When a drive in your array dies on Friday evening, the rebuild starts when a replacement drive is in your hand - not a moment before. The cheapest insurance in all of server hardware is the spare that is already in the building.
Every RAID level's protection story has the same hidden clause: once you replace the failed drive. Until a replacement is physically installed, a RAID 5 array runs with zero remaining protection, a RAID 6 or 10 array runs with half its margin gone, and the surviving drives work harder than at any other time in their lives - which is exactly when a second, tired drive decides to join the first. The gap between "drive failed" and "rebuild finished" is the most dangerous window in a server's life, and most of that window, in most offices, is spent waiting for a courier. This page is about deleting that wait: buy the spare with the array, match it correctly, label it, and keep it honest.
A failed drive is not itself the disaster - RAID was built for that day. The disaster is the second failure arriving while the array is still degraded, and probability is not on your side in that window. The surviving drives are usually the same age, from the same batch, having lived the same heat and workload as the one that just died - the failure was information about them, too. Meanwhile a rebuild, once started, reads every surviving drive end to end under full load, the hardest work of their lives. That is why the rebuild window is measured in worry: with large modern drives it runs many hours or days even when you start immediately. Every day spent sourcing a replacement is a day added to the front of that window, with no protection at all in the RAID 5 case. Shrinking the window is the whole game, and it is won in the purchasing step, not the emergency.
The rules for the shelf spare are few and strict:
Cost honesty: one extra drive adds a modest line to the server budget. Price it against four days of degraded running plus express-shipping panic, and it is the easiest approval in the whole purchase.
A cold spare sits on the shelf until a human swaps it in. A hot spare sits installed in a spare bay, powered, and the controller starts rebuilding onto it automatically the moment a member fails - no human, no wait, even at 3 AM on a holiday. If your chassis has a free bay and the controller supports it, the hot spare is the stronger design; the shelf spare is the fallback that still beats any courier. Either way: label it. Write on the drive and its box which server and array it belongs to, its model, and the date bought. In an emergency, an unlabelled drive invites the one catastrophic mistake - pulling or overwriting the wrong disk. And a rule with no exceptions: the shelf spare is not a general-purpose drive. The day someone borrows it "temporarily" for a desktop is the day it stops existing.
Two pieces of maintenance keep the insurance real. First, a spare that sat for two years should be proven before you trust it: check it once when bought (a full surface scan or at least a SMART long test on a spare machine or dock) and glance at it during yearly maintenance - an untested spare can be dead on the shelf, which you would otherwise discover mid-emergency. Second, when the spare is used, it stops being a spare: the array is whole again but the shelf is empty, and the failed drive's RMA replacement becomes the new spare when it arrives - or order a fresh one the same week. Put the spare on the paper plan with its serial number, and put a line in the maintenance calendar: "spare present and tested". Five minutes, once a year, for the whole benefit.
Spare planning sits inside the full hardware tour - watch the complete video before you install your server.
Watch Full VideoBecause the array runs unprotected (RAID 5) or half-protected (RAID 6/10) until the rebuild finishes, and sourcing a drive typically adds days of courier time to the front of that window. The surviving drives are the same age and batch as the one that failed, and the rebuild is the hardest work of their lives. A shelf spare starts the rebuild the same evening.
The same model as the array members where possible - equal or slightly larger true capacity, same interface and form factor, CMR recording, RAID-rated with an error-recovery limit (TLER/ERC/CCTL), and ideally from a different manufacturing batch than the array drives. Label it with the server and array it belongs to, and test it once when bought.
Yes, where the chassis and controller support it: a hot spare is installed and powered, and the controller begins rebuilding onto it automatically when a member fails - no waiting for a human at all. A labelled, tested shelf spare is still far better than ordering under pressure, which is the only real alternative.