcegraph are replicated. If a single instance of a service fails, that instance is restarted and removed from operation until it comes online again.
Updates are applied in a rolling fashion to each service such that a subset of instances are updated first while traffic continues to flow to the old instances. Once the health check determines the set of new instances is healthy, traffic is directed to the new set and the old set is terminated. By default, some database operations may fail during this time as migrations occur so a scheduled 10m downtime window is required.
Each service includes a health check that detects whether the service is in a healthy state. This check is specific to the service. These are used to check the health of new instances after an update and during regular operation to determine if an instance goes down.
Database migrations are handled automatically on update when they are necessary.

Database Migrations

By default, database migrations will be performed during application startup by a migrator init container running prior to the frontend deployment. These migrations must succeed before Sourcegraph will become available. If the databases are large, these migrations may take a long time.

In some situations, administrators may wish to migrate their databases before upgrading the rest of the system to reduce downtime. Sourcegraph guarantees database backward compatibility to the most recent minor point release so the database can safely be upgraded before the application code.

To execute the database migrations independently, follow the Kubernetes instructions on how to manually run database migrations. Running the up (default) command on the migrator of the version you are upgrading to will apply all migrations required by the next version of Sourcegraph.

Failing migrations

Migrations may fail due to transient or application errors. When this happens, the database will be marked by the migrator as dirty. A dirty database requires manual intervention to ensure the schema is in the expected state before continuing with migrations or application startup.

In order to retrieve the error message printed by the migrator on startup, you'll need to use the kubectl logs <frontend pod> -c migrator to specify the init container, not the main application container. Using a bare kubectl logs command will result in the following error:

Error from server (BadRequest): container "frontend" in pod "sourcegraph-frontend-69f4b68d75-w98lx" is waiting to start: PodInitializing

Once a failing migration error message can be found, follow the guide on how to troubleshoot a dirty database.

Troubleshooting

See the troubleshooting page.