I should note that these OSDs also drop out of the pool as part of their symptoms -
it's not clear to me at the moment if they drop out *because* of the memory, or if the
buffer_pool is growing large because it's buffering communications that aren't
getting to the cluster [and hence they drop out].