See description for explanation of the problem and possible workarounds.

Comment made by: mikerod

I'm confused by http://dev.clojure.org/jira/browse/CLJ-1544?focusedCommentId=36734&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-36734

Isn't the issue just that the "clj" source files also need to be included on the runtime classpath in order to be JITed when needed - since they were not AOTed?

i.e.

Why would the expectation be that "target" contains everything needed by the classpath when the AOT compilation was not done on everything? I don't get a failure when I add this step here.

The AOTed code looks fine. It just calls a `require` on the `myrecord` ns to JIT it. So I'm guessing I'm not seeing something else about this general problem statement.

I just spent quite a bit of time tracking down what I thought might be a variant of this problem. I've been trying to use Colin Fleming's new gradle-clojure plugin and was running into issues with the resulting shadow jar (the equivalent of an uberjar). At the time I believed it to be a problem in the gradle-clojure plugin, but it turned out to be a different issue. The shadow plugin was not preserving the last modified time on files extracted from dependencies, and it resulted in some source files looking newer than the class files. I suspect that Clojure was then recompiling the class, thinking it was out-of-date. This was nasty to track down and quite unexpected, but I can see the sense in the behavior now that I know what's going on. I'm not sure if anything should be done on the Clojure side, but it points to a problem with including the source and AOT files together--you really need to make sure the timestamps are kept intact and I can see that fact being easily overlooked. In this case, it was a plugin completely unrelated to Clojure that had to be fixed. I should also add that taking the infectious approach Stuart mentions is probably an issue in the Gradle and (possibly?) Maven environments, since there are separate plugins for packaging the uberjar.

FWIW, I have a fair amount of information in the ticket for gradle-clojure about the failure mode and the steps I went through to try and track down the problem: https://github.com/cursive-ide/gradle-clojure/issues/8. Also, I've put a pull request in on the shadow plugin to help keep the timestamps intact: https://github.com/johnrengelman/shadow/pull/260. There is also an issue in the shadow plugin describing the problem too: https://github.com/johnrengelman/shadow/issues/259.

Comment made by: sperber

This problem occurs within the compilation of a single library/project, so I don't think this can be solved by simple usages of Leiningen or Boot.

This is actually a runtime problem, not a compilation problem.

AOTed Clojure code cannot see on-the-fly Clojure classes (vars are fine), because of the rules of Java classloader delegation. The on-the-fly code is loaded by a DynamicClassLoader, which delegates up to the classpath loader for the AOTed code. This is a particular problem when AOT code wants to consume a protocol from source, because protocols necessitate a class-level (not just var-level) relationship.

The person who runs a particular Clojure app can solve this problem by making sure their own consumption of AOT compilation is "infectious", i.e. if you want to AOT-compile library A which uses library B, then you need to AOT-compile library B as well.

I think that attempts to have the Clojure compiler magically implement the "infectious" rule above will cause more problems than they solve, and that we should close this ticket and provide good guidance for tools like lein and boot.