Choosing Storage Tiers Without Overbuying
Flash for everything is a simple policy and an expensive one. A four-tier model, and the questions that place data into each.
The honest reason many organisations run all-flash for everything is that tiering requires knowing your data, and knowing your data is work. Flash for everything is a defensible policy — it removes a whole class of performance problem — but it is not free, and for estates with meaningful volumes of cold data it is often two to four times the cost of a tiered design that performs identically where performance is observable.
Tiering is worth the effort above roughly fifty terabytes, or wherever archive obligations run longer than five years. Below that the analysis costs more than it saves.
Four tiers, defined by access pattern rather than by medium
The useful distinction is not disk versus flash versus tape. It is how the data is accessed.
Tier 0 — latency-critical. Data where a millisecond is visible to a user or a transaction. Database indexes, transaction logs, virtual machine boot volumes, active working sets. This is NVMe flash territory and there is no sensible substitute.
Tier 1 — throughput-critical. Data read and written in bulk where sustained bandwidth matters more than per-operation latency: analytics inputs, media, build artefacts, active file shares. SAS or SATA flash serves this well; the endurance rating matters more than the latency figure.
Tier 2 — capacity. Data accessed occasionally, where retrieval within seconds is acceptable. Completed projects, older mailbox archives, secondary backup copies. High-capacity spinning disk remains substantially cheaper per terabyte, and for this access pattern the latency difference is invisible.
Tier 3 — retention. Data kept because it must be kept, accessed rarely or never, where retrieval within hours is acceptable. This is where tape remains genuinely unbeaten on cost per terabyte, on shelf life, and on the one property no online system provides.
The case for tape that has nothing to do with cost
Tape's cost advantage is real but well understood. Its more important property is that a cartridge on a shelf is not addressable by anything running on your network. Ransomware that reaches a backup server can encrypt every online and cloud-replicated copy it can authenticate to. It cannot reach a cartridge in a rack forty metres away.
That air gap is a structural property, not a configuration setting, which means it cannot be misconfigured away. Every serious retention design should include at least one copy that is offline in this physical sense. Tape is the most economical way to achieve it, and modern LTO generations hold enough per cartridge that the media handling burden is far lower than the format's reputation suggests.
The questions that place a dataset in a tier
Four questions, in order:
When was it last read? Not written — read. Most tiering analysis over-weights write recency, and write-once-read-never data is extremely common. If nothing has read a dataset in ninety days, it is not tier 0 whatever its owner believes.
What happens if retrieval takes an hour? If the answer is "nothing much", the data belongs in retention. If the answer is a specific operational or contractual consequence, name it and price it — that consequence is the budget for keeping the data hotter.
How long must it survive? Retention obligations frequently outlast the hardware, the format, and occasionally the vendor. Data that must survive fifteen years needs a format and a migration plan, not simply a large drive.
Can it be regenerated? Derived data — caches, rendered outputs, rebuildable indexes — often needs no retention tier at all. The cost of regenerating it is frequently lower than the cost of storing it for a decade.
Where tiering goes wrong
Two failure modes account for most disappointments. The first is tiering by file age alone, which reliably demotes the annual report that is read once a year and urgently. Access frequency over a rolling window is a better signal than age, and a small allow-list of never-demote datasets handles the rest.
The second is failing to test retrieval from the coldest tier. A retention tier nobody has restored from is a hypothesis, not a backup. Restore something real from tier 3 twice a year, time it, and record the result — including the tape drive's own condition, since drives are consumable and a drive that has not been exercised or cleaned is a single point of failure in front of every cartridge you own.
The pragmatic version
If a full analysis is out of reach, a three-line policy captures most of the benefit: active working set on flash, everything else on capacity disk, and one offline copy of anything with a retention obligation. That is not an optimal design, but it is a substantially cheaper one than flash-for-everything and it is meaningfully safer than online-only.