In T305966#7854160, @Krinkle wrote:

There's some history here that may be worth reading up on, as this is actually not the time time we've encountered this:

• T54777: user_properties table bloat
  • T54778: Clean up skin user preferences in user_properties table on Wikimedia wikis
  • T171643: Clean up skin properties

Thanks @Krinkle for the background. With this in mind, I would rather start from the assumption that this is not a new problem but an old problem (we can revisit that later if it proves incorrect).

According to T54777 in 2013 75% of the skin rows had an empty string as their value severely impacting the index and data size. An attempt to fix this was made in the same year (T54778) but it was declined due to concerns around data loss and only having one full-time DBA. The DBA at the time was quoted as saying "dropping rows like this has 0 effect on both disk space and performance due to fragmentation".

It was reopened in 2018 by @Niharika as these junk rows make data analysis difficult e.g. getting metrics around skins. As this task and related spikes show, this analysis problem is very real - I spent three days debugging a problem due to this.

The task remains open so the question is what to do next and if web sees the value in this (it seems there is value across product mostly for analysis purposes).

I spoke to @Ladsgroup about T54777 and T54778 and he says this is something that needs fixing but he doesn't have the capacity. He tipped me off that @Tgr might also be having a problem with this too, so perhaps there is an opportunity for growth team and web team to tackle this together? Presumably, that would involve writing a new script in http://github.com/wikimedia/mediawiki-extensions-WikimediaMaintenanceScripts and running it.