Fwd: Can I move block data directly?

[Azuryy Yu <az...@gmail.com>]

Hi Dear all,

There are some unbalanced data nodes in my cluster, some nodes reached more
than 95% disk usage.

so Can I move some block data from one node to another node directly?

such as: from n1 to n2:

1) scp /data/xxxx/blk_*   n2:/data/subdir11/
2) rm -rf data/xxxx/blk_*
3) hadoop-dameon.sh stop datanode (on n1)
4) hadoop-damon.sh start datanode(on n1)
5) hadoop-dameon.sh stop datanode (on n2)
6) hadoop-damon.sh start datanode(on n2)

Am I right? Thanks for any inputs.

Re: Can I move block data directly?

[Chris Embree <ce...@gmail.com>]

I know nothing.

It seems that circumventing normal operations could be very bad.  There was
an example of something similar at hadoop summit.  Some very experienced
contributors decided they should edit meta data..... they broke their
cluster.

Just say no !  ;)
On Jul 8, 2013 9:01 PM, "Azuryy Yu" <az...@gmail.com> wrote:

> Hi Harsh,
>
> I also do agree with you that this is crude. and balancer is the right way.
> I just want to slove the problem very quickly. and only a few nodes
> involved.
>
>
> Thanks.
>
>
>
>
> On Tue, Jul 9, 2013 at 8:50 AM, Harsh J <ha...@cloudera.com> wrote:
>
> > Eitan,
> >
> > The block to host mapping isn't persisted in the metadata. This is
> > also the reason why the steps include a restart, which will re-trigger
> > a block report (and avoid gotchas) that will update the NN of the new
> > listing at each DN. Thats what makes this method "crude" at the same
> > time - you're leveraging a behavior thats not guaranteed to be
> > unchanged in future.
> >
> > The balancer is the right way to go about it.
> >
> > On Mon, Jul 8, 2013 at 6:53 PM, Eitan Rosenfeld <ei...@gmail.com>
> wrote:
> > > Hi Azurry, I'd also like to be able to manually move blocks.
> > >
> > > One piece that is missing in your current approach is updating any
> > > block mappings that the cluster relies on.
> > > The namenode has a mapping of blocks to datanodes, and the datanode
> > > has, as the comments say, a "block -> stream of bytes" mapping.
> > >
> > > As I understand it, the namenode's mappings have to be updated to
> > > reflect the new block locations.
> > > The datanode might not need intervention, I'm not sure.
> > >
> > > Can anyone else chime in on those areas?
> > >
> > > The balancer that Allan suggested likely demonstrates all of the ins
> > > and outs in order successfully complete a block transfer.
> > > Thus, the balancer is where I'll begin my efforts to learn how to
> > > manually move blocks.
> > >
> > > Any other pointers would be helpful.
> > >
> > > Thank you,
> > > Eitan
> > >
> > > On Mon, Jul 8, 2013 at 2:15 PM, Allan <wi...@gmail.com> wrote:
> > >> If the imbalance is across data nodes then you need to run the
> balancer.
> > >>
> > >> Sent from my iPad
> > >>
> > >> On Jul 8, 2013, at 1:15 AM, Azuryy Yu <az...@gmail.com> wrote:
> > >>
> > >>> Hi Dear all,
> > >>>
> > >>> There are some unbalanced data nodes in my cluster, some nodes
> reached
> > more
> > >>> than 95% disk usage.
> > >>>
> > >>> so Can I move some block data from one node to another node directly?
> > >>>
> > >>> such as: from n1 to n2:
> > >>>
> > >>> 1) scp /data/xxxx/blk_*   n2:/data/subdir11/
> > >>> 2) rm -rf data/xxxx/blk_*
> > >>> 3) hadoop-dameon.sh stop datanode (on n1)
> > >>> 4) hadoop-damon.sh start datanode(on n1)
> > >>> 5) hadoop-dameon.sh stop datanode (on n2)
> > >>> 6) hadoop-damon.sh start datanode(on n2)
> > >>>
> > >>> Am I right? Thanks for any inputs.
> >
> >
> >
> > --
> > Harsh J
> >
>

Re: Can I move block data directly?

[Azuryy Yu <az...@gmail.com>]

Hi Harsh,

I also do agree with you that this is crude. and balancer is the right way.
I just want to slove the problem very quickly. and only a few nodes
involved.


Thanks.




