The multi-generational LRU

Your assumption is correct. But our "wildly different workloads" are only a fraction of all real-world workloads. AFAIK, we don't use Postgresql, and we do plan to reach out to the Postgresql community and see if they could help us with benchmarking.

And we don't use io_uring (yet). Jens Axboe, the io_uring maintainer helped us identify a regression in buffered I/O and verify the fix. Now there is an improvement compared with the mainline: https://lore.kernel.org/linux-mm/20210413065633.2782273-1...

By design, the multigenerational LRU shouldn't make any workload worse -- why should it make worse decisions after it gathered more information? But that's just the theory :)

The bottomline is we'll do our best to fix any regressions the community reports to us. Again, the email address is page-reclaim@google.com.