Default to automation
The evaluation of my thesis involved comparing my own Kubernetes CNI to several other CNIs with benchmarking tools like iperf3, wrk2 and netperf.
My test bed consisted of two directly connected cloud machines. On each machine I had to perform the following: (i) install my custom kernel, (ii) install Kubernetes (master/agent) and any dependencies (iii) configure a particular CNI. After doing this, I had to (i) deploy a particular set of pods, (ii) exec into those pods and run a particular script on each (iii) copy the results over to my host machine.
This was a super tedious process that would end up taking at least an hour for each CNI I was evaluating, not including any time spent painfully debugging some issue caused by my inevitable messing up of something somewhere.
I was naive enough to believe that I would only have to go through this process once. Of course, I actually ended up doing this around 10-20 times.
Because I needed to manually run some commands every 30-60 minutes to kick off the next experiment or setup another CNI, I ended up just sitting around watching Netflix or reading my book whilst waiting for each experiment to finish. At the time I told myself "I'm exhausted and this way I can relax a bit whilst still getting the results". But it wasn't relaxing at all: my brain was constantly switched on, and I was constantly checking on the experiment to make sure the results were inline with my expectations.
Looking back, this was the worst thing I could've done: I wasn't really doing anything "productive" in that time and I also wasn't relaxing. I probably lost between 50 and 100 hours doing this.
Ideally, I would've automated this process to require a single push of a button and then for all the results to be saved on my host machine automatically. This way, I could've gone and done something actually relaxing for a few hours (like playing golf or tennis). Or, I could've let the experiment run in the background whilst I work on another part of the project (or a completely different project).
The other big benefit is that when you want to change your experiment, you can do so very quickly and without much hassle. Near the deadline, we realised it would be nice to widen the scope of one of the benchmarks but couldn't because it would take too long to set everything up. If this was automated, we probably would have been able to.
I think the "idea" here is super obvious, but I think the challenge is being able to have the foresight to realise when to automate something and stop yourself from taking the path that is slightly easier in the short-term but much more tedious in the long term.
Going forward, I'm going to try to "default towards automation" rather than "default towards manual work" and see where that gets me. Plus, with ChatGPT and vibe-coding automating things has never been easier.