Prefer Brotli over Gzip

I see that Brotli was released in v4.1.0 - awesome, thanks for adding!

There’s one small issue. Chrome’s typical Accept-Encoding header looks like this:

Accept-Encoding:gzip, deflate, sdch, br

It seems that koa-send interprets the order as a preference, and if it sees gzip, serves that first. That means if both .gz and .br files are available, the Brotli variants will never get sent:

In instances where more than one compression algorithm is accepted, I think it’d be better to automatically send the smallest file.

[jtmthf]

@leebenson It happens. Anyways I looked into why chrome was not retrieving the brotli assets and it’s because brotli is https only. Here is the relevant chromium issue https://bugs.chromium.org/p/chromium/issues/detail?id=452335

[bredthau]

I too have run into this issue last weekend. Wanting to get it to work I spent some time analysing the issue.

First of all as @nickhstr mentioned ‘deflate’ shouldn’t be included in ctx.acceptEncodings(). What is the point of checking whether deflate would be a better match when the library doesn’t provide an option for serving deflate compressed data. The problem is that chrome sends br as the last entry of accepted encodings and ctx.acceptEncodings() seems to process the encodings in the order specified in the header (at least in the absense of priority specifications), thus always matching deflate instead of br.

Changing this however totally broke my website: The type() function for setting the correct MIME-Type in the face of encryption only deals with .gz, which means that .br files will be served as application/octet-stream, which chrome doesn’t like to much for html files (which it always tried to download) or css files (which weren’t applied correctly). Besides this problem the type function might also set the wrong MIME-Type when trying to serve actual .gz files, so the correct fix is to only remove the outer file extension when actually using the appropriate compression algorithm.

Having fixed both of these issues it works correctly. If there is interest I could provide a patch for this, however there is still a small caveat I couldn’t fix: Fixing the MIME-Type leads to the Unittests for brotli failing with a timeout. I haven’t looked too far into it, but it seems that the testing framework doesn’t comprehend brotli responses for text content.