hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From youknow...@heroicefforts.net
Subject Re: problems with theUnsafe
Date Sun, 10 Nov 2013 02:34:41 GMT

I don't have access to the system at the moment (SE 6 RHEL 64-bit).  Basically, invoking a
client Get operation causes the app server / JRE to core dump. I wasn't able to find any specific
IBM bug associated with this behavior, but their issue tracking is not exactly transparent.
 I don't think that this will work:
try { } catch(CoreDump) {  :D  }
Avro and others allow explicit avoidance:
if("false".equals(System.getProperty("hbase.disable.unsafe", "false"))
//attempt unsafe impl
//use pure impl
An alternative might be to whitelist java.vendor.
However, I think the explicit solution is better than trying to be too clever.
-----Original Message-----
From: "Ted Yu" <yuzhihong@gmail.com>
Sent: Saturday, November 9, 2013 8:28pm
To: "dev@hbase.apache.org" <dev@hbase.apache.org>
Subject: Re: problems with theUnsafe

bq. The IBM JRE supplies Unsafe, but apparently it doesn't work

Can you give us a bit more detail (in this case, UnsafeComparer) ? Is there
a BUG associated with it ?

If there is a known procedure / test by which we can tell whether
works or not, that would help make the decision of switching to pure Java


On Sat, Nov 9, 2013 at 4:24 PM, <youknowwho@heroicefforts.net> wrote:

> Our WebSphere Application Server started core dumping immediately after we
> merged our hbase feature branch to our mainline.  I've tracked this down to
> Bytes use of theUnsafe.  The IBM JRE supplies Unsafe, but apparently it
> doesn't work worth a damn.  I managed to temporarily work "around" the
> issue using some AOP wizardry.  Would it be possible to add some property
> check to the static initializer to allow one to explicitly opt for the pure
> Java comparer?
> http://grepcode.com/file/repo1.maven.org/maven2/org.apache.hbase/hbase/0.94.7/org/apache/hadoop/hbase/util/Bytes.java#Bytes.LexicographicalComparerHolder.getBestComparer%28%29
> thanks,
> -Jess
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message