Many of the hardest bugs I've seen are related to multithreading. Writing code while being mindful not to create buffer overflows is almost trivial in comparison to writing correct multithreaded code with fine-grained locking. I've experienced many cases in the past where someone thought a certain set of race conditions would "never" happen (as documented in FIXME type comments in the code) --- and it basically always does. "I didn't know another thread could be doing that..." is a common sentiment my coworkers have expressed whenever I find bugs of this nature.
And that's why my take for Hammer2 has been: sounds great, but there's not enough usage to shake out the bugs. DragonflyBSD is a 1% of the BSD community, which is like 5% of the total *nix deploys. ZFS has trillions of hours of drive-time. What does Hammer2 have? A great developer? Everyone knows one great developer isn't enough, there are bugs there and not enough usage to shake them out. He could be Actual Literal Jesus and he would still write a small minority of unpredictable, difficult-to-diagnose bugs that have to be sterilized out of the code.
(as a total nobody, on my part - it's just not good enough for even casual usage vs the competition here)
Pick whatever is more fun to use.
I don't understand the "notion" either - to me it reads that the probability is so low it happens for a pair of operations once every 500 million operations or so. If it is a race condition you might have the issue after doing just two operations.
> I am somewhat tickled by the notion that you might have a problem after deleting half a billion directory entries.
What's wrong with performing 500 million operations on a file system? There is an expectation of very high reliability when it comes to file systems.