Incorrect behaviour during netsplits? about horde

[frekw]

Oh and also, the source can be found here.

[derekkraan]

Hi, thanks again for the bug report.

This is a tricky situation, because this is not an actual netsplit. In fact, both nodes are still in the cluster, but it may seem temporarily that they are not (if you . So if you send an actual message from one side to the other, you'll notice that it arrives and the cluster recognizes again that it is intact.

If you try send(v(3), :hello) in the right pane, you'll see that it succeeds.

I am trying to use local_cluster and schism to add tests for these scenarios to Horde, so I'll leave this issue open until then.

[derekkraan]

I've added an initial test to ensure that the supervisor performs as expected during and after a netsplit. See #93.

[frekw]

Hi!

This is indeed a proper netsplit. The way I simulate it is by changing the cookie for one of the nodes and then using Node.disconnect; see https://github.com/frekw/horde-test/blob/master/lib/hord_test.ex#L26-L31.

This is usually how I simulate netsplits when writing tests using local_cluster as well!

However, I tried this exact test with #92 and then everything seems to be working as expected! 👍

When testing with #92 I did find another bug more closely related to what you describe, which is that if you change the cluster memberships for connected nodes it will fail to remove/restart processes running on nodes in the previous configuration (but not the new one). Not sure what the proper behaviour under those circumstances should be, but I think they should be treated the same way as a netsplit. Do you want me to file a different bug for that?

And lastly, thanks for all your hard work. This is a great library which fills a big gap in the Erlang/Elixir distribution story!