Notes I keep while looking after a handful of machines. Mostly storage, backups, and the small mistakes that cost me a weekend.
The monthly scrub on the big pool kept getting cancelled by the nightly send job. Neither task logged a conflict, they just took turns starving each other, and the scrub was quietly restarting from zero every month.
Moving the scrub to the first Sunday and giving the send job a hard window fixed it. Eleven hours, no contention. The lesson is that two jobs which each look fine in isolation can still be wrong together, and nothing will tell you.
I do a restore drill twice a year now. Pick a random snapshot, restore it to a scratch machine, boot it, check that the database comes up and the files are the size they should be. It takes an afternoon.
The last drill found that one dataset had been excluded by a filter I wrote in 2023 and forgot about. Two and a half years of backups that did not contain the thing I most wanted backed up.
Spent a week chasing read errors that moved between drives. Swapped two disks, reseated everything twice, read a lot of SMART output that said nothing was wrong. It was a backplane cable with a cracked latch.
Now I keep a spare set of cables in the same box as the spare drives, and when errors follow a slot rather than a disk I start there.
Rotation by size rather than by day. Anything that is only useful for debugging gets a week. Anything that answers "who changed this" gets a year, compressed, on a different machine from the one that wrote it.
The point is not disk space. It is that grepping four gigabytes of noise to find one line teaches you to stop looking.
No build step, no generator, no dependencies to update. I edit the HTML, I copy it across, it is done. Every time I have used something cleverer, the clever thing broke before the writing did.