LSM-tree, fractal tree, B-epsilon tree or similar ideas based stuff for on disk storage is what solves the problem. Pretty much no tradeoffs compared to classic approaches, these things are just better.
LSM trees are good for frequent updates on recent data. Frequent updates of old data are much worse on LSM trees since they trigger merges on the oldest (and typically largest) levels of the tree.
Some storage engines allow for in-place updates of records if the fields are fixed-length or if the updated variable-length value fits in the existing space. I've used some of those (e.g., MS SQL Server) to good effect in high-update scenarios.
Depends on your implementation of course, but updates shouldn't be triggering merges on the largest levels. They should be accumulated until there is enough of them to merge or there is some other reason to merge things, like scrabbing. Either way, it's still more efficient.
