> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.nvidia.com/nvcf/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/nvcf/_mcp/server. # 0.6.0 to 0.6.1 Upgrade The 0.6.1 release replaces the archived Bitnami Cassandra runtime with the official Apache Cassandra 5.0.8 image and an in-house Cassandra Helm chart. This removes the CVE-affected Bitnami runtime and its bundled components. Fresh installs need no special handling. Existing clusters that already hold data need a migration, because three things change at once: the runtime image, the Helm chart shape, and the on-disk data layout. A normal in-place Helm upgrade does not carry an existing Bitnami-based cluster onto the new stack: - The new Cassandra chart intentionally removes the `app.kubernetes.io/name` and `app.kubernetes.io/instance` labels from `volumeClaimTemplates.metadata.labels`. This is the only change to an immutable StatefulSet field. Kubernetes treats the entire volume claim template as immutable, so the StatefulSet has to be recreated. - The Bitnami image stored data nested one directory deeper than the official image expects (`/bitnami/cassandra/data/...` versus `/var/lib/cassandra/...`), so the new image does not find the old data by default. - The old and new node configuration must agree on the settings the existing data was written with. Because of this, an operator with an existing data-bearing cluster chooses a migration method. This guide describes the methods and their trade-offs; none is mandated over the others. These steps assume 0.6.0 is already deployed and healthy, that you have downloaded the 0.6.1 stack, and that you run `make` from the extracted `nvcf-self-managed-stack/` directory with your environment set. For an existing data-bearing cluster, do not run an unscoped 0.6.1 stack install before completing the Cassandra steps for your chosen migration method. This includes `make install`, `helmfile sync`, and equivalent sync commands. An unscoped sync attempts to patch the existing Cassandra StatefulSet, fails on its immutable fields, leaves the Cassandra Helm release in a failed state, and stops the later Helmfile stages. If an unscoped sync already failed, do not uninstall Cassandra or delete its PersistentVolumeClaims. Continue with the selector-scoped Cassandra install in your chosen migration method. That install replaces the failed release. Verify the migration before running the full stack install. ```bash export HELMFILE_ENV= # for example eks-cds-qa ``` ## Prepare the 0.6.1 configuration Return to the extracted 0.6.1 stack before running any 0.6.1 sync command. Reconcile the site-specific settings from the 0.6.0 deployment into the 0.6.1 configuration files: - `environments/$HELMFILE_ENV.yaml` - `secrets/$HELMFILE_ENV-secrets.yaml` Start from the files or templates included with 0.6.1 and carry forward the required values from 0.6.0. Do not replace the 0.6.1 files wholesale because available settings can change between releases. Preserve the existing registry, storage, endpoint, credential, and secret values unless this procedure explicitly instructs you to change them. Do not change the Cassandra superuser password or application role passwords during this migration. Preserve the Cassandra and OpenBao credential values from the 0.6.0 deployment. The 0.6.1 Cassandra chart does not create the old `cassandra-system/cassandra` Secret, so do not rely on that Secret after the upgrade. Rotate Cassandra credentials in a separate maintenance procedure. Confirm that both files exist before continuing: ```bash for file in \ "environments/${HELMFILE_ENV}.yaml" \ "secrets/${HELMFILE_ENV}-secrets.yaml"; do [ -f "$file" ] || { echo "missing required file: $file"; exit 1; } done ``` ## Fresh installs A new 0.6.1 install has no existing Cassandra data and needs no migration. Install the stack normally; Cassandra comes up on the official image with the in-house chart: ```bash make install HELMFILE_ENV="$HELMFILE_ENV" ``` The rest of this guide applies only when migrating an existing data-bearing cluster. ## Check Cassandra before migrating Confirm the current Cassandra deployment is healthy and record its state: ```bash kubectl -n cassandra-system get pods,pvc kubectl -n cassandra-system exec cassandra-0 -c cassandra -- nodetool status ``` All Cassandra pods should be `Running` and report `UN` (Up/Normal) in `nodetool status`, and the data PersistentVolumeClaims should be `Bound`. Record the keyspaces and a row count for a known table so you can verify data after the migration: ```bash kubectl -n cassandra-system exec cassandra-0 -c cassandra -- \ cqlsh -u -p -e "DESCRIBE KEYSPACES;" ``` Check the existing schema migration state: ```sql SELECT table_name FROM system_schema.tables WHERE keyspace_name = 'schema_migrations'; ``` For each table returned, run: ```sql SELECT version, dirty FROM schema_migrations.; ``` Every migration record must have `dirty` set to `false`. If any record is dirty, reconcile that partial migration before starting this upgrade. Take a backup before migrating. On every node, run `nodetool flush` and then `nodetool snapshot`, and copy the snapshots off-node. Do not delete Cassandra PersistentVolumeClaims during any migration method. ## Choose a migration method Three methods are supported. They are listed in no particular order; choose the one that fits your deployment and your tolerance for downtime and operational steps. | Method | Downtime | Data handling | Requires | |---|---|---|---| | Datacenter expansion | Effectively none | New nodes stream the data; old volumes untouched | Capacity to run both datacenters at once, datacenter-aware clients | | In-place volume adoption | Brief restart per node | The new pods reuse the existing volumes | A manual StatefulSet recreate; a subPath or a one-time layout relocation | | Backup and restore | Maintenance window | Snapshot the old cluster, load into a fresh one | Capacity for a parallel cluster during the copy | Datacenter expansion keeps the cluster online by adding the new stack as a second Cassandra datacenter and streaming data to it, then cutting clients over. It has the most moving parts and needs enough capacity to run both datacenters during the migration. In-place volume adoption reuses the existing data volumes. It is the smallest change in footprint, but it requires deleting and recreating the Cassandra StatefulSet (its volumes are retained) and pointing the new pods at the existing Bitnami on-disk layout. Backup and restore is the most conservative for data integrity: the existing cluster is left untouched until a fresh 0.6.1 cluster is verified. It needs a maintenance window and capacity to run both clusters during the copy. ## Method: datacenter expansion Add the 0.6.1 stack as a second Cassandra datacenter in the same cluster, replicate the data to it, move clients over, then remove the old datacenter. Because the new nodes bootstrap empty and receive their data by streaming, this method avoids the volume-layout and configuration-compatibility concerns of an in-place adoption. Validate this procedure in a non-production environment before running it against production. Perform the keyspace replication and decommission steps deliberately; an incorrect replication change or an early decommission can lose data. Two settings on the new datacenter's nodes decide whether the join and rebuild succeed: - `storage_compatibility_mode` must match the existing cluster. The existing Cassandra fleet runs `NONE` (full Cassandra 5.0 native format), and the 0.6.1 chart defaults the new nodes to `NONE` for this reason. If the new nodes run a different mode (the Apache Cassandra 5.0 default is `CASSANDRA_4`), gossip and schema exchange still succeed and the datacenters appear healthy, but the `nodetool rebuild` in step 4 fails: the streaming connection is rejected as version-incompatible. - Token allocation. When the existing cluster already holds a large schema, a node that joins with schema-aware token allocation (`allocate_tokens_for_local_replication_factor`) must reach schema readiness within a fixed window during bootstrap, or it aborts with a "Could not achieve schema readiness" error and restarts. If the new nodes fail to finish joining with that error, have them allocate tokens without the schema-aware option (random tokens are acceptable for a new, empty datacenter) and retry. 1. Stand up the new datacenter joined to the existing cluster. Deploy the 0.6.1 Cassandra nodes with the same cluster name as the existing cluster, a distinct datacenter name, seeds that reach the existing datacenter, and `auto_bootstrap: false` so the new nodes join without streaming during bootstrap. Do not run the schema-initialization or migration jobs yet. The schema already exists in the cluster, and the new nodes receive it through gossip. Disable both jobs in the new datacenter's values with `cassandra.hooks.initializeCluster.enabled: false` and `cassandra.hooks.migrations.enabled: false`. Keep them disabled until the new datacenter is included in the application and migration-state keyspaces in step 3. 2. Confirm both datacenters are visible and healthy from a node in either datacenter: ```bash kubectl -n cassandra-system exec cassandra-0 -c cassandra -- nodetool status ``` Every node should report `UN`. The new datacenter nodes appear with no or minimal load until the rebuild in step 4. 3. Extend replication for every application keyspace, `schema_migrations`, `system_auth`, `system_distributed`, and `system_traces` to include the new datacenter. Replicating `schema_migrations` preserves the version and dirty state that the 0.6.1 migration job uses to apply only pending migrations. For each keyspace: ```sql ALTER KEYSPACE WITH replication = { 'class': 'NetworkTopologyStrategy', '': , '': }; ``` 4. On each node in the new datacenter, stream the existing data across: ```bash kubectl -n cassandra-system exec -c cassandra -- \ nodetool rebuild -- ``` 5. Run the pending 0.6.1 schema migrations before moving clients. Keep cluster initialization disabled and enable only the migration hook: ```yaml cassandra: hooks: initializeCluster: enabled: false migrations: enabled: true ``` Apply the Cassandra release and wait for its migration job: ```bash set -euo pipefail make install HELMFILE_ENV="$HELMFILE_ENV" HELMFILE_SELECTOR=name=cassandra kubectl -n cassandra-system wait \ --for=condition=complete job/cassandra-migrations --timeout=10m kubectl -n cassandra-system logs job/cassandra-migrations ``` The logs must end with `All Cassandra migrations completed successfully`. Do not continue if the job failed or timed out. 6. Inventory application keyspace replication after the migration job: ```sql SELECT keyspace_name, replication FROM system_schema.keyspaces; ``` For each application keyspace that does not list both datacenters, update its replication: ```sql ALTER KEYSPACE WITH replication = { 'class': 'NetworkTopologyStrategy', '': , '': }; ``` Run the `system_schema.keyspaces` query again. Confirm that every application keyspace lists both datacenters before continuing. If a newly created keyspace already contains data, rebuild that keyspace on each new-datacenter node before continuing: ```bash kubectl -n cassandra-system exec -c cassandra -- \ nodetool rebuild -ks -- ``` 7. Confirm the migration state is clean: ```sql SELECT table_name FROM system_schema.tables WHERE keyspace_name = 'schema_migrations'; SELECT version, dirty FROM schema_migrations.; ``` Run the second query for each table returned by the first query. Each query must return one record with `dirty` set to `false`. This confirms that the cluster can receive later schema migrations. Do not move clients or remove the old datacenter while a migration is missing or dirty. 