starting...
loading...
running...
How to use
The rack in the 3D view is a live random experiment: every drive has a random lifetime, a failed sled slides out and blinks amber while it rebuilds, and if the redundancy runs out the whole array turns red (data lost) and a fresh rack takes its place. Real arrays almost never lose data, so use the Time-lapse preset or the speed slider to see it happen, and compare the observed loss rate with the Markov prediction. Try RAID 5 with 18 TB consumer drives, then switch to RAID 6.
The science
Failure model. A drive with annual failure probability AFR fails at a constant rate λ = −ln(1 − AFR) / 8766 h. A failed drive is rebuilt in an exponentially distributed time with mean MTTR. Mean time to data loss (MTTDL) is found from a continuous-time Markov chain whose states count degraded drives (or degraded mirror pairs); solving (−Q) t = 1 gives the mean time until the absorbing "data lost" state. The loss probability by year t is read from the matrix exponential exp(Q t).
RAID 5: MTTDL ≈ MTBF² / (N (N−1) MTTR) RAID 6: ≈ 2 MTBF³ / (N (N−1)(N−2) MTTR²) RAID 10: ≈ MTBF² / (N MTTR)
Read errors. Rebuilding RAID 5 means reading every surviving drive in full. With an unrecoverable read error rate b per bit, the chance of at least one error is 1 − exp(−b · bits read). For seven 18 TB consumer drives that is about 1e15 bits and ten expected errors, which is why large single-parity arrays are discouraged. RAID 6 can still repair one such error because a second parity remains.
Why more parity helps so much. Each extra layer of redundancy multiplies MTTDL by roughly MTBF / (N MTTR), a factor of thousands for real drives. Shortening the rebuild window helps quadratically for RAID 5 and RAID 10 and cubically for RAID 6, which is why a hot spare matters.
Checks. The Markov results are compared with the textbook closed forms in their valid regime, and with a seeded event-by-event Monte Carlo of 3000 rack lives in an accelerated stress case (the same random-drive code that drives the 3D rack). The last check confirms that the chance of loss within one MTTDL is 1 − 1/e.
Simplifications. Drive failures are independent and constant-rate: real drives show infant mortality and wear-out, batches fail together, and heat or a power surge hits many drives at once. The rebuild time is the same for every drive and does not slow down under load. Controller, software and human errors, and RAID write holes, are not modelled. RAID is not a backup: it protects against drive failure, not against deletion, ransomware or fire, so a copy kept elsewhere (the vault door beside the rack) is still needed.