Storage planning has at least four ledgers: active capacity, expected growth, writes over time, and recoverable copies. One large device may satisfy the first three while failing the fourth. Keep the backup boundary visible from the first calculation.
Key takeaways
- Reserve working free space instead of planning to the label.
- Separate retained data from temporary data.
- Estimate host writes over the ownership period with clear assumptions.
- A second folder on the same device is not a separate recovery copy.
Calculate the working ceiling
Start with formatted usable capacity, not the package label. Subtract the operating environment, applications, current active data, retained project history, expected growth, and a free-space reserve. Capacity conventions and formatting overhead differ, so copy the usable figure reported by the actual system rather than relying on a generic conversion.
The reserve is operational room, not wasted capacity. Temporary exports, caches, updates, and file reorganisation can demand space before old data is removed. Pick a reserve from the workload and document it as gigabytes plus a percentage of usable capacity.
Original asset: 30-day storage diary
| Data class | Start | Added in 30 days | Deleted in 30 days | Retained net |
|---|---|---|---|---|
| Active source files | 620 GB | 148 GB | 22 GB | 126 GB |
| Project cache | 96 GB | 310 GB | 286 GB | 24 GB |
| Exports | 184 GB | 92 GB | 71 GB | 21 GB |
| Documents | 28 GB | 4 GB | 1 GB | 3 GB |
| Total | 928 GB | 554 GB | 380 GB | 174 GB |
Net growth is 174 GB per month, but host writes are at least 554 GB from newly written files. Those are different planning lines. If this month is representative, twelve-month retained growth is 174 × 12 = 2,088 GB. Three-year retained growth is 6,264 GB before project archiving or changed workload.
The diary also exposes leverage: cache writes 310 GB but retains only 24 GB. Moving or trimming cache policy may change active-capacity needs, though it does not erase the write workload. Retention should follow business, legal, contractual, and recovery requirements set by the real operator; this example invents none.
Run a sensitivity row before choosing capacity. If net retained growth falls to 110 GB per month, three-year growth is 3.96 TB. If it rises to 240 GB, the same period needs 8.64 TB before reserves. The spread is larger than many active devices. That result may favour scheduled archive movement and a written retention rule over buying all forecast capacity in the primary machine on day one.
Estimate writes without pretending to know endurance
For 554 GB per month over 36 months, host writes equal 19,944 GB, or about 19.9 TB using decimal units. Add a 25% workload uncertainty: 19.944 × 1.25 = 24.93 TB. Compare that estimate with the real device’s documented endurance terms, warranty conditions, temperature limits, and supported use. Internal write amplification and wear management are device-specific and cannot be derived from host writes alone.
Daily averages hide burst days. 554 GB over 30 days averages 18.47 GB per day, yet one ingest day might write far more. If sustained write performance matters, record the heaviest ordinary day and how long writes continue. Capacity and endurance arithmetic do not establish sustained speed after caches fill.
Draw a line around failure
A backup copy needs a separate failure boundary and a tested recovery path. A second partition, folder, or mirrored copy inside one enclosure may reduce some inconvenience but can share power, controller, user-error, theft, and enclosure risks. Define which failures each copy is meant to survive.
Work backward from recovery. List essential datasets, acceptable data loss, acceptable restore time, copy frequency, retention points, encryption responsibility, and a restore test. Storage capacity for backup must include version history and change rate, not merely one copy of the current active set. Deletion and corruption can propagate if the design offers no retained earlier state.
Use the constraint-first framework to balance storage against the whole build. Large datasets can also raise the memory working set. Connector availability and physical mounting belong in the pre-purchase compatibility gate. Calculation rules are described in proposed editorial standards.