Andrew Purtell wrote:
Thanks Stack. I'll walk over your list of questions and see
if maybe one leads down the correct path!
One thing I can answer right away is that no storefile in
particular seems to be the bullet. It seems to me that after
a while heap pressure builds to a point where the
regionserver falls over, and in a place where the OOME does
not take it down. Indeed I do think that backporting the
OOME handling improvements to 0.18 branch would be helpful.
Lets figure whats up over on your cluster and roll a 0.18.2 to address
them, quickly.
Something I will do right away is disable blockcache. It's
use as I can see looking at our code is gratuitous.
Ok. In TRUNK we've been testing it and have fixed at least one bug.
Also, ok based on what you say what I am experiencing is
different from what's happening on jgray's cluster. There is
plenty of available VM and minimal swapping.
Ok. You have ganglia or something in place so you can see across time?
Weird thing about the jgray phenomeon seen last weds. was loads of mem
and cpu but crazy swap anyways.
St.Ack
