38 TB on Four Hosts
A customer wants 38 TB of usable capacity with 25% of it held free, on a four-host vSAN cluster. Each host has two 1.6 TB cache devices and six 3.84 TB capacity devices, all on the hardware compatibility list, and the cluster is licensed for all-flash. Their draft policy is: site disaster tolerance none, failures to tolerate 1, RAID-1 mirroring, thin provisioning, no deduplication or compression. They refuse to buy a fifth host this fiscal year and will not accept a capacity shortfall after go-live. They also want to know whether they can raise failure tolerance to 2 later simply by adding hosts. You are asked to supply a written verdict.
1. Does the draft policy meet the stated requirement? Show the capacity arithmetic.
Read the worked answer
Raw capacity: 4 hosts x 6 devices x 3.84 TB = 92.16 TB.
With RAID-1, failures to tolerate 1: roughly 50% efficiency, so 92.16 / 2 = 46.08 TB before metadata and object overhead. Budget a few percent for overhead and you have a practical figure in the low-to-mid 40s.
The requirement: 38 TB of data with 25% held free means 38 x 1.25 = 47.5 TB of usable capacity needed.
Verdict: no. The design misses by roughly 2-4 TB, and the shortfall appears after go-live, which is exactly the outcome the customer said they would not accept. Note how thin the margin looked at first glance: 46 TB against a 38 TB headline reads as comfortable until the 25% free-space rule is applied. Free space is not optional on vSAN - it governs rebuild behaviour and slack, and running at 92% is how clusters become unstable during a single disk replacement.
The arithmetic, not opinion, is the deliverable here. Present it as a table so the customer can audit each line.
2. What are the options, ranked by risk, and what is the answer on raising failure tolerance to 2 later?
Read the worked answer
Options, lowest to highest risk:
- Add a fifth host. 5 x 23.04 = 115.2 TB raw, RAID-1 gives about 57.6 TB, comfortably meeting 47.5 TB. Best resilience, blocked by their budget, so it must be priced against the next option rather than dismissed.
- Change bulk storage to RAID-5, failures to tolerate 1. Four data plus one parity is roughly 80% efficiency: 92.16 x 0.8 = 73.7 TB, meeting the requirement with margin. Correctness condition: RAID-5 with failures to tolerate 1 requires a minimum of four hosts, so this cluster qualifies exactly. The trade is real - after one host failure there is no redundancy during the rebuild, and rebuilding a 3.84 TB device on a busy cluster takes hours. Their risk appetite, not the arithmetic, decides this.
- Split policies by object. RAID-5 for bulk and cold data, RAID-1 for the small set that needs it. Often the best answer, and the only one that does not require a decision about all data at once.
- Reduce the requirement through deduplication and compression, if the data is compressible - unproven, so it should be measured, not assumed.
On failure tolerance 2: no. RAID-6 needs at least six hosts, and RAID-1 at failure tolerance 2 needs three copies on at least five hosts - at 33% efficiency, five hosts yield about 38 TB before overhead, which still fails the 47.5 TB test.
3. What single piece of information would most change your recommendation, and how would you word the verdict so the customer can act on it?
Read the worked answer
The measurement that matters most: the actual deduplication and compression ratio on their real data, sampled from an existing datastore rather than a vendor benchmark. If the effective ratio is 1.3:1 or better, RAID-1 at four hosts becomes viable and the whole argument about rebuild risk disappears. If it is 1.05:1, option 4 is dead and the decision narrows to "RAID-5 now" versus "fifth host in the next budget cycle." A close second is the twelve-month growth rate: if 38 TB is a year-one figure and the cluster grows 20% annually, the apparent headroom in the RAID-5 option evaporates inside a year.
Wording for the verdict: three sentences and a table. "Your current policy provides approximately 44-46 TB of usable capacity against a requirement of 47.5 TB. As written, the design will run short by 2-4 TB. Two compliant paths exist: switch the bulk of the data to RAID-5 within this cluster, accepting no redundancy during a rebuild, or add a fifth host and keep RAID-1. We recommend separating bulk and critical data onto different policies and reviewing again in twelve months." Name the risk in the same breath as the remedy; do not leave the customer to infer it.