Repository navigation
Fix the dashboard proxy's service names and volume ownership - #427
Conversation
nginx's resolver does not apply resolv.conf search domains, so the dashboard's short upstream names failed on Kubernetes and every API, gateway and worker route would answer 502. Name the services fully, with the namespace from the downward API. The bundled Postgres and Redpanda ran as non-root users with no fsGroup, so a root-owned volume (Longhorn, EBS) stopped them; kind's world-writable volumes hid both defects. Also note that release signatures need cosign 3. Part of #391
artyomsv
left a comment
There was a problem hiding this comment.
First review at 55ce553. Both fixes are right. nginx's resolver does not apply resolv.conf search domains, so the short names would answer 502 on /api, /gw and /wk while /healthz stays green; the fully qualified name with the namespace from the downward API fixes it, and SPIRE_POD_NAMESPACE is defined first in the env list, which Kubernetes needs for $(VAR) to expand. The image's env filter (^SPIRE_) admits the new variable without harm. fsGroup 999 and 101 match the runAsUser the chart already sets on those containers, with OnRootMismatch so a big volume is not walked on every start. deploy/k8s was re-rendered, render.sh checks both fixes and its self-test breaks both. The v0.1.0 claims check out: release run 43730758996 succeeded on the tag and the release carries spire-0.1.0.tgz. CI is green. I found no defect. One LOW.
Findings: 1 LOW. Reviewed head 55ce5537f5 as @artyomsv.
artyomsv
left a comment
There was a problem hiding this comment.
✅ APPROVED
Posted as a comment: GitHub does not let an account approve or request changes on its own pull request. The label carries the verdict.
Agent findings: 0 resolved, 1 open. CI: 13/14 checks passing. Head 55ce5537f5, reviewed by @artyomsv.
Open findings:
- [chart/L-1] LOW — #427 (comment)
An empty clusterDomain rendered upstreams ending in .svc.: and every API route would answer 502 again. The render now fails naming the value, with a self-test case. Part of #391
artyomsv
left a comment
There was a problem hiding this comment.
✅ APPROVED
Posted as a comment: GitHub does not let an account approve or request changes on its own pull request. The label carries the verdict.
Agent findings: 0 resolved, 1 open. CI: 13/14 checks passing. Head 9136e42684, reviewed by @artyomsv.
Open findings:
- [chart/L-1] LOW — #427 (comment)
Part of #391.
Found by the review of the first home install (stukans/monorepo#1462). The kind install hid both defects.
resolverdoes not apply the search domains from resolv.conf. So the short namesspire-orchestrator,spire-gatewayandspire-workerfail on Kubernetes. Every/api,/gwand/wkrequest would get a 502 while/healthzstays green. The upstreams are nowspire-<svc>.$(SPIRE_POD_NAMESPACE).svc.<clusterDomain>. The namespace comes from the downward API, so a rendered manifest that kustomize moves to another namespace still works. New value:clusterDomain(defaultcluster.local).fsGroup, so they cannot write to a volume owned by root (Longhorn, EBS). Both pods now setfsGroupwithfsGroupChangePolicy: OnRootMismatch.render.shcheck_first_installasserts both, and its self-test breaks both.deploy/k8swas re-rendered with Helm 3.21.3.deploy/README.mdsays release signatures need cosign 3 or later, because cosign 2 finds no signature. HISTORY records the realv0.1.0release.