[HDFS-user] NameNode crash - cannot start dfs - need help

Hi Matt,

If you want to keep your recent edits, you'll have to place an 0xFF at the
beginning of the most recent edit entry in the edit log. It's a bit tough to
find these boundaries, but you can try applying this patch and rebuilding:

https://issues.apache.org/jira/browse/hdfs-1378

This will tell you the offset of the broken entry ("recent opcodes") and you
can put an 0xff there to tie off the file before the corrupt entry.

-Todd

Thank you Todd.

It does indeed seem like a challenge to find a record boundary, but if I wanted to do it... here is how I did it in case others are interested in doing the same.

It looks like that value (0xFF) is referenced as OP_INVALID in the source file: [hadoop-dist]/src//hdfs/org/apache/hadoop/hdfs/server/namenode/FSEditLog.java.

Every record begins with an op code that describes the record. The op codes are in the range [0,14] (inclusive), except for OP_INVALID. Each record type (based on op code) appears to have a different format. Additionally, it seems that the code for each record type has several code paths to support different versions of the hdfs.

I looked in the error messages, and found the line number of the exception within the switch statement in the code (in this case, line 563). That told me that I was looking for an op code of either 0x00 or 0x09. I noticed that this particular code path had a record type that looked like this:
[# bytes: name]

[1:op code][4:int length][2:file system path length][?:file system path text]

All I had to do was find a filesystem path, and look 7 bytes before it started. If the op code was a 0x00 or 0x09, then this was a candidate record.

It would have been easier to just search for something from the error message (i.e. "12862" for me) to find candidate records, but in my case that was in almost every record. Additionally, it would have also been easier to just search for instances of the op code, but in my case one of the op codes (0x00) appears too often in the data to make that useful. If your op code is 0x03 for example, you will probably have a much easier time of it than I did.

I was able to successfully and quickly find record boundaries and replace the op code with 0xff. After a few records I was back to the NPE exception that I was getting with a zero length edits file:

2010-10-05 16:47:39,670 INFO org.apache.hadoop.hdfs.server.common.Storage: Edits file /mnt/name/current/edits of size 157037 edits # 959 loaded in 0 seconds.
2010-10-05 16:47:39,671 ERROR org.apache.hadoop.hdfs.server.namenode.NameNode: java.lang.NullPointerException
at org.apache.hadoop.hdfs.server.namenode.FSDirectory.addChild(FSDirectory.java:1081)
at org.apache.hadoop.hdfs.server.namenode.FSDirectory.addChild(FSDirectory.java:1093)
at org.apache.hadoop.hdfs.server.namenode.FSDirectory.addNode(FSDirectory.java:996)
at org.apache.hadoop.hdfs.server.namenode.FSDirectory.unprotectedAddFile(FSDirectory.java:199)
at org.apache.hadoop.hdfs.server.namenode.FSEditLog.loadFSEdits(FSEditLog.java:627)
at org.apache.hadoop.hdfs.server.namenode.FSImage.loadFSEdits(FSImage.java:1022)
at org.apache.hadoop.hdfs.server.namenode.FSImage.loadFSImage(FSImage.java:830)
at org.apache.hadoop.hdfs.server.namenode.FSImage.recoverTransitionRead(FSImage.java:378)
at org.apache.hadoop.hdfs.server.namenode.FSDirectory.loadFSImage(FSDirectory.java:92)

One hurdle down, how do I get past the next one?

(BTW, what if I didn't want to keep my recent edits, and just wanted to start up the namenode? This is currently expensive downtime; I'd rather lose a small amount of data and be up and running than continue the down time).

Thank you for your help,

Matthew

error there (by truncating a few edits from the end), then the later edits
in edits.new failed, because they depended on a path that should have been
created by "edits".
If you really want to do this, you can remove "edits.new", and replace
"edits" with a file containing hex 0xffffffeeff I believe (edits header plus
OP_INVALID)

-Todd

Oct 5, 2010, at 8:42 AM, Todd Lipcon wrote:

Thank you Ayon, Allen and Todd for your suggestions.

I was tempted to try to find the offending records in edits.new, but opted for simply moving the file instead. I kept the recently edited edits file in place.

The namenode started up this time with no exceptions and appears to be running well; hadoop fsck / reports a healthy filesystem.

Thank you,

Matthew

Hi Matthew,
Congratulations. Having a HDFS back is quite a relief and you were lucky enough
to not loose any files/blocks.
Another thing I ended up doing was to decommission the namenode machine from
being a data node. That is what had caused the namenode to get out of disk
space.
-Ayon





________________________________
From: Matthew LeMieux <mdl@mlogiciels.com>
To: hdfs-user@hadoop.apache.org
Sent: Tue, October 5, 2010 11:25:57 AM
Subject: Re: NameNode crash - cannot start dfs - need help

Thank you Ayon, Allen and Todd for your suggestions.

I was tempted to try to find the offending records in edits.new, but opted for
simply moving the file instead. I kept the recently edited edits file in
place.

The namenode started up this time with no exceptions and appears to be running
well; hadoop fsck / reports a healthy filesystem.

