Eight 12 TB drives give you 96 TB raw and four sensible ways to arrange them. What each level costs in capacity, what it buys in redundancy, and how long a rebuild really takes.
A storage node here is eight bays of 12 TB — 96 TB raw — with the operating system on its own 240 GB SSD, so the array holds nothing but your data. RAID 0, 5, 6 and 10 are supported, and the array is configured before handover: you tell us the level at order time.
This is how to pick a RAID level for a storage server.
| Level | Usable | Survives | Rebuild reads | Best for |
|---|---|---|---|---|
| RAID 0 | 96 TB | nothing | — | Scratch, cache, anything reproducible |
| RAID 5 | 84 TB | 1 drive | all 7 survivors | Archives you also back up elsewhere |
| RAID 6 | 72 TB | any 2 drives | all 6 survivors | The default for 12 TB drives |
| RAID 10 | 48 TB | 1 per mirror | one 12 TB mirror | Random I/O, databases, busy seedboxes |
Drive capacity is sold in decimal terabytes; your filesystem reports binary tebibytes. Nothing is missing — the units differ:
Subtract filesystem overhead on top of that, and reserve space if you run ZFS — a pool kept above roughly 80% full slows down noticeably.
A rebuild is not instant, and the array is degraded until it finishes. One 12 TB drive at a sustained 100 MB/s of rebuild throughput takes:
12,000,000 MB ÷ 100 MB/s = 120,000 s ≈ 33 hours
That is the optimistic case, on an idle array. Serving traffic while rebuilding halves it easily, which puts a real rebuild somewhere between one and three days.
Why that matters per level:
Drive datasheets quote an unrecoverable read error rate — commonly 1 in 10^14 bits for consumer drives, 1 in 10^15 for enterprise ones. A RAID 5 rebuild here reads about 84 TB, which is 6.7 × 10^14 bits. At the consumer figure that is a coin-flip-scale risk of hitting a URE mid-rebuild; at the enterprise figure it is much lower. Real drives usually do better than their spec, and a URE does not always kill the rebuild — modern controllers can drop the sector and continue. Treat it as the reason RAID 6 is the sensible default at this drive size, not as a prophecy.
Archive, backup target, cold catalogue → RAID 6. Capacity per euro matters, the data is written once and read rarely, and you want the array to survive a second failure during a three-day rebuild.
VOD origin or media library → RAID 6, plus an NVMe tier in front for the titles people actually watch. Sequential reads are what HDD arrays are good at; the hot 5% belongs on flash.
Seedbox, or anything with many small random reads and writes → RAID 10. You give up half the raw capacity and get roughly double the random write performance of a parity array, plus the fastest rebuild of the four.
Scratch space, transcode workspace, a cache you can refill → RAID 0, and only then. One drive failure loses the array. Eight drives fail eight times as often as one.
Every level here survives hardware failure. None of them survives rm -rf, a bad deploy, ransomware or a mistyped dd — all of which replicate to the parity in real time. If the data matters, it needs a copy on another machine, ideally in another location. Two nodes in two PoPs is a common layout; ask us for matched builds if you want them identical.
Not without destroying the array. Converting between levels in place is either unsupported or a multi-day operation with the data at risk throughout — plan to back up, rebuild and restore instead. This is why the level is chosen at order time.
Both work. Hardware RAID hides the array behind one block device and needs no tuning; ZFS gives you checksums, snapshots and send/receive at the cost of RAM and a controller in passthrough mode. If you want ZFS, say so at order time so the drives are not locked behind a RAID volume.
The OS lives on a separate 240 GB SSD, so an array failure never takes the system down with it — and reinstalling the OS never touches the array. If you want the boot device mirrored as well, ask for it in the quote.
On ext4 or XFS, plan to stay under about 90%. On ZFS, under about 80% — a pool past that point fragments and slows measurably. Size the array for the capacity you will have in a year, not the one you need this week.
Storage server specifications list the current builds. If you are pairing deep storage with a separate egress node, or replicating across PoPs, send a quote request with the layout you have in mind — and see how much bandwidth you actually need before you pick the port tier.
The BGP session is the easy step. What decides whether your route reaches anyone is the LOA, the IRR objects, and a ROA whose max length is not quietly authorising somebody else to originate your more-specifics.
A broken AAAA record does not take a site down, it makes it slower for everyone who has IPv6. The prefix you get, the static configuration to prefer, the three things that break it, and how to prove it works.
A stock kernel moves a few gigabits and then flattens. The bandwidth-delay arithmetic behind it, the sysctl ceilings to raise, the NIC queue work that follows, and the nginx settings that stop copying bytes.