pigeon.dev wiki

scaleway/pigeon-cluster v1.2.0

By pigeon.dev wiki

Every job in a pigeon-cluster fleet was forced onto the cluster's shared Private Network with no public IP (ADR-0145/0146), and that shared Private Network/Public Gateway had no lifecycle independent of the cluster as a whole — it was all-or-nothing.

jobs gains two new optional fields:

  • enable_private_network (default true, unchanged behavior). false makes this job a plain, undecorated compute-instance call — no private NIC, no dependency on the cluster's shared Private Network/Gateway at all, with its own public IPv4 as its only network path.
  • enable_ipv4 (default false) — reinstates the per-job public-IP override removed in v1.0.0. A job that keeps its Private Network attachment can now also request its own public IP; note this still doesn't help with inbound SSH reachability while the cluster's gateway is pushing a default route (the ADR-0146 finding), so it's primarily useful once enable_public_gateway = false or for outbound-only use. A job with enable_private_network = false always gets a public IP regardless of this field, since it would otherwise have no network path at all.

cluster_config gains two new optional fields, both defaulting true:

  • enable_private_network — tear down the cluster's shared Private Network (and its Gateway-Network attachment) once no job needs it anymore, independent of the Public Gateway.
  • enable_public_gateway — tear down the metered Public Gateway independent of the Private Network, e.g. once nothing needs internet/private-Object-Storage egress.

A new validation block requires both of the above to be true whenever enable_bastion is true, since the bastion is only reachable through the Gateway's PAT rule onto the shared Private Network.

The four previously-unconditional shared resources (scaleway_vpc_private_network.jobs, scaleway_vpc_public_gateway_ip.jobs, scaleway_vpc_public_gateway.jobs, scaleway_vpc_gateway_network.jobs) move to count-gated, each protected by a moved block (same precedent as compute-instance's own enabled kill switch, ADR-0126), so no existing caller's state breaks.

Purely additive — every new field defaults to reproducing today's exact resource graph for any caller that sets nothing new. The one real consumer, workloads/willowgraysen.com/terraform/pigeon-cli/cluster/, needs no required change.

See ADR-0149 for the full design discussion and rationale.

#254