Improve Transactional Performance

Question

Improve Transactional Performance

ecton opened this issue 2 years ago · comments

Changes in Nebari v0.5.3 that fixed actual ACID compliance have slowed BonsaiDb significantly. Since then, I have been working on a new storage layer that will improve BonsaiDb's transactional performance. Here is an overview of the timeline:

May 2022: Discovered File::flush doesn't call fsync, and learned about tmpfs. The summary is that after calling the correct method for fsync to happen, design decisions I had made early on caused BonsaiDb/Nebari's transactional writes to be quite slow. For light applications, the speed would still have been perfectly acceptable, but under any significant write load, the database would become a bottleneck much quicker than PostgreSQL or SQLite would.
May 2022: Updated Nebari with new transaction batching. This change significantly improved performance, but there were still two fsync operations per transaction. The only way I could see improving things would be changing my approach to how data was stored.
June 2022: While trying to measure and understand various file synchronization mechanisms performance, I discovered that SQLite on MacOS isn't actually ACID compliant.
July 2022: I wrote an overview of my goals of Sediment, a storage layer I am planning on sitting below Nebari, which BonsaiDb uses for the underlying database implementation.
August 2022: I get Sediment to the point of benchmarking, and I feel pretty good about its overall performance relative to other embedded stores. However, while preparing a new blog post, I went and did the same benchmark against PostgreSQL and discovered that PostgreSQL outperformed them all. Why? It turns out Write-ahead logging is the fastest way to get incoming writes to disk.
September 2022: I wrote my own WAL implementation, inspired in-part by sharded-log. Because I again used a new benchmarking implementation, I lost track of the performance of PostgresSQL. I knew I was outperforming sharded-log in my particular benchmark suite, and it was that progress that made me start a new blog post to let anyone following the BonsaiDb blog know what was going on.

While writing that post, I realized I needed to compare it against PostgreSQL. What I found shocked me -- PostgreSQLs much simpler single-writer-at-a-time WAL design outperformed my implementation and sharded-log significantly, even with a large number of threads all competing to write at the same time. I scrapped the blog post and began rewriting my implementation to be inspired by PostgreSQL instead.
October 2022: I finished my rewrite of OkayWAL, and I saw the mountain of work ahead of me to get everything tied back together. I was a bit burned out, and I needed a break.
December 2022: I began a rewrite of Sediment due to changing some of my goals with the format. Because the integration of a WAL made Sediment no longer a single-file database, I decided to simplify how Sediment works by utilizing multiple files.
January 2023: I published the first release of OkayWAL, and wrote a blog post introducing it. I also completed the Sediment rewrite that utilizes OkayWAL, and I'm very proud to have reached nearly 95% line coverage. And even better news, the performance is looking promising.

This ticket began as a look into improving the view indexing system, but because of the storage rewrite, it has morphed into an overall combined "refactor all the storage changes needed into one big release" ticket. The current to-do list includes:

The net result of these changes should look like this:

BonsaiDb's transactional insert performance will be competitive with PostgreSQL, thanks to utilizing a similarly-designed WAL.
BonsaiDb's databases will no longer require a compact operation to reclaim unused disk space, thanks to Nebari's new underlying format Sediment.
BonsaiDb's view system will require fewer internal trees and offer more efficient reduce operations.
BonsaiDb's document storage will be designed more efficient storage and much faster revision history retrieval.
BonsaiDb will be better equipped if the storage format needs to change in the future again.