epub: Honor SOURCE_DATE_EPOCH for dcterms:modified and dc:identifier

[bmwiedemann]

The EPUB content.opf embedded the build time as dcterms:modified and a random UUID as dc:identifier, making the .epub files differ between two builds of the same docs.

When SOURCE_DATE_EPOCH is set
(https://reproducible-builds.org/docs/source-date-epoch/), use it as the modification time, and derive the identifier from the project name and version instead of random data. Behavior without SOURCE_DATE_EPOCH is unchanged.

This PR was done while working on reproducible builds for openSUSE. Together with #2255, this makes our erlang-doc package build reproducible.

[josevalim]

Suggested change

	case System.get_env("SOURCE_DATE_EPOCH") do
	nil ->
	:calendar.universal_time()
	epoch ->
	case Integer.parse(epoch) do
	{seconds, ""} -> :calendar.system_time_to_universal_time(seconds, :second)
	_ -> :calendar.universal_time()
	end
	end
	with epoch when is_binary(epoch) <- System.get_env("SOURCE_DATE_EPOCH"),
	{seconds, ""} <- Integer.parse(epoch) do
	:calendar.system_time_to_universal_time(seconds, :second)
	else
	_ -> :calendar.universal_time()
	end

[josevalim]

Suggested change

	time =
	with epoch when is_binary(epoch) <- System.get_env("SOURCE_DATE_EPOCH"),
	{seconds, ""} <- Integer.parse(epoch) do
	:calendar.system_time_to_universal_time(seconds, :second)
	else
	_ -> :calendar.universal_time()
	end
	{{year, month, day}, {hour, min, sec}} = time
	{{year, month, day}, {hour, min, sec}} =
	with epoch when is_binary(epoch) <- System.get_env("SOURCE_DATE_EPOCH"),
	{seconds, ""} <- Integer.parse(epoch) do
	:calendar.system_time_to_universal_time(seconds, :second)
	else
	_ -> :calendar.universal_time()
	end

[josevalim]

I wonder if we should either always default to config.project/config.version, but that is likely not randomic enough OR add the source_date_epoch here. Basically, we always build the datatime first, and use it to generate the datetime and the UUID.

[bmwiedemann]

Adding SOURCE_DATE_EPOCH to the hash would mean we get extra diffs in the output for unrelated changes. That is OK, as it does not block verification with identical build inputs. But is there really an advantage in it?
Are there other relevant inputs to the epub output? Maybe those could be added to the hash as well.

[josevalim]

Good questions. I did a quick check and this is actually good. My suggestion is to always apply this new version, instead of conditionally.

[github-actions]

📦 Docs artifacts are ready: https://github.com/elixir-lang/ex_doc/actions/runs/30304750221/artifacts/8667987990

[josevalim]

💚 💙 💜 💛 ❤️