Hi Antonio, On Thu, Oct 17, 2024 at 11:10:02PM +0100, Antonio Ojea wrote: > On Thu, 17 Oct 2024 at 17:36, Pablo Neira Ayuso <pablo@xxxxxxxxxxxxx> wrote: > > > > On Thu, Oct 17, 2024 at 02:46:32PM +0200, Florian Westphal wrote: > > > Antonio Ojea <antonio.ojea.garcia@xxxxxxxxx> wrote: > > > > In the context of Kubernetes, when DNATing entries for UDP Services, > > > > we need to deal with some edge cases where some UDP entries are left > > > > orphaned but blackhole the traffic to the new endpoints. > > > > > > > > At high level, the scenario is: > > > > - Client IP_A sends UDP traffic to VirtualIP IP_B --> Kubernetes > > > > Translates this to Endpoint IP_C > > > > - Endpoint IP_C is replaced by Endpoint IP_D, but since Client IP_A > > > > does not stop sending traffic, the conntrack entry IP_A IP_B --> IP_C > > > > takes precedence and is being renewed, so traffic is not sent to the > > > > new Endpoint IP_D and is lost. > > > > > > > > To solve this problem, we have some heuristics to detect those > > > > scenarios when the endpoints change and flush the conntrack entries, > > > > however, since this is event based, if we lost the event that > > > > triggered the problem or something happens that fails to clean up the > > > > entry, the user need to manually flush the entries. You can still stick to the event approach, then resort to resync/reconcile loop when userspace gets a report that events are getting lost, ie. hybrid approach. > > > > We are implementing a new approach to solve this, we list all the UDP > > > > conntrack entries using netlink, compare against the existing > > > > programmed nftables/iptables rules, and flush the ones we know are > > > > stale. > > > > > > > > During the implementation review, the question [1] this raises is, how > > > > impactful is it to dump all the conntrack entries each time we program > > > > the iptables/nftables rules (this can be every 1s on nodes with a lot > > > > of entries)? > > > > Is this approach completely safe? > > > > Should we try to read from procfs instead? > > > > > > Walking all conntrack entries in 1s intervals is going to be slow, no > > > matter the chosen interface. Even doing the filtering in the kernel to > > > not dump all entries but only those that match udp/port/ip criteria is > > > not going to change it. > > We are not worried about being slow in the order of seconds, the > system is eventually consistent so there can always be a reasonable > latency. > Since we only care about UDP, losing packets during that period is not > desirable but is assumable. > My main concern is if constantly dumping all the entries via netlink > can cause any issue or increase resources consumption. > > > > > > > Also both proc and netlink dumps can miss entries (albeit its rare), > > > if parallel insertions/deletes happen (which is normal on busy system). > > > > > That is one of the reasons we want to implement this reconcile loop, > so it can be resilient to this kind of errors, we keep the state on > the API in the control plane, so we can always rebuild the state in > the dataplane (recreating nftables rules and delete conntrack entries > that does not match the current state) > > > > I wonder why the appropriate delete requests cannot be done when the > > > mapping is altered, I mean, you must have some code that issues > > > either iptables -t nat -D ... or nft delete element ... or similar. > > > > > > If you do that, why not also fire off the conntrack -D request > > > afterwards? Or are these publish/withdraw so frequent that this > > > doesn't matter compared to poll based approach? > > > > > > Something like > > > conntrack -D --protonum 17 --orig-dst $vserver --orig-port-dst 53 --reply-src $rserver --reply-port-src 5353 > > > > > > would zap everything to $rserver mapped to $vserver from client point of view. > > > > This is how it is implemented today and it works, but it does not > handle process restarts per example, or is not resilient to errors. > The implementation is also much more complex because we need to > implement all the possible edge cases that can leave stale entries It should also be possible to shrink timeouts on restart via conntrack -U which would be similar to the approach that Florian is proposing, but from control plane rather than updating existing UDP timeout policy.