Big Tech's quest to keep enterprises on board

So here's the thing: with the right software, storage really is just a whitebox commodity. Look at Caringo for object storage, as one example. Take your whitebox storage systems, set them to boot via PXE, and the master Caringo node simply hands out an operating system that gloms on to all the storage in the new node and adds it into the cluster. Easy peasy.

Now, Caringo is object storage. But what prevents this from being done with some sort of clustering scale-out storage solution that offers up SMB or NFS? Coho Data style tech but with a Caringo-like distribution to whitebox nodes. Nothing, really. You could even offer up iSCSI or FCoE from such a cluster.

Of course, why stop there? You can PXE deploy VMware, can't you? So you should be able to push a hypervisor with a config on it, VSAN and...oh. Hyperconverged! I hear the new Windows Server might have some laughingly horrible version of hyperconvergence built in, so if you hate yourself a lot you can do the same thing there too!

How long before Maxta figures out that this is the future and rolls together a KVM-based solution with their hypercovnerged software and gives you a self-borging HCI cluster that can also offer up SMB and NFS? Yottabyte? Cloudweavers? There are lots of players out there...someone's either done this already or will be soon. The tech exists, I've built a few versions myself.

Of course, then you have the problem of hardware. If you play this game you need to set a standardized model that you vet thoroughly, iterate once every 18 months or so as motherboards change and continually test. You need to establish redundant supply chains, buy enough through both to keep your distie accounts active and establish policies for retiring old nodes.

Contrary to what the article says, you don't need to hire extra bodies. You do need to be pee-in-jars obsessed with automated testing. Especially automated regression testing. When designing your next generation of node you need to buy between three and five different designs, varying the critical components (motherboard and HBA, if the HBA isn't on your motherboard) and you need to spend approximately three months running through all the testing. (If your test suite doesn't take three months, it isn't testing enough things.)

In today's world, your primary concerns are going to be around the HBA. Everything is about getting a good HBA that works in JBOD mode and allows great big huge queue depths. (Fortunately, most of the newer LSIs do at this point.) You need to make sure drives don't have a problem with your HBA (you can pretty much do this by checking to see if Dell has unique firmware for any of the drives they're selling with their PERCs) and you need to make sure there are no bizzarenesses with the motherboards. (Such as "you can only have one LSI card installed, or the PCI-E bus crashes due to ????????)

Every now and again you get a bug in the Intel drivers for the NICs. It's always the same bug. I'll save you a lot of time and tell you what the problem is. The problem is that if you have multiple Intel cards on a Trident-based switch where not all cards are running at the same speed (say you have links on the switch at 10GbE and one link at 1GbE) then the NIC drivers shit themselves and start passing traffic at something stupid like 200kib. Nobody seems to know why. The solution is to make all the links the same speed, or to put a "buffer" switch that isn't Trident-based in between the offending system and the Trident switch.

If you can cope with the above types of hardware issues (they should all come out in testing, if your test lab is designed to simulate all the edge cases of your production environment), you too can roll your own storage.

If you can find the software. Software that deploys via PXE and autoconfigures the nodes. If you cannot find a vendor selling you storage software that works in exactly that fashion and does precisely what you want, do not under any circumstances try to build your own storage at scale.

By "at scale" I mean more than about 5 nodes. Past 5 nodes, it's too much hassle to hand-hold storage. Trust me on this. Do not do.

On the other hand, if you have the software and the inclination towards automated testing that lets you cope with the hardware, you too can make glorious scaleable, self-healing pools of storage. Or hyperconvergence. Or both!

Enjoy!