An engineer I work with asked me for tips on what to read about database benchmarking. I told him I've learned a lot from reading Mark Callaghan's blog. Now that I think about it, articles and conference talks from Baron Scwhartz were also, or even more, fundamental early on when I was getting started.

I overheard - over-read, really - an internet discussion about database storage engines. The discussion was about what functionality is considered part of a storage engine, and what functionality is in the common parts of the database server. My first reaction was something like "how isn't this obvious?" Then I realized for a lot of the database functionality it isn't obvious at all and the answer really is that it could be either way.

We are pleased to announce that our paper Automated System Performance Testing at MongoDB will be presented at DBTest 2020. (A workshop in conjunction with SIGMOD/PODS.) It is available today on Arxiv.org.

This paper presents the framework we developed to do automated performance testing on realistic MongoDB clusters. We have used and evolved this system to run hundreds of benchmarks every day as part of our Continuous Integration system. In conjunction with publishing this paper, we have also finally open sourced the Python code to this framework. The framework is called Distributed Systems Infrastructure 2.0, or DSI for short.

At work we were discussing whether MongoDB will retry operations in some circumstances or whether the client needs to be prepared to do so. After a while we realized different participants in the discussion were discussing different retries.

So I sat down to get to the bottom of all the retries that can happen in MongoDB, and write a blog post about them. But after googling a bit it turns out someone has already written that blog post, so this will be a short post for me linking to other posts.

Retries by the driver

If you set retryWrites=true in your MongoDB connection string, then the driver will automatically retry some write operations for some types of failures. Ok, can I be more specific? Yes I can...

Previously in this series: Reading about Serializeable Snapshot Isolation.

Last week I took a deep dive into articles on Serializeable Snapshot Isolation. It ended on a sad note, as I learned that to extended SSI to a sharded database using 2PC for distributed transactions1 , there is a need to persist - which means replicate - all read sets in addition to all writes.

This conclusion has been bothering me, so before diving into other papers on distributed serializeable transactions, I wanted to understand better what exactly happens in SSI when a node (shard) fails. This blog doesn't introduce any new papers, just more details. And more speculation.

• 1for example, MongoDB

Previously in this series: Toward strict consistency in distributed databases.

Apparently, when Snapshot Isolation was invented, it was received with great enthusiasm, since it avoids all the anomalies familiar from the SQL standard. It was only as time went by that people discovered that SI had some completely new anomalies and therefore wasn't equivalent to serializeable isolation. (This I learned from the PostgreSQL paper below.)

July is vacation time in Finland. Which means it is a great time to read blogs and papers on consistency in distributed databases! This and following blogs are my reading notes and thoughts. Mostly for my own benefit but if you benefit too, that's great. Note also that my ambition is not to appear like an authority on consistency. I just write down questions I had and things I learned. If you have comments, feel free to write some below. That's part of the point of blogging this!

I was recently invited to speak at Fwdays Highload in Kyiv. This was my first ever visit to Ukraine, so I was excited to go and visit this large and beautiful European capital. Over a thousand years ago Vikings would row their boats through the rivers in Russia, and take the Dniepr southward to Kyiv and ultimately Turkey. It was exciting to travel in the footsteps of my forefathers.

My talk isn't really MongoDB specific, rather about an EC2 performance tuning project we did in 2017:

A month ago I published a quasi-academic paper, proposing 3 modifications to the Raft replication algorithm. I got some great review and feedback on the Raft mailing list. So based on that I have now updated the paper, hopefully to be much clearer than the first iteration.