GCP shared-service event capacity contract¶
Issue #1816 establishes one immutable catalog in
shifter/installation/capacity_profiles_gcp.py. A profile ID encodes a version
and participant envelope. The resolver projects that one object into four
consumers:
- Terraform: access-node minimum/maximum, Cloud SQL tier/HA/disk, Redis tier and memory, plus the capacity identity label.
- Helm: portal and guacd minimum replicas/resources/HPAs/PDBs, the fixed single-replica Guacamole client, process counts, absolute tunnel ceilings, WebSocket keepalive, termination grace, backend timeout, and connection drain.
- Gate: concurrency, hold and latency budgets, ready guacd minimums, and saturation ceilings.
- Drift: an allowlisted desired-state document compared with live provider and Kubernetes observations. Missing observations fail closed; raw cloud payloads and secrets are not emitted. Cloud SQL disk capacity is a minimum because the provider cannot shrink storage; drift accepts a retained larger disk after profile scale-down but still rejects an undersized disk.
The installation setting shared_service_capacity_profile is the only selector.
Provider, deployment, traffic, and gate profiles remain distinct concepts.
gcp-deploy merges the Helm projection into the generated values of every
rollout. The Helm chart is the only GCP renderer (ADR-007, ADR-070). Thus a
release can not keep the profile identity and silently restore the default
workload resources.
The p30 profile provisions five portal replicas, two guacd replicas, one
Guacamole client, regional db-custom-4-15360 Cloud SQL, 8 GiB Standard HA
Redis, and access-node autoscaling from three to eight nodes. Portal and guacd
HPAs may add headroom, but their minimum replicas are sized to carry the gate
without a scale-up race. The Guacamole client remains one replica to preserve
the token-affinity contract from issue #928.
Database connection accounting is explicit: Django retains its qualified
CONN_MAX_AGE=0 behavior, the Guacamole/MyBatis active JDBC pool is budgeted at
its fixed ten-connection limit, and the remaining Cloud SQL budget is reserved
for failover and operations. POSTGRESQL_ABSOLUTE_MAX_CONNECTIONS is a
Guacamole tunnel-concurrency guard; it is not misrepresented as the JDBC pool.
The strict gate reuses the range functional smoke's Identity Platform/TOTP,
target selection, Guacamole URL parsing, and wire protocol. A successful client
must observe ready, acknowledge a post-ready sync, and keep parsing the
tunnel for the full hold. Cloud Monitoring supplies restart, CPU, Cloud SQL,
Redis, and external HTTPS load-balancer evidence; bounded kubectl get and
gcloud compute backend-services get-health reads supply effective readiness
and backend health. No missing signal is converted to zero.
The binding architecture analysis is GCP shared-service event capacity preflight. The operator procedure is GCP event capacity profiles.