Swift object durability calculator

This tool can help to choose the required amount of replicas in a Swift cluster, based on the wanted object durability. If you need a lower risk to loose an object, you can increase the number of replicas for objects in your cluster. You can even use different number of replicas for different classes of data using Swift Storage Policies.

With the choosen parameters (see below), the calculated probability to loose one object of 5GB size in a year is: .

The disk failure rate per year corresponds to the MTBF of an disk. For example, a disk with a specified MTBF of 350.000 hours has a probability of nearly 2.5% to fail within one year. Most disk vendors specify a much higher MTBF, however in reality you will most likely see much higher disk failures per year.

The rebuild time depends on the amount of data stored on the disks, the disk size, network speed, partition power and total number of disks in the cluster.

The bit error rate is the probability for an unrecoverable error per bit read. This is also the reason why the probability to loose an object increases with the object size. Modern drives have a specified bit error rate as low as 1e-15 or even 1e-16.

Probability calculation

s Object size in bytes
t rebuild Rebuild time in hours
p AFR Annual Failure Rate. Probability for a disk failure in one year
p RFR = p AFR t rebuild 8760 Rebuild Failure Rate. Probability for a disk failure during rebuild
p BER Probability for an unrecoverable error per bit read
Replication (3 replicas):
p loss = ( 3 ⋅ p AFR ⋅ (1 - p AFR ) 2 ⏞ 1 out of 3 disks failed + ( 1 - p AFR ) 3 ⋅ p BER ⋅ 8 ⋅ s ⏞ Unrecoverable Read Error ) ⏟ 1st replica loss ⋅ ( 2 ⋅ p RFR ⋅ (1 - p RFR ) ⏞ 1 out of 2 disks failed + ( 1 - p RFR ) 2 ⋅ p BER ⋅ 8 ⋅ s ⏞ Unrecoverable Read Error ) ⏟ 2nd replica loss ⋅ ( p RFR + ( 1 - p RFR ) ⋅ p BER ⋅ 8 ⋅ s ⏞ Unrecoverable Read Error ) ⏟ 3rd replica loss
Erasure Code (ndata=6, nparity=3):
p loss = ( 9 ⋅ p AFR ⋅ (1 - p AFR ) 8 ⏞ 1 out of 9 disks failed + ( 1 - p AFR ) 9 ⋅ p BER ⋅ 8 ⋅ s ⏞ Unrecoverable Read Error ) ⏟ 1st fragment loss ⋅ ( 8 ⋅ p RFR ⋅ (1 - p AFR ) 7 ⏞ 1 out of 8 disks failed + ( 1 - p RFR ) 8 ⋅ p BER ⋅ 8 ⋅ s ⏞ Unrecoverable Read Error ) ⏟ 2nd fragment loss ⋅ ( 7 ⋅ p RFR ⋅ (1 - p AFR ) 6 ⏞ 1 out of 7 disks failed + ( 1 - p RFR ) 7 ⋅ p BER ⋅ 8 ⋅ s ⏞ Unrecoverable Read Error ) ⏟ 3nd fragment loss ⋅ ( 6 ⋅ p RFR ⋅ (1 - p AFR ) 5 ⏞ 1 out of 6 disks failed + ( 1 - p RFR ) 6 ⋅ p BER ⋅ 8 ⋅ s ⏞ Unrecoverable Read Error ) ⏟ 4th fragment loss