The ListenBrainz team has been working hard on moving our primary listen store from InfluxDB to TimescaleDB, and today at UTC 16:00 we’re going to make the switch.
We were asked on Twitter as to why we’re making the switch — and in the interest of giving a real world use case for switching, I’m writing this post. The reasons are numerous:
Openness: InfluxDB seems on a path that will make it less open over time. TimescaleDB and its dependence on Postgres makes us feel much safer in this regard.
Existing use: We’ve been using Postgres for about 18 years now and it has been a reliable workhorse for us. Our team thinks in terms of Postgres and InfluxDB always felt like a round peg in a square hole for us.
Data structure: InfluxDB was clearly designed to store server event info. We’re storing listen information, which has a slightly different usage pattern, but this slight difference is enough for us to hit a brick wall with far fewer users in our DB than we ever anticipated. InfluxDB is simply not flexible enough for our needs.
Query syntax and measurement names: The syntax to query InfluxDB is weird and obfuscated. We made the mistake of trying to have a measurement map to a user, but escaping measurement names correctly nearly drove one of our team members to the loonie bin.
Existing data: If you ever write bad data to a measurement in InfluxDB, there is no way to change it. I realize that this is a common Big Data usage pattern, but for us it represented significant challenges and serious restrictions to put simple features for our users into place. With TimescaleDB we can make the very occasional UPDATE or DELETE and move on.
Scalability: Even though we attempted to read as much as possible in order to design a scalable schema, we still failed and got it wrong. (I don’t even think that the docs to calculate scalability even existed when we first started using InfluxDB.) Unless you are using InfluxDB in exactly the way it was meant to be used, there are chances you’ll hit this problem as well. For us, one day insert speed dropped to a ridiculously low number per second, backing up our systems. Digging into the problem we realized that our schema design had a fatal flaw and that we would have drastically change the schema to something even less intuitive in order to fix it. This was the event that broke the camel’s back and I started searching for alternatives.
In moving to TimescaleDB we were able to delete a ton of complicated code and embrace a DB that we know and love. We know how Postgres scales, we know how to put it into production and we know its caveats. TimescaleDB allows us to be flexible with the data and the amazing queries that can be performed on the data is pure Postgres love. TimescaleDB still requires some careful thinking over using Postgres, it is far less than what is required when using InfluxDB. TimescaleDB also gives us a clear scaling path forward, even when TimescaleDB is still working on their own scaling roadmap. If TimescaleDB evolves anything like Postgres has, I can’t wait to see this evolution.
Big big thanks to the Postgres and TimescaleDB teams!