Effect of listener on existing connections?

From: <groups.broberg_at_gmail.com>
Date: Fri, 13 Jun 2008 06:40:24 -0700 (PDT)
Message-ID: <a8f8b4b1-adbb-4916-9b23-8d72f6d221ba@f36g2000hsa.googlegroups.com>

Hi all,

Oracle support is not giving us satisfactory results. Perhaps you can give some answers?

We've recently upgraded our system to Oracle, running on Solaris (sparc) 10 inside a ZFS zone (our previous system was Oracle running on sparc Solaris 8, and was running on that for the last 5 years). Since the upgrade 6 weeks ago, we've had two instances where our applications (running in the same O/S environment on a different node on the cluster) have locked up - existing connections to Oracle become unresponsive when executing SQL (with no error message - they just block), and attempts to create new connections are met with the error:

"ORA-03135: connection lost contact".

The first time this happened, the outage lasted for about 5 minutes, then it went away, and execution proceeded normally. The second time it happened it lasted for 25 minutes until we were able to intervene manually, and "fixed" the problem by restarting the instance & listener. During the time of the outage, there were no error messages in the alert.log, listener.log or /var/adm/messages. However, a few minutes after normal operation was restored (the first time), and right as we were restarting the instance (the second time), we saw these messages appear in the alert.log:

"WARNING: inbound connection timed out (ORA-3136)".

We also see these messages appearing with regularity in our listener log, during all times of operation (not just in proximity to the outage):

"WARNING: Subscription for node down event still pending"

We opened an SR with Oracle Support, but so far, I'm unimpressed with their response. They've told me nothing that I already found on Google from searching for those error messages - namely that we need to add some lines to our listener.ora & sqlnet.ora:

SQLNET.INBOUND_CONNECT_TIMEOUT = 0 We've made these changes, but I have low confidence that they will actually solve the problem (and I've told Oracle as much) for the following reasons:

  1. The SUBSCRIBE_FOR_NODE_DOWN_EVENT_LISTENER value is to address the issue of the listener locking up if you're not using ONS. However, how does a blocked listener explain the fact that apps with existing connections to the db become blocked? My understanding (and I could be wrong here) is that once you're connected to the instance, there is no further authentication that needs to be performed. Our in-house experiments also show that we can "kill -STOP" the listener and apps with existing connections continue to perform normally.
  2. The INBOUND_CONNECT_TIMEOUT value is to address the issue with clients that are "slow to authenticate", but the "WARNING: inbound connection timed out (ORA-3136)" message appears AFTER the crisis interval - in the recent case, it appeared 20 minutes after our app became blocked. I would expect to see it within 60 seconds of the block, as that's what the value currently is.
  3. All apps trying to establish new connections (including sqlplus, running on the same node as the instance) received the login error, not just "certain apps" as described in the Oracle tech note 274303.1. Why would an app like sqlplus, running on the same box as the server (which was the case here) need more than 1 minute to authenticate when logging in? Even our own apps shouldn't be taking long to authenticate.

Should this event occur again, I don't see how Support will be able to resolve it, as they haven't asked me for any additional info. I want to know from them what steps I need to take to gather information so that they can REALLY fix the problem, or give me an answer that unambiguously addresses the issue, rather than just googling on error numbers. So far my requests for a clear action plan from them have been met with the email equivalent of a blank, slack-jawed stare.

So, my basic questions are:

  1. Can listener unavailability cause existing client connections to become unresponsive?
  2. If the answer to #1 is "no", Is there some way I can escalate this issue within Support to get to a analyst who actually understands how the Oracle server works, and is capable of doing something other than typing search queries into metalink?


-S Received on Fri Jun 13 2008 - 08:40:24 CDT

Original text of this message