Posts

It is rewarding to work on software that people care about and use all around the clock. This constant usage means we can’t simply take the system offline for maintenance without upsetting users. Therefore, techniques that allow us to update the software seamlessly without downtime or compromising service quality are incredibly valuable. Most projects I’ve worked on use a relational database for persistence, and have some sort of migration tool like flyway or liquibase to make changes to the database schema. This post is about a particular kind of migration situation that, in my experience, most developers who work on such projects will encounter at some point in their career. They will want to apply a simple, and seemingly innocent migration, like adding a column to a table and it’ll cause some number of requests to fail, or maybe even a small outage. There are some tricks we can use here to reduce risk and automatically detect some patterns that cause this problem. ...

If you’re developing a transactional application backed by postgres, there’s a pretty cool trick you can use to check if you’re missing indexes that could potentially cause serious performance issues or even outages. In particular, I mean foreign keys where the referencing side of the constraint does not have an index. The idea is very simple, we can select all of the columns that take part in a foreign key, then remove the ones that take part in a complete index, and the remainder should be the empty set, or possibly match a known allowlist. I think this is a valuable addition to the test cases for your database migrations, or if you can’t easily do that, maybe in your CI/CD pipeline. ...

I worked for a few years in the intersection between data science and software engineering. On the whole, it was a really enjoyable time and I’d like to have the chance to do so again at some point. One of the least enjoyable experiences from that time was to deal with big CSV exports. Unfortunately, this file format is still very common in the data science space. It is easy to understand why – it seems to be ubiquitous, present everywhere, it’s human-readable, it’s less verbose than options like JSON and XML, it’s super easy to produce from almost any tool. What’s not to like? ...

For tests that need to touch the database, it is generally a really good idea to roll back transactions. That way, you can run lots of tests in parallell or in any arbitrary order and the tests won’t interfere with each other. But sometimes, that just isn’t possible. One reason for this could be that the code base handles transactions in a way that makes it really hard to get a handle on them in the right place, or it could be a legacy code base where everything is running with auto-commit or some other explanation. Either way, it is very important to be able to isolate effectful tests from each other both to get good performance out of building on multi-core machines and to avoid flaky tests or tests that break when they run in a different order. ...

There are 2 configuration options that every OLTP application that uses postgres should set, in order to protect the database from high load: statement_timeout idle_in_transaction_session_timeout These can both be set by client configuration and require no special permissions to set, and are easily overridden locally for transactions that have different requirements. They can be a bit scary to retrofit to existing applications, but we can activate two postgres extensions to help us measure our queries to find safe values to set: ...