When Did You Last Test a ZFS Restore? Get the Checklist and Learn About Your Quarterly Health Check
In this issue:
🔹 Notes from SNIA SDC ‘26: what’s next for storage.
🔹 Is your AI-generated ZFS config right? What we check first →
🔹 A quarterly DR drill for ZFS, plus a one-page checklist →
🔹 New guides: ZFS misconfigurations, RAIDZ expansion, and ZFS vs Ceph replay →
🔹 FreeBSD EoL dates and why LTS matters →
Our team attended the SNIA Developer Conference September 28 to 30 in Santa Clara. Many sessions returned to the same point: drives keep getting faster with lower latency, meaning more of the cost sits in the software and interfaces that drive them. This will continue to shape how ZFS systems are built and run over the next few years.
This issue brings together what we learned at SDC 2026 and two practical steps we recommend every customer take this quarter: have your ZFS configuration reviewed and test your disaster recovery restore procedure. We also share resources to help your team put both into practice.
From SNIA SDC ’26 in Santa Clara
What’s Next for Storage
Telling the drive what you are writing. A Day 1 session, “Characterizing and Emulating FDP SSDs with WARP,” looked at how to test Flexible Data Placement (FDP) SSDs in a vendor-agnostic way and understand why write amplification occurs.
Filesystems can now pass placement IDs to the drive, marking which blocks are likely to have similar lifetimes and which are not. This can significantly improve garbage collection by aligning placement groups with different flash cells, but this shifts the work of identifying the different types of data from the firmware to the host.
A small share of your data can cause most of your write amplification. The correct write policy can reduce competition between data types for over-provisioning space, improving performance while reducing wear.
Even sequential writes can reach the media out of order after the various layers of drivers, firmware, and hardware reorder them, despite the filesystem’s best efforts to keep them aggregated.
Memory has become the most expensive part of a server. The talk titled "Rethinking Memory" noted that memory used to be about 30% of server cost and is now more than 50%. Most VMs never use half the RAM allocated to them. AMD MEXT can tier cold RAM to fast NVMe, allowing the underutilized RAM to be reallocated. When tested on a chip design workload, this allowed two jobs that required 1 TB of memory each to be run concurrently on a system with only 1 TB of RAM, saving 8 hours of runtime.
Per-request overhead adds up. On Day 3, a session on batched I/O explained that as storage gets faster, the fixed software cost of each request takes a larger share of the total processing time. Batching raised Gen5 NVMe throughput by 127%, reaching 13,850 MiB/s, and reduced CPU use from 100% to 77%. It only works if applications support submitting batches and every layer of the OS stack supports batching end-to-end. Care must be taken to ensure that I/O batch completions maintain the strict ordering required for data safety.
What It Means for ZFS Users
Over the next five years, more of your storage performance bottlenecks will come from the software stack and memory bandwidth limits. ZFS sits in that stack and it will need to continue to evolve to keep up with high-performance flash media, while not losing its considerable capabilities to manage spinning rust. Expert tuning of ZFS to take maximum advantage of the hardware will offer increasing gains.
Missed SNIA SDC ’26? Explore the sessions and watch the available recordings →
Next, our team heads to the FreeBSD Vendor Summit, November 5–6 in San Jose. Stay tuned for takeaways from the event. Follow us →
Test Your Restores Every Quarter
Attacks are moving faster than ever before. CrowdStrike's 2026 Global Threat Report puts average eCrime breakout time at 29 minutes, and the fastest on record at 27 seconds. In Veeam's 2025 ransomware research, only 10% of affected organizations recovered more than 90% of their data. Risk also comes from inside: in July 2025 an AI coding agent at Replit deleted a production database during a code freeze.
The standard DR test once a year was planned for a slower-moving threat environment. We recommend a quarterly drill. Restore a dataset from a replicated snapshot onto an isolated host and time it against your RTO. Our one-page checklist walks through each step with the ZFS commands.
October Tech Guides
More teams now get their pool layouts, tunables, and replication scripts from an AI assistant. Some of that output is fine. Some of it was written for an older OpenZFS release, a different OS, or a different workload. ashift and vdev layout are locked in when the pool is created, changing them means having to rebuild the entire pool.
We review configurations the same way no matter who wrote them. The review starts with the settings you cannot change later and the snapshot and replication setup you rely on for recovery. We ensure each change has a documented reason and clear criteria for when to revisit it.
Get your ZFS configuration checked by Klara engineers, regardless of who wrote it, and receive a written report with fixes ranked by risk.
RAIDZ Expansion in Practice: What It Solves and What It Doesn't
Add capacity one disk at a time, without downtime. Learn why expansion won’t fix random IOPS or change parity, and what to plan for before reflow starts.
Biggest ZFS Misconfigurations, Part 2
An idle SLOG. Useful caching bypassed. A special vdev without redundancy. Spot configuration mistakes that waste hardware, hurt performance, or put data at risk.
WEBINAR REPLAY
Klara was joined by Clyso to discuss when a workload needs scale-out storage and when scale-up ZFS is the better fit, and what each choice means for performance and operations.
Why LTS Matters
Enterprise systems often outlive their OS support windows. Our latest article shows how LTS for FreeBSD closes that gap and lets you migrate on your own schedule.
The EoL Clock
14.4-RELEASE: security support ends December 31, 2026.
15.1-RELEASE: supported until March 31, 2027.
If those dates apply to systems you cannot upgrade this quarter, LTS by Klara keeps them patched while you plan the move.
New Upstream Fixes & Features
FreeBSD
Restored Compatibility for Longer XDR Strings — A fix to xdr_string removes an imposed SunRPC size limit that broke applications which deliberately use XDR with longer strings. This restores compatibility for those applications without changes to their data formats.
ZFS
Easier Reproduction of ZFS Test Failures— ztest now supports a platform-agnostic PRNG with either a randomly generated seed or one saved from a previous run. Developers can reuse the seed to repeat the same sequence of random choices, which helps reproduce certain failures and retest fixes more quickly. Execution timing can still vary between runs.
Faster storage still depends on sound configuration and a recovery plan that works. Review the choices behind your ZFS environment, test your restores, and use the results to plan your next changes.
Need Help?
Get your ZFS configuration reviewed by Klara’s engineers every quarter with our Quarterly ZFS Health Check.
Klara Inc., 286 South Bend Road East, Hamilton, Ontario L9a2c5, Canada