Just FYI: We have worked around these issues by running regular (scripted andthus controlled) vaccuums on all tables but the active ones and adding L2 ZFScaching (l2arc). I hope to get back to this again soon.