One of my biggest headaches is that the pipeline cache never seemed to work predictably (e.g. not pulling cached data that was available, it uses zip so the permissions are messed up, etc), so I ended up writing my own caching system. It was important for my use case because full runs could take 2 to 3 hours.

Then there is the YAML config which is too convoluted. It took me several attempts to get things right.

What works better, IMO, is to have each CI step as a script or a Make target that can run anywhere (CI or dev PC) and in the CI configuration have a single line in each step that invokes the script.