<snip>
This sounds an awful lot like a problem we saw awhile back when
upgrading from JUNOS 4.x to 5.x. At some point (I don't remember
exactly when, but the details should be in the case notes of
PR.19592) Juniper implemented a change that which makes their box
compliant with RFC 2858. However, when speaking BGP with a
non-RFC-compliant box (such as a Cisco running something like
12.0(15)S) the session flaps continuously in the manner you
describe because the other box expects the NEXT_HOP attribute to
be present in every update message.
attribute in an UPDATE message if only the MP_REACH_NLRI attribute is
present. A workaround is to only use family inet unicast instead of
multicast or any on all BGP sessions to those cisco routers or upgrade
all of the cisco routers."
<snip>
Anyone have a Cisco BugID for this?