Ceph

sometimes clusters have pgs that are degraded for long periods of time. this forces the mon to retain lots of old osdmaps, even when they don't provide any new or useful information about past pg locations that we'll need to search to find missing objects. the past_intervals metadata can also grow very large, slowing peering.

i think what we want to do is add a pg_stat_t field that controls how many past osdmaps we want, independent of last_epoch_clean. in some cases, the primary pg can move that value forward because the past intervals are full of redundant information. alternatively, we can explicitly include a set of osds and collapse the older past_intervals into that set so that the interval data (and old osdmaps) can be trimmed.