Rewrite/deploy manager

[Tanker2020]

Summary

Ports the Python oper8 deploy manager layer to Go. Provides the DeployManager interface (the single point of contact between operator components and the Kubernetes cluster) along with a fully in-memory DryRunDeployManager for unit testing, and owner-reference helpers.

Files changed:

• rewrite/deploymanager/deploymanager.go — DeployManager interface + types
• rewrite/deploymanager/ownerref.go — OwnerRef, ApplyOwnerRef
• rewrite/deploymanager/dryrun.go — DryRunDeployManager
• rewrite/deploymanager/dryrun_test.go — 19 table-driven tests
• rewrite/.golangci.yml — shared lint config for all rewrite PRs
• .github/workflows/pr2-deploy-manager.yml — CI

Depends on: PR-1 (dag/ package) — stacked branch, but the deploy manager package itself does not import dag/. PR-2 can be reviewed independently of PR-1.

---

What was ported

Python	Go equivalent
deploy_manager/base.py — DeployManagerBase, DeployMethod	deploymanager/deploymanager.go
deploy_manager/kube_event.py — KubeWatchEvent, KubeEventType	deploymanager/deploymanager.go (WatchEvent, EventType)
deploy_manager/owner_references.py — update_owner_references	deploymanager/ownerref.go (OwnerRef, ApplyOwnerRef)
deploy_manager/dry_run_deploy_manager.py — DryRunDeployManager	deploymanager/dryrun.go

---

Design decisions

Abstract base class → Go interface

Python used DeployManagerBase(abc.ABC) with @abc.abstractmethod decorators. Go uses a plain interface. No base struct, no embedded types. Every implementation must satisfy all six methods — the compiler enforces it.

(success, changed) → (changed bool, err error)

Python returned two booleans: success (did the operation not error) and changed (did the cluster state change). Go collapses these: if there is an error, the operation failed; the single return value is changed. Callers use the standard if err != nil pattern.

watch_objects generator → <-chan WatchEvent

Python's watch_objects was a generator/iterator. Go returns a <-chan WatchEvent that receives events until the passed context.Context is cancelled, at which point the channel is closed. The caller ranges over it naturally.

DryRunDeployManager cluster store

Python used a nested defaultdict. Go uses an explicit type alias clusterStore = map[string]map[string]map[string]map[string]map[string]any (keyed: namespace → kind → apiVersion → name → object). All mutations are protected by a single sync.RWMutex making the implementation safe under -race.

disable() → Delete()

Python named the delete method disable(). The Go interface names it Delete() to match Kubernetes/controller-runtime conventions and make the semantics unambiguous.

Deep copy via JSON round-trip

Get() and List() return deep copies so callers cannot accidentally mutate the store. The copy is done via json.Marshal + json.Unmarshal — simple, correct for map[string]any, no external dependency.

strict_resource_version / generate_resource_version not ported

Python's DryRunDeployManager had optional strict_resource_version and generate_resource_version modes used for testing optimistic concurrency. These are not needed for the current component/controller tests and are left for a future PR if required.

---

Test coverage (19 tests)

Area	Tests
Deploy	create (changed=true), idempotent re-deploy (changed=false), field update (changed=true)
Get	not found (nil, nil), found, returns deep copy
Delete	existing (changed=true), non-existent (changed=false)
List	all objects, label selector filtering
SetStatus	sets status, not-found error
Watch	receives ADDED event on deploy, receives DELETED event on delete, channel closes on context cancel
OwnerRef	stamps reference, idempotent (no duplicate), cross-namespace skipped, Deploy with manageOwnerRefs=true

All tests run with -race.