^^ I think the travis rendering looks better

@GWicke Do you think I should try to use https://www.npmjs.com/package/png-diff for the png comparision?

@GWicke: Any comment on @Physikerwelt's question?

In T94379#1162658, @Physikerwelt wrote:

Do you think I should try to use https://www.npmjs.com/package/png-diff for the png comparision?

That looks like a good option (if we do not supply it an output stream that writes to a file).

I ran the tests locally, and they fail for me (in the sense that the PNGs' texts differ). However, the svg components of the response match. Perhaps it would be acceptable to compare only those and make sure the png component is a non-empty string?

On second thought, I wonder if png-diff would help at all -- the figures in the description are clearly different, so png-diff is bound to give us a mismatch.

Update: when running the tests inside a container, all of the tests pass. Both my host and the container are on Ubunty Trusty, but from the more visible differences, I'm using the Oracle JRE, while the container installs the OpenJDK one.

@mobrovac: Can you fix the travis config?
This should be an open question. So if you have no idea on how to fix that, that would also be an interesting information.

If a solution for T78579 was found, the problem might be resolved automatically.

@Physikerwelt a PR proposing disabling exact PNG checks can be found here.

I think we should reclassify the importance o this bug for two reasons.

1. The problem is also preverlent in production (https://en.wikipedia.org/api/rest_v1/media/math/render/png/ed35a9e1b18047599437224378bcf37e505d5aa3)
2. The png image generated by vagrant and in production looks even worse now (click on the link above)

See https://github.com/wikimedia/MathJax-node/pull/18 for more details