pp reworks

TomYum works out pp with two libraries, can route each mode to one of them, and can redo every score's pp from the admin panel when a library changes.

The pp libraries

	Library	Used by	Pinned in
Vanilla	ppy's ppy.osu.Game.Rulesets.*	TomYum (stable scores) and ppy's score processor (lazer, the website)	src/TomYum.Performance/TomYum.Performance.csproj, deploy/score-processor/Dockerfile (PROCESSOR_REF)
Relax / Autopilot	akatsuki-pp-rs	the pp-sidecar (osu! Relax: Akatsuki's osu_2019 formula)	sidecars/pp-sidecar/Cargo.lock

Vanilla pp has to match between TomYum and the processor, so the newest vanilla version is the one the processor's newest commit uses.

Routing Relax and Autopilot to the sidecar

By default TomYum works out every mode in-process with ppy's calculator. To give Relax and Autopilot akatsuki-pp-rs's pp, run the pp-sidecar and route those modes to it (the Performance settings):

performance:
  routes:
    - { modes: [vanilla],          backend: official }
    - { modes: [relax, autopilot], backend: akatsuki }
  backends:
    official: { type: inprocess-lazer }
    akatsuki: { type: http, url: "http://pp-sidecar:9000/pp", timeoutMs: 2000, fallback: official }

Mode tokens are a group (vanilla, relax, autopilot, all), a short name (rx!catch, …) or a number (0–11). With fallback, TomYum uses the official calculator while the sidecar is down; without it, the score gets 0 pp. TomYum refuses to start if a route names a backend that isn't defined.

The production stack doesn't include the sidecar, and reads no YAML. Add the service and the routes in a docker-compose.override.yml next to docker-compose.yml:

services:
  pp-sidecar:
    build:
      context: ../../sidecars/pp-sidecar
    restart: unless-stopped
  server:
    environment:
      TOMYUM_PERFORMANCE__ROUTES__0__MODES__0: vanilla
      TOMYUM_PERFORMANCE__ROUTES__0__BACKEND: official
      TOMYUM_PERFORMANCE__ROUTES__1__MODES__0: relax
      TOMYUM_PERFORMANCE__ROUTES__1__MODES__1: autopilot
      TOMYUM_PERFORMANCE__ROUTES__1__BACKEND: akatsuki
      TOMYUM_PERFORMANCE__BACKENDS__akatsuki__TYPE: http
      TOMYUM_PERFORMANCE__BACKENDS__akatsuki__URL: http://pp-sidecar:9000/pp
      TOMYUM_PERFORMANCE__BACKENDS__akatsuki__TIMEOUTMS: "2000"
      TOMYUM_PERFORMANCE__BACKENDS__akatsuki__FALLBACK: official

Then docker compose up -d --build pp-sidecar server. After changing which backend a mode uses, start a pp rework for those modes (below), so older scores get the new pp too.

The pp updater

deploy/pp-updater is a small Python service on the host (systemd tomyum-pp-updater, as root: it drives Docker, so no container needs Docker's socket). Every 6 hours, or on Check now, it compares the pins with the newest versions; Admin → pp rework → pp libraries shows both. Update asks it to:

• Relax/Autopilot: cargo update -p akatsuki-pp, build and restart the pp-sidecar, check its /version, commit Cargo.lock, then start a rework of Relax and Autopilot.
• Vanilla: move the processor to its newest commit and TomYum's packages to its ppy version, check TomYum builds, build and restart server and score-processor, commit the pins, then start a rework of the vanilla modes.

A failed build puts the pins back and leaves the running containers alone; the error shows in the panel. Commits are local: push them yourself if you keep a fork.

Install it (set ExecStart to your checkout's updater.py first):

sudo cp deploy/pp-updater/tomyum-pp-updater.service /etc/systemd/system/
sudo systemctl enable --now tomyum-pp-updater
sudo python3 deploy/pp-updater/updater.py check      # a check by hand

It finds the checkout from its own path; TOMYUM_DIR, COMPOSE_DIR, RUN_AS (the user that builds and commits; by default the checkout's owner) and CHECK_HOURS change that. The Relax/Autopilot update expects a pp-sidecar service, as added above.

pp reworks

Started after an update, or by hand (Admin → pp rework: the modes and a reason). While one runs:

• its modes are paused: TomYum doesn't recompute their ranks or change their in-game pp, and the website's ranking page shows a copy of the stats from before;
• new scores in those modes wait for their pp: not a best, not in totals;
• every passed score's new pp is worked out a batch at a time (it resumes after a restart, and scores made meanwhile join in);
• every page shows a banner (the modes, why, how far), and the admin panel shows the progress.

Then it switches over at once: the new pp, each player's best per map, everyone's totals. Vanilla modes then wait for ppy's processor to redo the website's scores; then ranks and in-game pp follow and the rankings are live again. Cancelling keeps the old pp.

For a one-off recalculation from the command line, see tools/RecalcPp on updating.