inttest/ct3: fix travis-ci breakage

[ghost]

No description provided.

[ghost]

Let's see if this works better on travis-ci.

[ghost]

Before:

{error_logger,{{2015,1,1},{13,48,39}},
    "Can't set long node name!\nPlease check your configuration\n",[]}

After:

{error_logger,{{2015,1,2},{21,22,11}},
  \"Protocol: ~tp: the name ct_rt3@localhost seems to be in use by another Erlang node\",
[\"inet_tcp\"]}

[ferd]

What happens if we generate a random name? I'm wondering if the issue is actually the name or something regarding their network and instances or something entirely different.

[ghost]

@BanzaiMan mentioned travis-ci/travis-cookbooks@31b08f9

[ghost]

Modified Makefile to test and gather info.

[ghost]

Failed:

• -name ct3

Succeeded:

• -sname ct3
• -name ct3@127.0.0.1
• -name ct3@::1
• -sname ct3@::1

[ferd]

So that's a sign about long names and short names. Erlang considers a domain to be long if it's got a period in it (.) and short otherwise. Or rather, it perceives a host with a . in it to not be short, and to accept mostly anything submitted for a long name:

https://github.com/erlang/otp/blob/maint/lib/kernel/src/net_kernel.erl#L1244-L1268

So 127.0.0.1 is forced long, but ::1 is either. I'm guessing that's why it fails to set the short name with a 127.0.0.1 address, while ct3@::1 works.

For the auto-detection, we'd have to look at the output of {inet_db:gethostname(),inet_db:res_option(domain)} which would explain fairly quickly why it fails.

[ferd]

So you see the problem in the output: {inet_db:gethostname(), inet_db:res_option(domain)} -> {"localhost",[]} That domain value is not set. It can be set by hand with https://github.com/erlang/otp/blob/maint/lib/kernel/src/inet_db.erl#L211 , but in practice, it should be done by the system: https://github.com/erlang/otp/blob/172e812c491680fbb175f56f7604d4098cdc9de4/lib/kernel/src/inet_config.erl#L250-L261

https://github.com/erlang/otp/blob/172e812c491680fbb175f56f7604d4098cdc9de4/lib/kernel/src/inet_config.erl#L231-L245

I'm guessing this bit here has a single name found and ends up not setting both values. Not sure why that happens exactly.

[ghost]

Before I test the container-based workers, test binding to 127.0.0.1.

[ferd]

Well that seems to work. If we can rewrite tests to use that, that would probably be best so our build status is accurate again. In any case I'm not sure we should depend on a properly configured host file when possible to avoid, given that may turn out contributors will also have a broken config on their local machine.

[ghost]

It's a last resort, and it has problems like being IPv4-specific and having to remember that any and all distributed nodes we start use @127.0.0.1.

[ghost]

Now, let's try container-based workers.

[ferd]

That seemed to go well too.

[ghost]

Yes, @BanzaiMan's suggestion to use container-based workers fixed it.
I'd prefer to go with this, considering it worked correctly the way it's supposed to and the name looks sane (ct_run -noshell [...] -name test@testing-worker-linux-docker-f95833d6-3233-linux-5.prod.travis-ci.org).

[ghost]

Cleaned up branch and commit.

[ferd]

Sweet, thanks for the fix. Merging.