Next server update: 29.03.2021 - Preparation log

The next server update has been scheduled for the 29.03.2021 somewhere between 09.00 - 11.00 CET (= 08.00-10.00 UTC/GMT).

It will be a minor patch, so the downtime itself shouldn't be more than 5 minutes.

What is planned so far:

• Rating wipe for TMM 2v2
• Update the lobby server to v1.7.0:
  • TMM 2v2 players initial rating will be copied from Global rating
  • Adds support for including randomly generated maps in matchmaker map pools
  • Fixed coop leaderboards
  • Fixed a bug where matchmaker games would fail to launch when both players on a team were in a clan together that had non-ascii characters in the clan tag.
  • Fixed detection of some units for unlocking achievements. It should now be possible to unlock the "build czars" achievement.
  • Fixed a bug that would skip some validity checks when playing featured mods. Phantom games with uneven teams should no longer be rated.
• Update the replay server to v1.1.2:
  • All (new) replays will now also be stored in the new format with Zstandard compression
  • The replay server will mark replays as available once they are written to disk. (The corresponding client feature will follow.)

What will probably not come:

• Move from MySQL to MariaDB
  • We had a lot of progress, yet this needs more serious testing as it could break stuff
• Implement ZFS tuning as researched by Mozilla/Let's Encrypt (* requires MariaDB)

Update:

Added:

• TMM 2v2 Rating wipe

Where can I complain? After gathering months of reliable 2v2 data, we are now going to throw it all away for seemingly no reason? I get that it was planned from the start, but makes precisely zero sense atm.

@Blodir talk to @FtXCommando

@Blodir In addition here is a list comparing tmm2v2 rating with global rating order by descending rating difference:

rating_diff_tmm2v2_to_global.csv

"TMM 2v2 players initial rating will be copied from Global rating"

Copied and used for the initial rating, or global now = TMM? I am confused by this. People are saying that the TMM and global will now accumulate one and the same, but do not know if this is random, opinion banter verse fact.

Your global rating is used to define your mu value. You then get 150 sigma added to your sigma value with a hard cap of 250 sigma.

This means your shown rating will be your global rating, but roughly 500 points lower than what it shows in global lobbies.

Global games have no impact on your tmm rating except from allowing a more refined initial prediction of your baseline skill when the system has no data on you. So basically, it impacts where you should be placed in the distribution of players prior to actually playing the matchmaker.

So I took a little bit of time to check out the .csv file and gather more insight on the relationship between global and 2v2 rating.

First thing I had to do was clean the data to remove all the people with extreme deviation as it's impossible to really factor them into the distribution as the system itself is pretty unsure as to where to place them. I decided to use people that had <200 deviation in both tmm and global as they represented reliable focal points for both environments. This left me with about 350 data points out of the total 2300, which was honestly a surprisingly small amount.

Anyway, here's the information I got for the rating distribution:

It does seem like tmm rating is more centered at a proper distribution than global rating, though the difference between the two is not so extreme that it warrants throwing out the global data as unreliable. In fact, we can test the reliability with a simple regression, which I did.

Overall, global rating has a p-value of <.001 and about 70% of your tmm rating can be explained via your global rating performance.

What does that mean for FAF? It means there is really no reason to throw away the information of global rating in predicting player competency, particularly since we include the safeguards of an additional sigma increase alongside artificially set trueskill values for your first 10 games. In fact, with such a low p-value, it would basically be incompetent to ignore the data when trying to place a player transitioning between game environments.

What this tells me is that it is without question the proper decision to base all tmm ratings off of (at least) global rating; it probably gets even more accurate for something like a 4v4 matchmaker since 4v4 is a popular global game mode. The only real question is whether it makes sense to apply the reset to players that already have been placed by the system. The logic for taking that extra step revolves around maintaining as much consistency between the two systems to prevent future unintended idiosyncrasies. It was also early enough in the lifetime of the matchmaker that the loss in information was not a massive destabilization of the system, especially since it was done (or intended to be done) in a controlled manner.

So the TMM rating was reset but instead of being based on global it's just 0, what happend and what is the plan now? Another reset or?

Fix the underlying problem, deploy server patch, reset rating once more

Alright, sounds good. Perhaps we should make a news post about it? People might get angry to have their rating reset once more.