Sangeeta Misra wrote:
> I think one open design discussion was how does ILB's load balancer
> handle key failures...
>
> One such case( are there others that we need to consider?) is handling
> of idle or inactive session in stateful loadbalancing ( which can
> result from client or the server dying during the session)
For the ILB NAT tabkles presumably we'd just need the normal NAT
mechanisms (look for some TCP FIN/RST sequences, and also a timer).
We don't plan a non-NAT stateful for phase I, right?
When we do that we'd need some timer unrelated to the NATting.
> I am assuming one way to solve this would be to implement a timer
> related task that gets rid of idle session ( should the timer value be
> configurable?) Also I assume LB sends a TCP reset to servers so they
> can clean up resources associated with the session. Would you folks agree?
You might want a different timer value for UDP NAT than for TCP NAT.
When the TCP closes (seeing the FINs, or a RST) then there isn't any
need to send additional packets. When TCP state is removed by the timer
it might be useful to send a RST, but it would make sense to check what
other LBs/NATs do here.
Erik
