You are viewing a plain text version of this content. The canonical link for it is here.
Posted to hdfs-user@hadoop.apache.org by felix gao <gr...@gmail.com> on 2011/04/12 16:46:31 UTC
Question regarding datanode been wiped by hadoop
What reason/condition would cause a datanode’s blocks to be removed? Our
cluster had a one of its datanodes crash because of bad RAM. After the
system was upgraded and the datanode/tasktracker brought online the next day
we noticed the amount of space utilized was minimal and the cluster was
rebalancing blocks to the datanode. It would seem the prior blocks were
removed. Was this because the datanode was declared dead? What is the
criteria for a namenode to decide (Assuming its the namenode) when a
datanode should remove prior blocks?
Re: Question regarding datanode been wiped by hadoop
Posted by Ayon Sinha <ay...@yahoo.com>.
If you've only lost a few partitions on a data node and no loss of complete
files due to replicated blocks, then I'd wipe the dfs.data.dir partitions and
rebalance. It can get time-consuming to find the exact reason why the data
blocks got removed.
-Ayon
See My Photos on Flickr
Also check out my Blog for answers to commonly asked questions.
________________________________
From: felix gao <gr...@gmail.com>
To: hdfs-user@hadoop.apache.org
Sent: Tue, April 12, 2011 8:05:43 AM
Subject: Re: Question regarding datanode been wiped by hadoop
From the timestamp point of view, the only directory seems to be modified and
removed is the Current directory under dfs.home.dir. However, the storage file
under dfs.home.dir is untouched since the datanode started.
On Tue, Apr 12, 2011 at 8:02 AM, felix gao <gr...@gmail.com> wrote:
The xml files have not been changed for more than two months, so that should not
be the reason. Even the in_use.lock is more than a month old. However, we did
shut it down few days ago and restarted it afterward. Then the second shutdown
might not be clean.
>
>
>
>On Tue, Apr 12, 2011 at 7:52 AM, Ayon Sinha <ay...@yahoo.com> wrote:
>
>The datanode used the dfs config xml file to tell the datanode process, what
>disks are available for storage. Can you check that the config xml has all the
>partitions mentioned and has not been overwritten during the restore process?
>> -Ayon
>>See My Photos on Flickr
>>Also check out my Blog for answers to commonly asked questions.
>>
>>
>>
>>
>>
>>
________________________________
From: felix gao <gr...@gmail.com>
>>To: hdfs-user@hadoop.apache.org
>>Sent: Tue, April 12, 2011 7:46:31 AM
>>Subject: Question regarding datanode been wiped by hadoop
>>
>>
>>
>>What reason/condition would cause a datanode’s blocks to be removed? Our
>>cluster had a one of its datanodes crash because of bad RAM. After the system
>>was upgraded and the datanode/tasktracker brought online the next day we noticed
>>the amount of space utilized was minimal and the cluster was rebalancing blocks
>>to the datanode. It would seem the prior blocks were removed. Was this
>>because the datanode was declared dead? What is the criteria for a namenode to
>>decide (Assuming its the namenode) when a datanode should remove prior blocks?
>
Re: Question regarding datanode been wiped by hadoop
Posted by felix gao <gr...@gmail.com>.
>From the timestamp point of view, the only directory seems to be modified
and removed is the Current directory under dfs.home.dir. However, the
storage file under dfs.home.dir is untouched since the datanode started.
On Tue, Apr 12, 2011 at 8:02 AM, felix gao <gr...@gmail.com> wrote:
> The xml files have not been changed for more than two months, so that
> should not be the reason. Even the in_use.lock is more than a month old.
> However, we did shut it down few days ago and restarted it afterward. Then
> the second shutdown might not be clean.
>
>
> On Tue, Apr 12, 2011 at 7:52 AM, Ayon Sinha <ay...@yahoo.com> wrote:
>
>> The datanode used the dfs config xml file to tell the datanode process,
>> what disks are available for storage. Can you check that the config xml has
>> all the partitions mentioned and has not been overwritten during the restore
>> process?
