What shingled recording actually is
A conventional (CMR) drive writes each track next to its neighbour with a small guard gap, so any track can be rewritten in place at any time. SMR removes the gaps: tracks overlap like roof shingles, because the write head is physically wider than the read head. Reads still work fine — the readable strip of each track is untouched — but a write to one track now clobbers part of the track under it. So an SMR drive cannot rewrite a single sector. It has to read the whole overlapping band (tens to hundreds of megabytes), modify it, and write the band back out in order.
Vendors do this because overlapping tracks buy roughly 10–20% more capacity from the same platters. On the drive-managed SMR (“DM-SMR”) units sold to consumers, firmware hides the band mechanics behind a normal block interface. Incoming random writes land in a small conventional cache region first, and the drive destages them into shingled bands during idle time. As long as your workload is bursty, you never notice.
Why it collapses under sustained writes
The problem is the word idle. Keep writing and the CMR cache region fills; the drive must then do read-modify-write on entire bands while still accepting your data. Throughput falls off a cliff — from 150–200 MB/s to single-digit MB/s is normal, with individual commands stalling for tens of seconds while the drive shuffles bands. Benchmarks that run for a minute look fine. Workloads that run for a day do not.
A ZFS resilver is close to the worst case imaginable: many hours of sustained, partly random writes across the whole surface of the new drive, with the array timing every response. In 2020, users rebuilding onto the DM-SMR WD Reds watched resilvers projected in hours stretch into a week, and worse: under sustained load some of those drives returned IDNF (sector ID not found) errors after long internal stalls, at which point ZFS reasonably concluded the drive was dying and faulted it out of the pool — mid-rebuild, on a degraded array. That failure mode is why the rule below has no exceptions for RAID and ZFS. Hardware RAID controllers are even less patient: many drop any drive that takes more than about eight seconds to answer.
The 2020 controversy: SMR hidden in NAS lines
None of this would have been a scandal if the drives had been labelled. They were not. In early 2020, users and then journalists (Blocks & Files broke it) established that Western Digital had quietly moved the 2–6 TB WD Red — a product line marketed specifically for NAS and RAID use — to DM-SMR, without a word on the datasheet. The affected models were WD20EFAX, WD30EFAX, WD40EFAX and WD60EFAX. Seagate and Toshiba were then found to have undisclosed DM-SMR in desktop lines (Barracuda Compute, P300), though not in their NAS-branded drives.
The fallout mostly fixed the disclosure problem: all three vendors published lists of which models use which recording technology, WD split the line into Red (still SMR) and Red Plus / Red Pro (CMR), and WD settled a US class action over the affected Reds in 2021. The lasting lesson is not “WD bad” — it is that recording technology is a spec vendors will quietly change within a product line, at the same model name and price, so you verify the exact model number, not the brand or the marketing tier.
How to verify a specific model
Since 2020 the datasheet is the ground truth: every current WD, Seagate and Toshiba datasheet carries a “recording technology” line reading CMR or SMR. Pull the PDF for the exact model number on the listing — not the product family page — and read that line. If a datasheet does not state the recording technology at all, treat the drive as SMR and move on.
Some patterns that speed up triage:
- WD Red 2–6 TB: EFAX suffix is SMR; the older EFRX and the Red Plus EFZX are CMR.
- Enterprise and datacentre families — Seagate Exos, WD/HGST Ultrastar, Toshiba MG — are CMR across the board, except SKUs explicitly sold as host-managed SMR for hyperscalers (e.g. the archive-class Exos M SMR variants). You will not buy those by accident; they need OS support to work at all.
- 2.5″ drives above 1 TB are almost universally SMR. Assume SMR until proven otherwise.
- Desktop bargain lines (Barracuda Compute, WD Blue, Toshiba P300) mix both; the capacity point matters, so check every time.
The buying rule
If the drive will ever sit in a RAID group, a ZFS pool, or anything else that resilvers: CMR only, no exceptions. SMR is acceptable for exactly one job — a single drive holding sequential, write-once cold data (backup target you fill slowly, media archive), where the extra terabyte per spindle is the whole point. And never buy any drive whose recording technology you could not confirm; unstated means SMR, priced as CMR.
Every drive in our SATA and SAS rankings is an enterprise CMR model — that is a deliberate filter, not an accident of what is cheap. If you are sizing an array, the RAID planner will show you resilver time and rebuild risk with real CMR throughput numbers — the numbers SMR quietly takes away from you.