osu! diffcalc / perf deploy 2024Q4

[peppy]

osu! diffcalc / perf deploy 2024Q4

Published to osu! diffcalc / perf deploy 2024Q4 · Issue #30 · ppy/osu-infrastructure

community awareness

• News post


deployment scope

There has been a bit of discussion over whether we should be deploying “all” changes (including the major combo scaling change) or a limited subset. I am going with the full deployment for the following reasons:

• I want to get the changes out as soon as possible and with minimal overhead. Running the process twice will take longer in both realtime and babysitting hours from my schedule.
• If anything goes wrong, we will have a chance to assess and react before it impairs users’ game experience. During the reprocessing we will have things like rank history paused, so it should not have any lasting effects if we need to temporarily pause or restart the run.
• Even if not deploying the full set of changes, it has been confirmed that virtually every osu! score’s pp value is going to change by at least 1 pp due to a slider lenience change. If this was not the case, doing a limited deploy (where we could verify that pp values didn’t change) may have had value.

preparation

• Figure out what will still be using the old tables, and how much we need to consider them (check we're writing pp values back from the new processor at least).
• By extension, ensure that all stable scores are given the chance to preserve in order for the highest pp score to be considered by the UserTotalPerformanceProcessor.
• Inactive players still have non-updated pp values.
Note that these are outdated in both the old and new tables (see forum thread / Inactive users' score PP values are not reprocessed when they become active again osu#28864), peppytodo: reply in email thread when fixed
• Ensure we have a queue setup to allow score-statistics to process stable pp values
• Ensure stable is able to return values to the client from new pp values (yes, see https://github.com/peppy/osu-web-10/commit/8e05bb0794acba0f560fb1a492e3893e8ca7f421)

pre-run

• Merge Remove combo scaling from Aim and Speed from osu! performance calculation osu#16280 (and other pending review items)
• Remove estimations where score data is available for osu! difficulty calculations osu#27691
• Disable global rank history updates
• Disable global rank updates(?) while things are bound to be fluctuating

main

• Update beatmap processor to latest SR calculation version
• Run difficulty calculation across all ranked beatmaps
osu.Server.DifficultyCalculator all -m 0 -ac -c 4 (run per ruleset)
• Stop osu-performance from updating pp values


This is currently deployed and runnning on kubernetes, one instance per ruleset. Setting replicas to zero should be a safe move, and means we can rollback if required. Once we’re happy with things, removing the config.

• Start osu-score-statistics-processor updating pp values-
Set envvar WRITE_LEGACY_SCORE_PP=1 on kubernetes
Change RunOnLegacyScores flag on performance processor, deploy with https://github.com/ppy/osu-queue-score-statistics/actions/runs/11547127429
• Merge osu!stable change to bring in diffcalc changes

post-run

• Start reprocess of all scores
. This should probably happen in multiple rounds, favouring the score which matter first. For the old infrastructure, we’ve run this on only active users initially:
ScoreStatisticsProcessor performance scores sql “<active users>”
(see command)

This could be followed by a secondary exhaustive run:
ScoreStatisticsProcessor performance scores all
(see command)
. Ensure we set REFRESH_STORES_PERIODICALLY correctly.
• Check whether Classic scores set on some beatmap appear to have its total score heavily inflated. osu#29772 is fixed
• Run diffcalc on all non-ranked beatmaps

future

• Enable realtime processing for unhandled mod combinations