Lean iptables updates #324

dajudge · 2022-07-26T12:38:15Z

Why

We were repeatedly expecting connectivity issues when scaling larger clusters up/down and believe it's related to the way kilo applies changes to the iptables when nodes are added/removed: in many cases that resulted in many rules (or even entire chains) being dropped and recreated during scaling.

Idea

After taking a closer look at the iptables rules created by kilo we figured we have identified two different classes of the same:

some rules need to be applied strictly in order and always at the end of a chain (e.g. the FORWARD to KILO-IPIP and DROP rule in the INPUT chain - also e.g. the MASQUERADE rule at the end of the KILO-NAT chain)
other rules don't have a strict ordering requirement (apart from having to be in front of the rules that need to be at the end of a chain where applicable)

How

So we started preparing this patch that puts the different rules in separate slices and

first applies the order-sensitive/append-only rules using the existing mechanics
then inserts the 2nd class of rules each one at the beginning of the respective chain if they're not already there - regardless of order
removes any rules of the 2nd class that are not required anymore

chain rules are treated as order-sensitive/append-only which does not matter a lot beyond being applied in the first wave of changes

Results

That way we were able to accomplish the following:

the first class of rules gets only applied once because they don't change after the first reconciliation (it's only the chains and the static rules at the end of some chains)
for updates of the 2nd class of rules (they tend to be per ip/subnet) already existing rules never get dropped unless they need to be removed

We were able to confirm that this reduces the number of mutating iptables operations after the initial iptables configuration to the absolute minimum required.

squat · 2022-07-26T14:22:18Z

In general I think this is an important priority for Kilo. Another way I've considered solving this in the past is by using ipsets and atomically swapping out the IP addresses in the sets so that Kilo does not have to change rules so often. This means that whenever new Nodes or Peers join a cluster, this would instead require an "instantaneous" ipset operation rather than changing iptables rules. This would bring other performance benefits with it as the evaluation of a firewall rule containing ipset is faster than evaluating many independent iptables rules.

clive-jevons · 2022-07-26T15:00:28Z

In general I think this is an important priority for Kilo. Another way I've considered solving this in the past is by using ipsets and atomically swapping out the IP addresses in the sets so that Kilo does not have to change rules so often. This means that whenever new Nodes or Peers join a cluster, this would instead require an "instantaneous" ipset operation rather than changing iptables rules. This would bring other performance benefits with it as the evaluation of a firewall rule containing ipset is faster than evaluating many independent iptables rules.

I do really like the sound of that approach. However, how would you feel about us starting with the implementation in this PR, so that our immediate problem can be solved (we verified that it does indeed solve our problems on one of our test clusters 😁 ), and then have a stab at doing a PR using ipsets at some point later? 😬

dajudge · 2022-07-27T07:55:28Z

pkg/iptables/iptables.go

@@ -105,7 +127,14 @@ func NewIPv6Rule(table, chain string, spec ...string) Rule {
 	return &rule{table, chain, spec, ProtocolIPv6}
 }

-func (r *rule) Add(client Client) error {
+func (r *rule) Prepend(client Client) error {
+	if err := client.Insert(r.table, r.chain, 1, r.spec...); err != nil {


@clive-jevons should we maybe implement & use a InsertUnique here?

I think the intention of the Client interface was to only shadow the existing API of the ipsets client? As such the InsertUnique semantics should probably be part of the API here? Could be wrong though - some guidance by @squat or @leonnicolas would be good here, I think 😬

If you look a couple of lines down the Append() function is implemented via AppendUnique() - hence the question.

I think this is relevant when kilo was terminated abruptly without having a chance to clean up and then spins up again - without the check for Exists() the Insert() would probably cause an error?

Very fair point 😉 I'd still be interested in the feedback / guidance from the maintainers on this 😬
But absolutely agree with the necessity of the functionality - where it ends up being implemented 👍

You're right re the interface - it's implemented in an external dependency which doesn't mirror AppendUnique() for Insert(). So my suggestion here would be to implement this here while at the same time making a PR in the dependency to implement InsertUnique() and later switch to it once it's there. 🤔

Progress here: it's been merged to main. Once there is a new release we can make use of the new function.

Great work!

Hmm we can probably just pin to the latest commit on main rather than wait for a release, no?

have done that and replaced iptables.Insert() with iptables.InsertUnique()

dajudge · 2022-07-27T09:26:44Z

The idea of using ipsets sounds great! I think this PR here might even be a nice step towards that direction by introducing the distinction between the two iptables rules classes:

the append rules seem to be the catch-all kind of rules that have to be in place statically and I don't think ipsets make much sense here?
the prepend rules seem to be the ip-/subnet-specific ones that would benefit from a move to ipsets

The PR at hand already introduces a carrier datastructure RuleSet to carry more semantic information than a []Rule slice - which sounds like a great start to later put more semantics in for ipsets - while in this iteration still relying on the proven approach of using iptables.

So all in all I think this PR could even be considered as a nice evolutionary step towards an ipsets-based implementation.

