As Steven knows, my <comment> field contains quite a bit of information: credits, notes, reviews (songs and albums). The problem is that with all of that information in MB's database for 180k+ songs, memory usage for MB balloons >1GB! My collection also brought mpd to its knees no matter how much memory I allocate it. However, mpd has a feature that allows me to choose which of the supported fields it would scan into its db. Since I don't query anything in the <comment> field, I removed it from mpd's list of fields to import and now mpd is nice and snappy. Furthermore, I use a Qt frontend called Cantata that permits me to configure it to read <comment> from the playing track.
I hope there could be a similar option to MB in the future: exclude fields like <comment> from being imported to its db but also allow it to read that tag from the playing track.