Deep Dive into Neon storage engine that enables serverless Postgres

This is a very cool overview! How much data ends up being cached in the postgres compute tier in the steady state? How much does query performance suffer after cold start?

Also the video in the post doesn't play for me (and is a bit too dense for its size), I had to download it to my computer to see it. A couple images might be more clear in conveying the key points.

Will you ever consider using EBS volumes or another mechanism which supports random writes?

Can you comment on main differences between this and Aurora?

Neon optimizes it's storage architecture to make any point-in-time lookup of pages cheap, whereas Aurora only maintains the latest version of a page. This means that Neon only stores the differences between split branches, while Aurora stores a whole copy of the data - or at least, that's what you pay for.

Other changes are that Neon still (mostly) uses Postgres' XLog storage format, whereas Aurora does not, and that the source code for Neon's storage engine is fully open source and supports PostgreSQL 15, while Aurora is closed-source and doesn't.

I’m CEO of Neon. Happy to answer questions.