Hacker News Reader

> The problem here would be the "filesystem" (read: the distributed service): I don't have much experience with Ceph,

I think at that scale you would want a ceph expert on staff as a full time salaried position.

For an organization that has 10PB now and can project a growth path to 15, 20, 25PB in the future, you should talk with management about creating a vacant position for that role, and filling it.

> EDIT: btw, we let the dead drives "rot": replacing them would cost more, and the failure rate is not that bad, so they stay in the machine, and we disable them in fstabs, configs, etc.

I am a huge advocate of hosting stuff yourself on bare metal you own, but this is a ridiculous statement. Any drive in that class should come with a 3 or 5 year warranty. And the manual labor and hassle time to replace one (you have hundreds of thousands of dollars of storage and no ready to go cold spares on a shelf?!?!) is infinitesimal.

OK, clarification: most of our fleet has a LOT of supermicro machines, where it's impossible to identify the drive unless by serial number. There's on UID light, the machines needs to go offline, plus some 10 screws needs to come out to open the chassis, 4 more by drives.

The amount of downtime this would generate for a single machine plus the operation cost doesn't worth the hassle unless the machine loses a significant chunk of drives.

Chassis built for mass storage usually have lever caddies backed on to hotplug SAS backplanes.

One would think indeed, but not the early FatTwins.

If the colo is far and there’s plenty of headroom, it might not justify much urgency.

That's what remote hands are for. Yes, you batch the replacements up, but this is exactly what remote hands are for.

Costs less to leave them alone, and go once a year for a trash run. Cattle not pets, no trips to the vet. Don’t waste money diagnosing / fixing.

'cattle not pets' is not a valid argument when you're the owner and operator of the bare metal hardware. Do you also recommend that ISPs not replace failed fans in core and edge routers and optical transport systems? Let things with dual power supplies run for six months on one failed power supply?

Also you've clearly never interacted with cattle, sheep, goats, llamas or alpacas, which absolutely do get things like veterinary care and vaccinations. Large animal vet is a whole specialty and they spend lots of time working on animals other than horses. No trips to the vet???

I’ve been both farmer of black angus beef cattle (on 750 acres) right down to castrating steer, and founder/owner of the world’s largest VDN (14 international data centers) right down to pulling drives.

For what it’s worth, meat packers buy dead cattle and don’t ask questions. But this is a well known metaphor, and I’m pointing out by that metaphor, no trips to the vet. As a cattle farmer, I’d argue it holds true if you’re big enough they’ve got tags not names: the vet comes to you and only if you think you’ve got a herd problem instead of an individual problem.

As for the HN angle: these are contrarian and objection-inspiring policies that let us wholesale video delivery to/through other CDNs while making a profit.

This doesn't make sense to me. I work for a CDN with tens of thousands of servers in over a hundred data centers. We are always working to improve our turnaround time on repairing servers, even though we have thousands. Hard drive failure is one of the leading causes of server failures. Dead servers means diminished capacity, and capacity is what pays our bills.

Farmers absolutely have a vet who takes care of the cattle. I am not sure what you are on about.

The whole point of cattle-vs-pets is you are supposed to treat all the servers the same, not that you have to treat all of them poorly.

Less is more.

Another contrarian view — particuarly suitable for large VDN content like media, not small CDN content like html/js — is that one doesn’t need to be in over a hundred data centers, one needs to be in the key exchanges: you don’t have to be at the ends of every spoke if you pick the right hubs.

Agree swapping bad drives is a reasonable use of smart hands when done in batches, as no diagnosis is needed. I’d advocate considering extending that practice to the servers themselves. Math works if you find local tech repo/refurb shops that take gear (w/o drives) to bulk refurb & resell. The other way is to get your OEM to provide aliveness-as-a-service.

Anything so you don’t have to do manual labor, ideally ever.

You can also get units like this direct from Western Digital/HGST. We have a system with 3 of their 4U60 units, and they weren't all that expensive. Ordering direct from HGST, we only paid a small premium on top of the cost of the SAS drives.

This is the answer that worked for us storing petabytes a decade ago.

We collaborated with OEMs and also shared/compared notes with Backblaze on rackable mass storage for commodity drives.

Backblaze published a series of iterations of designs of multi-drive chassis, and one of the OEMs would make them for other buyers as well. If you’re doing this route, read through those for considerations and lessons learned.

Performance was > 10x better than enterprise solutions. A policy to “leave dead disks dead” aka “let them rot” as said elsewhere in this thread kept maintenance cheap.

The secret sauce part making this viable for commercial online storage hosting (we hosted video) was we used disks as JBOD with an in-house meta index with P2P health awareness to place objects redundantly across disks, chassis, racks, colocation providers, and regions.

Don't buy HPE gear. Qualify the gear with sample units from a few competing vendors and you'll see why.

I’ve done qualification on hpe and various competing vendors and honestly haven’t seen dramatic differences in terms of performance and failure rates. From my experience the biggest difference was with vendor support services rather than the actual hardware. I’d be curious to hear more about your qualification experience with this particular vendor if you’d be willing.

When we set up user content storage of images and mp3s for Last.fm back in 2006ish we used MogileFS (from the bradfitz LJ perl days) running on our own hardware. 3/4/5/6u machines stuffed full of disks. I still think it's an elegant concept – easy to grok, easy to debug, easy to reason about. No special distributed filesystem to worry about.

Don't take this as an endorsement of the MogileFS perl codebase in 2021, but worth considering this style of storage system depending on your precise needs.

MinIO is an option as well and would allow you to transition from testing in S3 to your own MinIO cluster seamlessly.

I wonder if anyone can comment if they have experience running minio at scale. It would be a pleasant surprise if a “simple” minio cluster could handle such workload

Does it scale that far?

Likely not that "gracefully". But Ceph absolutely does, and has an S3 gateway.

Using Ceph like S3 could be a bit tricky if all of that 10pb is very small files.

Redhat did an interesting series of blog posts about Ceph and getting it to 1 billion objects

https://www.redhat.com/en/blog/scaling-ceph-billion-objects-...

I'd be interested to learn more about your HDFS usage and your experience at that scale. Would you be willing to have a chat? If so, my email is in my profile.