1 Commits

Author SHA1 Message Date
shankar0123 30b251ea13 feat(agent): per-target deploy mutex serializes concurrent deploys to the same target
Phase 2 of the deploy-hardening I master bundle. Closes the agent-side
race window where two concurrent renewals against the same target ID
(typical: two SAN entries renewing in the same window) would otherwise
collide on the connector's temp-file path or run the reload command
against itself.

The Agent struct grows a sync.Map of *sync.Mutex keyed on target ID;
targetDeployMutex(targetID) lazy-init's one on first acquisition.
executeDeploymentJob acquires the mutex before connector.DeployCertificate
and releases via defer at function exit — the lock spans the full
Deploy duration including PreCommit (validate), atomic-rename, PostCommit
(reload), and post-deploy verify (Phases 4-9).

Granularity per frozen decision 0.5: one mutex per target ID, NOT per
(target, cert) pair. Cert deploy throughput is operator-grade
tens-per-minute; coarse serialization simplifies reasoning about
reload-side race windows. Mutexes live for the agent's lifetime —
target IDs are bounded so no janitor needed (~16 bytes per entry).

Empty TargetID (defensive — should never happen for deploy jobs)
bypasses the lock to avoid a singleton serialization point pulling
all targetless work onto a shared mutex.

Tests (5 named cases in cmd/agent/deploy_mutex_test.go):

- TestAgent_ConcurrentDeploysToSameTarget_Serialize — race-detector
  smoke; 10 goroutines acquire same target's mutex; max-in-flight
  asserts == 1
- TestAgent_DifferentTargetIDs_ParallelizeIndependently — per-target
  granularity proof
- TestAgent_EmptyTargetID_ReturnsNilMutex — defensive contract
- TestAgent_TargetMutex_IsStable — sync.Map LoadOrStore returns same
  pointer across calls
- TestAgent_TargetMutex_RaceLookup — race-free under N=50 concurrent
  lookups for same key

go test -race -count=1 green; gofmt + go vet + golangci-lint v2.11.4
all 0 issues against my new code (pre-existing import-grouping drift
in agent_test.go / main.go / verify*.go is unrelated to this change
and not caught by `go fmt ./...` which CI uses).

Phase 3 next: ValidateOnly method on target.Connector interface;
default impl returns ErrValidateOnlyNotSupported across all 13
connectors.
2026-04-30 14:32:40 +00:00