[j-nsp] 10.0 or 10.4?

[Chuck Anderson]

On Tue, Mar 15, 2011 at 11:27:22PM +0100, Daniel Roesen wrote:
> On Tue, Mar 15, 2011 at 11:13:33PM +0100, magno wrote:
> > I can confirm, 10.4 is the first JUNOS release developed with a new
> > methodology.
> > 
> > This would allow Juniper to catch much more bugs before releasing the code
> > than in the past.
> 
> Such as 10.4R2 not showing transceivers on 16-port 10GE MPC in "show
> chassis hardware" anymore, but 10.4R1 did? :-)

Works for me, 16-port MPC, 10.4R2.

> Our RANCID caught that on on of our lab MXes within less then one hour
> after upgrading the box. Systest didn't - or it wasn't deemed a
> show-stopper for release.

Still hasn't happened to me.  I'm also using RANCID, and I've been 
running 10.4R2 on MX960s with Trio cards (including 16-port) since 
March 4.

> Other than that, our lab and field trials of 10.4R1 + 10.4R2 went
> astonishingly well. Except abovementioned bug we didn't run into any
> issue yet (knocking on wood - broad MPC deployment ahead).

On a third MX960 where we additionally run BGP/RSVP/MPLS/L3-VPN (the 
first two only do OSPF/VRRP/bridging/IRB), 10.4R2 core dumped when 
enabling GRES.  We've had to leave GRES/NSR/NSB disabled until a fix 
comes out for that.

> We're currently waiting for R3 to have that bug fixed and another
> general sweep of bugfixes (and hopefully no new issues).

We have a few other issues with 10.4R2/Trio:

1. VRRP got stuck on one irb only on one link, leaving hosts on that 
link on that bridge-domain unable to ping or otherwise use the VRRP 
default gateway.  The hardware forwarding records in the MPC were 
somehow programmed incorrectly.

2. AE was looping packets, as in a single-armed MX960 with only a 
single AE interface enabled (and no non-AE links) was looping packets 
back out the way they came in.  Not sure yet what caused this, but we 
can't enable our redundant MX960 until this is resolved.

But on the whole, it is much better than the 10.2 we came from.  
Though I'm looking forward to the day when Virtual Chassis comes to MX 
which will allow us to eliminate spanning tree and perhaps sidestep 
many of these weird looping issues.