From: Alex Loddengaard
Sent: Thursday, July 08, 2010 11:39 AM
Subject: Re: rebalancing replciation help
Consider setting dfs.balance.bandwidthPerSec to something as high as 20971520 for the balancer and the setrep. You can do this by supplying -D at the command line.
Your strategy for getting data onto the 5 nodes is correct: balance and setrep. Just understand these things take time.
Hope this helps.
On Wed, Jul 7, 2010 at 4:09 PM, Arun Ramakrishnan wrote:
I have more than a specific question. I am going to layout the steps I have taken. Please comment on what I can do better.
I was trying to to add 5 nodes to my existing 10 node cluster and also increase the replication factor from 2 to 3.
I thought I don't have to run the balancer cause it would most likely put the new replicas into the new nodes.
There are about 500k blocks.
I wanted to get it all stabilized(replication and balancing) within 24 hours. Its more than 24 hours now and fsck reports 30% under replication. Is there a way to force hdfs to use balance/replicate more aggressively.
It would be great if someone explained what/when things happen to blocks in the context of
3) Restarting cluster with a higher/lower replication factor.
A few questions and a few issues here.
1) When you restart the cluster with a higher than previous replication value. Does it also apply to existing blocks or only to new blocks being created ?
2) Does the balancer take into account under replication of blocks or does it blindly start moving existing blocks to reach threshold ?
A very specific problem . I am having this strange problem where the -setrep hangs on one particular block for hours. Is this because its corrupt ?. But, fsck said its healthy.