...but then look at my just sent mail "JobTracker does not start when
specifying a Windows path for hadoop.tmp.dir": If you specify a Windows
path for "hadoop.tmp.dir" then the JobTracker does not start.
Michael Stack wrote:
Great. Thanks. I added it to the wiki.
St.Ack
Holger Stenzhorn wrote:
Ok, here is my first shot at it:
--- SNIP ---
3.1. Set the hadoop tmp directory property, hadoop.tmp.dir (For
Windows/Cygwin users: If you do not include this variable in the site
file then the value /tmp/hadoop-${user.name} from hadoop-default.xml
is used. But since this does not work properly for Windows/Cygwin you
have to specify a Windows-formatted path for hadoop.tmp.dir such as
c:\tmp\hbase)
--- SNIP ---
Cheers,
Holger
Michael Stack wrote:
Holger Stenzhorn wrote:
Perhaps you could add to the "10 minutes" page at point 3.1 an
example for Unix and Windows/Cygwin to make the different path
handling a bit clearer.
Sounds like a good idea to me Holger. Would you mind taking a first
cut at it since you've had first hand-experience of where the doc.
is lacking?
St.Ack
Cheers,
Holger
Michael Stack wrote:
Holger Stenzhorn wrote:
Hadoop RunJar looks to contain the only use of hadoop.tmp.dir
other than test and contrib code. It does the following: 'new
File(new Configuration().get("hadoop.tmp.dir"));' using
java.io.File. In hbase, its all Path and hadoop FileSystem.
So, yes, content of hadoop.tmp.dir is treated differently.
Could that be changed? ...meaning, wouldn't it be sensible to
change HBase (to use Path etc.) so it behaves like Hadoop?
You mean, can hadoop be changed? (Its hbase that's drunk the
kool-aid; its the one that goes via hadoop FileSystem for all
filesystem accesses whereas the hadoop use of hadoop.tmp.dir
goes via java.io.File).
Nope. I mean the other way around: Can you change the HBase code
so it goes via the Hadoop filesystem? ...or would this imply a
performance drop for HBase?
It already does (See my comments above).
What happens if you clean up your local directory removing
${hadoop.tmp.dir} and retry?
Again, if I use "c:\tmp\hadoop-holste" as value for
"${hadoop.tmp.dir}" it works: The log reflects just the situation
when this variable is not set in "hbase-site.xml" - then it does
not work.
I also forgot to mention in my last my that it is not just
"${hadoop.tmp.dir}" set to "/tmp/hadoop-${user.name}" that causes
the problem but also if it is set to "/tmp/hadoop-holste".
Did you try my suggestion of removing anything that might once
have been a value for hadoop.tmp.dir?
Testing here on cygwin using local filesystem and all defaults, it
all just works fine. I'm disturbed by the state you seem to be in
where the .META. table is not being deployed AND there is no
exception saying why. The only think I can think is that
previous runs have left a horked -ROOT- region on the filesystem.
Good on you Holger,
St.Ack
Thanks,
St.Ack
Did you just change the value of hadoop.tmp.dir from
/tmp/hadoop-${user.name} to /tmp/hadoop-holste on cygwin and
it then worked?
No. I changed it to changed the value "c:\tmp\hadoop-holste".
Also, I tried to do the very same (as described in my
intitial mail) on Ubuntu also *without* specifying
"hadoop.tmp.dir".
There it works wihtout problems...
Yeah. I know things work out-of-the-box on linux/macosx. Jim
has been doing a bunch of work trying to get the same to
happen on cygwin (10Minutes has the latest fruit of his work).
As requested by Michel Stack I also attach my log (turned to
debug) when starting the server.
Thanks for the log. Comments interlaced.
Cheers,
Holger
07/11/02 13:40:09 DEBUG conf.Configuration:
java.io.IOException: config()
at
org.apache.hadoop.conf.Configuration.<init>(Configuration.java:144)
at
org.apache.hadoop.hbase.HBaseConfiguration.<init>(HBaseConfiguration.java:30)
at org.apache.hadoop.hbase.HMaster.doMain(HMaster.java:3043)
at org.apache.hadoop.hbase.HMaster.main(HMaster.java:3097)
...
