Skip to main content

Rob Pankow Rob Pankow

vSAN ESA vs OSA: What the Express Storage Architecture Actually Requires From Your Hardware

Oct 5, 2026  |  15 min read

Last edited: Oct 19, 2026

vSAN ESA vs OSA: What the Express Storage Architecture Actually Requires From Your Hardware

“What is the main difference between ESA and OSA?”

Typed almost exactly that way, that question pulls more readers to a single community forum thread than any vendor page on the subject does. The answers it gets back are feature lists. The Express Storage Architecture is newer. It does RAID-5 and RAID-6 at something close to RAID-1 performance, which removes the write penalty that pushed most clusters onto mirroring in the first place. It compresses data at the top of the stack, before the data crosses the network.

All of that is true. None of it is the difference that shows up on a purchase order.

Here is the one that does. The Original Storage Architecture was a software layer you could lay over hardware you largely already owned, within a fairly wide set of choices. The Express Storage Architecture is closer to a hardware specification with software attached. Adopting it is not a software upgrade you schedule into a maintenance window. It is a procurement event, and that reframes the decision that follows it.

This post compares the two architectures on the criteria that decide what you have to buy. It deliberately does not cover hardware certification and compatibility-list mechanics after a platform change, which is a different question answered in the compatibility matrix after a vSphere exit. The question here is narrower: what does ESA demand of the hardware while you are still on VMware® vSphere, and what does that demand open up?

The five criteria that separate the two architectures

Comparisons of ESA and OSA usually wander because they mix capability claims with hardware consequences. Setting the criteria before the comparison keeps the two apart. Five things actually differ in a way that changes a bill of materials.

  1. What devices the architecture will accept. The media class, the interface and the endurance grade it is willing to run on.
  2. How those devices are organised. Whether the host groups drives into tiers with distinct jobs, or treats them as one pool.
  3. What the network has to carry. Both raw throughput and what crosses the wire before it gets there.
  4. Where the data services run, and what they cost. Compression, checksums and parity have to execute somewhere, and that somewhere is host CPU.
  5. How prescriptive the hardware choice is. Whether you build to a broad certification list or select a defined node profile.

Everything else, including the headline parity performance, follows from these. Take them in order and the architectures separate cleanly.

The two architectures on those five criteria

Table 1: vSAN OSA and ESA compared on the criteria that change a bill of materials.

CriterionOSAESAWhat the difference costs you
Accepted devicesHybrid or all-flash; SAS, SATA or NVMe behind a controllerAll-NVMe TLC only, no hybrid, no SAS or SATA tierExisting SAS and SATA capacity cannot carry forward
Device organisationDisk groups: a dedicated cache device plus a capacity tierOne flat storage pool per host, every device serves both rolesCache-device failure domains disappear, but so does reuse of a mixed estate
Network floor10 GbE is workable for all-flash25 GbE is the practical floor, particularly for paritySwitch ports and NICs become part of the project
Data servicesDeduplication and compression applied per disk group, lower in the stackCompression per object at the top of the stack, before the wireMore capability, paid for in host CPU and memory headroom
Hardware freedomBuild broadly to a certification listSelect a defined node profile with fixed drive counts and classesThe node, not the component, becomes the unit of purchase

Read down the last column and the pattern is consistent. Every ESA improvement is real, and every one of them is purchased by narrowing what hardware qualifies.

The device rule is the sharpest edge. ESA will not take SAS or SATA, and it will not run hybrid. For a cluster already on all-flash NVMe, that is a non-event. For the large population of clusters still running a hybrid or SAS all-flash estate, and that population is far larger than the marketing suggests, it means the drives in the chassis today do not travel. Not as a cache tier, not as cold capacity, not at all.

The disk-group change is the one most often reported as pure simplification, and mostly it is. Losing the cache device removes a genuinely awkward failure domain, where one device going down took its whole disk group with it. What gets less attention is the second-order effect on sizing. Under OSA, a chunk of each disk group’s flash was spent on cache and returned no usable capacity. Under ESA every device contributes capacity, so the usable-per-raw arithmetic changes, and so does the drive count you need to hit a target. That is a sizing exercise to redo from scratch, not to carry across.

The networking floor is the item that most often turns a storage project into a network project. Parity under ESA moves more small traffic between hosts than mirroring did, and the architecture is designed on the assumption that the fabric is not the constraint. Teams that budgeted for drives and discovered mid-project that the top-of-rack switches were also in scope are a recurring pattern, and the switches are rarely the cheaper half.

What an ESA node profile actually commits you to

The certification model is where the two architectures differ most in practice, and it is the part least visible from a feature comparison.

Under OSA, certification was largely componentwise. You picked a server, checked the controller, the drives and the firmware against the list, and you had considerable latitude in how you assembled it. Clusters were routinely built from hardware already on the floor, or from a server line a team had standardised on years earlier.

