I think the bigger question is why they're using a C++ Thrift server. Haskell is good at transforming data - they should be able to use an OS socket and do the message decoding/encoding in pure Haskell.
Because it's already there and it already works just fine.
It's so frustrating to see this sort of comment on nearly every post about a language outside the mainstream. It's a religious argument and not a productive one.
Well clearly it didn't work just fine if its threading model was interacting poorly with that of their Haskell code. If there's a mismatch between those two components then that will continue to bite them in the future, and it's legitimate to ask whether moving the boundary would be a better way to solve the problem.
When they realized the bug the path forward probably seemed like patching and not a Haskell re-write, and that's a reasonable mistake to make.
They seemed to have succeeded in creating a culture of coders/hackers that were opinionated/empowered sufficiently to use their tool of choice, and to persuade others of the advantages.
This doesn't really happen in large-cap companies (though does happen to empowered teams, but even then, only somewhat - thinking finance). And speaks to advantages.