InfraPlaybook publishes engineering guides for the people who carry the pager: sysadmins running small virtualization estates, and the managed service providers who inherit them. Every article answers a question an operator actually typed into a search box at two in the morning.
Editorial mission
One editorial rule carries the whole publication: every figure on this site is traceable to something you can check yourself. A number here is either measured on hardware we describe, quoted from vendor documentation we link to, or derived from arithmetic we show in full. A number that is none of those does not get published.
That rule costs us articles. It is why there is no “2026 MSP salary survey” on this site assembled from three other blog posts averaged together, and why our per-seat pricing article publishes its formula rather than just its conclusion. If you can reproduce the number, you do not have to trust us — which is the point.
Lab and benchmark methodology
What gets benchmarked, and what does not
Not everything. Benchmarking a licensing change is meaningless; benchmarking a storage reclaim is not. We put hardware behind a claim when the disputed question is whether something actually recovers capacity, changes throughput, or completes inside a workable maintenance window.
Most articles on this site carry no benchmark, and we say so rather than implying otherwise. Where an article does publish measured results, the test host and the measurement date appear in that article, directly beside the results table.
The test host
The ZFS discard and capacity-reclaim measurements published in Proxmox Disk 100% Full but the VM Has Free Space were taken on a single physical host:
| Component | Specification |
|---|---|
| CPU | Dual Intel Xeon Silver 4210R |
| Memory | 128 GB DDR4 ECC |
| Pool | Four 1.92 TB enterprise NVMe drives, ZFS RAID-10 |
| Guest disk | Presented to the guest as a zvol |
| Guest OS | Windows Server 2022 Datacenter, 500 GB virtual disk |
| Workload | 220 GB of NTFS file churn written and deleted before each run |
| Measured | 28 August 2026 |
That is one host, described in the singular on purpose. We do not call it “our labs”, and we do not imply that articles without a results table were tested on it.
Three kinds of claim, labelled as such
- Measured. Produced on the host above, with the commands that produced them printed in the article —
zfs list,qm config,qm set,fstrim -avand the rest — so you can re-run them against your own pool. - Documented. Sourced from vendor or first-party documentation and linked inline. Licensing, retention windows and product behaviour fall here. Where vendor documentation is ambiguous, we say that it is ambiguous instead of resolving it for them.
- Modelled. Our per-seat pricing article builds a cost model from stated inputs. It is a model, not a survey, and it says so on the page. We publish the formula and every input so you can substitute your own figures; we do not present a model’s outputs as observed industry data.
Corrections
Errors are corrected in place rather than quietly deleted, and material corrections are noted in the article. If you can show that a figure here is wrong, that is a contribution and it will be treated as one — tell us.
About the editor
Daniel Hayes — Senior Infrastructure Architect, and the editor of InfraPlaybook.
Daniel has spent more than fourteen years designing and operating virtualization and cloud infrastructure for small-business and MSP environments. Since 2014 that work has centred on VMware-to-Proxmox migration: 400+ migrations planned and executed across 12,000+ managed endpoints, with a working specialism in storage controller driver remediation, Microsoft 365 retention behaviour, and the cost arithmetic underneath MSP pricing.
The articles here come out of that work. The failure modes described in them are ones that cost somebody a weekend first.