HCI Go/No-Go for a Hospital Refresh
A regional hospital runs 42 VMs on a five-year-old three-tier stack: two database VMs for the patient portal, 18 application servers, 14 VDI desktops for clinicians, and eight file/print servers. The SAN array is at 78% capacity and shows 18 ms peak write latency during the 07:00–09:00 medication round; the vendor refresh is due in six months. The CIO has asked whether hyperconverged infrastructure can replace the stack and meet a 4 ms write-latency target, but the database team insists on shared block storage and the VDI team expects boot storms of 60 desktops in 10 minutes. The network team can provide two 10 GbE ports per node and a 10 GbE interswitch link. The hospital has a single data center room with 12 free rack units and 6 kW of spare power. The CIO wants a go/no-go recommendation in two weeks, not a product demo.
1. What workload evidence must be collected before accepting or rejecting the 4 ms write-latency target?
Read the worked answer
Collect per-VM and per-datastore evidence, not an array average:
- IOPS and latency by workload: peak IOPS, read/write ratio, block size, queue depth, and latency sensitivity for the two database VMs, the VDI boot storm, and the file/print servers. The 18 ms is a symptom; determine whether it is array contention, queue depth, or a single noisy VM.
- Write path: with RF2 every write is acknowledged after two copies are written, so inter-node network latency and SSD endurance matter. A 4 ms write SLO is end-to-end and must be proven with a PoC using the real workload.
- VDI boot storm: 60 desktops in 10 minutes is a burst; model the IOPS and the read/write mix, not just steady-state.
- Capacity and growth: 78% array utilization today plus refresh headroom.
If the evidence shows a few large sequential streams or a latency-sensitive workload with no tolerance for distributed commits, reject or redesign. If it shows many mixed random workloads, HCI is a credible candidate but still needs a measured PoC.
2. How should the database team's shared-block requirement be addressed without abandoning the HCI model?
Read the worked answer
Treat 'shared block' as a requirement to validate, not an automatic blocker:
- Nutanix Volumes can present iSCSI block storage to VMs from the cluster, which can satisfy an application that insists on shared block. It adds a storage layer and should be tested for latency and multipath behavior.
- Native AHV storage is usually better: run SQL Server on AHV with Always On availability groups, using separate virtual disks and application-level replication. That avoids the Volumes layer and keeps data local.
- Clarify the real requirement: is it shared block, or shared storage with HA and consistent backups? If it is the latter, native HCI plus snapshots and replication may meet it.
- Check the support matrix for the database version, cluster type, and Volumes configuration; do not assume all clustered databases are supported.
Recommend a PoC that runs the database on native storage and, if needed, a second test with Volumes, then compare latency and operational complexity.
3. State two conditions that would justify a no-go recommendation and explain the reasoning.
Read the worked answer
Two credible no-go conditions:
- A small number of high-throughput sequential streams dominate. If the workload needs sustained multi-GB/s single-stream throughput, a distributed HCI cluster may not match a dedicated array and fewer, larger nodes; the network and rebuild traffic can also interfere. Size and test before saying no.
- A hard requirement for external-array-only features or unsupported apps. Examples include a vendor-certified SAN-only configuration, strict external encryption/key management, or a latency SLO that cannot tolerate distributed write commits. In a hospital, clinical app certification matters more than architecture preference.
A third condition is physical: 12 free rack units and 6 kW may limit node count or drive a different model. The recommendation should be 'no-go for this stack as specified' or 'conditional go after PoC', not an ideological answer.