iiinova — where your digital self keeps evolving

Do you know when your last scrub completed? zpool status tank | grep scan Check the DR pool too. That's the copy you'll restore from. Link to download the checklist is in the comments. #OpenZFS #DisasterRecovery #Storage

100

Klara Inc @klarainc.bsky.social · 09/10/2026

A SLOG accelerates exactly one thing: synchronous writes. Async writes never touch it. Watch zpool iostat -r under real load before buying. If sync=disabled makes you fast, that is ZFS telling you to buy the SLOG. Part 2 linked below.

120

Klara Inc @klarainc.bsky.social · 08/10/2026

One week to go: ZFS vs. Ceph webinar 📅 October 15, 11 AM EDT When does distributed storage make sense, and when does it add unnecessary complexity? Join Allan Jude and Dan van der Ster live to compare ZFS and Ceph across performance, scalability, and operations. Register: na3.hubs.ly/y0CQsP0

010

Klara Inc @klarainc.bsky.social · 07/10/2026

OpenZFS 2.3 shipped RAIDZ expansion, exactly what small pools have wanted for a decade. Just know that old blocks keep their original data-to-parity ratio, so the capacity arrives as data is rewritten. What else people get wrong: Part 2, linked below.

120

Klara Inc @klarainc.bsky.social · 06/10/2026

Storage appliances and embedded fleets live for a decade. A FreeBSD stable branch carries 4 years of security support. Ten-year products on four-year branches: that mismatch is exactly the gap LTS was invented to close. Article linked below.

120

Klara Inc @klarainc.bsky.social · 06/10/2026

Now available for download: the ZFS Disaster Recovery Drill Checklist. Run a 90-day restore drill: restore from a replicated snapshot, measure RTO and RPO, verify protections, and document the results. ⬇️ Download the checklist. Link in the first comment. #ZFS #DisasterRecovery #DataStorage

100

Klara Inc @klarainc.bsky.social · 05/10/2026

In 2023, kernel.org cut LTS kernel support from 6 years to 2. Too much backporting, not enough funded people. The honest fine print of LTS: it is engineering someone has to fund, not a policy you declare. Why that matters for FreeBSD: linked below.

100

Klara Inc @klarainc.bsky.social · 02/10/2026

RHEL 7 shipped in 2014. With Extended Life Cycle Support, security coverage runs into 2028. Fourteen years on one major version is not an edge case Red Hat tolerates. It is a product Red Hat sells. The case for FreeBSD LTS: article linked below.

120

Klara Inc @klarainc.bsky.social · 02/10/2026

Ubuntu LTS: 5 years standard, 5 more with ESM, now 5 more with the Legacy add-on. 15 years of coverage, expanded because demand kept growing. Nobody talked enterprises into LTS. They pulled vendors into it. Why FreeBSD needs the same: linked below.

120

Klara Inc @klarainc.bsky.social · 01/10/2026

ZFS or Ceph? The answer depends on workload and design constraints. Join us live on October 15 for a practical look at when scale-out makes sense, when scale-up is the better fit, and what each choice means for performance and operations. Register now: na3.hubs.ly/y0CQsP0

010

Klara Inc @klarainc.bsky.social · 01/10/2026

Enterprises don’t buy operating systems. They buy time. Time between forced changes, time to certify, time to amortize hardware. Twenty years of Linux LTS proved the model. What it takes to bring the same guarantee to FreeBSD: article linked below.

100

Klara Inc @klarainc.bsky.social · 30/09/2026

[3/3] The results showed the impact: Gen5 throughput increased by 127%, reaching 13,850 MiB/s, with CPU utilization dropping from 100% to 77%. That wraps up three days at SDC '26. Thanks to SNIA and everyone who shared their work and conversations with us this week. #SDC2026 #SNIA #DataStorage

000

Klara Inc @klarainc.bsky.social · 30/09/2026

[2/3] One session looked at using batched I/O to overcome that overhead and saturate Gen5 NVMe devices with reduced CPU overhead. Applications need to submit batches, and every layer of the stack has to understand batching.

100

Klara Inc @klarainc.bsky.social · 30/09/2026

[1/3] Final day at SNIA SDC '26 in Santa Clara, with one more technical takeaway from our team. As storage gets faster, the fixed software overhead of each I/O request becomes a larger part of the total cost. 👇

120

Klara Inc @klarainc.bsky.social · 29/09/2026

[3/3] Using scatter/gather with offsets allows the data to be passed with the required alignments. More from the SDC ’26 floor tomorrow. Stay tuned. #SDC2026 #SNIA #NVMe #Storage

000

Klara Inc @klarainc.bsky.social · 29/09/2026

[2/3] AMD MEXT provides a mechanism to tier cold RAM to fast NVMe, but would benefit from better priority queuing in the NVMe protocol. 2. Another talk looked at how to avoid memory copies in network protocols, where data can be misaligned in pages behind protocol headers.