WDYT?

dajudge · 2022-07-28T14:09:54Z

We just ran a small test against a cluster running this PR (plus the metrics PR for observability) and this is what it looks like in grafana 😍

The two iptables related charts display the following metrics:

sum(rate(kilo_iptables_operations_total[5m])) by (operation)
sum(rate(kilo_iptables_operations_total{operation!~"(List|Exists|ListChains)"}[5m])) by (chain)

We also didn't observe any disconnects during autoscaling as far as we can tell.

squat · 2022-10-20T10:30:31Z

@dajudge what's the status of this PR? Is it ready for a final review so we can merge and maybe include in a Kilo 0.6.0? 😻

dajudge · 2022-10-20T11:00:41Z

@dajudge what's the status of this PR? Is it ready for a final review so we can merge and maybe include in a Kilo 0.6.0? heart_eyes_cat

Yes. We've been running patched version of kilo with this PR included since July.

We're currently running it on over half a dozen kubernetes clusters with dozens of nodes in total. We have not encountered any issues and it's even playing nicely with kube-router ever since (we had severe problems with that before).

From our perspective this PR is ready to go! 😍 🚀 👍

clive-jevons · 2022-12-20T12:26:02Z

@squat <- any chance of getting this merged and into a release? 😬 🙏 Anything we can do to help move this forward?

squat · 2023-03-28T06:36:01Z

@clive-jevons thanks for your patience!
If you can help resolve the merge conflict on the PR then I would be very happy to merge this bad boy!

dajudge · 2023-03-28T13:31:00Z

rebased / merge conflicts resolved.

running it on a test cluster now.

looks good at first glance. 👍

squat · 2023-03-28T13:35:24Z

thanks @dajudge

squat

Thanks for being patient @dajudge @clive-jevons <3

dajudge marked this pull request as ready for review July 26, 2022 15:32

dajudge commented Jul 27, 2022

View reviewed changes

Alex Stockinger and others added 11 commits March 28, 2023 15:00

Split iptables rules into append and prepend rules

9815e2e

fix existing tests

03420b7

add iptables tests which include prepend rules

6b7336c

fix iptables test storage usage

f7fc148

add test to check reconcile behaviour

68c4f12

Reconcile prepend rules

bd96aeb

Make usage of RuleSet prettier

86db06d

Properly implement Insert for fake enc

bbf5768

Make Rule.Prepend() behave like Rule.Append() regarding uniqueness

2c35343

Implement Insert() for metricsClientWrapper

11a5e64

Use iptables.InsertUnique() instead of iptables.Insert()

9f8ac50

squat approved these changes Mar 28, 2023

View reviewed changes

squat merged commit 12ad275 into squat:main Mar 28, 2023

dajudge deleted the lean-iptables-updates branch March 28, 2023 16:43

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Lean iptables updates #324

Lean iptables updates #324

dajudge commented Jul 26, 2022

squat commented Jul 26, 2022

clive-jevons commented Jul 26, 2022

dajudge Jul 27, 2022 •

edited

Loading

clive-jevons Jul 27, 2022

dajudge Jul 27, 2022

clive-jevons Jul 27, 2022

dajudge Jul 27, 2022 •

edited

Loading

squat Aug 17, 2022

dajudge Aug 31, 2022

squat Aug 31, 2022

squat Aug 31, 2022

dajudge Sep 15, 2022 •

edited

Loading

dajudge commented Jul 27, 2022 •

edited

Loading

dajudge commented Jul 28, 2022 •

edited

Loading

squat commented Oct 20, 2022

dajudge commented Oct 20, 2022 •

edited

Loading

clive-jevons commented Dec 20, 2022

squat commented Mar 28, 2023

dajudge commented Mar 28, 2023

squat commented Mar 28, 2023

squat left a comment

Lean iptables updates #324

Lean iptables updates #324

Conversation

dajudge commented Jul 26, 2022

Why

Idea

How

Results

squat commented Jul 26, 2022

clive-jevons commented Jul 26, 2022

dajudge Jul 27, 2022 • edited Loading

Choose a reason for hiding this comment

clive-jevons Jul 27, 2022

Choose a reason for hiding this comment

dajudge Jul 27, 2022

Choose a reason for hiding this comment

clive-jevons Jul 27, 2022

Choose a reason for hiding this comment

dajudge Jul 27, 2022 • edited Loading

Choose a reason for hiding this comment

squat Aug 17, 2022

Choose a reason for hiding this comment

dajudge Aug 31, 2022

Choose a reason for hiding this comment

squat Aug 31, 2022

Choose a reason for hiding this comment

squat Aug 31, 2022

Choose a reason for hiding this comment

dajudge Sep 15, 2022 • edited Loading

Choose a reason for hiding this comment

dajudge commented Jul 27, 2022 • edited Loading

dajudge commented Jul 28, 2022 • edited Loading

squat commented Oct 20, 2022

dajudge commented Oct 20, 2022 • edited Loading

clive-jevons commented Dec 20, 2022

squat commented Mar 28, 2023

dajudge commented Mar 28, 2023

squat commented Mar 28, 2023

squat left a comment

Choose a reason for hiding this comment

dajudge Jul 27, 2022 •

edited

Loading

dajudge Jul 27, 2022 •

edited

Loading

dajudge Sep 15, 2022 •

edited

Loading

dajudge commented Jul 27, 2022 •

edited

Loading

dajudge commented Jul 28, 2022 •

edited

Loading

dajudge commented Oct 20, 2022 •

edited

Loading