(I originally wrote this as a StackExchange answer and it occurred to me that I should probably submit it here!)

The order of files in a tar archive might appear unimportant, but it can actually make a significant difference.

I was just archiving a large number of highly-redundant feed files on an NFS filesystem, and compressed the tar file with "xz -9e" for maximum compression.  For a directory containing 18.8 GB of feed files, the "*.tar.xz" compressed archive ended up being 144.5 MB.  Excellent compression ratio!

After this, I created a SquashFS filesystem of the same directory of feed files, removed the original files, loopback mounted the SquashFS filesystem, and created another compressed archive exactly the same way as before.  This time, the "*.tar.xz" compressed archive containing the same 18.8 GB of feed files ended up being 93.3 MB, over 35% smaller than the original archive!

Investigating this unexpected difference, I discovered that NFS was returning the filenames in random order, while SquashFS returns the filenames in sorted order by default.  This ended up creating a tar file which was better optimized for compression, grouping similar data closer together, allowing xz to perform substantially better.  I could have achieved the same result from the NFS filesystem the first time by adding "--sort=name" to the GNU tar command, if I had realized how much it would help.

Moral of the story?  GNU tar should really default to "--sort=name" instead of "--sort=none" to archive the files in a nicer order and potentially optimize the tar file for better compression!