#752 — investigate cpx21/cpx31 availability in EU DCs
Concrete proof gathered against the live Hetzner Cloud API on 2026-05-04.
GET /v1/server_types LISTS cpx11/cpx21/cpx31/cpx41 with full EU prices in
fsn1/nbg1/hel1, but POST /v1/servers rejects every order for those SKUs in
those DCs with:
{"error":{"code":"invalid_input",
"message":"unsupported location for server type"}}
Probed all 6 (SKU × DC) combinations end-to-end via real POST + immediate
DELETE. cpx22 + cpx32 were also probed as a sanity check and returned
ORDERED. The /v1/server_types price entry is misleading: Hetzner advertises
prices for every (SKU, location) pair regardless of orderability.
Conclusion: NO SKU bump-back. cpx22 + cpx32 (PR #744) remain the floor.
README + variables.tf docstrings now carry the durable reproducer so future
engineers don't re-attempt cpx21/cpx31.
#753 — kubectl retry / LKG observer reliability
/tmp/autopilot.sh updated (script lives outside the repo, on the VPS):
• Every kubectl call carries --request-timeout=8s so a hung TLS handshake
surfaces as a fast empty rather than a 30s+ stall.
• Last-known-good (LKG) state held across transient flakes: hr/cert/nodes
no longer flip to "0/0 nodes=0" on a single failed poll.
• Only 3 consecutive transients count as a real failure; below the
threshold the observer prints "hr=<LKG> (transient N/3)".
UI side: the wizard's StatusPill / ApplicationPage drive off SSE from
catalyst-api (useDeploymentEvents.ts), not direct kubectl polling, so no UI
change needed. catalyst-api itself uses client-go (helmwatch / phase1_watch),
not exec kubectl, so its observer is not subject to the same shell-out flake.
Co-authored-by: hatiyildiz <269457768+hatiyildiz@users.noreply.github.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
05065b66d6