[infinispan-dev] Key set, value set and size within transaction

"이희승 (Trustin Lee)" trustin at gmail.com
Thu Oct 21 10:57:46 EDT 2010


I spotted it too while investigating the behavior of SizeCommand and
EntrySetCommand.

Map.size() doesn't need to be accurate, but we can get accurate enough
value by just subtracting the number of removed entries in the current
transaction.  SizeCommand therefore can be optimized.

KeySetCommand is a different beast.  We need to be accurate in the same
transaction:

    tx.begin();
    cache.put("newKey", "newValue");
    assert cache.keySet().contains("newKey");

The current KeySetCommand implementation uses
getKeySetWithinTransaction() and it creates a new set as you pointed
out.  We could write a wrapper set to minimize memory consumption.  Same
applies to EntrySetCommand.

However, if someone ever passes the keySet or entrySet to other thread,
it will be a disaster.  Perhaps we should document it clearly?

Manik Surtani wrote:
> Guys
> 
> I just noticed this in the codebase:
> 
> http://fisheye.jboss.org/browse/Infinispan/branches/4.2.x/core/src/main/java/org/infinispan/commands/read/AbstractLocalCommand.java?r=2518#l43
> 
> This seems to have come in as a part of changeset 2518 - http://fisheye.jboss.org/changelog/Infinispan/?cs=2518 - for ISPN-679 (Cache.values does not work correctly).
> 
> Please be careful when doing stuff like this - the new method added to AbstractLocalCommand effectively creates a new collection the size of the entire data container and loads it with (almost) all keys.  An Infinispan node running close to capacity will almost certainly die with an OOM if you do this.  It will even OOM if you are not necessarily close to capacity (under half your heap used) and you have > 1 thread concurrently invoking this method.  
> 
> I presume this is to make Cache.keySet(), Cache.values() and Cache.size() transactionally accurate, but remember that this comes at a huge cost.  Which is why we chose to go with a best-effort approach.  See Javadocs on Cache [1].  The relevant bit states:
> 
> "Other methods such as Map.size() provide an approximation-only, and should not be relied on for an accurate picture as to the size of the entire, distributed cache. Remote nodes are not queried and in-fly transactions are not taken into account, even if Map.size() is invoked from within such a transaction."
> 
> Could we please discuss why we have added this?
> 
> Cheers
> Manik
> 
> [1] http://docs.jboss.org/infinispan/4.2/apidocs/org/infinispan/Cache.html
> --
> Manik Surtani
> manik at jboss.org
> Lead, Infinispan
> Lead, JBoss Cache
> http://www.infinispan.org
> http://www.jbosscache.org
> 
> 
> 
> 
> 
> _______________________________________________
> infinispan-dev mailing list
> infinispan-dev at lists.jboss.org
> https://lists.jboss.org/mailman/listinfo/infinispan-dev

-- 
Trustin Lee - http://gleamynode.net/


-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 290 bytes
Desc: OpenPGP digital signature
Url : http://lists.jboss.org/pipermail/infinispan-dev/attachments/20101021/6232fd94/attachment.bin 


More information about the infinispan-dev mailing list