On Tue, Jul 9, 2013 at 8:50 AM, Harsh J <ha...@cloudera.com> wrote:

> Eitan,
>
> The block to host mapping isn't persisted in the metadata. This is
> also the reason why the steps include a restart, which will re-trigger
> a block report (and avoid gotchas) that will update the NN of the new
> listing at each DN. Thats what makes this method "crude" at the same
> time - you're leveraging a behavior thats not guaranteed to be
> unchanged in future.
>
> The balancer is the right way to go about it.
>
> On Mon, Jul 8, 2013 at 6:53 PM, Eitan Rosenfeld <ei...@gmail.com> wrote:
> > Hi Azurry, I'd also like to be able to manually move blocks.
> >
> > One piece that is missing in your current approach is updating any
> > block mappings that the cluster relies on.
> > The namenode has a mapping of blocks to datanodes, and the datanode
> > has, as the comments say, a "block -> stream of bytes" mapping.
> >
> > As I understand it, the namenode's mappings have to be updated to
> > reflect the new block locations.
> > The datanode might not need intervention, I'm not sure.
> >
> > Can anyone else chime in on those areas?
> >
> > The balancer that Allan suggested likely demonstrates all of the ins
> > and outs in order successfully complete a block transfer.
> > Thus, the balancer is where I'll begin my efforts to learn how to
> > manually move blocks.
> >
> > Any other pointers would be helpful.
> >
> > Thank you,
> > Eitan
> >
> > On Mon, Jul 8, 2013 at 2:15 PM, Allan <wi...@gmail.com> wrote:
> >> If the imbalance is across data nodes then you need to run the balancer.
> >>
> >> Sent from my iPad
> >>
> >> On Jul 8, 2013, at 1:15 AM, Azuryy Yu <az...@gmail.com> wrote:
> >>
> >>> Hi Dear all,
> >>>
> >>> There are some unbalanced data nodes in my cluster, some nodes reached
> more
> >>> than 95% disk usage.
> >>>
> >>> so Can I move some block data from one node to another node directly?
> >>>
> >>> such as: from n1 to n2:
> >>>
> >>> 1) scp /data/xxxx/blk_*   n2:/data/subdir11/
> >>> 2) rm -rf data/xxxx/blk_*
> >>> 3) hadoop-dameon.sh stop datanode (on n1)
> >>> 4) hadoop-damon.sh start datanode(on n1)
> >>> 5) hadoop-dameon.sh stop datanode (on n2)
> >>> 6) hadoop-damon.sh start datanode(on n2)
> >>>
> >>> Am I right? Thanks for any inputs.
>
>
>
> --
> Harsh J
>

Re: Can I move block data directly?

[Harsh J <ha...@cloudera.com>]

Eitan,

The block to host mapping isn't persisted in the metadata. This is
also the reason why the steps include a restart, which will re-trigger
a block report (and avoid gotchas) that will update the NN of the new
listing at each DN. Thats what makes this method "crude" at the same
time - you're leveraging a behavior thats not guaranteed to be
unchanged in future.

The balancer is the right way to go about it.

On Mon, Jul 8, 2013 at 6:53 PM, Eitan Rosenfeld <ei...@gmail.com> wrote:
> Hi Azurry, I'd also like to be able to manually move blocks.
>
> One piece that is missing in your current approach is updating any
> block mappings that the cluster relies on.
> The namenode has a mapping of blocks to datanodes, and the datanode
> has, as the comments say, a "block -> stream of bytes" mapping.
>
> As I understand it, the namenode's mappings have to be updated to
> reflect the new block locations.
> The datanode might not need intervention, I'm not sure.
>
> Can anyone else chime in on those areas?
>
> The balancer that Allan suggested likely demonstrates all of the ins
> and outs in order successfully complete a block transfer.
> Thus, the balancer is where I'll begin my efforts to learn how to
> manually move blocks.
>
> Any other pointers would be helpful.
>
> Thank you,
> Eitan
>
> On Mon, Jul 8, 2013 at 2:15 PM, Allan <wi...@gmail.com> wrote:
>> If the imbalance is across data nodes then you need to run the balancer.
>>
>> Sent from my iPad
>>
>> On Jul 8, 2013, at 1:15 AM, Azuryy Yu <az...@gmail.com> wrote:
>>
>>> Hi Dear all,
>>>
>>> There are some unbalanced data nodes in my cluster, some nodes reached more
>>> than 95% disk usage.
>>>
>>> so Can I move some block data from one node to another node directly?
>>>
>>> such as: from n1 to n2:
>>>
>>> 1) scp /data/xxxx/blk_*   n2:/data/subdir11/
>>> 2) rm -rf data/xxxx/blk_*
>>> 3) hadoop-dameon.sh stop datanode (on n1)
>>> 4) hadoop-damon.sh start datanode(on n1)
>>> 5) hadoop-dameon.sh stop datanode (on n2)
>>> 6) hadoop-damon.sh start datanode(on n2)
>>>
>>> Am I right? Thanks for any inputs.



-- 
Harsh J

Re: Can I move block data directly?

[Azuryy Yu <az...@gmail.com>]

Hi Eitan,

yup, The namenode has a mapping of blocks to datanodes, which keeping in
memory.
and you are also right, DN also keeps block structure in memory.

but you've noticed, I restart the DN after I moved block data manually.
then during the data node's start,

Block Scanner can scan the block local files, then construct it in the
memory, then, send block report to the NN, when NN received block report
from this data node, NN update <blockID, datanode storageID> mapping.

so I am right and tested it. but as Harsh mentioned, we'd stop both src and
target data node, then move block data manually, then start these two peer
data nodes.



