Drive health as a percentage, never "Unknown"
Turning vendor-specific SMART counters across SATA, NVMe, eMMC and SAS into one drive-health percentage in named bands that is never "Unknown" - one formula in one place - and four corrections that stopped a warm afternoon, a bad cable, an unread SAS counter or a missing eMMC estimate being graded as failure.
Status
Shipped, 19–20 September 2026.
The problem
A buyer's first question about a second-hand machine is how much life is left in the drive. SMART data answers it, but not in a form anyone can act on: raw attribute counters, vendor-specific meanings, and different fields on SATA, NVMe, eMMC and SAS.
The contract
The operator's requirement, and the rule that shaped every decision:
A drive health figure is a percentage, in named bands, and it is never "Unknown".
If health cannot be measured, the row says what could not be measured and what to do about it — never a blank and never a guess. "Not measurable — behind a RAID/Intel RST controller" plus "set the storage mode to AHCI in the BIOS, then press Rescan" is an answer; "Unknown" is not.
What was built
One formula, in one place, with a formatter per app so the station, the asset page and the reports word the same drive identically. A drive that reads 74% on the bench must not read "Caution" in one place and "Fair" in another.
Measured during the capture — one timed smartctl JSON read, or
mmc-utils for eMMC — and shown on the asset page and spec rows,
in the kiosk, and as a Drive health column next to Battery
health in the reports.
show the captured health, not an unprivileged probe. The kiosk was re-probing drives without privilege and getting worse answers than the capture already held. Show what was actually measured.
Four corrections that define the bands
stop grading a drive Bad for a warm afternoon or a bad cable. Temperature and interface CRC errors were dragging the grade down. A drive that is warm is not a drive that is failing, and a CRC error is usually the cable.
grade a SAS disk from its own logs, and never claim a counter we did not read. SAS drives report error counters differently; the naive read produced zeroes, and a zero read as "no errors" rather than "not read".
keep an eMMC's pre-EOL warning when it reports no life estimate. eMMC exposes both a life-used estimate and a pre-EOL warning. Some report the warning and not the estimate; discarding the warning because the percentage was missing threw away the more urgent signal.
put the action in the report cell, and a SMART FAILED drive first. A report cell that says "Caution" leaves the reader to work out what to do. It carries the action. And a drive whose SMART self-assessment has failed sorts to the top, because that is the one to look at.
The rule this exercise taught
write down the one change the capture makes to a drive.
The capture is read-only with exactly one exception, and rather than leave that implicit it is documented. A tool that claims to be read-only and is not, in one place, is worse than one that says where.
bring the station check and the kiosk preview up to what the screen says. Three surfaces had drifted apart.
How it was proved
test-drive-health.py and test-drive-health.sh, driving the real formula
against fixtures from real drives, plus an on-station check so the
figure can be verified on hardware rather than only in tests.
What was deliberately not built
No prediction of remaining lifetime in days or months. The drive reports life used; converting that into a forecast would be inventing a number.
Open questions
drives hidden behind a RAID/Intel RST controller have no lsblk
entry and therefore no health. They are reported in hiddenStorage with the
reason and the fix. Whether such a machine should be certifiable at all
remains an owner decision.
A published copy. Commit references and internal identifiers have been removed and the operator is not named; the engineering, the counts and the stated limits are unchanged.