Re: (ITS#4240) back-meta loses "connection" with remote backend

[pfnguyen@hanhuy.com]

> OK, I'm digging thru your log; I see an anonymous search getting to a
> glued database; the search is propagated to a back-meta that sends one
> search locally and two presumably remotely.  The local search gives no
> result, while one of the remote searches in one case gets an entry and
> in the other none.  Then the client unbinds.  I don't see any
> significant evidence of an erroneous behavior in the proxy; it does
> garbage-in garbage-out.  I've uploaded an annotated version of your logs
> at <ftp://ftp.openldap.org/incoming/2.3.12-disconnecting-meta.tar.gz>

In terms of timing out in the query, the response to the query is under 
1 second, whether by direct query or back-meta (when it is working).

> I note that you set very low size/time limits (12 entries, 15 seconds);
> how long does it take to the remote servers to return the entry if
> queried directly?  Can you check what happens if you increase those time
> limits?  I note that if there were any relevant issue in the
> communication between the proxy and the remote server we would have seen
> it in the logs, since both libldap and back-meta log at TRACE, ARGS and
> ANY, which means always; if anything serious goes on, it's logged at
> ANY, so it appears regardless of the log level you set; if anything else
> unusual happens, it's logged at TRACE, while ARGS allowed us to track
> what was going on with your ops.  I really think something may be going
> on with the remote server; I mean something like a too long lasting
> connection gets "lazy" so that the proxy sees it as valid while it
> doesn't actually work.

I will try running 2.3.13 and see if this resolves my problem, 
otherwise, I will try the idle-timeout feature available in head (if the 
problem isn't solved, and 2.3.14 comes out without the feature 
incorporated).

> Let me note that if you try to run current HEAD code, it contains a new
> feature (ITS#4115) that allows to configure an idle-timeout at the proxy
> side of a connection, so that long lasting connections are closed and
> reopened by the proxy (the reverse of the idletimeout); I think you
> might even be able to use back-meta from HEAD with 2.3.13 code.
> 
> Would you mind trying that?  Also, note that 2.3.13 fixed an issue with
> handling of dead connections in both back-ldap and back-meta (ITS#4189),
> so it might definitely be worth upgrading, at least for testing, given
> the type of issues you're facing.