Migrating 8,000+ VMs in production: Yardi’s move to Talos Director

Yardi has moved more than 8,000 virtual machines onto Talos Director, without re-engineering the applications that run on them. The racks stayed, the applications kept their addresses, and only the operating system changed.

Yardi moved 8,000+ VMs to Talos Director

Yardi, the parent company of Sidero Labs, ran its infrastructure on a commercial virtualization platform for years.

Over the past few months, Yardi has been moving the estate onto Talos Director and Talos Hypervisor, one application group at a time, while the business kept running.

More than 8,000 virtual machines have moved so far, without re-engineering the applications that run on them. This means Yardi has already run into the issues and found the answers that other teams ask about: how a VM actually moves, what was hard, and what to plan for.

For teams mapping their own estate, this is what Yardi replaced and with what at Yardi:

BEFOREAFTER
VMware ESXi hostsTalos Hypervisor on Talos Linux
vCenter and vSphere managementLive migration, DRS and HA in Talos Director
vSphere networking (vDC, port groups)Director networks and security groups
NSX micro-segmentationSecurity groups and network policies in Talos Director
Tanzu Kubernetes clustersKubernetes clusters on Talos Linux, managed in Talos Director, powered by Omni
Storage arraysThe same arrays, presented to Talos Director as storage pools

1. Before production moved

When Yardi’s cloud operations team started planning the move, the first thing they settled on was what to convert first, based on resource consumption, whether a service had native high availability, and its uptime requirements. At each site, internal and test systems were migrated first. Production moved once the team felt comfortable at each site, one application group at a time.

The production conversions were run by the cloud operations engineers and the database administrators, through the normal change-management process.

Every Director host runs Talos Hypervisor: there is no package manager, so every machine runs the same image, and nothing drifts; the OS upgrades as a single atomic image that rolls back if anything goes wrong; and with no SSH or shell, there is no interactive access to defend.

2. Moving a single machine

The standard move was deliberately simple: simple procedures are the most reliable, and they fit the Talos philosophy.

The application owner got at least a day’s notice. In the agreed window, the virtual machine was shut down cleanly on the old platform, and its disks were converted onto a host running Talos Hypervisor. The guest operating system was verified, with the old drivers removed, and the boot configuration repaired if needed. The VM then started up with the same IP addresses and network connectivity. The old copy stayed where it was, powered off: if something went wrong, the team turned it back on.

Downtime was the time required to convert the disks, so that was where the team spent most of its effort, including reading straight from the storage array where possible.

Some of the database servers ran to many terabytes and couldn’t take hours of downtime. For those, Yardi’s team used what Talos Director calls a scheduled cutover: the VM is converted while it stays online, then cut over at a time the team sets, with only a brief final sync. Downtime is meaningfully reduced, from hours to tens of minutes. The cold move stayed the default, since it behaved more predictably at this volume.

3. The exceptions

Some workload types came with conditions attached. Clustered SQL Server databases moved one replica at a time, so the application saw a routine failover rather than an outage. Servers behind load balancers had their pool membership checked after the move. Older Windows machines were booted on a conservative disk setting first to avoid a blue screen.

Kubernetes never went through the converter. Each cluster had a new one built alongside it on the new platform; the applications were deployed and cut over, and the old cluster was retired. Talos Director manages those clusters in the same console as the VMs, with Omni’s cluster management built in.

4. Hardware and the people running it

Nothing much changed in the racks. The servers, cabling, network and storage all stayed, and only the operating system was replaced. A host was emptied on the old platform, redeployed using the bare-metal provisioning in Talos Director, and enrolled into a cluster. Additional capacity was bought where the migration needed it, but the bulk of the estate was reused on the same commodity rack servers that had been running the old platform.

For administrators who had spent years on the old platform, placement, high availability, migration, storage policies, templates and lifecycle management all kept their names and their meaning. Diagnosis runs through a central console. However, there were differences that had to be learned:

  • The muscle memory of working in a new system
  • The micro-segmentation model
  • Omni and the Kubernetes tooling, in place of the tools the previous vendor supplied
  • Upgrading and rolling back by reboot, rather than patching a running system

5. Plan for the seam

The hardest day-to-day part was the in-between, when applications had parts on both platforms. Every phased migration has a period of coexistence, and everything that spans it has to account for both halves.

Micro-segmentation, the load balancers, the monitoring, and any tool that resolves which host a virtual machine runs on all have to read from two inventories instead of one. The seam between the old micro-segmentation solution and the new one was the single biggest challenge of the program.

6. What changed

Virtual machines and Kubernetes run on the same hosts, on one platform and one operating system. The hosts are designed to prevent drift from their intended configuration, and an upgrade is a reboot with a rollback available. Backup and micro-segmentation are part of Talos Director rather than separate products to buy and integrate.

Other things are simply different. Losing the shell changed how people troubleshoot, and some of the engineers still miss it. Some performance tuning, such as handling CPU contention, had to be refined. The Talos Director console doesn’t yet cover every workflow the team was used to. They keep a running list, and the product team continues to add them.

Yardi’s migration is still underway, and the applications on it have carried on as normal:

"Performance and uptime have stayed stable throughout. Applications run on Talos Director the way they ran on the platform they left."

Neil Levien, Director of Cloud Services, Yardi 

7. Beyond Yardi

Yardi is one estate. The goal for Talos Director is a like-for-like replacement for the core VM operations most teams rely on, and so far, Yardi is the first company where we can confirm a move without re-engineering.

Director is built around the infrastructure teams already run: storage pools on existing NFS, iSCSI or Fibre Channel arrays, and standard VLANs on existing switches rather than overlay networks. If an estate depends on a software-defined storage layer or on overlay networks, that functionality is on the roadmap.

Talos Director, including the conversion tooling described here, reaches general availability in January 2027.

The most direct way to see how your own estate would move is to try Director yourself. Talos Director is free for six months, with places offered on a first-come, first-served basis. The trial runs in Sidero cloud, and a sales engineer helps you get set up and stays available for questions.

Find out more about the 6-month free trial and reserve your spot.


VMware, ESXi, vCenter, vSphere, Tanzu and NSX are trademarks or registered trademarks of Broadcom Inc. and/or its subsidiaries. Sidero Labs and Talos Director are not affiliated with, sponsored by, or endorsed by Broadcom Inc. or VMware.