Re: how to build a database from scratch

Volker Hetzer wrote:
> That may be, but when starting from scratch, one has to start somewhere.
> And while ACID is important, before you can guarantee data integrity
> you have to store the data first, in memory or wherever. So, IMHO
> the data access comes first, then joins, then (with going multiuser)
> locking, then everything else.

  The fact that you believe it is possible to retro-fit recovery and support for concurrent operations to a simple B-Tree implementation reveals how little you know about how complex recovery and concurrency are to implement.

  Locking is the very tiniest of your challenges.

  Consider how your B-tree implementation copes with:

1. The computer crashing in the middle of a page split.
2. The computer crashing in the middle of a page I/O.
3. How to reverse an update operation (how to reverse an update that caused a split).
4. How to manage two writers who's insertions split the same internal page.
5. What does a reader see when they access a page a writer is modifying?

   Never mind interactions with the log, the back-up and recovery system, the non-index data storage, buffers, and what not. Some of these problems have simple solutions. Most don't. Solving properly means modifying the basic B-Tree algorithms in profound ways. Received on Tue Dec 05 2006 - 15:43:02 CST