I wonder if an optimizing regex-compiler could transform the set of regexes into a single finite state machine that’s more efficient than running individual regex recognizers in parallel. The optimal solution feels NP-hard, but I wonder if one could get sufficient improvements with some heuristic optimizations, like extracting common clauses, etc.
Maybe @burntsushi has some insights about this?
However, when you're talking about thousands of regexes, that's going to be a very large automaton. Probably impractically large. It's not NP-hard. It's "just" impractical.
Using heuristics to whittle down the set of possible regexes to match---likely using literals extracted from each regex---is exactly what you want to do for a problem like this. If all you have are literals, an Aho-Corasick automaton is feasible to build for 90,000 entries. Aho-Corasick, in my experience, doesn't really start to break down until you eclipse 1,000,000 entries.