We recently got access to a 4 GB dump of real terminal sessions (all public), and we've been extracting measurements from it for Ghostty so we can optimize real world distributions of terminal escape sequences rather than synthetic benchmarks. So many positive results, read on!
On straight IO performance (the time to read and process all 4GB of data, "process" meaning parser and execute all escape sequences, this isn't plain text data!) we've improved somewhere between 20 and 30%. Ghostty was already fast, now it's even faster, and based on real sequences.
On rendering, we improved the frame build times by anywhere from 2x to 5x (I tweeted about this recently).
On correctness, there was one crash and one integrity violation (runtime safety checks in debug builds). And identified about a dozen invalid behaviors on the edges that we never heard anyone report. Those are all fixed now.
I want to also note that a majority of this work (all but the renderer work) was done by one of Ghostty's maintainers, not by me. Excellent work all around.
I believe they plan on blogging about it. If not, I might. In any case, a longer form blog about specific discoveries and improvements we targeted will follow. I just wanted to share some of the exciting news!