Thanks all for your answer!
I had this impression that TLA+ spec would be way shorter than the actual program. I think that come from Newcombie's comments on this paper introducing TLA+, and it also comes from Leslie's mention on his video course that TLA+ could be short abstraction.
If I understand this correctly, Leslie's point is TLA+ may be or not be shorter than the program, since they're not comparable, but the resulting program would definitely be short than the one without TLA+ thinking applied.
I believe the above statement is true, I have heard it many times from multiple sources, the video course, the Practical TLA+ book, the papers. But I don't ever see a concrete example of how a long program is shorten and better designed by applying TLA+ abstraction. Do you mind point me to one such example ?