I sometimes think that we are optimizing our tools and our process using the wrong objective function. We put a lot of effort into reducing overall build times

I sometimes think that we are optimizing our tools and our process using the wrong objective function. We put a lot of effort into reducing overall build times, but I would argue that the average time to find an error (if there is one) is more important than the average time to get confidence that the change is good. In my custom build tool-chain I try to compile and run the unit tests first -- even running the static analysis unit checkout -- before I compile the whole application. NASA have a similar (but more refined) approach -- they test each change in widening ripples through the dependency graph.