Distributed Storage 2026. A (controversial) technical guide

Another year alive, another year of watching the distributed storage industry outdo itself in commercial creativity. If 2024 was the year everyone discovered they needed “storage for AI” (spoiler: it’s the same old storage, but more expensive), 2025 was the year MinIO decided to publicly immolate itself while IBM and Red Hat continue to bill millions for putting their logo on Ceph.

Hold on, the curves are coming.

The drama of the year: MinIO goes into “maintenance mode” (read: abandonment mode).

If you haven’t been following the MinIO soap opera, let me give you some context. MinIO was the open source object storage that everyone was deploying. Simple, fast, S3 compatible. You had it up and running in 15 minutes. It was the WordPress of object storage.

Well, in December 2025, a silent commit in the README changed everything: “This project is currently under maintenance and is not accepting new changes.” No announcement. No migration guide. No farewell. Just a commit and good bye.

The community, predictably, went up in flames. One developer summed it up perfectly: “A silent README update just ended the era of MinIO as the default open-source S3 engine.”

But this didn’t come out of the blue. MinIO had been pursuing an “open source but don’t overdo it” strategy for years:

2021: Silent switch from Apache 2.0 to AGPL v3 (no announcement, no PR, no nothing)
2022-2023: Aggressive campaigns against Nutanix and Weka for “license violations”.
February 2025: Web console, bucket management and replication removed from the Community version
October 2025: Stop distributing Docker images
December 2025: Maintenance mode

The message is clear: if you want MinIO for real, pay up. Their enterprise AIStor product starts at €96,000/year for 400 TiB. For 1 PB, we are talking about more than €244,000/year.

The lesson? In 2025, “Open Source” without Open Governance is worthless. MinIO was a company with an open source product, not a community project. The difference matters.

In the meantime, Ceph continues to swim peacefully.

While MinIO was self-destructing, Ceph was celebrating its 20th stable release: Tentacle (v20.2.0), released in November 2025. The project accumulates more than 1 exabyte of storage deployed globally on more than 3,000 clusters.

The most interesting thing about Tentacle is FastEC (Fast Erasure Coding), which improves the performance of small reads and writes by 2x to 3x. This makes erasure coding finally viable for workloads that are not pure cold file. With a 6+2 EC profile, you can now achieve approximately 50% of the performance of replication 3 while using only 33% of the space.

For those of us who have been hearing “erasure coding is slow for production” for years, this is a real game changer.

Other Tentacle news:

Integrated SMB support via Samba Manager
NVMe/TCP gateway groups with multi-namespace support
OAuth 2.0 authentication on the dashboard
CephFS case-insensitive directories (finally)
ISA-L replaces Jerasure (which was abandoned)

The Crimson OSD (based on Seastar for NVMe optimization) is still in technical preview. It is not production ready, but the roadmap is promising.

The numbers that matter

Bloomberg operates more than 100 PBs in Ceph clusters. They are a Diamond member of the Ceph Foundation and their Head of Storage Engineering is on the Governing Board. DigitalOcean has 54+ PBs in 37 production clusters. CERN maintains 50+ PBs in more than 10 clusters.

And here’s the interesting part: ZTE Corporation is among the top 3 global contributors to Ceph and number 1 in China. Its TECS CloveStorage product (based on Ceph) is deployed in more than 320 NFV projects worldwide, including China Mobile, izzi Telecom (Mexico) and Deutsche Telekom.

The telco sector is Ceph’s secret superpower. While enterprise vendors sell you expensive appliances, telcos are running Ceph in production on a massive scale.

The vendor circus: paying for your own free software

This is where it gets interesting (and I would even go so far as to say a little absurd).

IBM Fusion: two flavors, same goal

IBM has two products under the Fusion brand:

IBM Fusion HCI: Uses IBM Storage Scale ECE (the old GPFS/Spectrum Scale). Proprietary parallel file system with distributed erasure coding. Hyperconverged appliance that scales from 6 to 20 nodes.
IBM Fusion SDS: Uses OpenShift Data Foundation (ODF), which is… drum roll… Ceph packaged by Red Hat.

Yes, you read that right. IBM sells you Ceph under a different name. Red Hat packages it as ODF, IBM puts it in Fusion SDS, and you pay the “enterprise” premium.

IBM Fusion HCI performance claims are impressive on datasheets: 90x acceleration on S3 queries with local caching, performance equivalent to Databricks Photon at 60% of the cost, etc.

But here’s the question no one asks: how much of this is technology and how much is simply well-sized hardware with correct configuration?

Because GPFS/Storage Scale is an excellent parallel file system. No one disputes that. But it’s also proprietary, it ties you to IBM, and the licensing cost can be astronomical. And when IBM sells you Fusion SDS… they’re selling you Ceph under their brand name.

Red Hat: the open source middleman

Red Hat Ceph Storage continues to be the enterprise distribution of choice. They offer 36 months of production support plus an optional 24 months of extended support. The product is fine, I’m not going to lie.

But the price reflects the “Red Hat tax”. You are paying for:

Tested and certified packages
24/7 enterprise support
Predictable life cycles
OpenShift integration

Is it worth it? It depends. If your organization needs a support contract to sleep easy, probably yes. And we’d be happy to sell it to you, but if you have the technical equipment to operate Ceph upstream, you may not need it.

SUSE: the one who jumped ship

Plot twist of the year: SUSE completely exited the Ceph enterprise market. Their SUSE Enterprise Storage (SES) product reached end of support in January 2023. After acquiring Rancher Labs in 2020, they pivoted to Longhorn for Kubernetes-native storage.

If you were an SES customer, you are now an orphan. Your options are to migrate to Red Hat Ceph Storage, Canonical Charmed Ceph, community Ceph, or find a specialized partner to help you.

