I've always thought b-trees should be the starting "default" for everything, because their speed is far more consistent, you get extra features, and you don't have to make decisions in advance about allocating extra memory or worry whether the keys aren't well-distributed enough that you'll get horrible collisions. And how often do the extra few lookups for a tree materially affect the speed of your software?

Hash tables to me aren't a go-to -- they're an optimization I reach for only when I have a real performance reason where I have to use them, and my data is highly predictable so I have confidence how to set them up. But honestly, that just doesn't happen that often. I swear, I think I've used Bloom filters more often than hash tables in my life.