Feature request: Support env _ METHOD_

[ThommyH]

Hi,
thank you for your great work on this project. Like many, we faced issues with kiwigrid/k8s-sidecar loop getting stuck and hence seek for alternatives.
We use it as part of grafana, since it is the default.
For feature matching with kiwigrid we would need the flag _ METHOD_ which allows to run the container once and then exit. In grafana, the same image for init container (fetching datasources) and sidecar (fetching dashboards) is used.
In order to replace it with kopf-k8s-sidecar one need to make a lot of changes and hacks there due to the missing flag here.

[bergerx]

Just came here to create this exact issue just now and found it created 2 minutes ago :)

We deploy grafana as a dependency to kube-prometheus-stack chart, and here is what i changed to make this work in our stack (other users may need to use extraRoleRules instead of extraClusterRoleRules if not using searchNamespace: ALL as below) :

diff --git a/clusteraddons/monitoring-stack.yaml b/clusteraddons/monitoring-stack.yaml
index 4e087cc..e95fcbe 100644
--- a/clusteraddons/monitoring-stack.yaml
+++ b/clusteraddons/monitoring-stack.yaml
@@ -70,10 +70,26 @@ spec:
         enabled: true
         sidecar:
+          image:
+            # Workaround for connection hanging bug in kiwigrid/k8s-sidecar see these issues for details:
+            # https://github.com/kiwigrid/k8s-sidecar/issues/85
+            # https://github.com/grafana/helm-charts/issues/18
+            repository: omegavveapon/kopf-k8s-sidecar
+            tag: 1.1.1
           dashboards:
             searchNamespace: ALL
           datasources:
             searchNamespace: ALL
+          rbac:
+            # This is a workaround for using the omegavveapon/kopf-k8s-sidecar
+            # instead of kiwigrid/k8s-sidecar. omegavveapon/kopf-k8s-sidecar is
+            # implemented with kopf, and kopf requires patch verb on any
+            # resources it watches, even if its needed just to watch, 
+            # see https://github.com/zalando-incubator/kopf/issues/318
+            extraClusterRoleRules:
+            - apiGroups: [""] # "" indicates the core API group
+              resources: ["configmaps", "secrets"]
+              verbs: ["patch"]
         plugins:
           - natel-discrete-panel
           - grafana-piechart-panel

Here is the status with this config:

[OmegaVVeapon]

@bergerx @ThommyH I see, I suppose we could do that but it sorta defeats the purpose of using an operator, which uses a control loop, AKA "a non-terminating loop that regulates the state of a system."

If we run one loop and then kill the process, then it's obviously not a control loop...

Does the chart allow you to specify a different image for the init container and for the actual running sidecar?

[bergerx]

I checked it already before coming here, it doesn't.

[bergerx]

Seems like our best bet is to go back to kiwigrid/k8s-sidecar and provide a PR there.

[OmegaVVeapon]

I've commented on the issue open on the Grafana helm chart since it seems like most of the action is happening there.

In the meantime, I'm trying to see if I can fit this use case in kopf-k8s-sidecar but it might prove challenging.

[OmegaVVeapon]

After trying the latest version of kiwigrid sidecar merged to the Grafana Helm chart today (1.10.0) and seeing that it hung after a couple of hours, I decided to give this a try...

To my surprise, it wasn't that difficult to add a separate codepath for the LIST METHOD using kopf's embedding functionality.

Tentative PR in #16

Have a few details to figure out but the core functionality is 100% working for both LIST and WATCH.

Will leave the pod running overnight just to check I didn't somehow affect its stability.

[OmegaVVeapon]

The PR has been merged. Closing.