Most of a management domain build is forgiving. NTP servers change. DNS records get fixed. Somebody tightens a password policy in month four and nobody remembers it was ever different.

A few decisions don't work that way. They're unsupported to change after deployment, or fixable only by rebuilding, and they tend to get settled in the first hour of a design workshop by whoever happens to be in the room.

Storage Type and the ESA/OSA Fork

Storage type can't change after a host is commissioned. It's a hard rule in the commissioning prerequisites, and it sounds obvious until you watch how it gets decided: the hardware was ordered months ago, the quote said NVMe, and somebody assumed that meant ESA.

ESA versus OSA isn't a setting you revisit, either. There is no in-place path from OSA to ESA. A cluster built as OSA because the hosts were one generation too old gets rebuilt or migrated onto new hardware later, and "later" arrives with a budget attached. On VCF 9 every new vSAN cluster needs a vLCM image regardless of architecture, so the hardware compatibility conversation has to finish before the installer runs, not alongside it.

CPU generation belongs in the same conversation. The VCF 9 installer blocks Skylake-SP outright. Cascade Lake-SP is deprecated: the installer lets you continue, and the support goes away at VCF 10. Hosts like that will deploy cleanly today, and whether that's acceptable depends on how many years the hardware has to live, which is a budget question and not an installer question.

Subnet Sizing

VCF doesn't support changing the subnet mask on an infrastructure network after deployment, and expanding one later is disruptive. So the /27 that matched the host count on the whiteboard becomes the thing you explain to the network team in year two. Size for the growth you don't have yet. Addresses are usually cheaper than change windows.

One caveat on that: confirm it for your target release before quoting it to a client, but I wouldn't design as though it had been relaxed.

One Address Family on vSAN Traffic

VCF 9.1 doesn't support dual-stack vSAN VMkernel networking. IPv4 or IPv6, one of them. This catches teams partway through an IPv6 initiative who reasonably assume the cluster can straddle the transition. On vSAN traffic it can't, so the answer has to exist before the VMkernel adapters do.

How Much Resilience the Management Services Get

The management services deployment has a size choice, and the small option has a sharp edge. In a Simple deployment there's a single control plane node. Lose it and the answer is a full redeploy and a restore from backup. There's no HA at that tier.

In a lab, fine. In production it's only fine if someone has actually restored from that backup at least once, and in my experience the restore test is the thing that gets cut when the schedule tightens. Whether a Simple deployment can be converted in place to the HA layout is something I'd find in the docs for the exact target version before betting on it.

A decision nobody wrote down as a decision still gets made. It just gets made by default.

Reversible, But You'll Feel It

  • Forward and reverse DNS. Pre-deployment validation in the VCF Installer blocks the install entirely on failure, with no partial proceed. The fix is cheap. Finding out at install time is not.
  • NTP. The most common blocker at host commissioning, and it has to be configured on the host before commissioning starts.
  • Network pool timing. The pool must exist before hosts are commissioned, and overlapping IP ranges fail pool creation with NETWORKPOOL_WITH_OVERLAPPING_IPPOOLS_FOUND.
  • Appliance sizing. It changed substantially between 9.0.x and 9.1.x, and VCF Operations (fleet management, collector, analytics) is no longer optional. A sizing sheet built on 9.0.x numbers and used for a 9.1.x deployment is working from the wrong table.

Requirements change between builds. Pull the page for your exact target build before a design goes to paper.

Questions I Put on the Table Before Anyone Orders Anything

Is every host's CPU generation on the supported list for the VCF version we're deploying, and for how long after that?

Who decided ESA or OSA, and was the vLCM image compatibility checked before or after that decision?

What's the growth assumption behind each infrastructure subnet, and who signed off on it?

Has anyone ever restored the management backup into a clean environment?

Which address family is vSAN traffic going to use, for the life of the cluster?

If you're about to run a design workshop and want a second opinion on which decisions are really one-way doors, send me a DM. It's a short conversation and it tends to save a long one later.