101

Klara Inc @klarainc.bsky.social · 29/09/2026

[1/3] Day 2 at #SNIA SDC ’26 is underway in Santa Clara, and our team has two more technical takeaways to share. 1. Memory tiering: Memory used to account for around 30% of the cost of a server. Now it accounts for more than 50%, yet most VMs never use half of the RAM allocated to them. 👇

100

Klara Inc @klarainc.bsky.social · 29/09/2026

Migrate on your terms. Stay supported on the release you run now. Upgrade when migrating is the plan, not the emergency. LTS by Klara: na3.hubs.ly/y0CD_V0 #FreeBSD

030

Klara Inc @klarainc.bsky.social · 28/09/2026

[4/4] 3. Even when you write data sequentially, and your filesystem aggregates to improve it further, the various driver and hardware layers can reorder things and result in access patterns that do not resemble sequential at the media. More updates to come. #SDC2026 #SNIA #DataStorage #SSD #Storage

010

Klara Inc @klarainc.bsky.social · 28/09/2026

[3/4] 2. A small fraction of your data can contribute the majority of your write amplification. Selecting the best write policy can reduce how often two data types are competing for over-provisioning space.

110

Klara Inc @klarainc.bsky.social · 28/09/2026

[2/4] Three takeaways from the session: 1. Using placement IDs, a filesystem can share information about which blocks are likely to have similar lifetimes and which will have very different lifetimes. This improves garbage collection by shifting some of the work from the firmware to the host.

110

Klara Inc @klarainc.bsky.social · 28/09/2026

[1/4] Day 1 updates from the floor at SNIA SDC ’26, happening this week in Santa Clara. “Characterizing and Emulating FDP SSDs with WARP” looked at how to understand and test Flexible Data Placement (FDP) in a reliable, vendor-agnostic way to determine why write amplification occurs. 👇

100

Klara Inc @klarainc.bsky.social · 28/09/2026

For years, zfs set dedup=on was the most expensive command in storage. When the dedup table outgrows RAM, every write turns into random table reads. Fast Dedup in OpenZFS 2.3 changes the math, but you still have to do the math. Part 1 of ZFS misconfigurations: linked below.

110

Klara Inc @klarainc.bsky.social · 25/09/2026

InnoDB writes 16K pages. A default ZFS dataset uses 128K records, so every page update becomes a read-modify-write. That is 8x amplification you get by not configuring anything. One command fixes new writes: zfs set recordsize=16K tank/mysql. More ZFS fixes linked below.

110

Klara Inc @klarainc.bsky.social · 24/09/2026

ZFS random IOPS scale with vdevs, not disks. One 12-disk RAIDZ2 has roughly the random I/O profile of a single disk. The same 12 disks as six mirror pairs have six vdevs pulling I/O. Layout is a day-one decision; fixing it later means rebuilding. Full article linked below.

100

Klara Inc @klarainc.bsky.social · 23/09/2026

ashift is set when a ZFS vdev is created and can never be changed. A 512e drive that lied about its sector size means write amplification for the life of the pool. Check yours: zdb -C tank | grep ashift. The biggest ZFS misconfigurations and how to fix them: linked below.

110

Klara Inc @klarainc.bsky.social · 22/09/2026

FreeBSD 13 is EoL. 15.0-RELEASE follows on September 30. Your upgrade window does not have to be upstream’s. Keep the branch you run patched while you plan the move. na3.hubs.ly/y0CD_V0 #FreeBSD

020

Klara Inc @klarainc.bsky.social · 21/09/2026

Quarterly maintenance windows: 4 chances a year to touch every machine. A FreeBSD branch ships 2 minors a year, each supported only 3 months past its successor. Miss one window and part of the fleet slides out of support. The full cadence math: article linked below.

130

Klara Inc @klarainc.bsky.social · 18/09/2026

Upgrade one server and a 99% success rate sounds great. Run it across 200 machines and it means two expected failures every cycle. At scale, upgrade failures are not a risk. They are a line item. The hidden costs of upgrading FreeBSD in production: article linked below.

110

Klara Inc @klarainc.bsky.social · 17/09/2026

Best practice for FreeBSD upgrades: bectl create pre-15.1 before you touch anything. A ZFS boot environment makes rollback a reboot, not a restore-from-backup. More practices that lower the production upgrade bill: article linked below.

152

Klara Inc @klarainc.bsky.social · 17/09/2026

In our next live session, we’ll compare ZFS and Ceph, where each architecture fits, and what the choice means for performance, failure recovery, and operations. Join Klara’s Allan Jude and Clyso’s Dan van der Ster on October 15 for a technical discussion. Register now: na3.hubs.ly/y0CQsP0

020