HOME INFRASTRUCTURE · MEASURED
The unglamorous half of running your own server
A cold spare is the machine on the shelf, unplugged, for the day the running one dies. This is about that half: the backups, the restores, and the settings that decide whether a power cut costs you an afternoon or a year of photographs. Every number here comes out of a script that is published next to it.
360x the time, 6.8% more saved
What the slowest backup setting buys over the fast one, on data a home server actually holds.
A factor of ninety, off by default
On macOS, asking the drive to genuinely flush costs that much. SQLite does not ask unless you tell it to.
Every script published
Corpus builder, benchmark, chart generator, raw JSON. Rerun it and disagree with the data.
What has been measured so far
Two questions, settled by running the thing. Each page carries the question, the script, the machine, and the raw output.
Backup compression
Thirteen settings across five kinds of data on a 282 MB corpus you can rebuild from public URLs. Photographs and video compress by nothing at all, which changes the whole argument — and the restore column, which nobody prints, is wider than the ratio column.
Read the measurement →SQLite durability
Home Assistant, Immich, Paperless and Miniflux all keep state in SQLite and all pick journal_mode and synchronous for you. Thirty-six configurations measured, including the macOS flag that makes synchronous=FULL mean what its name implies — at ninety times the cost.
Backup window calculator
Enter your data mix, your window and your link speed. It applies the measured per-type ratio and throughput to your numbers and tells you which setting finishes before morning, and whether the CPU or the network is what you are waiting for.
Open the calculator →A few results, without the article around them
Measured on an Apple M4 with 10 cores. Absolute figures move with the hardware; the orderings should not. Raw output and scripts.
METHOD
A measurement you cannot rerun is an opinion with a decimal point
The corpus builder, the benchmarks, the chart generator and every line of JSON output are published under CC0. If your hardware disagrees with these numbers, that result is worth more than another one that agrees.
Two things worth knowing before you trust any of this
Both are set out in full on the about page, which is where a site should put the parts that reflect badly on it.