Thank you,

Matthew

On Oct 5, 2010, at 10:09 AM, Todd Lipcon wrote:
It's unclear whether you're getting the error in "edits" or "edits.new". From
the above, I'm guessing maybe "edits" is corrupt, so when you fixed the error
there (by truncating a few edits from the end), then the later edits in
edits.new failed, because they depended on a path that should have been created
by "edits".
If you really want to do this, you can remove "edits.new", and replace "edits"
with a file containing hex 0xffffffeeff I believe (edits header plus OP_INVALID)

-Todd

Oct 5, 2010, at 8:42 AM, Todd Lipcon wrote:

Hi Matthew,
"(BTW, what if I didn't want to keep my recent edits, and just wanted to start
up the namenode? This is currently expensive downtime; I'd rather lose a small
amount of data and be up and running than continue the down time). "
This was exactly my use-case as well. I chose small data loss over spending
hours on end trying to get past the exceptions.
Try this:
rename the 4 files under /mnt/name/current to something like *.corrupt

then copy over the 4 files from /mnt/namesecondarynode/current
Make sure you have enough space on the namenode box.
Try starting the namenode. It worked for me. I was at the same place as you only
a week ago.
-Ayon





________________________________
From: Matthew LeMieux <mdl@mlogiciels.com>
To: hdfs-user@hadoop.apache.org
Sent: Tue, October 5, 2010 9:58:53 AM
Subject: Re: NameNode crash - cannot start dfs - need help

Thank you Todd.

It does indeed seem like a challenge to find a record boundary, but if I wanted
to do it... here is how I did it in case others are interested in doing the
same.


It looks like that value (0xFF) is referenced as OP_INVALID in the source file:
[hadoop-dist]/src//hdfs/org/apache/hadoop/hdfs/server/namenode/FSEditLog.java.

Every record begins with an op code that describes the record. The op codes are
in the range [0,14] (inclusive), except for OP_INVALID. Each record type (based
on op code) appears to have a different format. Additionally, it seems that the
code for each record type has several code paths to support different versions
of the hdfs.

I looked in the error messages, and found the line number of the exception
within the switch statement in the code (in this case, line 563). That told me
that I was looking for an op code of either 0x00 or 0x09. I noticed that this
particular code path had a record type that looked like this:
[# bytes: name]

[1:op code][4:int length][2:file system path length][?:file system path text]

All I had to do was find a filesystem path, and look 7 bytes before it started.
If the op code was a 0x00 or 0x09, then this was a candidate record.

It would have been easier to just search for something from the error message
(i.e. "12862" for me) to find candidate records, but in my case that was in
almost every record. Additionally, it would have also been easier to just
search for instances of the op code, but in my case one of the op codes (0x00)
appears too often in the data to make that useful. If your op code is 0x03 for
example, you will probably have a much easier time of it than I did.

I was able to successfully and quickly find record boundaries and replace the op
code with 0xff. After a few records I was back to the NPE exception that I was
getting with a zero length edits file:

2010-10-05 16:47:39,670 INFO org.apache.hadoop.hdfs.server.common.Storage: Edits
file /mnt/name/current/edits of size 157037 edits # 959 loaded in 0 seconds.
2010-10-05 16:47:39,671 ERROR org.apache.hadoop.hdfs.server.namenode.NameNode:
java.lang.NullPointerException
at
org.apache.hadoop.hdfs.server.namenode.FSDirectory.addChild(FSDirectory.java:1081)

at
org.apache.hadoop.hdfs.server.namenode.FSDirectory.addChild(FSDirectory.java:1093)

at
org.apache.hadoop.hdfs.server.namenode.FSDirectory.addNode(FSDirectory.java:996)
at
org.apache.hadoop.hdfs.server.namenode.FSDirectory.unprotectedAddFile(FSDirectory.java:199)

at
org.apache.hadoop.hdfs.server.namenode.FSEditLog.loadFSEdits(FSEditLog.java:627)
at
org.apache.hadoop.hdfs.server.namenode.FSImage.loadFSEdits(FSImage.java:1022)
at
org.apache.hadoop.hdfs.server.namenode.FSImage.loadFSImage(FSImage.java:830)
at
org.apache.hadoop.hdfs.server.namenode.FSImage.recoverTransitionRead(FSImage.java:378)

at
org.apache.hadoop.hdfs.server.namenode.FSDirectory.loadFSImage(FSDirectory.java:92)


One hurdle down, how do I get past the next one?

(BTW, what if I didn't want to keep my recent edits, and just wanted to start up
the namenode? This is currently expensive downtime; I'd rather lose a small
amount of data and be up and running than continue the down time).

Thank you for your help,

Matthew

On Oct 5, 2010, at 8:42 AM, Todd Lipcon wrote:

Hi Matt,

No 2nd copy of the edits file/another entry in dfs.name.dir?

No second copy. There will be from now on, but that doesn't get me out of the current hole.

I'm not too concerned with recent edits. I'd be much happier losing recent edits if I could just get the thing to start!

So, the question is, how do I tell the name node to just start up no matter what?