ESA node profiles work the other way around. A profile is a defined configuration: a drive count, a drive class and endurance grade, a CPU floor, a memory floor and a network specification, validated as a unit. You select the profile that matches your workload band, and the components inside it are substantially fixed. The unit of purchase moves from the component to the node.

That has three consequences worth planning for.

The drive count is not yours to trim. A profile specifies how many devices the host carries. Buying fewer and planning to grow later is not the same flexibility it was under OSA, where adding a disk group was a normal expansion path.

The CPU and memory floors are real floors. Compression at the top of the stack, per-object parity and checksums all consume host cycles. A profile’s CPU and memory minima exist because the data services need that headroom, and a host sized for OSA will not necessarily clear them. This is the line item teams most often miss, because it is not a storage line item.

The network specification travels with the profile. The 25 GbE expectation is not advisory. It is part of what the profile was validated against, and under-provisioning it undermines the parity performance that was the reason to move.

Taken together, this is why the move reads as a procurement event rather than an upgrade. You are not enabling a feature. You are specifying, quoting and buying a set of all-NVMe nodes, and the timeline is set by lead times rather than by change windows.

The quote is for all-NVMe hardware. The real question is what it should run

This is the point where the decision quietly changes shape, and it is worth being precise about why.

Up to here, the question has been framed as “OSA or ESA.” But working through the hardware bar produces something more consequential than an answer to that question. It produces a quote for a set of all-NVMe TLC nodes with a 25 GbE fabric, a CPU floor and a memory floor. That is a general-purpose, high-performance hardware platform. vSAN ESA is one of several things it could run, and because the hardware is being bought either way, the real decision is what the hardware runs rather than which vSAN architecture it runs.

Start with the capacity arithmetic, because it determines how many of those drives you actually buy.

Raw capacity required per 100 TB of usable capacity, comparing vSAN ESA RAID-5 and RAID-6, simplyblock erasure coding, vSAN OSA RAID-1 mirroring, and three-way replication
Figure 1: Raw capacity to buy per 100 TB usable. Shorter bar means fewer drives on the purchase order.

Two readings matter here, and the second is the one usually left out.

The first is that ESA’s efficiency gain over OSA is genuine and substantial. Moving from RAID-1 mirroring at roughly 2:1 to RAID-5 at roughly 1.25:1 is a real reduction in drives purchased for the same usable capacity, and it is the strongest argument in ESA’s favour. Anyone presenting ESA as marketing noise is not being straight with you.

The second is that ESA’s ratio is not unique to ESA. Simplyblock’s erasure coding lands at 1.25:1 at FTT1 and 1.5:1 at FTT2, which is the same order of efficiency on the same drives. By contrast, three-way replication, the default resilience model in several open-source storage layers, sits at roughly 3:1 in practice. These figures come from simplyblock’s own capability comparison, based on publicly available information as of September 2026, and they are not independently verified. Treat them as a sizing starting point and measure your own.

So efficiency is not the thing that separates the options on this hardware. What separates them is what else comes with the layer, and what the layer locks you into.

Table 2: What the same all-NVMe nodes can carry, and the condition that makes each choice correct.

PathWhat stays the sameWhat changesChoose it when
ESA node profiles running vSANOperating model, tooling, skillsHardware refreshed, cluster re-sized, capacity model redoneThe vSphere estate is staying put for the hardware’s life
Same nodes running simplyblock under KubeVirt or Red Hat® OpenShift® VirtualizationDrives, fabric, chassisHypervisor and storage layer both changeA platform move was already on the roadmap within that window
Same nodes running an open-source replicated layerDrives, fabric, chassisCapacity model roughly doubles at three-way replicationPerformance is genuinely not the primary goal

The middle row is the one the hardware makes newly available, and it is worth being concrete rather than promotional about it.

Simplyblock runs hyperconverged, with storage and compute co-located on the same nodes, which is the deployment model that matches what an ESA-class node profile already looks like. It also runs disaggregated over an NVMe/TCP or NVMe/RoCE fabric when separating the tiers is the better fit, and that choice can change later without changing vendors. The drives that would satisfy an ESA profile back a Kubernetes-native pool directly. The 25 GbE fabric specified for ESA parity traffic is more than the storage layer needs.

Two further points belong here because they change the arithmetic rather than the pitch.

Backing devices are not restricted to local NVMe. Any whole Linux block device can join the pool, including LUNs an array you already own presents to the storage nodes, which makes an existing array into capacity rather than into something to decommission. Two caveats travel with that claim and both matter. Non-NVMe device support is newer than the NVMe path, so local NVMe remains the production performance recommendation and array-backed pools are better framed as a transition or asset-sweating option than as a greenfield design. And simplyblock formats the devices it takes, so capacity is reused but the data on it still has to be migrated. Consuming an array’s capacity, yes. Re-presenting existing LUNs in place, no.

