10 Blog

Founder Log 1: Warnings 54 → 0, CPU About Halved

Two weeks of backend work on Hetk: build warnings went to 0, and every new one now stops the build. Server CPU fell by about half, and calendar reads got faster.

By Andrei Reinus

Founder log is a short note on what I worked on, with the numbers behind it. This is the first one. It covers two things in Hetk’s backend, the part that runs your syncs: the build warnings, and the server’s CPU.

Warnings: 54 → 0

A warning is the compiler pointing at code that works but could hide a mistake. Most of Hetk’s backend warnings were fixed in code; the last 54, in payment and calendar-writing code, were reviewed one by one and kept on purpose, so none are open now. Any new warning now stops the build, so a possible mistake gets looked at before it can reach your calendar.

CPU: about half

In early October I measured where the server spends its time. Most of it went to two jobs: reading your calendars from Google, Microsoft and Apple, and reading Hetk’s own record of every copy it has made, at the start of each sync.

The fixes:

  • Microsoft calendars are read in pages of 200, as compressed responses, and parsed with less work.
  • Connections to Google are reused, and Microsoft calls moved to HTTP/2.
  • Hetk’s record of copies is now stored grouped by user, so one user’s rows come back in one read.
  • A sync no longer counts those rows before it reads them.
Before (9–10 Oct)After (16–17 Oct)
Server CPU, hourly average17.6%9.2%
Server CPU, busiest hour80%66%
Google calendar read, average4.2 s1.7 s
Microsoft calendar read, average11.6 s8.0 s
Memory clean-ups per hour (.NET gen 0)2,942362

The database also spent 70% less CPU on reading the record of copies.

Average CPU fell by 48%. That is the “about halved” in the title, and it needs two caveats. It is one night compared with one night, not a month. And fewer syncs ran in the second window (4,112 against 5,045), so part of the drop is less work, not cheaper work. Per sync, the saving is closer to a third.

What it means for you

A Google calendar read now takes less than half as long, and a Microsoft one about a third less. That leaves room to grow on the same server.

Most of this work was done by AI coding agents, with me reading the results and making the calls. I wrote about where that line sits in We Build Hetk With AI. We Don’t Run Hetk With AI.