Updated recommendations

A lot has happened. Some months ago the osu! ppv2 formulae were updated. This update was the largest update to ppv2 in years. Not only did this break the pp values that Tillerinobot claimed for certain plays, but it also messed with the recommendation algorithm.

Meanwhile I was working on general technical improvements. The largest improvement is that Tillerinobot now completely runs on Kubernetes. Another technical improvement was the so called response-queue. This fixes an ancient "overheating" problem where the bot would sometimes just die because it was trying to do too much at the same time. Since this could happen at any time the team came up with an ad-hoc workaround which just preemptively restarted the bot once a day. Of course this was not very popular with the osu! staff since it spawned a wall of "F"s in the osu chat every time it happened. The fix is a massive restructuring of Tillerinobot - the kind that you need do to software that was never meant to be big but somehow became big.

Finally (and much too late) the ppv2 formulae in Tillerinobot were adjusted. Luckily, the difficulty parameters are available from the osu! API nowadays. This made the update much easier than in earlier days. Unfortunately, there are rare deviations which we cannot explain. For now, it's close enough.

What about the recommendations update

Since pp is very important for recommendations the pp update made recommendations much less accurate. Unfortunately, for many the accuracy dropped by so much that the recommendations were no longer useful. The answer was to re-calibrate the entire model. This is the fifth installment of the "gamma" model and is now available. This should be giving much better recommendations especially if you play a lot of newer maps. You can switch to the newer model by adding "gamma5" to your command like so:

!r gamma5

As usual, this can be combined with other filters, e.g. "!r gamma5 DT LEN<=60 AR<=9.7 AR>=9.5 OD<=9.5 -HR -HD" for the pleb pp shower.

gamma5 is now available for Patrons. Once we've gathered a bit of feedback, we'll make it generally available and if everything goes well, we'll use it as the default model soon thereafter.