Pure Storage and NetApp: the other side of the ring

Pure Storage has created a category called “Unified Fast File and Object” (UFFO) with its FlashBlade family. Impressive hardware, consistent performance, premium price. Its FlashBlade//S R2 scales up to 60 PB per cluster with 150 TB DirectFlash Modules.

NetApp StorageGRID 12.0 focuses on AI with 20x throughput improvements via advanced caching and support for more than 600 billion objects in a single cluster.

Both are proprietary solutions that compete directly with Ceph RGW in the S3-compatible object storage space. The performance is excellent, but the lock-in is total.

The awkward question: do you really need to pay that premium?

This is where I put on my cynical engineer’s hat.

Ceph upstream is extremely stable. It has 20 releases under its belt. Ceph Foundation projects include IBM, Red Hat, Bloomberg, DigitalOcean, OVHcloud and dozens more. Development is active, the community is strong, and documentation is extensive.

So why would you pay IBM or Red Hat to package it?

Legitimate reasons:

Your organization requires a compliance support contract or internal policy
You do not have the technical staff to operate Ceph
You need predictable and tested upgrade cycles.
The cost of downtime is higher than the cost of the license.

Less legitimate reasons:

“It’s what everyone does.”
“No one was fired for buying IBM/Red Hat.”
“The vendor told us that open source is not supported.”
No one in your team has bothered to learn Ceph for real.

The real problem is knowledge. Ceph has a steep learning curve. Designing a cluster correctly, understanding CRUSH maps, tuning BlueStore, optimizing placement groups… this requires serious training and practical experience.

But once you have that knowledge, you have full control over your infrastructure. You don’t depend on vendor release cycles. You don’t pay for “rights to use” software that is GPL’d. You are not orphaned when the vendor decides to pivot to another market (right, SUSE?).

Manufacturer’s claims vs. reality

One of the things that irritates me most about this industry is the number of marketing claims that are accepted without question.

“Our product is 90x faster” – Compared to what? On what workload? With what competitor configuration?

“Performance equivalent to [competitor] at 60% of cost” – Does this include licensing cost? Support? Training? Additional staff?

“Enterprise-grade” – what exactly does that mean? Because Ceph upstream is also “enterprise-grade” at CERN, Bloomberg, and hundreds of telecos.

The reality is that distributed storage performance is highly dependent on:

Correct cluster design (failure domains, placement groups)
Appropriate hardware (25/100GbE network, NVMe with power-loss protection)
Operating system configuration (IOMMU disabled for peak performance, CPU governors)
Workload specific tuning (osd_memory_target, bluestore settings)

A well-designed Ceph cluster operated by people who know what they are doing can match or beat many proprietary solutions. The Clyso benchmark achieved 1 TiB/s with 68 Dell PowerEdge servers. IBM/Ceph demonstrated over 450,000 IOPS on a 4-node cluster with 24 NVMe per node.

Can you get these numbers with Ceph upstream? Yes, if you know what you’re doing.

Can you get them by buying an appliance? Also, but you will pay considerably more.

The smart move: control over your infrastructure

After 15 years in this industry, my conclusion is simple:

The smart choice for most organizations is Ceph upstream with independent specialist support and training.

Why?

Governance of foundation: Ceph is a project of the Linux Foundation with open governance. MinIO can’t happen.
Active community: Thousands of contributors, regular releases, bugs fixed quickly.
No lock-in: It’s your cluster, your configuration, your code. If tomorrow you decide to change support partner, you lose nothing.
Real TCO: The software is free. You invest in appropriate hardware and knowledge (training and technical support).
Total flexibility: You can upgrade when you want, not when the vendor releases the next packaged version.

The “but” is obvious: you need the technical knowledge to operate this. And this is where many organizations fail. They deploy Ceph without proper training, configure it wrong, have performance issues, and conclude that “Ceph doesn’t work” when the problem was the implementation.

Where to get this knowledge?

Ceph is complex. But there are clear paths:

The official documentation is extensive and much improved. The Ceph blog has excellent technical deep-dives.

Cephalocon is the annual conference where you can learn from those who operate Ceph at full scale (Bloomberg, CERN, DigitalOcean).

Structured training with hands-on labs is the most efficient way to build real competence. You don’t learn Ceph by reading slides; you learn by breaking and fixing clusters.

L3 technical support from people who live Ceph every day gets you out of trouble when things get complicated in production. Because they will. At SIXE, we’ve spent years training technical teams in Ceph and providing L3 support to organizations that have decided to take control of their storage infrastructure. Our Ceph training program covers everything from basic architecture to advanced operations with real hands-on labs. And if you already have Ceph in production and need real technical support, our specialized technical support is designed for exactly that.

In addition, we have just launched a certification program with badges in Credly so that your team can demonstrate their competencies in a tangible way. Because in this industry, “be Ceph” doesn’t mean the same thing to everyone.

Conclusions for 2026

MinIO is dead for serious use. Look for alternatives. Ceph RGW, SeaweedFS, or even the OpenMaxIO fork if you are brave.
Vendors sell you your own free software with markup. There are cases where it makes sense to pay for it. There are many more where it does not.
Ceph upstream is mature and production-ready. Bloomberg, DigitalOcean, CERN and 320+ telecoms projects can’t all be wrong.
The true cost of distributed storage is knowledge. Invest in training and quality technical support, not in software licenses that is already free.
Control over your infrastructure has value. Ask SUSE SES customers how it went when the vendor decided to drop the product.
Project governance matters as much as technology. Foundation > company with open source product.

2026 looks interesting. FastEC is going to change the erasure coding equation. AI and ML integration is going to keep pushing for more performance. And vendors are going to keep trying to convince you that you need their premium packaging… rightly or wrongly.

You decide. This is the only important thing.