Ryunix wrote:
Hi Phil,
"testOnBorrow tests the connection before only when it is borrowed from
the pool. I am not a Hibernate expert, but to understand exactly what
is going on here, you need to determine when connections are being
borrowed from the pool vs being reused by your application or Hibernate
without being returned to the pool. If a connection is reused without
being returned and re-borrowed from the pool, dbcp cannot validate it."
That what we are currently suspecting of, we suspect the error occur in the
sequence of action as below :
1. Application open a new connection but failed to close it.
2. That particular connection is recognized by dbcp as 'active connection',
and hence cannot be cleared by idle object evictor.
3. as time goes by, the connection hit the oracle time out.
4. at the next webservice call the application used that particular
connection (which already closed), and hence the error occur.
How does the application get the connection in 4? If it gets it from
the pool, it will be validated before the application gets it.
do you think the above suspicion valid?, I was having the impression when
the item '3' took place, the dbcp could validate the connection before
return it to my application, but as you highlighted, 'if the connection is
never returned to the pool, it will not be validated'. but it is possible
for the dbcp to give that connection to my application (since it is still
accounted under 'activeConnection', though it is not in used)?
No, dbcp will not hand out multiple handles to the same connection. If
a client has the connection checked out, it will not be given to another
client.
Phil
Phil Steitz wrote:
Ryunix wrote:
Hi Carman and Phil,
thanks for the effort put for this issue.
please excuse my lengthy explanation on the problem which lead me my
question.
Our application provides webservices to be used by various application in
our client's environment and generally works fine. However, if there is a
firewall between our application and the database, some problem occur. at
certain webservice call, it will throw the error message 'no more data to
read from socket', and we found out that this is because of the firewall
will interrupt inactive connection between the app server and the DB
after
120 minute.
To resolve this, we suggested to update the idle time setting in the
oracle
to 90, with the hope that the oracle will terminate the connection before
the firewall, and let the 'testOnBorrow' of dbcp to validate the
connection
before it is used. However, this does not resolve our problem.
:(:confused:.
testOnBorrow tests the connection before only when it is borrowed from
the pool. I am not a Hibernate expert, but to understand exactly what
is going on here, you need to determine when connections are being
borrowed from the pool vs being reused by your application or Hibernate
without being returned to the pool. If a connection is reused without
being returned and re-borrowed from the pool, dbcp cannot validate it.
Our next action was to enable the 'removeAbandoned' of the dbcp,
suspecting
that the unclosed unconnection was the one contributing to the whole
issue.
Even though enabling this feature does not resolve our problem, it does
gives us some clue on the issue. My removeAbandoned does not take effect
because it did not meet the condition where numActive > maxActive -3,
hence
I tested in my local with a low 'maxActive' setting and found out that
there
is a function which indeed create unclosed connection on certain
occasion.
This had finally led me to the finding that when the poolable connection
is
closed, i am not able to obtain a new poolable connection within the same
webservice invocation.
It is important to me to know how to obtain the new connection as we
would
like the webservice to be able to obtain a new connection if the current
connection is not active.
As long as each time your web service needs a connection it is using
BasicDataSource's getConnection(), testOnBorrow=true should ensure that
the returned connection is valid. It should not make a difference, but
you could also set testOnReturn to true to see if this improves the
situation. That property triggers validation before connections are
returned to the pool.
Here is how things are supposed to work from the dbcp perspective (as of
1.2.2):
0) application calls getConnection. If there is an idle connection in
the pool, the connection is returned. If testOnBorrow is true, the
connection is validated first. If validation fails, the connection is
destroyed and getConnection (actually pool's borrowObject) tries again.
If all of the idle connections are "bad" (as could be happening in your
case), they will all be examined and destroyed in this process and a new
one will be created, validated and returned.
1) application calls close on the PoolableConnection that it got from
getConnection. If testOnReturn is set to true, the connection is
validated before being returned to the pool. If validation fails, it is
destroyed.
If removeAbandoned is set to true, before getConnection tries to
retrieve a connection from the pool in 0), it checks
(getNumIdle() < 2) && (getNumActive() > getMaxActive() - 3) ) . If
this is true, abandoned connections are destroyed. Whether or not a
connection is considered "abandoned" is determined by whether or not it
has been used (checked out, checked in, prepared a statement, etc.) more
recently than the removeAbandonedTimeout.
What version of dbcp and pool are you running?
Phil
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