>>
>> -Ayon
>> See My Photos on Flickr <http://www.flickr.com/photos/ayonsinha/>
>> Also check out my Blog for answers to commonly asked questions.<http://dailyadvisor.blogspot.com>
>>
>>
>> ------------------------------
>> *From:* felix gao <gr...@gmail.com>
>> *To:* hdfs-user@hadoop.apache.org
>> *Sent:* Tue, April 12, 2011 7:46:31 AM
>> *Subject:* Question regarding datanode been wiped by hadoop
>>
>> What reason/condition would cause a datanode’s blocks to be removed? Our
>> cluster had a one of its datanodes crash because of bad RAM. After the
>> system was upgraded and the datanode/tasktracker brought online the next day
>> we noticed the amount of space utilized was minimal and the cluster was
>> rebalancing blocks to the datanode. It would seem the prior blocks were
>> removed. Was this because the datanode was declared dead? What is the
>> criteria for a namenode to decide (Assuming its the namenode) when a
>> datanode should remove prior blocks?
>>
>
>
Re: Question regarding datanode been wiped by hadoop
Posted by felix gao <gr...@gmail.com>.
The xml files have not been changed for more than two months, so that should
not be the reason. Even the in_use.lock is more than a month old. However,
we did shut it down few days ago and restarted it afterward. Then the
second shutdown might not be clean.
On Tue, Apr 12, 2011 at 7:52 AM, Ayon Sinha <ay...@yahoo.com> wrote:
> The datanode used the dfs config xml file to tell the datanode process,
> what disks are available for storage. Can you check that the config xml has
> all the partitions mentioned and has not been overwritten during the restore
> process?
>
> -Ayon
> See My Photos on Flickr <http://www.flickr.com/photos/ayonsinha/>
> Also check out my Blog for answers to commonly asked questions.<http://dailyadvisor.blogspot.com>
>
>
> ------------------------------
> *From:* felix gao <gr...@gmail.com>
> *To:* hdfs-user@hadoop.apache.org
> *Sent:* Tue, April 12, 2011 7:46:31 AM
> *Subject:* Question regarding datanode been wiped by hadoop
>
> What reason/condition would cause a datanode’s blocks to be removed? Our
> cluster had a one of its datanodes crash because of bad RAM. After the
> system was upgraded and the datanode/tasktracker brought online the next day
> we noticed the amount of space utilized was minimal and the cluster was
> rebalancing blocks to the datanode. It would seem the prior blocks were
> removed. Was this because the datanode was declared dead? What is the
> criteria for a namenode to decide (Assuming its the namenode) when a
> datanode should remove prior blocks?
>
Re: Question regarding datanode been wiped by hadoop
Posted by Ayon Sinha <ay...@yahoo.com>.
The datanode used the dfs config xml file to tell the datanode process, what
disks are available for storage. Can you check that the config xml has all the
partitions mentioned and has not been overwritten during the restore process?
-Ayon
See My Photos on Flickr
Also check out my Blog for answers to commonly asked questions.
________________________________
From: felix gao <gr...@gmail.com>
To: hdfs-user@hadoop.apache.org
Sent: Tue, April 12, 2011 7:46:31 AM
Subject: Question regarding datanode been wiped by hadoop
What reason/condition would cause a datanode’s blocks to be removed? Our
cluster had a one of its datanodes crash because of bad RAM. After the system
was upgraded and the datanode/tasktracker brought online the next day we noticed
the amount of space utilized was minimal and the cluster was rebalancing blocks
to the datanode. It would seem the prior blocks were removed. Was this
because the datanode was declared dead? What is the criteria for a namenode to
decide (Assuming its the namenode) when a datanode should remove prior blocks?
Re: Question regarding datanode been wiped by hadoop
Posted by Marcos Ortiz <ml...@uci.cu>.
El 4/12/2011 10:46 AM, felix gao escribió:
>
> What reason/condition would cause a datanode’s blocks to be removed?
> Our cluster had a one of its datanodes crash because of bad RAM.
> After the system was upgraded and the datanode/tasktracker brought
> online the next day we noticed the amount of space utilized was
> minimal and the cluster was rebalancing blocks to the datanode. It
> would seem the prior blocks were removed. Was this because the
> datanode was declared dead? What is the criteria for a namenode to
> decide (Assuming its the namenode) when a datanode should remove prior
> blocks?
>
1- Did you check the DataNode´s logs?
2- Did you protect the NameNode´s dfs.name.dir and the dfs.edits.dir ´s
directories?
On these directories, the NameNode stores the file system image and the
second is where the edit log or journal is written. A good practice for
these directories is to have them on RAID 1 or RAID 10 to guarantize the
consistency of your cluster.
Any data loss in these directories (dfs.name.dir and dfs.edits.dir)
will result in a loss of data in your HDFS. So, the second good practice
is to have a secondary NameNode to setup in any case that the primary
NameNode fails.
Another thing to keep in mind, is that when the NameNode fails, you have
to restar the JobTracker and the TaskTrackers after that the NameNode
will be restarted.
Regards
--
Marcos Luís Ortíz Valmaseda
Software Engineer (Large-Scaled Distributed Systems)
University of Information Sciences,
La Habana, Cuba
Linux User # 418229
Re: Question regarding datanode been wiped by hadoop
Posted by Matthew Foley <ma...@yahoo-inc.com>.
Here's another hypothesis. You'll have to check your namenode logs to see if it is the correct one.
In a healthy cluster, when one datanode goes down, 10 minutes later the namenode will "notice" it, and mark all its blocks as under-replicated (in namenode memory). It will then generate replication requests to other holders of those blocks, to get back to the required replication count for each. In a large cluster, this process can take as little as a few minutes, because there are many candidate senders and receivers for the copy operations.
When the datanode comes back a whole day later and sends its Initial Block Report, which we'll assume still has all the blocks it did before, since the disks were not corrupted, then all those blocks will be marked as OVER-replicated. The namenode will generate delete requests for one replica for each of those blocks.
Now, I would have expected those delete requests to be randomly distributed across all nodes holding replicas of those blocks. But there is some indication in the code that all the deletes may go to the new excess source, especially if the number of blocks is small (say, under one thousand). I'm not sure whether this is a sufficient explanation or not.
If the block deletes were ordered by the namenode, they may be in the namenode logs, with a prefix "hdfs.StateChange: BLOCK*", like this:
11/04/12 10:00:28 INFO hdfs.StateChange: BLOCK* NameSystem.chooseExcessReplicates: (127.0.0.1:49946, blk_-44847565016350316_1001) is added to recentInvalidateSets
11/04/12 10:00:28 INFO hdfs.StateChange: BLOCK* ask 127.0.0.1:49946 to delete blk_-44847565016350316_1001
Obviously you have to be logging at INFO level for these block-level manipulations to be in your namenode logs.
The datanode then echoes it with log lines like:
11/04/12 10:00:29 INFO datanode.DataNode: Scheduling block blk_-44847565016350316_1001 file .../current/finalized/blk_-44847565016350316 for deletion
11/04/12 10:00:29 INFO datanode.DataNode: Deleted block blk_-44847565016350316_1001 at file .../current/finalized/blk_-44847565016350316
Your log extract below shows that the datanode was logging at INFO level. See if your namenode was too, and if you can show that it generated a bunch of delete requests shortly after your repaired datanode came up.