-Matthew

Have you tried getting rid of the edits.new file completely (by renaming it to
something else)?
-Ayon





________________________________
From: Matthew LeMieux <mdl@mlogiciels.com>
To: hdfs-user@hadoop.apache.org
Sent: Tue, October 5, 2010 10:14:21 AM
Subject: Re: NameNode crash - cannot start dfs - need help

No second copy. There will be from now on, but that doesn't get me out of the
current hole.


I'm not too concerned with recent edits. I'd be much happier losing recent
edits if I could just get the thing to start!

So, the question is, how do I tell the name node to just start up no matter
what?

-Matthew

We had almost exact problem of namenode filling up and namnode failing at this
exact same point. Since you have created space now you can copy over the
edits.new, fsimage and the other 2 files from your
/mnt/namesecondarynode/current and try restarting the namenode.
I believe you will loose some edits and probably some blocks of some files but
we could recover most of our files.
-Ayon





________________________________
From: Matthew LeMieux <mdl@mlogiciels.com>
To: hdfs-user@hadoop.apache.org
Sent: Tue, October 5, 2010 8:16:15 AM
Subject: NameNode crash - cannot start dfs - need help

The namenode on an otherwise very stable HDFS cluster crashed recently. The
filesystem filled up on the name node, which I assume is what caused the crash.
The problem has been fixed, but I cannot get the namenode to restart. I am
using version CDH3b2 (hadoop-0.20.2+320).

The error is this:

2010-10-05 14:46:55,989 INFO org.apache.hadoop.hdfs.server.common.Storage: Edits
file /mnt/name/current/edits of size 157037 edits # 969 loaded in 0 seconds.
2010-10-05 14:46:55,992 ERROR org.apache.hadoop.hdfs.server.namenode.NameNode:
java.lang.NumberFormatException: For input string: "12862^@^@^@^@^@^@^@^@"
at
java.lang.NumberFormatException.forInputString(NumberFormatException.java:48)
at java.lang.Long.parseLong(Long.java:419)
at java.lang.Long.parseLong(Long.java:468)
at
org.apache.hadoop.hdfs.server.namenode.FSEditLog.readLong(FSEditLog.java:1355)
at
org.apache.hadoop.hdfs.server.namenode.FSEditLog.loadFSEdits(FSEditLog.java:563)
at
org.apache.hadoop.hdfs.server.namenode.FSImage.loadFSEdits(FSImage.java:1022)
...

This page (http://wiki.apache.org/hadoop/TroubleShooting) recommends editing the
edits file with a hex editor, but does not explain where the record boundaries
are. It is a different exception, but seemed like a similar cause, the edits
file. I tried removing a line at a time, but the error continues, only with a
smaller size and edits #:

2010-10-05 14:37:16,635 INFO org.apache.hadoop.hdfs.server.common.Storage: Edits
file /mnt/name/current/edits of size 156663 edits # 966 loaded in 0 seconds.
2010-10-05 14:37:16,638 ERROR org.apache.hadoop.hdfs.server.namenode.NameNode:
java.lang.NumberFormatException: For input string: "12862^@^@^@^@^@^@^@^@"
at
java.lang.NumberFormatException.forInputString(NumberFormatException.java:48)
at java.lang.Long.parseLong(Long.java:419)
at java.lang.Long.parseLong(Long.java:468)
at
org.apache.hadoop.hdfs.server.namenode.FSEditLog.readLong(FSEditLog.java:1355)
at
org.apache.hadoop.hdfs.server.namenode.FSEditLog.loadFSEdits(FSEditLog.java:563)
at
org.apache.hadoop.hdfs.server.namenode.FSImage.loadFSEdits(FSImage.java:1022)
...

I tried removing the edits file altogether, but that failed
with: java.io.IOException: Edits file is not found

I tried with a zero length edits file, so it would at least have a file there,
but that results in an NPE:

2010-10-05 14:52:34,775 INFO org.apache.hadoop.hdfs.server.common.Storage: Edits
file /mnt/name/current/edits of size 0 edits # 0 loaded in 0 seconds.
2010-10-05 14:52:34,776 ERROR org.apache.hadoop.hdfs.server.namenode.NameNode:
java.lang.NullPointerException
at
org.apache.hadoop.hdfs.server.namenode.FSDirectory.addChild(FSDirectory.java:1081)

at
org.apache.hadoop.hdfs.server.namenode.FSDirectory.addChild(FSDirectory.java:1093)

at
org.apache.hadoop.hdfs.server.namenode.FSDirectory.addNode(FSDirectory.java:996)
at
org.apache.hadoop.hdfs.server.namenode.FSDirectory.unprotectedAddFile(FSDirectory.java:199)



Most if not all the files I noticed in the edits file are temporary files that
will be deleted once this thing gets back up and running anyway. There is a
closed ticket that might be
related: https://issues.apache.org/jira/browse/HDFS-686 , but the version I'm
using seems to already have HDFS-686 (according
to http://archive.cloudera.com/cdh/3/hadoop-0.20.2+320/changes.html)

What do I have to do to get back up and running?

Thank you for your help,

Matthew