## Problem The GoProxy egress-check init container only handled `x86_64` and `aarch64`. On the `xu4` node, `uname -m` is `armv7l`, so the init container exited with `unsupported architecture: armv7l` and the Deployment stayed in CrashLoopBackOff. ## Fix - select the upstream GoProxy v15.3 `proxy-linux-arm-v7.tar.gz` asset for `armv7l`; - verify it with its pinned SHA-256 checksum; - document ARMv7 support. This PR is limited to the GoProxy workload and does not change the registry-DNS or connectivity-tester PRs. Validation: `git diff --check`, architecture-to-asset mapping for `x86_64`, `aarch64`, and `armv7l`, and SHA-256 verification of the downloaded upstream ARMv7 release asset. Reviewed-on: #19 Co-authored-by: hermes <rudolfss13+hermes@gmail.com>
K3s Cross-Node Connectivity Tester
Functional Requirements
- Run one tester pod on each of the three target nodes, with required hostname anti-affinity preventing more than one tester pod on any node.
- Use parallel StatefulSet replicas with stable internal names:
k3s-connectivity-0,k3s-connectivity-1, andk3s-connectivity-2. - Probe every other replica over HTTP by its internal Kubernetes DNS name, such
as
k3s-connectivity-1.k3s-connectivity.connectivity-test.svc.cluster.local. - Expose
/healthz,/identity, and/checkon port8080. - Keep
/healthzlocal and fast; it must not perform DNS or peer requests. - Run peer checks only when
/checkis requested and report source/peer pod and node names. - Report
all_reachable: trueonly when every peer responds successfully. - Report
all_cross_node: trueonly when every successful peer response comes from a different node. - Support HTTPS through the mounted
k3s-connectivity-tlsSecret. - Pull the multi-architecture registry image for
amd64,arm64, andarm/v7nodes. - Leave pods Pending rather than co-locating them when insufficient eligible
nodes are available; nodes with a disk-pressure
NoScheduletaint are not eligible.
Transparent egress check
egress-check.yaml adds a client and Squid sidecar in one pod. The client
checks CHECK_URL_1 and CHECK_URL_2 every 30 seconds; the URLs must use HTTP,
share a domain, and have different paths. The init container installs iptables
rules in the shared pod network namespace. Client HTTP traffic is redirected
to Squid, while direct TCP and UDP traffic other than DNS is rejected.
This is transparent HTTP proxying. HTTPS cannot be transparently proxied by this configuration because it would require TLS interception and a trusted CA. Use an explicit HTTPS proxy or add a separately managed TLS interception design if HTTPS endpoints are required.
The client prints each URL and returned status to the container log, then repeats both requests every 30 seconds. Change the two URL environment variables in the Deployment before applying it.
GoProxy egress check
goproxy-egress-check.yaml adds a second client/proxy POC in the
egress-check-goproxy namespace. The client checks the same two URLs every 30
seconds and uses the GoProxy sidecar at 127.0.0.1:3128 through the HTTP proxy
environment variables. The sidecar uses the multi-architecture alpine:3.22
image and an init container downloads the matching pinned GoProxy v15.3 binary
for amd64, arm64, or armv7 with SHA-256 verification.
The init container selects the release asset from the runtime architecture, so
this workload is not tied to a single Kubernetes node platform. GoProxy runs as
proxy http -t tcp -p :3128.
This variant uses explicit HTTP proxying rather than iptables interception. The
client has no proxy bypass list, and the proxy and client run in the same pod so
each request is sent to GoProxy before leaving the pod. GoProxy logs are
available with kubectl logs -n egress-check-goproxy deploy/egress-check-goproxy -c goproxy; client results are available from the client container.