Find patterns across the fleet

Compare rates, not just counts

Ten failures in one thousand units (1%) mean something very different from ten failures in twenty units (50%). A rate divides the count by the size of the group. That group size is called the denominator. Also check exposure time, meaning how long each group was observed. Groups observed for different lengths of time are not directly comparable.

Choose cohorts (groups of units) that answer one specific question. Model, firmware, operating profile, location and service history can all affect a comparison. Mixing very different equipment may hide a problem. Splitting into very small groups may make random chance look dramatic.

Decide whether you are counting affected units or repeat events. One miner that restarts ten times is one affected unit and ten restart events. Also check how each group was chosen. A pilot group of only recently repaired units starts with a different service history than the general fleet. That stays true even if both are observed in the same week.

A pattern suggests where to investigate

When faults cluster in one row, look at the conditions that row shares. The cluster does not tell you which condition is responsible. Equipment age, configuration and maintenance history can also cluster by location.

Use the pattern to narrow the investigation, and write down the other possible explanations. Fleet analysis is strongest when three things agree. They are the pattern across many units, a believable way the failure could happen, and what you observe on individual machines.

Coordinate, then verify

Give the incident one owner. Send specific questions to the network, cooling, electrical or vendor specialists. Keep one shared timeline, and do not let people make changes without coordinating. Confirm recovery in affected groups that represent the whole problem. Record any units that did not recover, instead of hiding them in an average.

Worked example: more failures, lower proportion

Over the same fictional week, Group A has 8 affected units out of 200 and Group B has 12 out of 1,200. Their affected proportions are 4% and 1%.

Group B has more affected units in total. Group A has a higher proportion of its units affected. Which number matters depends on the decision. The total tells you how much repair work to staff for. The proportion tells you where to look first for a cause. Before blaming hardware, check exposure time, profile and how units were chosen for each group.