Machine learning stuff
On Monday morning, I flipped the switch to start serving our ML search ranking model from Amazon SageMaker, which also means we’re training and deploying a new model every day. We’re still evaluating it, but it looks pretty comparable to what we had running previously. It’s a bit hard to say at the moment, because the deadline for self assessment was at the end of January, so recent search traffic has been heavily skewed towards searching for that sort of thing.
I also made some changes to make it easier to train and deploy a model using a branch of the scripts, but now I need to write up some documentation on how to actually do that, as it’s all in my head currently.
I’m also planning to write a memo on using Learning to Rank with SageMaker, because that was pretty tricky to figure out.
We regularly review our dev docs, each source file starts with a header like this:
--- owner_slack: "#govuk-searchandnav" title: How to debug underperforming search section: Monitoring layout: manual_layout parent: "/manual.html" last_reviewed_on: 2020-02-04 review_in: 3 months ---
owner_slack channel gets notified when
last_reviewed_on + review_in is the current date or older. However, when reviewing documentation like that, people often just make sure that the current content is still correct, and if so, bump the
last_reviewed_on date. Something we don’t often do is review groups of documentation for consistency or missing information.
It’s possible that the problem could be addressed by some process change. A lot of documentation is assigned to the in-hours support team (well, their slack channel), and they usually have other stuff going on to put much effort into reviewing docs. Maybe splitting up the load amongst other teams more, and having pages about similar topics all get reviewed at once, would help.
The RFC got accepted, and I’ve been working through the implementation. I’ve got steps 1 to 4 done. Now I’m getting to the steps which require changes to Whitehall and Content Publisher. As Content Publisher is under active development, I think the time has come to have a chat with some developers on its team.
Splunk for rookies
I attended a two-hour “Splunk for rookies” workshop in the office, given by a visiting Splunk employee (splunker?).
First thoughts: the Splunk query language is far nicer than the Kibana query language; I was able to actually figure out stuff for myself without needing to read documentation first. I can also see why people like this over Grafana, where you’re writing queries in the language of the underlying data store.
Second thoughts: they really emphasise how the data you give to Splunk is analysed at query time, rather than being indexed into a database or anything like that. But given that you tell Splunk what the format of your data is (eg, an Apache access log), isn’t that needlessly slow? The splunker presented it as a universal good, solving all sorts of problems with databases, but of course neglected to mention how it compared to the advantages of databases. There are reasons we don’t use flat files, parsed on demand, for everything.
I’ve decided that my new game, to fill the void left by Call of Cthulhu, will be Apocalypse World. The first game will be next week: Saturday the 15th.
I’ve not played Apocalypse World before, and now I’m running it, so this’ll be interesting.
I made a few improvements to memo.barrucadu.co.uk:
There are still some things to do, particularly on the mobile friendliness front: font size in
<pre> tags is really small, for reasons I’ve not yet figured out, and I need to make margin notes appear in the main body of text. I also need to either pick a new syntax highlighting theme for dark mode, or find a theme which works for both. The current colours mostly work, but there are a few syntax elements which are hard to read.
I’m planning to introduce dark mode and mobile improvements to www.barrrucadu.co.uk as well.
I had a go at making bookdb.barrucadu.co.uk more mobile friendly… but quickly gave up.
- Old CSS, new CSS
- How To Add CSS Dark Mode To A Website
- How I Work From Anywhere in the World
- What makes a good runbook?
- What to do when you don’t trust your data anymore
- Building an Effective Test Pipeline in a Service Oriented World
- Creating Sigils
- Structural typing and first-class case expressions
- My productivity app for the past 12 years has been a single .txt file
- Issue 197 :: Haskell Weekly
- #02 - NixOS Weekly
- This Week in Rust 324
- The TTY demystified
- Linux PTY - what powers docker attach functionality
- Lessons learned from writing ShellCheck, GitHub’s now most starred Haskell project