It turns out we can't avoid the regex hostname checks by setting a zero-width range on auto_expand_replicas, we would have to really get rid of it. That is possible, but it seems more error prone than is worthwhile as there are a variety of places that change the replica count, do some work, then change it back. Having multiple code paths through all of that seems less ideal.

Moving forward I think our solution will be partially to know that we can't have a large ban list, and that it should go away as soon as reasonable. Also potentially T399900 can monitor the situation and remind us of the above. Plausibly we could also upstream some code to improve the performance, we suspect a medium sized lru cache for compiled regex automatons could improve performance. Or even a cache over (pattern, string) -> match if it was appropriately sized would potentially be a large improvement.