I finally finished my long support shift on Tuesday.
On Friday I put live an A/B test for the new (new) search query, which performed pretty well for the handful of queries we manually checked. But ultimately the only way to know if it’s good or not is to put it in front of a lot of users and see what the aggregate metrics say. We also want to be careful about making search worse at the moment, as people are probably relying on GOV.UK a bit more currently than they normally would, so only 1% of traffic is getting the new query.
I made the first step towards bringing back search spelling suggestions. This is a feature which did exist, but got lost when we changed how search worked a while back. There are two bits of work in bringing this back: the backend work to fetch spelling suggestions from search-api (which this PR achieves), and some frontend / design work to figure out how to present suggestions to the user. Our search interface is pretty cluttered on mobile, so the design bit is going to be the hard part. You can see it in action by going to /search/all.json?keywords=funance and looking at the
Possible future work here is making it smarter than just correcting a single word at a time. For example, if you search for “drving loicence” it suggests “drving licence”, and if you search for that it suggests “driving licence”. It would be better to go straight to the final suggestion, without needing the intermediate search.
I did some investigation into how we handle provisioning new machines, as we get a burst of errors whenever we switch on a new search machine. Ultimately it turns out that our health checks, which determine if a machine can be added to a load balancer’s active pool, aren’t very good for search; they only check the machine is running, not that the app is running.
I started looking into how to better handle searches with quoted fragments. These four queries all give different results if you search for them in Google:
"marzipan shoes"(least results)
marzipan shoes(most results)
The behaviour seems to be that quoted fragments are exactly matched against the content (and are required), whereas the unquoted fragments are fuzzily matched (and aren’t required). In contrast, GOV.UK search treats cases 2, 3, and 4 all the same: we only do something different if the entire query is quoted. I don’t like that, so I’m working on a way to make it more Google-y, as I suspect that is what other people expect too.
I read Downward to the Earth (by Robert Silverberg)00Which is my newest book, and the 700th entry in bookdb.
. It was about colonialism and about what makes humanity different to animals, set on an alien world which was once humanity’s possession, but was returned to its native inhabitants. It was really good.
I’ve played a lot of Rise to Ruins this week. It’s difficult, but I’m gradually getting better I think.