Four years later, I still have not gotten the hang of
telegraf, I'm still running my own home-made
detailed system and process monitoring
reporting to InfluxDB running container-lessly in
lexicon
and I feel the time is up for moving these services
into the Kubernetes cluster. Besides keeping them
updated, what I'm most looking forward is leveraging
the cluster's infrastructure to expose these services
(only) over HTTPS with automatically renewed SSL certs.
In fact, when I started building detailed process monitoring
I knew nothing about
telegraf,
influxdb,
grafana
or even Raspberry Pi computers.
It was back in 2017, when pondering whether to build my next
PC around an
Intel Core i7-6950X
or an
AMD Ryzen 5 1600X, that I started looking into measuring
CPU usage of a specific process. I wanted to better see and
understand whether more (but slower) CPU cores would be a
better investment than faster (but fewer) CPU cores.
At the time my PC had a
AMD Phenom II X4 965 BE C3
with 4 cores at 3.4GHz, and I had no idea how often those CPU
cores were all used to their full extent. To learn more about
the possibilities (and limitations) of fully
multi-threading CPU-bound applications, I started running
top commands in a loop and dumping lines in .csv files to
then plot charts in Google Sheets. This was very crude, but
it did show the difference between rendering a video in
Blender (not multi-threaded) compared to using the
pulverize tool to
fully multi-thread the same task:
This early ad-hoc effort resulted in a few scripts to
measure per-proccess CPU usage, overall CPU with thermals,
and even GPU usage.