Matteo Bertozzi created HBASE-6992:
--------------------------------------
Summary: Coprocessors semantic issues: post async operations,
helper methods, ...
Key: HBASE-6992
URL: https://issues.apache.org/jira/browse/HBASE-6992
Project: HBase
Issue Type: Brainstorming
Components: Coprocessors
Affects Versions: 0.94.2, 0.92.2, 0.96.0
Reporter: Matteo Bertozzi
Assignee: Matteo Bertozzi
Discussion ticket around coprocessor pre/post semantic.
For each rpc in HMaster we have a pre/post operation that allows a coprocessor
to execute some code before and after the operation
* preOperation()
* my operation
* postOperation()
This is used for example by the AccessController to verify if the user can
execute or not the operation.
Everything is fine, unless the master operation is asynchronous (like
create/delete table)
* preOperation()
* executor.submit(new OperationHandler())
* postOperation()
The pre operation is still fine, since is executed before the operation and
need to throw exceptions to the client in case of failures...
The post operation, instead, is no longer post... is just post submit. And if
someone subscribe to postCreateTable() the notification can arrive before the
table creation.
To "solve" this problem, HBASE-5584 added pre/post handlers and now the
situation looks like this:
{code}
client request client response
| |
+------+-- submit op --+----+--- (HMaster)
pre op post op
(executor) +---- handler ----+
pre handler post handler
{code}
Now, we've two types of pre/post operation and the semantical correct are
preOperation() and postOperationHandler()
since the preOperation() needs to reply to the client (e.g AccessController
NotAllowException) and the postOperatioHandler() is really post operation.
postOperation() is not post... and preOperationHandler() can't communicate with
the client.
The AccessController coprocessor uses the postOperation() that is fine for the
sync operation like addColumn(), deleteColumn()... but in case of failure of
async operations like deleteTable() we've removed rights that we still need.
I think that we should get back just to the single pre/post operation but with
the right semantic...
Other then the "when is executed" problem, we've also functions that can be
described with other simpler functions
for example: modifyTable() is just a helper to avoid multiple
addColumn()/deleteColumn() calls
but the problem here is that modifyTable() has its own pre/post operation() and
if I've implemented the pre/post addColumn I don't get notified when I call
modifyTable(). This is another problem in the access controller coprocessor
In this case I'm not sure what the best solution can be... but in this way,
adding new helper methods means breaking the coprocessors, because they don't
get notified even if something is changed...
Any idea, thoughts, ...?
--
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
