InfoQ Homepage News Yelp Migrates Cassandra to EKS by Moving EBS Volumes.

Nov 18, 2026 4 min read.

Yelp has moved its Cassandra fleet from legacy self-managed Kubernetes clusters to Amazon EKS by re-pointing existing EBS volumes at new pods, rather than streaming data between nodes. A blog post by the site reliability team details a migration. It covered about 90 production clusters. There was zero downtime for customers and no data loss. The largest cluster, a 54-pod user-data store, took about 12 hours end to end.

Recent findings suggest that yelp had already moved its stateless workloads to EKS through its PaaSTA deployment layer, but Cassandra stayed behind. Kubernetes has no built-in way to migrate StatefulSets and their persistent volumes across clusters. The usual approach is to add nodes to the fresh cluster and shut down the old ones. This would temporarily double compute capacity but also lead to data transfer costs. For the largest clusters, streaming alone could have taken weeks. Yelp also considered a regional failover and rebuild, and a new Cassandra datacenter in EKS, but rated both as less transparent to customers.

Yelp's "Rolling PV Transplant" relies on the fact that an EBS volume is bound to an Availability Zone, not to a Kubernetes cluster, and that a PersistentVolume is effectively a pointer to a volume ID. The team pre-creates PVs and PVCs in EKS that reference the existing volume IDs, then creates an identical StatefulSet there, whose pods wait in ContainerCreating. The automation then deletes one pod at a time on the legacy cluster, waits for the volume to detach, and lets the EKS pod claim it. The pod rejoins the ring with the same data and token range, and the process repeats for the next pod. Cassandra's replication tolerates the few minutes each node is offline.

Yelp set a skipReconciliation: true flag on the custom resources in both clusters. This was done because two operators managed the same logical ring during the migration. The flag helped prevent them from conflicting over pod state. The flag was cleared on the EKS side at the end.

Recent findings suggest that across roughly 1,000 pods, about 1,000 TiB moved by pointer rather than over the wire. No node was ever decommissioned from the ring, so rollback was the same operation in reverse. Yelp used it a handful of times, for a missing subnet definition and a certificate mismatch, and restored pods within minutes. After a week of manual dev runs and a week of automation, one engineer ran the production rollout in about three weeks, migrating three to six clusters on a typical day.

The team had Claude generate the orchestration script, roughly 800 lines of bash, from its validated manual runbook, rather than asking it to design the procedure. The script needed minor adjustments after two days of testing.

According to recent reports, the approach relies on a previous Container Storage Interface migration by Yelp's Compute Infrastructure team. This work gave the Data Infrastructure team confidence in the volume detach and attach process. Yelp also used the technique for Kafka clusters, and states it applies to any AZ-pinned stateful workload on Kubernetes. The legacy clusters are now retired.

According to recent reports, the one failure specific to this technique was a stale VolumeAttachment object left behind by the AWS CSI controller after pod deletion. The EKS pod sat in Pending with no error, and deleting the object on the legacy cluster fixed it. Yelp advises building explicit timeouts for stalled detach and attach operations into any automation. Two other findings were unrelated to the mechanism. Backups failed with IAM AccessDenied errors for 48 hours because a new -eks role was not yet being assumed by scheduled jobs, and the alert only filed a ticket instead of paging. Pausing repairs for the migration also exposed one cluster whose repairs had been silently disabled for three months, leaving accumulated data inconsistencies. Teams planning a similar move should audit repair state, backup IAM roles and alert severity before starting.

Yelp's approach is not the only way to move Cassandra between EKS clusters. Monday.com's engineering team describes a different one in a post on zero-downtime Cassandra migration between EKS clusters. K8ssandra-managed clusters had to move without any downtime or data loss. So, the team created a workflow to add new Cassandra data centers, safely transfer data, and remove the old ones. That is the streaming model Yelp considered and rejected. The new cluster is brought up alongside the old one and receives data over the network, and the original is retired only afterwards. The team believes success relied more on control, automation, and handling failures than on speed. They automated long tasks like rebuilds and planned the migration in stages based on operator behavior.

The two migrations differ in what they trade away. Yelp's disk transplant skips data transfer and extra compute. It allows for a rollback by swapping the pointer. However, it relies on EBS volumes being linked to an Availability Zone, not a cluster. Monday.com's approach does not rely on that property, and it keeps the old environment intact until the end. The cost is a longer migration and the capacity to run both sides at once, which Yelp calculated would have taken weeks for its largest clusters.

Despite the different mechanics, the post-migration lessons are similar. Monday.com devotes part of its write-up to verifying that backups work after the migration, along with monitoring, dashboards and alerts. Yelp's own backups failed on an IAM role mismatch within 48 hours of its first production clusters landing in EKS. Teams planning either approach should treat those checks as part of the migration initiative, not as cleanup afterwards.

Presented by: David Iparraguirre - Senior Product Marketing Manager at Datadog, and Jessica Yang - Product Manager at Datadog.

Recent findings suggest that A round-up of last week’s content on InfoQ sent out every Tuesday. Join a community of over 250,000 senior developers. View an example.

β†’ Read original article