First of all, ContactPoint HTTP probing is done in parallel, so the first and second would happen pretty close to each other, then the failure happened.

Second, Depends on the RequiredContactPointNr setting and ContactWithAllContactPoints setting.

• If RequiredContactPointNr is 2 and ContactWithAllContactPoints is false, the third failure is ignored and cluster formed immediately after the first 2 succeeded.
• If RequiredContactPointNr is 2 and ContactWithAllContactPoints is true, the third failure blocks cluster formation until that entry is removed from either being stale or removed from the next round of Discovery resolution.

Join decider ticks, discovery resolution ticks, and HTTP probing ticks happens in their own timers.

• Decider ticks and HTTP probing tick timer interval is set to contact-point.probe-interval,
• decider ticks runs on the BootstrapCoordinator,
• HTTP probing runs in the HttpContactPointBootstrap,
• HTTP probing have a failure timeout of 3s

Discovery resolution tick timer interval is set to contact-point-discovery.interval, it is run in the BootstrapCoordinator

the third failure blocks cluster formation until that entry is removed from either being stale or removed from the next round of Discovery resolution.

Question I'm really asking is that if a node somehow dies shortly after being successfully pinged, before a cluster is formed, is it possible to form a cluster under the RequiredContactPointNr due to a stale entry?

I guess this is where our and scala implementation diverge with this update.
In the scala implementation, there is no stale contact point timeout setting, possible dead node entry is being guarded and pruned by a TTL that is the same as the probe-failure-timeout value. If a HTTP probe died without updating its entry in the contact point list, the coordinator is guaranteed to remove the entry at the same interval as the probe timeout interval.

Here, we're lenghtening that value so there is actually a possibility that there will be a race condition where a contact point is actually dead but its not being pruned out of the contact point list, if that makes any sense.

possible dead node entry is being guarded and pruned by a TTL that is the same as the probe-failure-timeout value. If a HTTP probe died without updating its entry in the contact point list, the coordinator is guaranteed to remove the entry at the same interval as the probe timeout interval.

So we should redesign this feature then - no need for a separate stale value, just always slide the timeout with the interval at all times (i.e it's never a hard timeout, it's always a +n value.) That'd be even simpler.

We'd need to remove 1.5.26 from NuGet then
Nevermind, I somehow crossed Akka.Management and Akka.NET, again... geeze

ok, done