8. Move clients to the new datacenter. Point application traffic at the new datacenter using `LOCAL_QUORUM` against the new datacenter name, and confirm reads and writes succeed there. 9. Remove the old datacenter from replication, then decommission it. Alter every keyspace again to drop the old datacenter from the replication map, then decommission the old-datacenter nodes and remove the old stack. ## Method: in-place volume adoption Reuse the existing data volumes with the new stack. The StatefulSet must be recreated because its shape is immutable, and the new pods must be pointed at the Bitnami on-disk layout. The StatefulSet recreate step deletes the StatefulSet object but keeps its pods and PersistentVolumeClaims. Do not delete the PersistentVolumeClaims. Confirm the StatefulSet's `persistentVolumeClaimRetentionPolicy.whenDeleted` is not `Delete` before deleting it. The chart supports adopting the existing volume with `persistence.subPath: data`, which surfaces the Bitnami nested layout at the paths the official image expects. A one-time relocation script is also available to convert the volume to the official layout so no subPath is needed afterward; use one or the other. This method is validated for the standard NVCF Cassandra configuration. It also depends on the new nodes reading the data the old fleet wrote, which requires the 0.6.1 `cassandra.yaml` to match the settings the old data was written with. The chart carries the settings the default NVCF deployment needs. If your Bitnami deployment ran a customized `cassandra.yaml`, verify those settings carry over to 0.6.1 before you commit to this method: a missed setting can prevent the node from starting on the existing data. When in doubt, use datacenter expansion or backup and restore, which do not depend on on-disk configuration compatibility. 1. In `environments/.yaml`, add `persistence.subPath` under the existing `cassandra` section. Do not replace the rest of the file: ```yaml cassandra: persistence: subPath: "data" ``` This is the only migration-specific override required for the standard NVCF Cassandra configuration. The chart already defaults `podSecurityContext.fsGroup` to `999`, `cluster.name` to `cassandra`, and the image to the 0.6.1 runtime. If the existing deployment customized its cluster name or Cassandra configuration, carry those matching values forward. 2. Recreate the Cassandra StatefulSet so the new-shape StatefulSet is created and re-adopts the existing volumes: ```bash kubectl -n cassandra-system delete statefulset cassandra --cascade=orphan kubectl -n cassandra-system delete pod cassandra-0 make install HELMFILE_ENV="$HELMFILE_ENV" HELMFILE_SELECTOR=name=cassandra ``` For a multi-node cluster, the StatefulSet controller recreates `cassandra-0` and then automatically replaces the remaining pods in reverse ordinal order, one at a time. Do not delete the remaining pods manually. In another terminal, monitor the automatic rollout: ```bash kubectl -n cassandra-system rollout status statefulset/cassandra \ --timeout=30m ``` If the rollout stalls, inspect the pod status, events, and Cassandra ring before taking manual action. 3. Confirm the new pod comes up on the existing data: ```bash kubectl -n cassandra-system exec cassandra-0 -c cassandra -- nodetool status ``` The node should report `UN`, and the row count for the known table you recorded earlier should match. ## Method: backup and restore Leave the existing cluster in place and load its data into a fresh 0.6.1 cluster. 1. Snapshot the existing cluster: on every node, run `nodetool flush` then `nodetool snapshot`, and copy the snapshots off-node. 2. Deploy a fresh 0.6.1 Cassandra cluster on new volumes and let the init and migration jobs create the schema. 3. Restore the data into the new cluster with `sstableloader`, or place the snapshot sstables and run `nodetool refresh` per table. 4. Verify row counts and application connectivity against the new cluster, then decommission the old cluster. ## Verify the migration After any method, confirm the Cassandra cluster is healthy on the 0.6.1 image: ```bash kubectl -n cassandra-system get pods kubectl -n cassandra-system exec cassandra-0 -c cassandra -- nodetool status ``` All Cassandra pods should be `Running` and report `UN`. Confirm the known table row count matches what you recorded before the migration, and confirm the control plane is healthy: ```bash kubectl get pods -A --no-headers | awk \ '$4 != "Running" && $4 != "Completed" { print } $4 == "Running" { split($3, ready, "/") if (ready[1] != ready[2]) print }' ``` This command should print nothing. It prints pods that are not `Running` or `Completed`, and `Running` pods whose ready-container count does not match their total-container count. Finally, query each table in the `schema_migrations` keyspace as described in the pre-migration check. Every record must have `dirty` set to `false` before you consider the upgrade complete.