[
https://issues.apache.org/jira/browse/HBASE-6651?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Hiroshi Ikeda updated HBASE-6651:
---------------------------------
Attachment: HBASE-6651.patch
Added a patch.
Added a new interface SharedMap, and added its implementations and unit tests,
and made HBaseClient and HTablePool use SharedMap instead of PoolMap. The API
of HTablePool is marked as “stable” and some constructors of HTablePool uses a
static nested class PoolMap.PoolType, and I’m not sure how to add Deprecated
marker partially to PoolMap, so I did nothing about it.
I don’t have environment to run the whole test of HBase, and except tests added
for SharedMap I only confirm that Eclipse can compile the patched code (where
the projects are imported as “existing maven projects” from the trunk). I hope
it will work well.
I purged ThreadLocal base logic from HBaseClient (that logic is terrible and I
never believe that works well). Instead, I complicate the Round-Robin based
logic to prefer to give idle connections to requesters. I think, this is the
same reason why the ThreadLocal based logic makes high performance. Moreover,
in order to find idle objects in the pool, SharedMap has different methods from
PoolMap, and I rewrote and fixed many parts of HBaseClient. On the other hand
fixing HTablePool to SharedMap is straightforward and I think I have no
additional explanations about it.
> Thread safety of HTablePool is doubtful
> ---------------------------------------
>
> Key: HBASE-6651
> URL: https://issues.apache.org/jira/browse/HBASE-6651
> Project: HBase
> Issue Type: Bug
> Components: Client
> Affects Versions: 0.94.1
> Reporter: Hiroshi Ikeda
> Priority: Minor
> Attachments: HBASE-6651.patch, sample.zip, sample.zip,
> sharedmap_for_hbaseclient.zip
>
>
> There are some operations in HTablePool to access to PoolMap in multiple
> times without any explict synchronization.
> For example HTablePool.closeTablePool() calles PoolMap.values(), and calles
> PoolMap.remove(). If other threads add new instances to the pool in the
> middle of the calls, the new added instances might be dropped.
> (HTablePool.closeTablePool() also has another problem that calling it by
> multple threads causes accessing HTable by multiple threads.)
> Moreover, PoolMap is not thread safe for the same reason.
> For example PoolMap.put() calles ConcurrentMap.get() and calles
> ConcurrentMap.put(). If other threads add a new instance to the concurent map
> in the middle of the calls, the new instance might be dropped.
> And also implementations of Pool have the same problems.
--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira
