Message ID | cover.1729252615.git.eclipse7@gmx.net |
---|---|
Headers | show |
Series | Make fate tests succeed with zlib-ng | expand |
On 10/18/2024 1:02 PM, Alexander Strasser via ffmpeg-devel wrote: > Alexander Strasser (2): > Reapply "tests/fate: disable compression for zlib-based codecs" > fate: Make it possible to have alternative reference files I want to add this here, since I brought it up on IRC. I think this entire saga has been pretty silly, and a workaround for what is fundementally a bad test. Testing the exact contents and size of encoded output here is fundementally the wrong thing to test and provides no value as a test. I'd argue it provides negative value. For a lossles encoder, especially one that uses external deps, but this applies to internal ones too, almost no value is provided by ensuring that its output is bit-exact - I'd actually consider it negative value. It's testing the wrong thing (that output remains correct, decodable, metadata, etc.) and only ensures no changes to anything can ever be made without a reference update. - Derek
Quoting Derek Buitenhuis (2024-10-18 16:55:58) > On 10/18/2024 1:02 PM, Alexander Strasser via ffmpeg-devel wrote: > > Alexander Strasser (2): > > Reapply "tests/fate: disable compression for zlib-based codecs" > > fate: Make it possible to have alternative reference files > > I want to add this here, since I brought it up on IRC. > > I think this entire saga has been pretty silly, and a workaround for > what is fundementally a bad test. Testing the exact contents and size > of encoded output here is fundementally the wrong thing to test and > provides no value as a test. I'd argue it provides negative value. > > For a lossles encoder, especially one that uses external deps, but this > applies to internal ones too, almost no value is provided by ensuring > that its output is bit-exact - I'd actually consider it negative value. > It's testing the wrong thing (that output remains correct, decodable, > metadata, etc.) and only ensures no changes to anything can ever be > made without a reference update. +1