Where do hard drives fail mechanically (heads, spindle, vibration, helium), and which public signals see it?

The most useful thing the opened sources say about head crashes is that they are threshold events with a mechanism behind them, not a slow decay. In the Berkeley simulation the same 2.5-inch drive survived 300 G and failed at 400 G for a 0.5 ms positive shock, and survived 1,500 G but failed at 1,600 G for a 2.0 ms negative shock; the two cases failed in different ways, one by head slap after the air bearing lost force and the other by brute inertia. The pulse width mattered because of resonance: a 1,000 Hz shock near the disk's 1,043 Hz first mode produced contact after the pulse. That is a model of one drive, so it tells you what kind of event to look for, not how many drives it has killed.

Vibration can stop a drive without breaking it. The 1.8-inch test saw throughput fall to zero from 300 to 1,000 Hz below 15 g and come back when the shaking stopped, and the authors tied the low-frequency cases to a suspension flexure mode and the high-frequency cases to the disk's resonance. For anyone reading logs, that points to timeouts or resets that clear on their own and track the physical environment. The sources do not give how common this is in racks; the Seagate wiki only says vibration may affect fly height, and counts high-fly writes as one indicator.

Helium turns a gas-tightness problem into a vibration problem. In HGST's measurements flutter excitation was far smaller in helium than in air (for example 1.09 against 20.7 for mode (0,0)), yet flutter rose quickly as air replaced helium and was near its air level after about 80% air. If that behaviour carries over to a leaking drive, the loss of margin comes early, not late, but the paper tested controlled mixtures in one drive and does not measure leak rates. Backblaze's field notes are thin: one drive out of its helium fleet showed a SMART 22 of 94 to 99 and kept working, and its helium AFR of 1.06% against 1.61% for air-filled drives at similar drive days is described by the author as a prediction, with different models and years behind it. Seagate describes its helium counter as pass or fail, so a fall in it may come as a trip rather than a trend.

Spindle faults have the plainest counters and the least public data. The documented counters are retries in the last eight spin-ups and a count of mechanical start failures, and neither source set gives a failure rate for spindle or bearing faults. The honest summary is that a mechanical fault leaves traces in a small set of counters and in timeouts, but no opened source says how many drives fail each way or how many of them warn first. The earlier notes on this site cover SMART's wider limits and the field studies that do measure drive failure rates.