[
https://issues.apache.org/jira/browse/HBASE-3766?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13018582#comment-13018582
]
Karthick Sankarachary commented on HBASE-3766:
----------------------------------------------
{quote}About the patch, I see you increment in HTable's constructor, but if the
user is reusing a conf object then the counter becomes wrong right?{quote}
Let's say that a certain conf object is used to create 2 HTable instances. As a
result, the reference count on the underlying connection object will be
incremented twice, which is the desired behavior. Then, when you close one of
the HTables, it will decrement the reference count on the connection, but not
close it, since it is still in-use. Later, when you close the other HTable, the
reference count will drop to zero, at which point, the underlying connection
gets deleted. On the other hand, if you use different conf objects to create
the HTables, they will end up having separate reference counts. Please correct
me if I'm misunderstood the question.
{quote}Also the counter itself isn't accessed in a synchronize way (as multiple
threads using different HTables would need) so this could pose a problem.{quote}
Good catch. I revised the patch such that the reference count is accessed
atomically.
> Manage Connections Through Reference Counts
> -------------------------------------------
>
> Key: HBASE-3766
> URL: https://issues.apache.org/jira/browse/HBASE-3766
> Project: HBase
> Issue Type: Improvement
> Components: client
> Affects Versions: 0.90.2
> Reporter: Karthick Sankarachary
> Assignee: Karthick Sankarachary
> Priority: Minor
> Fix For: 0.90.3
>
> Attachments: HBASE-3766.patch, HBASE-3766.patch
>
> Original Estimate: 1h
> Remaining Estimate: 1h
>
> As of now, the onus is on the developer to explicitly close connections once
> they're done using them. Furthermore, since connections are managed based on
> the identity of the {{Configuration}} object, one is forced to clone the
> configuration object in order to be able to clean it up safely (for a case in
> point, see {{HTablePool's}} constructor). As a matter of fact, this issue has
> been well-documented in the HConnectionManager class:
> {quote}
> But sharing connections makes clean up of {{HConnection}} instances a little
> awkward. Currently, clients cleanup by calling
> {{#deleteConnection(Configuration, boolean)}}. This will shutdown the
> zookeeper connection the {{HConnection}} was using and clean up all
> {{HConnection}} resources as well as stopping proxies to servers out on the
> cluster. Not running the cleanup will not end the world; it'll just stall the
> closeup some and spew some zookeeper connection failed messages into the log.
> Running the cleanup on a {{HConnection}} that is subsequently used by
> another will cause breakage so be careful running cleanup. To create a
> {{HConnection}} that is not shared by others, you can create a new
> {{Configuration}} instance, pass this new instance to
> {{#getConnection(Configuration)}}, and then when done, close it up by doing
> something like the following:
> {code}
> Configuration newConfig = new Configuration(originalConf);
> HConnection connection = HConnectionManager.getConnection(newConfig);
> // Use the connection to your hearts' delight and then when done...
> HConnectionManager.deleteConnection(newConfig, true);
> {code}
> {quote}
> Here, we propose a reference-count based mechanism for managing connections
> that will allow {{HTables}} to clean up after themselves. In particular, we
> extend the {{HConnectionInterface}} interface so as to facilitate reference
> counting, where, typically, a reference indicates that it is being used by a
> {{HTable}}, although there could be other sources.
> To elaborate, when a HTable is constructed, it increments the reference count
> on the connection given to it. Similarly, when it is closed, that reference
> count is decremented. In the event there are no more references to that
> connection, {{HTable#close}} takes it upon itself to delete the connection,
> thereby sparing the developer from doing so.
--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira
