Deciphering
Glyph
( )
Crouching Database, Hidden Transaction

Tue 07 October 2003

Today I actually discovered BSDDB. I can't believe I missed this. It's an efficient, in-process, open-source, simple, transactional database which does almost everything I have ever wanted a database to do. For free.

I read the transaction store documentation at sleepycat's site, and discovered that bsddb does things I hadn't even considered - for example, it is possible to transactionally move a non-database file using the bsddb API. I realize that this may not mean much to anyone but me, but I was so moved by that possibility that I nearly cried.

This discovery is both exciting and terrifying. Allen and I ran into this together, pair programming and attempting to reconcile the highly single-process logic in the current Quotient with the highly multi-process logic in the new QQ (Quotient Queues) module we're integrating. After looking at one particular function that was obviously fragile and could lose data at several points, we decided we needed to solidify and centralize our transaction processing into a single place. Since we knew that BSDDB had "some transaction support", we figured we could use it for what we needed.

We got more than we bargained for. One of the first things that we discovered that we had previously mis-read the documentation: we believed that BSDDB was single-process, based on the fact that their documentation talked about multi-reader access only being available for non-transactional data stores, and other areas referred to "threads of control".

It turns out that it's perfectly usable from multiple processes simultaneously. "Thread of control" actually means "process, thread, or other encapsulation of a program counter" the way they use it in their documentation.

So, on the one hand, it will make the incredibly arduous task of making our central data store 100% reliable much, much easier than it previously was. Rather than being an ongoing task with many threads left hanging, it will be almost completely done when we finish this refactoring. On the other, this increases the amount of changes we have to get done for this release - by saturday. This is a much larger snag than I expected to hit at this point, considering that we'd already gotten through a lot of the "hard stuff" - figuring out how to make multi-process communication both transactional and observable from the user interface. (Luckily, much of that work won't be wasted since we will be using it to manage transactions across multiple machines.)