Another week down, and I’m now on a project. I’m working with another dev to make our big monolith more scalable by reducing the load on one of the larger database tables. So I’ve been thinking about API boundaries and database partitioning schemes and refactoring, with an eye to potentially moving this whole subsystem off into a separate service with its own database once it’s untangled.
The team I’m on owns a lot of the core functionality which the rest of GoCardless builds upon, and also that code is a tangled mess. So it feels pretty similar to being on the Platform Health team at GDS, which I think was a great team to start on because I got exposed to all sorts of different areas of the stack pretty early on, which helped me build a good mental model of how GOV.UK worked. Hopefully my start on the Financial Orchestration team at GoCardless will be just as fruitful.
No books this week.
Sadly, my Tuesday night game has come to an end.
It turned out that nobody was really that enthusiastic about the campaign itself, we were all just happy to be playing a game. Which can be fine, but there were also a few issues we had with the game and so it became a problem.
The group’s decided to give D&D a go, but since I’m in two OSR games right now (a Stars Without Number game every other Saturday and a Whitehack game every other Sunday), I decided that that would be a bit too much D&D for me and so bowed out.
No significant changes this week, but I wrote a memo on how DNS works.
I switched my personal finance dashboard over to using Prometheus, via promscale, a time-series database built on top of postgres, and I give the Prometheus queries I’m using for my key metrics in my personal finance memo.
My dashboard used to use InfluxDB 1. But InfluxDB 1 doesn’t have a very good query language, so I had to do quite a lot of pre-processing of my hledger data in the daily ETL run. Rather than just dump my transaction data and upload it, I instead ended up generating a new metric for each sort of query I wanted.
This made the dashboard really simple, but the script pretty complicated. Sometimes it would even time out uploading the data to InfluxDB, and I’d have to go tweak the upload batch size. It was kind of a pain.
I wanted to change it, and Prometheus seemed the best choice.00There’s also InfluxDB 2, which has a much more powerful query language than InfluxDB 1, but I don’t use it for anything else. So I’d be learning a brand new database technology just for one dashboard I check maybe once a fortnight.
But how do you get historic data into Prometheus? At the time, it didn’t support backfilling old data (and it only got that feature pretty recently), and while Prometheus does support fetching data from InfluxDB 1, I couldn’t get that to work.
So I shelved the idea, and decided to just live with my suboptimal script.
…until this week, when it started timing out again. Enough was enough. Time to return to the Prometheus documentation and get this working.
This time I found promscale. It has an API to load a bunch of data with arbitrary timestamps, and to delete data. Just what I need. I also rewrote my script in python, by consuming hledger’s CSV output, rather than using it as a Haskell library as my old script did.
And now it’s so much simpler.