CBG-5838: wait for the local replicator state in ISGR upsert tests

[torcolvin]

CBG-5838

TestRequireReplicatorStoppedBeforeUpsert failed in CI with a 400 on the upsert that follows the stop:

Messages:   	Response status 400 "Bad Request" (expected 200 "OK")
            	for PUT <http://127.0.0.1/db/_replication/replication1> : {"error":"Bad Request","reason":"Replication must be stopped before updating config"}

Cause

GetReplicationStatus falls back to the cfg's target state while no replicator has published status, so a replication reports running before anything has started. The wait for running returned on its first poll.

RefreshReplicationCfg reads the cfg, then starts assigned replications while holding activeReplicatorsLock. The test carried on with that start still in flight, stopped the replication, and the start landed afterwards on the pre-stop cfg. The upsert then found a running replicator. The 20.7 ms the failing PUT took is it waiting on activeReplicatorsLock while that start held it.

Fix

Report starting for that stub. The target state stays the answer for every other state, because stopped and error are true when no replicator has reported, and running is not.

Two things already agree with this:

• PUT _replicationStatus/{id}?action=start answers starting through transitionStateName whenever the current state differs from the target.
• UpsertReplication gates on stopped, so the rejection of a config change during a start is unchanged.

A wait for running now needs either a local replicator in that state or a status document, which only a running replicator writes. No start can be in flight when the wait returns, because the refresh holds activeReplicatorsLock across Start and the status read needs the read lock.

Behaviour changes

• GET _replicationStatus and the _cluster response report starting instead of running for a replication that no replicator has published status for.
• ?activeOnly=true returns starting as well as running, so a replication does not drop out of the listing while it starts. The spec description of the parameter is updated to match.
• PUT _replicationStatus/{id}?action=start answers starting instead of running in that window.
• A caller polling for running now waits for a replicator instead of returning at once.

The status document is still the only cross-node channel, so a replication whose status write fails, or whose document expired, reads starting on other nodes until the next successful publish. That trades a silent false running for a visible false starting.

Tests

TestReplicationStatusBeforeReplicatorStarts adds a replication assigned to the local node, never runs a refresh, and asserts the reported status. It fails on the old code with expected: "starting", actual: "running" and has no timing dependence.

Verified by delaying RefreshReplicationCfg between its cfg read and the start, which widens the window deterministically: the endpoint reported running 3/3 before the change and starting 3/3 after, and TestRequireReplicatorStoppedBeforeUpsert passes 5/5 under that delay at 8.1 s per run. Without the delay it passes 100/100 under -race. The full rest/replicatortest package, the db replication tests and the rest replication tests pass.

unassigned is added to the ISGRReplicationState enum in the API spec. The code already returned it.

🤖 Generated with Claude Code

[github-actions]

Redocly previews

• Admin API
• Public API
• Metric API
• Diagnostic API

[Copilot]

Copilot review overview

🟡 Changes recommended

The empty target state is treated as running during startup but remains uncovered and incorrectly reported by the new fallback.

Review effort: Balanced
Findings: 1

Open (1)

• Handle empty target state as running in active-only filtering · New

What changed in this PR

Updates ISGR status reporting so pending replications report starting instead of prematurely reporting running.

Changes:

• Reports assigned, not-yet-started replications as starting.
• Includes starting replications in activeOnly results.
• Updates tests and OpenAPI documentation.

File	Description
docs/​api/​components/​schemas.yaml	Documents the unassigned state.
docs/​api/​components/​parameters.yaml	Clarifies activeOnly behavior.
db/​sg_replicate_cfg.go	Corrects pending replication status handling.
db/​sg_replicate_cfg_test.go	Tests pre-start and active-only status behavior.

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[factory-droid]

Droid finished @torcolvin's task —— View job

---

The status fallback and activeOnly behavior changes look correct and are backed by focused tests for the startup window and reconnecting behavior. One small OpenAPI wording mismatch remains: the activeOnly parameter description should reference the response field status, not state.