Cheers,
--Matt
On Apr 12, 2011, at 9:30 AM, felix gao wrote:
mapred.local.dir is under /hadoop/mapred defs.data.dir is /hadoop./dfs. The logs showing
011-04-11 14:34:10,987 INFO org.apache.hadoop.ipc.Server: IPC Server listener on 50020: starting
2011-04-11 14:34:10,987 INFO org.apache.hadoop.ipc.Server: IPC Server handler 0 on 50020: starting
2011-04-11 14:34:10,988 INFO org.apache.hadoop.ipc.Server: IPC Server handler 1 on 50020: starting
2011-04-11 14:34:10,988 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: dnRegistration = DatanodeRegistration(had41.xxx:50010, storageID=DS-922075132-69.170.130.173-50010-1297386088418, infoPort=50075, ipcPort=50020)
2011-04-11 14:34:10,988 INFO org.apache.hadoop.ipc.Server: IPC Server handler 2 on 50020: starting
2011-04-11 14:34:11,021 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: DatanodeRegistration(69.170.130.173:50010<http://69.170.130.173:50010/>, storageID=DS-922075132-69.170.130.173-50010-1297386088418, infoPort=50075, ipcPort=50020)In DataNode.run, data = FSDataset{dirpath='/hadoop/dfs/current'}
2011-04-11 14:34:11,021 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: using BLOCKREPORT_INTERVAL of 3600000msec Initial delay: 0msec
2011-04-11 14:34:15,545 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: BlockReport of 110651 blocks got processed in 4493 msecs
2011-04-11 14:34:15,545 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: Starting Periodic block scanner.
2011-04-11 14:34:15,692 INFO org.apache.hadoop.hdfs.server.datanode.DataNode: Deleting block blk_-9213431071914219029_3395500 file /hadoop/dfs/current/subdir7/subdir26/blk_-9213431071914219029
then followed by all the blocks in the dfs/current directory. Seems to me hadoop just want to invalidate all the blocks in that box by deleting all of it.
On Tue, Apr 12, 2011 at 9:17 AM, Harsh J <ha...@cloudera.com>> wrote:
Apart from ensuring all that others here have said, are your
mapred.local.dir and dfs.data.dir pointing to the same directory by
any chance? If that so happens, the tasktracker could potentially wipe
out the datanode directories when restarted.
On Tue, Apr 12, 2011 at 8:16 PM, felix gao <gr...@gmail.com>> wrote:
> What reason/condition would cause a datanode’s blocks to be removed? Our
> cluster had a one of its datanodes crash because of bad RAM. After the
> system was upgraded and the datanode/tasktracker brought online the next day
> we noticed the amount of space utilized was minimal and the cluster was
> rebalancing blocks to the datanode. It would seem the prior blocks were
> removed. Was this because the datanode was declared dead? What is the
> criteria for a namenode to decide (Assuming its the namenode) when a
> datanode should remove prior blocks?
--
Harsh J
Re: Question regarding datanode been wiped by hadoop
Posted by felix gao <gr...@gmail.com>.
mapred.local.dir is under /hadoop/mapred defs.data.dir is /hadoop./dfs. The
logs showing
011-04-11 14:34:10,987 INFO org.apache.hadoop.ipc.Server: IPC Server
listener on 50020: starting
2011-04-11 14:34:10,987 INFO org.apache.hadoop.ipc.Server: IPC Server
handler 0 on 50020: starting
2011-04-11 14:34:10,988 INFO org.apache.hadoop.ipc.Server: IPC Server
handler 1 on 50020: starting
2011-04-11 14:34:10,988 INFO
org.apache.hadoop.hdfs.server.datanode.DataNode: dnRegistration =
DatanodeRegistration(had41.xxx:50010,
storageID=DS-922075132-69.170.130.173-50010-1297386088418, infoPort=50075,
ipcPort=50020)
2011-04-11 14:34:10,988 INFO org.apache.hadoop.ipc.Server: IPC Server
handler 2 on 50020: starting
2011-04-11 14:34:11,021 INFO
org.apache.hadoop.hdfs.server.datanode.DataNode: DatanodeRegistration(
69.170.130.173:50010,
storageID=DS-922075132-69.170.130.173-50010-1297386088418, infoPort=50075,
ipcPort=50020)In DataNode.run, data =
FSDataset{dirpath='/hadoop/dfs/current'}
2011-04-11 14:34:11,021 INFO
org.apache.hadoop.hdfs.server.datanode.DataNode: using BLOCKREPORT_INTERVAL
of 3600000msec Initial delay: 0msec
2011-04-11 14:34:15,545 INFO
org.apache.hadoop.hdfs.server.datanode.DataNode: BlockReport of 110651
blocks got processed in 4493 msecs
2011-04-11 14:34:15,545 INFO
org.apache.hadoop.hdfs.server.datanode.DataNode: Starting Periodic block
scanner.
