HiveDeployReliability ledger
COVERAGE (mongodb × aws)2026-10-15T13:01Zgate version v1

21_percona_eks_hardened#2

FAILED
bucket:AGENTclass:three-faults-one-hardened-render

evidence:

Three agent-side faults in one hardened render, each one the deterministic module gets right: (1) a pmm block without an image — the operator's CRD validates it even with pmm disabled (two apply deaths, recorded as the 26-character multierror header while the cause sat a line below); (2) a default-deny NetworkPolicy whose only egress was a pod-selector rule on 443, so the operator could not reach the cluster API and crash-looped (the ladder added an ip_block route at attempts 3-4); (3) a requireTLS cluster without net.tls.allowConnectionsWithoutCertificates, so every certificate-less client — the probe included — was refused; attempt 5 switched to preferTLS, which does not help. The ladder exhausted five attempts and the harness banked the row at once (the 20#2 give-up fix held: no blind redeploy).

  1. date2026-10-15T13:01Z
  2. deploy statusfailed (8440s)
  3. gate––––healthz=— connection=— logs=— metrics=—
  4. teardown cleanTrue

resolution:

Fixed as rules: a pmm block without an image is refused before plan (8931dc8f); a terraform multierror carries its bullet lines into the repair signature (98453bc2); a default-deny policy must leave the operator an ip_block route (01717149); a TLS cluster must admit certificate-less clients, with a matching runtime verdict (c83f80c2). Scenario 21 passed on its 30 Oct re-run.

Recorded reason (scrubbed)

ladder exhausted (auto-repair gave up)