# Your harddrive is probably full

A short July 2026 post from the author of Marginalia Search, starting from his own `df` output and ending somewhere else entirely. His root drive has 17 GB free out of 0.47 TB usable, which is 3%. The 12 TB drive alongside it has 140 GB free, which sounds like a lot and is about 1%.

He guesses yours is full too, and checked: an informal Mastodon poll of 81 respondents put roughly half of them above 75% full. That is a self-selected sample from one person's followers and he presents it as such, not as a measurement. The stronger evidence is his own history — he started around 80 MB in the '90s, has doubled his storage many times since, is now at tens of terabytes, and they are still mostly occupied. Every scale of drive he has ever owned has been the kind of full where you have to go find things to delete. Cleanup software exists, and has existed for a long time, which is its own quiet proof that this is a problem people actually have.

## Two explanations, one of them enough

The entropic argument comes first: there are far more ways for a drive to be full than empty, so if you make roughly random changes of state without thinking about storage, you drift toward full. He accepts this and says it is incomplete.

The part it misses is timing. A full disk is not a problem until it is too full to accept the next thing. By the time that happens the disk is so cluttered that adjudicating every file on it is more work than anyone will do, so you clean up just enough to buy time and go back to what you were doing. The cycle never resolves because the pain arrives only at the point where fixing it properly has become expensive.

He generalizes this without stretching it. Software doesn't get optimized until it is too slow, so software tends to be somewhat slow. Technical debt accumulates until working in the code hurts enough to force a refactor, or in less experienced hands a rewrite, so most codebases are messy. Road networks get expanded once they are unbearably congested. Diets get attention once the pants stop fitting. And the personal one: he went from a full-time job to self-employment with nobody to assign him work, and feels just as busy. His schedule is another drive that stays full.

## The reconciliation

That leaves a contradiction, since waiting until the pain point makes the fix much larger, and yet premature optimization is a thing people are right to warn against. His proposed way out is to weaponize [Jevons paradox](https://en.wikipedia.org/wiki/Jevons_paradox) — impose a real constraint, deliberately smaller than the resources you actually have, and optimize against that instead. In personal finance this is uncontroversial and has a name, a budget. Elsewhere it keeps surprising people.

The two concrete versions are the useful part:

> Let me tell you, if you deploy and dogfood your software on a raspberry pi, then it will be fast on a Threadripper too. If you can navigate a codebase in vim in an 80x25 terminal, then you can navigate it in a powerful modern IDE as well.

The closing line is the one to keep: resources and capabilities feel like they let you do more, but often they let you do the same thing at greater cost.

## Related

This is the philosophy behind [[marginalia-systemd-migration]], written by the same author three days earlier — a search engine running on one box with two CPUs and their own RAM banks, where NUMA placement matters because the index processes are bandwidth-bound. Dropping docker compose for systemd was partly a constraint chosen on purpose rather than a limitation endured, and the pattern of running production somewhere tight is exactly what this post argues for.

The constraint is also becoming involuntary. [[hold-on-to-your-hardware]] documents the supply side: AI data centers absorbing DRAM, NAND and HDD output, Western Digital's and Kioxia's 2026 production sold out, consumer buyers being the segment that gets cut first. The usual answer to a full drive — buy a bigger one — is getting harder and more expensive, which makes the budget argument less of a discipline exercise than it was when the post was written.