Fix CI and flaky tests

[mopp]

Background

The CI fails on the main branch.

What I did

• Fix GitHub Actions
  • Use setup-beams
    • https://github.com/actions/setup-elixir is archived.
    • https://github.com/erlef/setup-elixir is redirected to https://github.com/erlef/setup-beam
  • Use ubuntu-20.04 because ubuntu-latest does not have old Erlang/OTP and Elixir versions
    • References:
      • https://repo.hex.pm/builds/otp/ubuntu-20.04/builds.txt
      • https://repo.hex.pm/builds/otp/ubuntu-22.04/builds.txt
      • Fix incompatibility Elixir/Erlang OTP error in GitHub CI telemetry_metrics#104
• Fix flaky tests
  • See the comment on the tests.

How to test

I confirmed that the CI was passed at the forked repository.
See mopp#2

[mopp]

We need to define pair to avoid the error below.

Error: Requested Elixir / Erlang/OTP version (1.10.4 / 24.0) not found in version list (did you check Compatibility between Elixir and Erlang/OTP?).Elixir and Erlang/OTP compatibility can be found at: https://hexdocs.pm/elixir/compatibility-and-deprecations.html

[mopp]

These randomly failed.
See https://github.com/mopp/telemetry_metrics_statsd/actions/runs/5868819319/job/15912473463?pr=2

TelemetryMetricsStatsd.udp_error/2 uses handle_cast/2 internally.

telemetry_metrics_statsd/lib/telemetry_metrics_statsd.ex

Lines 413 to 415 in f9229e1

	def udp_error(reporter, udp, reason) do
	GenServer.cast(reporter, {:udp_error, udp, reason})
	end

TelemetryMetricsStatsd.get_udp does NOT use the process.
It just selects randomly one UDP socket from the ETS table.

telemetry_metrics_statsd/lib/telemetry_metrics_statsd.ex

Lines 392 to 403 in f9229e1

	def get_udp(pool_id) do
	# The table can be empty if the UDP error is reported for all the sockets.
	case :ets.lookup(pool_id, :udp) do
	[] ->
	Logger.error("Failed to publish metrics over UDP: no open sockets available.")
	:error
	udps ->
	{:udp, udp} = Enum.random(udps)
	{:ok, udp}
	end
	end

So, it can return another UDP socket before completing the handle_cast/2.
Then, the log does not appear and the test fails.
Sometimes it returns the same UDP socket and retries the function which is given to eventually.
It passed in the case.

I use TelemetryMetricsStatsd.get_pool_id/1 to wait for the processing because it uses handle_call/2 internally.

[arkgil]

Ah, good catch. The reason this started flaking is because we introduced the default pool size of 10 after this test was written.

It should all work if we set the pool_size of the reporter to 1 in this test. Then the UDP socket we get from get_udp/1 will only be different when the reporter reopens a socket, which means it processed the error.

[mopp]

@arkgil Thanks for your review!
Got it. I missed the pool_size option.

I reverted the commit and set the pool_size to 1.
It passed. (See mopp#4)

Please check the latest commits 🙏

[mopp]

@arkgil Hi 👋 Please review and merge 🙏

[arkgil]

Thanks so much @mopp! Left one comment regarding the flaky test.

[arkgil]

Ah, good catch. The reason this started flaking is because we introduced the default pool size of 10 after this test was written.

It should all work if we set the pool_size of the reporter to 1 in this test. Then the UDP socket we get from get_udp/1 will only be different when the reporter reopens a socket, which means it processed the error.