That is very cool! When you end up doing reads, do you compute all the pending writes along the path from root to leaf, and then run queries on the projection?

For index-seeks, yes.

For key-seeks, it's a normal tree walk to the leaf (unless an INSERT is found at a higher level, iirc)

In practice, the blocks are so huge that any tree is going to have at most depth 3, and the root page is likely in cache, so 1 or 2 s3 fetches (= roughly 200ms, let's say)

I've been toying with the idea of putting the root page (or maybe the top 2 levels of pages) in DynamoDB, which would make them fast even if they weren't in cache.