TL;DR: we want to start failing the LB HC if a node is tainted with ToBeDeletedByClusterAutoscaler.
This field might need refinement, but currently is deemed our best way of understanding if
a node is about to get deleted. We want to do this only for eTP:Cluster services.
The goal is to connection draining terminating nodes
Historically, IptablesRulesTotal could have been intepreted as either
"the total number of iptables rules kube-proxy is responsible for" or
"the number of iptables rules kube-proxy rewrote on the last sync".
Post-MinimizeIPTablesRestore, these are very different things (and
IptablesRulesTotal unintentionally became the latter).
Fix IptablesRulesTotal (sync_proxy_rules_iptables_total) to be "the
total number of iptables rules kube-proxy is responsible for" and add
IptablesRulesLastSync (sync_proxy_rules_iptables_last) to be "the
number of iptables rules kube-proxy rewrote on the last sync".
This required fixing a small bug in the metric, where it had
previously been counting the "-X" lines that had been passed to
iptables-restore to delete stale chains, rather than only counting the
actual rules.
Rather than having GetNodeAddresses() return a special magic value
indicating that it matches all IPs, add a separate method to check
that. (And have GetNodeAddresses() just return the IPs as expected
instead.)
Both proxies handle IPv4 and IPv6 nodeport addresses separately, but
GetNodeAddresses went out of its way to make that difficult. Fix that.
This commit does not change any externally-visible semantics, but it
makes the existing weird semantics more obvious. Specifically, if you
say "--nodeport-addresses 10.0.0.0/8,192.168.0.0/16", then the
dual-stack proxy code would have split that into a list of IPv4 CIDRs
(["10.0.0.0/8", "192.168.0.0/16"]) to pass to the IPv4 proxier, and a
list of IPv6 CIDRs ([]) to pass to the IPv6 proxier, and then the IPv6
proxier would say "well since the list of nodeport addresses is empty,
I'll listen on all IPv6 addresses", which probably isn't what you
meant, but that's what it did.
Now that the endpoint update fields have names that make it clear that
they only contain UDP objects, it's obvious that the "protocol == UDP"
checks in the iptables and ipvs proxiers were no-ops, so remove them.
The APIs talked about "stale services" and "stale endpoints", but the
thing that is actually "stale" is the conntrack entries, not the
services/endpoints. Fix the names to indicate what they actual keep
track of.
Also, all three fields (2 in the endpoints update object and 1 in the
service update object) are currently UDP-specific, but only the
service one made that clear. Fix that too.
This commit did not actually work; in between when it was first
written and tested, and when it merged, the code in
pkg/proxy/endpoints.go was changed to only add UDP endpoints to the
"stale endpoints"/"stale services" lists, and so checking for "either
UDP or SCTP" rather than just UDP when processing those lists had no
effect.
This reverts most of commit aa8521df66
(but leaves the changes related to
ipvs.IsRsGracefulTerminationNeeded() since that actually did have the
effect it meant to have).
Today, the health check response to the load balancers asking Kube-proxy for
the status of ETP:Local services does not include the healthz state of Kube-
proxy. This means that Kube-proxy might indicate to load balancers that they
should forward traffic to the node in question, simply because the endpoint
is running on the node - this overlooks the fact that Kube-proxy might be
not-healthy and hasn't successfully written the rules enabling traffic to
reach the endpoint.
For some reason we were calculating the available nodeport IPs at the
top of syncProxyRules even though we didn't use them until the end.
(Well, the previous code avoided generating KUBE-NODEPORTS chain rules
if there were no node IPs available, but that case is considered an
error anyway, so there's no need to optimize it.)
(Also fix a stale `err` reference exposed by this move.)
In addition to actually updating their data from the provided list of
changes, EndpointsMap.Update() and ServicePortMap.Update() return a
struct with some information about things that changed because of that
update (eg services with stale conntrack entries).
For some reason, they were also returning information about
HealthCheckNodePorts, but they were returning *static* information
based on the current (post-Update) state of the map, not information
about what had *changed* in the update. Since this doesn't match how
the other data in the struct is used (and since there's no reason to
have the data only be returned when you call Update() anyway) , split
it out.
The unit tests were broken with MinimizeIPTablesRestore enabled
because syncProxyRules() assumed that needFullSync would be set on the
first (post-setInitialized()) run, but the unit tests didn't ensure
that.
(In fact, there was a race condition in the real Proxier case as well;
theoretically syncProxyRules() could be run by the
BoundedFrequencyRunner after OnServiceSynced() called setInitialized()
but before it called forceSyncProxyRules(), thus causing the first
real sync to try to do a partial sync and fail. This is now fixed as
well.)
In the dual-stack case, iptables.NewDualStackProxier and
ipvs.NewDualStackProxier filtered the nodeport addresses values by IP
family before creating the single-stack proxiers. But in the
single-stack case, the kube-proxy startup code just passed the value
to the single-stack proxiers without validation, so they had to
re-check it themselves. Fix that.
This rule was mistakenly added to kubelet even though it only applies
to kube-proxy's traffic. We do not want to remove it from kubelet yet
because other components may be depending on it for security, but we
should make kube-proxy output its own rule rather than depending on
kubelet.
Some of the chains kube-proxy creates are also created by kubelet; we
need to ensure that those chains exist but we should not delete them
in CleanupLeftovers().
We currently invoke /sbin/iptables 24 times on each syncProxyRules
before calling iptables-restore. Since even trivial iptables
invocations are slow on hosts with lots of iptables rules, this adds a
lot of time to each sync. Since these checks are expected to be a
no-op 99% of the time, skip them on partial syncs.
iptables-restore requires that if you change any rule in a chain, you
have to rewrite the entire chain. But if you avoid mentioning a chain
at all, it will leave it untouched. Take advantage of this by not
rewriting the SVC, SVL, EXT, FW, and SEP chains for services that have
not changed since the last sync, which should drastically cut down on
the size of each iptables-restore in large clusters.
Back when iptables was first made the default, there were
theoretically some users who wouldn't have been able to support it due
to having an old /sbin/iptables. But kube-proxy no longer does the
things that didn't work with old iptables, and we removed that check a
long time ago. There is also a check for a new-enough kernel version,
but it's checking for a feature which was added in kernel 3.6, and no
one could possibly be running Kubernetes with a kernel that old. So
the fallback code now never actually falls back, so it should just be
removed.
The proxies watch node labels for topology changes, but node labels
can change in bursts especially in larger clusters. This causes
pressure on all proxies because they can't filter the events, since
the topology could match on any label.
Change node event handling to queue the request rather than immediately
syncing. The sync runner can already handle short bursts which shouldn't
change behavior for most cases.
Signed-off-by: Dan Williams <dcbw@redhat.com>
Part of reorganizing the syncProxyRules loop to do:
1. figure out what chains are needed, mark them in activeNATChains
2. write servicePort jump rules to KUBE-SERVICES/KUBE-NODEPORTS
3. write servicePort-specific chains (SVC, SVL, EXT, FW, SEP)
This moves the FW chain creation to the end (rather than having it in
the middle of adding the jump rules for the LB IPs).