On cost: simplyblock is priced like the platform it runs on, by the CPU capacity of your worker nodes, counted the same way your OpenShift or Kubernetes subscription counts them, and data volume does not change the price. Disaster recovery, replication, S3 backup, encryption, QoS and support for both VMs and containers are included rather than sold separately, and control-plane and infrastructure nodes are not counted. The software itself, with every feature, is free to download; the subscription adds the commercial license and enterprise support recommended for production.

The honest version of the third row: Ceph and the layers built on it remain a good option when performance is not the main goal. The reason it appears in the table with a doubled capacity model is not that it is a bad choice, but that at three-way replication the same usable capacity needs materially more of the hardware you are about to buy, and that is a number worth having in front of you before the order goes in rather than after.

None of this argues that ESA is the wrong call. If the vSphere estate is staying for the life of the hardware, ESA on a matched node profile is a coherent, well-engineered decision and the efficiency gain over OSA is real. The argument is narrower and, I think, harder to dismiss: the hardware bar ESA imposes is the moment the platform question is cheapest to ask, because you are specifying new nodes anyway. Asking it eighteen months later, with the nodes racked and the capacity model already committed, costs considerably more.

Questions and Answers

What is the main difference between vSAN ESA and OSA?

The architectural difference is that OSA organises drives into disk groups with a dedicated cache device and a capacity tier, while ESA puts every device into a single flat storage pool where all devices serve both roles. The practical difference is that ESA accepts only all-NVMe TLC flash and expects a 25 GbE fabric, so it is far more prescriptive about hardware. OSA was a software layer you could put on hardware you largely had; ESA is close to a hardware specification with software attached.

Can I run vSAN ESA on my existing hosts?

Only if they are already all-NVMe TLC with sufficient CPU and memory headroom and a 25 GbE fabric. Hybrid clusters cannot, and SAS or SATA all-flash clusters cannot, because ESA will not take those devices in any role. Hosts sized for OSA also frequently miss the CPU and memory floors that ESA’s data services assume, and that is the requirement most often discovered late.

Is ESA more space-efficient than OSA?

Yes, and meaningfully so. RAID-5 under ESA delivers roughly 1.25:1 raw to usable at FTT1 against roughly 2:1 for OSA’s RAID-1 mirroring, at performance close to mirroring, which is what makes parity practical where it previously was not. That gain is the strongest argument for ESA. It is worth knowing that the same efficiency band is available from erasure coding in Kubernetes-native storage layers on the same drives, so efficiency alone does not settle which layer should run on the hardware.

If I have to buy all-NVMe nodes anyway, what else could they run?

The same nodes can run a Kubernetes-native storage layer under KubeVirt or Red Hat OpenShift Virtualization. Simplyblock is designed for exactly that hardware profile: it runs hyperconverged on the nodes themselves, reaches a comparable 1.25:1 efficiency at FTT1, uses the 25 GbE fabric the node profile already specifies, and includes replication, disaster recovery, S3 backup, encryption and QoS rather than selling them as separate entitlements. If a platform move was already on your roadmap, buying ESA-class hardware is the cheapest moment to decide which way it goes.

Does moving to ESA mean my existing storage array is stranded?

Not necessarily, though it is stranded as far as vSAN is concerned, since ESA takes only local NVMe. Under simplyblock, LUNs the array presents to the storage nodes can join the pool as capacity, with erasure coding, snapshots, QoS and replication supplied by simplyblock rather than by the array’s own control plane. Two caveats apply: local NVMe remains the production performance recommendation, so this is better treated as a transition or asset-sweating path, and simplyblock formats what it takes, so the capacity is reused but the data still has to be migrated.

How does ESA change my network requirements?

25 GbE should be treated as the floor rather than a recommendation, particularly if you intend to use RAID-5 or RAID-6. Parity generates more inter-host traffic than mirroring, and the architecture assumes the fabric is not the bottleneck. Budget switch ports and NICs as part of the project rather than discovering them during the pilot, because the network half is frequently the more expensive one.

You may also like:

Simplyblock Replaces Your VMware and Database Architecture
Simplyblock Replaces Your VMware and Database Architecture

The VMware + database stack was never designed for modern workloads. Here's how simplyblock and PostgreSQL replace it with a decoupled, API-driven, Kubernetes-native data architecture.

NVMe Storage Cost Optimization in 2026: Erasure Coding, Thin Provisioning, and Compute Efficiency
NVMe Storage Cost Optimization in 2026: Erasure Coding, Thin Provisioning, and Compute Efficiency

NVMe drives deliver the performance Kubernetes stateful workloads need, but triple replication and thick provisioning multiply their cost fast. Here is a practical breakdown of erasure coding economics, thin provisioning, and how sub-millisecond latency reduces compute waste.

KVM vs VMware: Architecture, Cost Model, and the Exit Question
KVM vs VMware: Architecture, Cost Model, and the Exit Question

KVM and VMware ESXi solve the same problem with opposite models: an open hypervisor inside the Linux kernel versus a proprietary integrated suite. How they differ in architecture, operations, and cost, and what a migration involves.