On Mon, Jul 8, 2013 at 9:23 PM, Eitan Rosenfeld <ei...@gmail.com> wrote:

> Hi Azurry, I'd also like to be able to manually move blocks.
>
> One piece that is missing in your current approach is updating any
> block mappings that the cluster relies on.
> The namenode has a mapping of blocks to datanodes, and the datanode
> has, as the comments say, a "block -> stream of bytes" mapping.
>
> As I understand it, the namenode's mappings have to be updated to
> reflect the new block locations.
> The datanode might not need intervention, I'm not sure.
>
> Can anyone else chime in on those areas?
>
> The balancer that Allan suggested likely demonstrates all of the ins
> and outs in order successfully complete a block transfer.
> Thus, the balancer is where I'll begin my efforts to learn how to
> manually move blocks.
>
> Any other pointers would be helpful.
>
> Thank you,
> Eitan
>
> On Mon, Jul 8, 2013 at 2:15 PM, Allan <wi...@gmail.com> wrote:
> > If the imbalance is across data nodes then you need to run the balancer.
> >
> > Sent from my iPad
> >
> > On Jul 8, 2013, at 1:15 AM, Azuryy Yu <az...@gmail.com> wrote:
> >
> >> Hi Dear all,
> >>
> >> There are some unbalanced data nodes in my cluster, some nodes reached
> more
> >> than 95% disk usage.
> >>
> >> so Can I move some block data from one node to another node directly?
> >>
> >> such as: from n1 to n2:
> >>
> >> 1) scp /data/xxxx/blk_*   n2:/data/subdir11/
> >> 2) rm -rf data/xxxx/blk_*
> >> 3) hadoop-dameon.sh stop datanode (on n1)
> >> 4) hadoop-damon.sh start datanode(on n1)
> >> 5) hadoop-dameon.sh stop datanode (on n2)
> >> 6) hadoop-damon.sh start datanode(on n2)
> >>
> >> Am I right? Thanks for any inputs.
>

Re: Can I move block data directly?

[Eitan Rosenfeld <ei...@gmail.com>]

Hi Azurry, I'd also like to be able to manually move blocks.

One piece that is missing in your current approach is updating any
block mappings that the cluster relies on.
The namenode has a mapping of blocks to datanodes, and the datanode
has, as the comments say, a "block -> stream of bytes" mapping.

As I understand it, the namenode's mappings have to be updated to
reflect the new block locations.
The datanode might not need intervention, I'm not sure.

Can anyone else chime in on those areas?

The balancer that Allan suggested likely demonstrates all of the ins
and outs in order successfully complete a block transfer.
Thus, the balancer is where I'll begin my efforts to learn how to
manually move blocks.

Any other pointers would be helpful.

Thank you,
Eitan

On Mon, Jul 8, 2013 at 2:15 PM, Allan <wi...@gmail.com> wrote:
> If the imbalance is across data nodes then you need to run the balancer.
>
> Sent from my iPad
>
> On Jul 8, 2013, at 1:15 AM, Azuryy Yu <az...@gmail.com> wrote:
>
>> Hi Dear all,
>>
>> There are some unbalanced data nodes in my cluster, some nodes reached more
>> than 95% disk usage.
>>
>> so Can I move some block data from one node to another node directly?
>>
>> such as: from n1 to n2:
>>
>> 1) scp /data/xxxx/blk_*   n2:/data/subdir11/
>> 2) rm -rf data/xxxx/blk_*
>> 3) hadoop-dameon.sh stop datanode (on n1)
>> 4) hadoop-damon.sh start datanode(on n1)
>> 5) hadoop-dameon.sh stop datanode (on n2)
>> 6) hadoop-damon.sh start datanode(on n2)
>>
>> Am I right? Thanks for any inputs.

Re: Can I move block data directly?

[Allan <wi...@gmail.com>]

If the imbalance is across data nodes then you need to run the balancer.

Sent from my iPad

On Jul 8, 2013, at 1:15 AM, Azuryy Yu <az...@gmail.com> wrote:

> Hi Dear all,
> 
> There are some unbalanced data nodes in my cluster, some nodes reached more
> than 95% disk usage.
> 
> so Can I move some block data from one node to another node directly?
> 
> such as: from n1 to n2:
> 
> 1) scp /data/xxxx/blk_*   n2:/data/subdir11/
> 2) rm -rf data/xxxx/blk_*
> 3) hadoop-dameon.sh stop datanode (on n1)
> 4) hadoop-damon.sh start datanode(on n1)
> 5) hadoop-dameon.sh stop datanode (on n2)
> 6) hadoop-damon.sh start datanode(on n2)
> 
> Am I right? Thanks for any inputs.