Building a New Database Management System in Academia
published on 2023/06/16
When we started the Peloton project we decided to fork Postgres and then cut out the parts that we wanted to rewrite. Postgres' code is beautiful. It's well-documented. It's portable. It's a textbook implementation of a relational DBMS. But it is a bit dated and the overall architecture has some issues. The first problem that we encountered was that we had to convert it from ANSI-95 C to modern C11 to make it work with our new storage manager. My PhD student Joy Arulraj did this with some summer interns in about a month (see his C Postgres fork on Github). We then spent another month converting its runtime architecture from a multi-process, shared-memory model to a single-process, multi-threaded model. We deemed that this was necessary to support better single-node scale-up now and eventually go distributed in the future. One surprising thing that we found was that using Postgres' WIN32 code is easier to convert to pthreads than the Linux-specific portions of the code.