1. Efficient random-access storage. If compression is supported, it is done on a per-member database.
2. Efficient all-at-once compression (block/solid compression). This is either handled explicitly by file format (RAR and 7z, for example), or the archive doesn't support compression (tar) and itself is processed through an external compressor.
On the first category, I would argue Zip has clear superiority over most other file formats. It supports all metadata fields for both DOS/Windows and Unix operating systems, and has very wide support amongst everything. Zip does not support the second category of archival formats.
In the second category, 7z (I don't know about RAR, but I suspect it holds true) only supports DOS attributes and modified dates, making it totally insufficient for backing up a lot of kinds of Unix file trees. tar is generally the de facto standard for Unix and naturally supports all metadata. I would argue gzip is long obsolete; zstd is much faster (it can nearly compete with xz too for file size).
tar files don't really support random access for individual members, lacking a central directory, though uncompressed files can be seeked rather rapidly anyway as the tar program just skips over members (each member encodes its length). Compressed tar files make this worse, since the entire archive must be uncompressed in memory just to find anything. This is a large reason that tar is mostly restricted to being used when extracting individual members is not a generally useful operation. If that is a useful operation in your datasets, go for Zip.
But something like OS install media I would expect to already have compression applied to payload. Eg. Rpm and deb files probably use gzip. That wouldn't compress too well. So maybe just .iso would be fine.
Rar itself seems to be pretty common in certain enthusiast communities where people tend to use windows. For example I still see it when people are passing Android ROMs and unlocking tools around on web forums. And very few other places.