2011-04-11 14:34:15,692 INFO
org.apache.hadoop.hdfs.server.datanode.DataNode: Deleting block
blk_-9213431071914219029_3395500 file
/hadoop/dfs/current/subdir7/subdir26/blk_-9213431071914219029
then followed by all the blocks in the dfs/current directory. Seems to me
hadoop just want to invalidate all the blocks in that box by deleting all of
it.
On Tue, Apr 12, 2011 at 9:17 AM, Harsh J <ha...@cloudera.com> wrote:
> Apart from ensuring all that others here have said, are your
> mapred.local.dir and dfs.data.dir pointing to the same directory by
> any chance? If that so happens, the tasktracker could potentially wipe
> out the datanode directories when restarted.
>
> On Tue, Apr 12, 2011 at 8:16 PM, felix gao <gr...@gmail.com> wrote:
> > What reason/condition would cause a datanode’s blocks to be removed?
> Our
> > cluster had a one of its datanodes crash because of bad RAM. After the
> > system was upgraded and the datanode/tasktracker brought online the next
> day
> > we noticed the amount of space utilized was minimal and the cluster was
> > rebalancing blocks to the datanode. It would seem the prior blocks were
> > removed. Was this because the datanode was declared dead? What is the
> > criteria for a namenode to decide (Assuming its the namenode) when a
> > datanode should remove prior blocks?
>
> --
> Harsh J
>
Re: Question regarding datanode been wiped by hadoop
Posted by Harsh J <ha...@cloudera.com>.
Apart from ensuring all that others here have said, are your
mapred.local.dir and dfs.data.dir pointing to the same directory by
any chance? If that so happens, the tasktracker could potentially wipe
out the datanode directories when restarted.
On Tue, Apr 12, 2011 at 8:16 PM, felix gao <gr...@gmail.com> wrote:
> What reason/condition would cause a datanode’s blocks to be removed? Our
> cluster had a one of its datanodes crash because of bad RAM. After the
> system was upgraded and the datanode/tasktracker brought online the next day
> we noticed the amount of space utilized was minimal and the cluster was
> rebalancing blocks to the datanode. It would seem the prior blocks were
> removed. Was this because the datanode was declared dead? What is the
> criteria for a namenode to decide (Assuming its the namenode) when a
> datanode should remove prior blocks?
--
Harsh J
Re: Question regarding datanode been wiped by hadoop
Posted by Boris Shkolnik <bo...@yahoo-inc.com>.
One thing to consider.. If the node was down for a day all of its blocks could’ve been replicated to other datanodes.
When machine is brought back , these blocks become overreplicated and NameNode decides to delete them.
You should check the logs of both DataNode and Namenode to see if it could be the case.
Boris.
On 4/12/11 7:46 AM, "felix gao" <gr...@gmail.com> wrote:
What reason/condition would cause a datanode’s blocks to be removed? Our cluster had a one of its datanodes crash because of bad RAM. After the system was upgraded and the datanode/tasktracker brought online the next day we noticed the amount of space utilized was minimal and the cluster was rebalancing blocks to the datanode. It would seem the prior blocks were removed. Was this because the datanode was declared dead? What is the criteria for a namenode to decide (Assuming its the namenode) when a datanode should remove prior blocks?