Work

Fourteen years, four places, no company names: two of them asked, and the other two I would rather not advertise either way.

2023 - now

On my own, from Tartu.

Measurement work for small operators: idle and low-duty power, scheduler and idle-state behaviour, acceptance tests for new hardware batches. A job runs two to six weeks. What the client gets is a short report with the numbers in it, the scripts that produced those numbers, and one afternoon of me explaining both. Flat rate, €1 900 a week, and I do not take anything shorter than two weeks because the first week always goes on learning the fleet.

2021 - 2023

Operations, a hosting provider, 912 machines.

I owned the fleet’s idle power budget, which nobody had owned before and which turned out to be the cheapest line on the electricity bill to move. Changing the default frequency governor across the 2019-era batch dropped median idle draw by 4.8 W a node - about 4.4 kW across the fleet, and roughly the same again in the cooling that no longer had to remove it. I also wrote the twenty-minute acceptance check that every new batch goes through before it gets a hostname: thermals, idle-state residency, one disk, one network card, and a deliberate yank of the second power cord.

The less pleasant half of those two years was learning how often a «dead» machine is a perfectly healthy machine with a counter being read wrong.

2016 - 2021

Build and release, an outsourcing development shop, about forty people.

I was the person who made the test suite stop taking 51 minutes. It now takes 12, and almost none of that came from clever parallelism - it came from deleting three sleeps, caching one dependency set properly, and refusing to let anybody add a step that polls. I ran the on-call rota for the last two years and wrote the handover notes that are, I am told, still in use.

2012 - 2016

Network operations, a regional telecom, around thirty thousand subscribers.

Night shifts, a wall of graphs, and the single most useful habit I have: read counters as differences, never as totals. A drop counter showing 630 000 means nothing at all until you know whether it moved in the last ten minutes. I have since watched that one mistake produce four confident, wrong diagnoses, two of them mine.

What a job looks like

1. Week one: baseline. I measure what you have, at the wall and in the kernel, and I write down the machine’s exact state while I do it. No changes.
2. Week two: one change at a time. Each change gets its own window, its own five repeats and its own median. If a change is worth less than the measurement noise, it goes in the report as worth nothing.
3. Last day: handover. The scripts, the raw samples, and a page that says what to re-run in six months to check whether it stuck.

I work with read access and a terminal on one machine of your choosing. I do not need, and will not ask for, anything else.

What I turn down

• Machines I am not allowed to touch at all. Looking at graphs from a distance is somebody else’s trade.
• «Make it faster» with no number attached. Bring a number, even a wrong one, and we have something to argue about.
• Jobs whose real output is a dashboard. I can build you one, badly, and you will not look at it twice.
• Anything that has to be finished before I have seen the hardware.

If none of that rules you out, the contact page says what to put in the first mail. Current availability is on the now page.