Flake: test_publish_after_identify_still_works does not wait for GossipSub mesh GRAFT

[acul71]

Description

tests/core/pubsub/test_gossipsub_identify_aware_publish.py::test_publish_after_identify_still_works flakes on CI (e.g. tox (3.13, core) on PR #1203, job 95452018160).

Failure:

AssertionError: Normal publish (after identify) was not delivered.

(from trio.TooSlowError inside wait_for_pubsub_payload, 10s cap)

Same commit passed tox (3.10/3.11/3.12, core); failure is timing-dependent, not health-monitor related.

Root cause

The test waits for pubsub stream and SUBSCRIBE RPC (wait_for_peer, wait_for_subscription) then publishes. That is necessary but not sufficient for GossipSub delivery.

After subscribe() before connect(), GossipSub.join() creates an empty mesh[topic]. Peers are added to the mesh and receive GRAFT on the heartbeat loop, which:

1. Starts with await trio.sleep(heartbeat_initial_delay) (default 0.1s)
2. Then runs every heartbeat_interval (100s in this test file via _HEARTBEAT_DISABLED = 100 to avoid mcache sliding on slow CI)

If connect/identify completes after the first 0.1s heartbeat, the next mesh repair is ~100s later. The test only waits 10s for the payload → flake. If identify finishes before 0.1s, the first heartbeat GRAFTs the peer and the test passes.

Relevant code:

• Test: tests/core/pubsub/test_gossipsub_identify_aware_publish.py (test_publish_after_identify_still_works)
• Heartbeat delay: libp2p/pubsub/gossipsub.py (heartbeat() → trio.sleep(heartbeat_initial_delay))
• Mesh fill on heartbeat: mesh_heartbeat() / emit_graft

What is already event-driven (good)

• Pubsub.wait_for_peer / wait_for_subscription — trio.Event + fail_after cap
• wait_for_pubsub_payload — blocks on async for subscription, 10s is only a safety cap

What is still time-driven (problem)

• Mesh membership after subscribe-before-connect relies on heartbeat timing, not an event the test waits on.
• Same file still has await trio.sleep(4) in test_three_nodes_publish_before_full_mesh (explicit time wait).

Proposed fix

1. Add a test helper (or use existing patterns) to wait until peer B is in gossipsub.mesh[topic] on both sides (or until GRAFT has been emitted for that topic), event-based with a reasonable timeout — not a longer sleep.
2. Use it in test_publish_after_identify_still_works after the existing wait_for_peer / wait_for_subscription calls and before publish.
3. Optionally replace trio.sleep(4) in the three-node test with the same mesh-ready wait.

Alternative (narrower): trigger an immediate mesh maintenance pass when a new gossipsub peer subscribes to a topic we are already subscribed to, instead of waiting for the next heartbeat — but that is a production-code change; test-side mesh wait is the minimal fix.

Related

• Identify-aware publish tests and prior de-flake: test_publish_before_identify_completes flakes on Windows CI (identify-aware publish timing) #1406 / test(pubsub): de-flake identify-aware publish waits #1444
• Not related to health monitoring (Health monitoring (continues closed PR #915) #1203 / Follow-up: health monitor ping protocol, real-network tests, docs, and GUI #1453)

Are you planning to do it yourself in a pull request?

Yes