For instance, consider this Gemfile(5):
If you run bundle cache under C Ruby, bundler will retrieve the version of nokogiri for the "ruby" platform. If you deploy to JRuby and run bundle install, bundler is forced to check to see whether a "java" platformed nokogiri exists.
Even though the nokogiri gem for the Ruby platform is technically acceptable on JRuby, it has a C extension that does not run on JRuby. As a result, bundler will, by default, still connect to rubygems.org to check whether it has a version of one of your gems more specific to your platform.
This problem is also not limited to the "java" platform. A similar (common) problem can happen when developing on Windows and deploying to Linux, or even when developing on OSX and deploying to Linux.
If you know for sure that the gems packaged in vendor/cache are appropriate for the platform you are on, you can run bundle install --local to skip checking for more appropriate gems, and use the ones in vendor/cache.
One way to be sure that you have the right platformed versions of all your gems is to run bundle cache on an identical machine and check in the gems. For instance, you can run bundle cache on an identical staging box during your staging process, and check in the vendor/cache before deploying to production.
By default, bundle cache(1) bundle-cache.1.html fetches and also installs the gems to the default location. To package the dependencies to vendor/cache without installing them to the local install location, you can run bundle cache --no-install.