Asciidoctor core (Ruby) relies on https://github.com/tigris/open-uri-cached which cache content on disk. I don't like this approach since it creates "temporary" files in an arbitrary location (by default /tmp/open-uri-#{Process.uid}).
In addition, the library does not automatically clean/remove/invalidate cache. As a result, a long running process could lead to a disk saturation.

I think we should define a default cache size and use an in-memory cache. It might be interesting to configure a global cache or tie a cache to a document.

Interestingly, the cache is actually used between executions since the uid is stable (for instance, when running the CLI multiple times). It's also not possible to disable the cache in a program once enabled:

Asciidoctor.convert("image::http://localhost:5000/cc-zero.svg[opts=inline]", { 
  safe: 'safe', 
  attributes: {'allow-uri-read' => '', 'cache-uri' => ''} 
})

Asciidoctor.convert("image::http://localhost:5000/cc-zero.svg[opts=inline]", { 
  safe: 'safe',
  attributes: {'allow-uri-read' => ''}
})

The second call will use the cache despite the fact that cache-uri is absent. The reason is that open-uri/cache was loaded (and cannot or at least is not unloaded)

I think the URI/URL caching layer in Asciidoctor Ruby could be improved by making it pluggable. We integrated open-uri/cache in the early days to provide a quick win, but it's clearly a simple implementation.

I think the URI/URL caching layer in Asciidoctor Ruby could be improved by making it pluggable

An abstract layer on content read (file/URI) would be my dream ✨

We integrated open-uri/cache in the early days to provide a quick win, but it's clearly a simple implementation.

The most straightforward approach is to provide a JavaScript compatible implementation of open-uri/cache.
Opal cannot unload a module (or I least I don't know how) once loaded so the behavior is consistent with Asciidoctor Ruby.

I think I will store the cache in memory with a default value of 16Mb and use a simple implementation.