The pre-recorded video of this presentation will become available within the next few hours.
There was a time when the world could afford only a handful of open source relational database management systems. You could have any database, as long as it was MySQL or Postgres. These systems took 30 years to develop and test, and enjoyed widespread adoption.
And then something happened. It almost passed us by. 2018 drew a line in the sand, and marked the end of an era.
Darkest Before the Dawn
The foundations of database design were rocked both by the discovery of fsyncgate in the Postgres community, and the “slow train coming” of storage fault research from UW-Madison. Not only was the Linux kernel page cache no longer trustworthy, but the write-ahead log designs of most databases (take your pick!) were found, at least in the research, to be fundamentally broken, even to the extent of a single disk sector fault cascading into global cluster data loss.
Mission-critical applications were moving to proprietary cloud databases, and our favorite open source databases were left behind, stuck in the twilight of single-node availability and manual failover. We were yet to face the fallout. And then came the dawning of a new era.
First Gleam of Dawn
The first ray of light was FoundationDB, already pioneering the deterministic simulation testing of distributed databases, to inject faults, to find and replay rare bugs for increased developer velocity, and to test in months what had before taken decades.
Next, Linux underwent a revolution in asynchronous I/O, saved from the complexity of kernel-bypass techniques such as SPDK and DPDK by Jens Axboe's io_uring.
Full Light of Day
Finally, safer languages such as Rust and Zig rose up to power the next 30 years of systems, making it easier to write correct code, bringing gamechangers like comptime, and reviving the lost art of static memory allocation, for predictable performance in the extreme.
In this talk, we will look at these pivotal moments, and how they influenced our design decisions for TigerBeetle, the distributed financial accounting database for mission-critical safety and performance, developed under the Apache 2.0 open source license.
Taken together, these advances in database design will unlock an abundance of new open source database management systems, tailored to their domain. The best is yet to come.