follow-up todos for maglev

[borkmann]

RELEASE BLOCKERS

@brb

• Add benchmarking test for maglev.GetLookupTable() and optimize the function.
• Simplify maglev.GetLookupTable() as per Add Maglev consistent hashing to kube-proxy replacement #13131 (comment).
• Preallocate UpsertService buffer (Add Maglev consistent hashing to kube-proxy replacement #13131 (comment)).
• perf & implementation quality of hash for agent
  • compare against siphash 2-4 wrt performance
  • have native implementation of hash alg and remove vendoring
• Add cilium bpf lb maglev get $SVC_ID.
• CI: K8sServicesTest Checks service across nodes Tests NodePort BPF Tests with direct routing Tests NodePort with Maglev Tests Maglev backend selection #13224

@borkmann

• Alias all --node-port-* agent flags to a better name if they are not strictly tied to node-port only (Add Maglev consistent hashing to kube-proxy replacement #13131 (comment))
• Remove old k8s srv disable flag

  cilium/daemon/cmd/daemon_main.go

  Lines 334 to 336 in 28d9889

  	flags.Bool(option.DisableK8sServices, false, "Disable east-west K8s load balancing by cilium")
  	option.BindEnv(option.DisableK8sServices)
  	flags.MarkDeprecated(option.DisableK8sServices, "This option is no longer supported and will be removed in v1.9")

• Rename Helm options for LB
• openapi capital spelling on nodeport options in general (Add Maglev consistent hashing to kube-proxy replacement #13131 (comment)).

1.10 or 1.9 best-effort

@brb:

• OpenOrCreateUnpinned -> CreateUnpinned (Add Maglev consistent hashing to kube-proxy replacement #13131 (comment)).
• Add struct for args of pkg/maps/lbmap.go:UpsertService().
• Potentially missing bits from V1.8.1 maglev #13189
• Extend test_maglev(): 1) remove a non-selected pod and check that the same backend is selected (EDIT: 1) is flaky due the possible 1% changes in the backends configuration) 2) remove the selected backend and check that different backend is selected.

@borkmann:

• Kernel: implement flexible max_elems for arraymap so htab can have flexible sized inner map (this can then be realised via Cilium-specific service annotations)
• Maglev nodeport test also for XDP layer (test: extend XDP coverage for Maglev #13847)

@brb / @borkmann:

• migrate from single slot / u16*N value size to 4 backends per slot with max_elems of N/4
• Implement annotations for defining M: maglev: Allow to configure lookup table size for each service individualy via service annotations #14396
• CI: K8sServicesTest Checks service across nodes Tests NodePort BPF Tests with direct routing Tests NodePort with Maglev Tests NodePort #13261
• CI: K8sServicesTest Checks service across nodes Tests NodePort BPF Tests with direct routing Tests NodePort with Maglev Tests NodePort #13839
• pretty print maglev map
• Doc update to state testing on little-endian nodes
• Strategy for draining old backends (Support graceful termination in kube-proxy replacement #10217, also check upstream k8s)
• Reduce nodeport from warn to info when kernel not newest
• annotation for snat/dsr for specific services

[joestringer]

@brb to follow up to determine which of these items should block the v1.9 release vs. which ones should not.

[borkmann]

Unmarked from release-blocker given #13821 is out and the rest in here are best-effort and can land in later 1.9 or 1.10.

[gaffneyd4]

Can the task Strategy for draining old backends be marked as complete by this PR?

If so to clarify, does this mean that for those using bpf-lb-algorithm: maglev, these issues (10217), (14844) are solved as of v1.11?

[brb]

Can the task Strategy for draining old backends be marked as complete by #17716?

cc @aditighag