resident memory ratio drops to zero on mere updates of existing items of single node with no upr consumers at all

This is the problem I've showed to you.

It's single cluster_run node (32-bit, but it shouldn't matter). Single bucket with name "default". Bucket quota set to 900 megs. I've uploaded 600k docs of 1k each using this:

./generate-keyset.rb -k 600000 -v 1024 | GOMAXPROCS=8 \time maxi

(generate-keyset can be found in my fork of mc-loader)

After everything is loaded and persisted I just do slow updates of same items (using same values):

./generate-keyset.rb -k 600000 -v 1024 | pv -L 3000000 | GOMAXPROCS=8 \time maxi

(pv slows down flow of data into maxi giving me about 3k of mutations per second)

And after couple passes of this I see drop of resident ratio. In my previous runs where I used even smaller rate (1k instead of 3k) I actually saw hard oom alert after some time.