fix(enterprise): mark nodes from unhealthy coordinators as lost #13123
Conversation
|
This stack of pull requests is managed by Graphite. Learn more about stacking. |
Instead of removing the mappings of unhealthy coordinators entirely, mark them as lost instead. This prevents peers from disappearing from other peers if a coordinator misses a heartbeat.
58801cf to
a8ac205
Compare
There was a problem hiding this comment.
This LGTM, but I'm relying on reading code to convince myself that it solves the high level use case of sending LOST updates to peers when a coordinator fails heartbeats.
In addition to the unit test on the heartbeats subcomponent, I'd like to see a test in a similar vein to TestPGCoordinatorSingle_MissedHeartbeats where we simulate a second coordinator entirely by DB calls, make it miss its heartbeats, and then verify that the first coordinator sends out a LOST update.
enterprise/tailnet/pgcoord_test.go
Outdated
| defer coordinator.Close() | ||
|
|
||
| agent := test.NewPeer(ctx, t, coordinator, "agent") | ||
| defer agent.Close(ctx) |
There was a problem hiding this comment.
I think the agent on the real coordinator is superfluous to this test. Just need to test the agent update on the fake coordinator, then have it miss heartbeats.
Fixes #13041
Instead of removing the mappings of unhealthy coordinators entirely,
mark them as lost. This prevents peers from disappearing from
other peers if a coordinator misses a heartbeat.