by Luis Eduardo Muñoz Guerrero
ABSTRACT
Archives with a fixed replacement budget must choose which drives to replace. We compare a greedy risk
density policy with an age-based policy out of sample, treating budgeted replacement as a 0/1 knapsack
problem. Backblaze drive statistics for two quarters of 2019 were used. The first quarter, with 115,231 distinct
drives, estimated failure rates per model; 106,218 drives remained in the analysis population. The second
quarter, sealed until the plan was frozen, gave the outcome. The budget was 5% of capacity. The greedy policy
captured 2.18 failures per 1000 terabytes (TB), against 1.21 for the age policy: a difference of 0.97 failures per
1000 TB (95% interval 0.43 to 1.53), positive under the rule fixed in advance. Post hoc, without the half-terabyte
or the six-terabyte class the difference is inconclusive. On first-quarter risks, the greedy value matched the
exact optimum at four of five budgets and was within 0.003% at the fifth. On one operator’s data, out-of
sample selection by estimated risk per terabyte captured more failures per terabyte than age-based replacement
at the primary budget; post hoc, the advantage depends on the mix of drive classes and was not detected against
ranking by model rate.
![]()


