[SURE-8169] [BUG] high memory usage for the aks.(*aksOperatorController).onClusterChange about rancher

[moio]

Reproducer

Prerequisites

• Rancher 2.8.3 with a publicly-accessible Internet address (eg. in AKS)
• an Azure account
• its subscription id
• a resource group in that Azure account

Setup

• import any number of AKS clusters as managed clusters
  • add Azure cloud credentials to Rancher (full docs). TL;DR:
    • run:

      export SUBSCRIPTION_ID=XXXXXXXX-YYYY-ZZZZ-AAAA-BBBBBBBB
      export GROUP=st-rg
      
      az ad sp create-for-rbac \
        --scope /subscriptions/$SUBSCRIPTION_ID/resourceGroups/$GROUP \
        --role Contributor

    • note the JSON output
    • go to ☰ -> Cluster Management -> Cloud Credentials -> Create -> Azure
    • paste appId from the JSON output to Client ID, password to Secret, $SUBSCRIPTION_ID to Subscription ID
  • import the clusters as as managed clusters ☰ -> Cluster Management -> Import Existing -> Azure AKS

Test (of base version)

In one terminal with kubeconfig pointing to the upstream cluster run:

while true
do
	for cluster in $(kubectl get clusters.management.cattle.io --no-headers -o custom-columns=name:.metadata.name | grep -v local)
	do
	  kubectl patch clusters.management.cattle.io ${cluster} --type='json' -p='[{"op": "replace", "path": "/status/serviceAccountTokenSecret", "value":""}]'
  done
done

This script forces Rancher to continuously execute a core function of AKS cluster deployment (generateAndSetServiceAccount). This emulates the reported environment where clusters are frequently deployed and redeployed without the necessity of really redeploying AKS clusters, which is time intensive.

In another terminal with kubeconfig pointing to the upstream cluster run:

while true; do
	for pod in $(kubectl -n cattle-system get pods -l app=rancher --no-headers -o custom-columns=name:.metadata.name); do
		kubectl exec -n cattle-system ${pod} -c rancher -- curl -s http://localhost:6060/debug/pprof/heap -o heap
		kubectl cp -n cattle-system -c rancher ${pod}:heap ./heap

		go tool pprof -top --trim=false --show_from=generateAndSetServiceAccount ./heap | grep kubernetes.NewForConfig
	done
done

This script monitors one important source of memory usage by generateAndSetServiceAccount (kubernetes.NewForConfig).

Expected output is either:

1. a continuous stream of ShowFrom expression matched no samples, which means no memory was allocated or
2. ShowFrom expression matched no samples interleaved with lines like the following:

         0     0%  7.72%     7.51MB  5.79%  k8s.io/client-go/kubernetes.NewForConfig

The fourth number indicates memory usage. It should increase only slightly, if at all, after several minutes. Importantly, stopping the first script in the other terminal should make usage go down to zero in some minutes.

Actual output is:
ShowFrom expression matched no samples interleaved with lines like the following:

         0     0%  7.72%     7.51MB  5.79%  k8s.io/client-go/kubernetes.NewForConfig

The fourth number indicates memory usage. It is increasing, and stopping the first script in the other terminal does not make usage go down to zero.

Test (of patched version)

When the expected behavior above is reproduced:

• stop both scripts
• swap the Rancher image with the patched one: kubectl set image -n cattle-system deployment/rancher rancher=rancher/rancher:v2.8.3-debug-45509-1
• wait for clusters to come back as fully available in the UI, refresh the homepage to double check that is the case
• re-run the scripts. Expected behavior should be observed