The default settings for the penciller_cache_size and cache_size were unscientific estimates of what would be helpful. The defaults do now lead to these processes carrying very large memory footprints.
Volume testing Riak recently, it was found that:
higher vnode counts are less detrimental than previously thought, and as a result the default recommendation should be to run a larger ring size (and therefore have more leveled processes per compute node);
higher cache_sizes have little/no impact on performance (reducing from 28K penciller cache size to 16K doesn't seem to reduce performance).
It would seem to be better to adjust defaults to reflect the fact there may be more penciller processes per node in the future.
Volume testing of https://github.com/basho/riak_kv/issues/1801 shows that higher vnode counts per node should be the default - as the main downside (2i query performance) has been addressed.
The default settings for the penciller_cache_size and cache_size were unscientific estimates of what would be helpful. The defaults do now lead to these processes carrying very large memory footprints.
Volume testing Riak recently, it was found that:
It would seem to be better to adjust defaults to reflect the fact there may be more penciller processes per node in the future.