Isolated classloader context for poly test is unusual/incorrect

[seancorfield]

See https://github.com/seancorfield/poly-classloader-bug for repro case and instructions.

This is blocking our migration at work. Although we can add an explicit require to get around this initial failure, other tests subsequently fail for classloader-related issues so it seems like solving this initial problem might address some larger issues.

Although the generated project for the repro uses an older version of Polylith, this behavior continues to exist in the latest release.

[furkan3ayraktar]

I quickly checked couple of things and leaving them here for reference.

This is how Polylith creates a new ClassLoader:

polylith/components/common/src/polylith/clj/core/common/class_loader.clj

Lines 67 to 76 in 36b7556

	(defn create-class-loader [paths color-mode]
	(try
	(let [url-array (into-array URL (map path->url paths))
	^URLClassLoader class-loader (url-classloader url-array ext-classloader)]
	(.loadClass class-loader "clojure.lang.RT")
	(eval-in* class-loader '(require 'clojure.main))
	class-loader)
	(catch Exception e
	(println (str (color/error color-mode "Couldn't create classloader for paths: ") paths " "
	(color/error color-mode e))))))

As you can see, after loading clojure.lang.RT only (require 'clojure.main) is run.

I followed what Clojure actually does when it starts and found this:
https://github.com/clojure/clojure/blob/5ffe3833508495ca7c635d47ad7a1c8b820eab76/src/jvm/clojure/main.java#L37-L41

...

final static private Symbol CLOJURE_MAIN = Symbol.intern("clojure.main");
final static private Var REQUIRE = RT.var("clojure.core", "require");
final static private Var MAIN = RT.var("clojure.main", "main");

...

public static void main(String[] args) {
    RT.init();
    REQUIRE.invoke(CLOJURE_MAIN);
    MAIN.applyTo(RT.seq(args));
}

The difference is the RT.init() call before requiring clojure.main. Below is the entire init() and doInit() function it calls in the clojure.lang.RT class.

https://github.com/clojure/clojure/blob/5ffe3833508495ca7c635d47ad7a1c8b820eab76/src/jvm/clojure/lang/RT.java#L466-L501

...

static public void init() {
	doInit();
}

private static boolean INIT = false; // init guard
private synchronized static void doInit() {
	if(INIT) {return;} else {INIT=true;}


	Var.pushThreadBindings(
			RT.mapUniqueKeys(CURRENT_NS, CURRENT_NS.deref(),
			       WARN_ON_REFLECTION, WARN_ON_REFLECTION.deref()
					,RT.UNCHECKED_MATH, RT.UNCHECKED_MATH.deref()));
	try {
		Symbol USER = Symbol.intern("user");
		Symbol CLOJURE = Symbol.intern("clojure.core");


		Var in_ns = var("clojure.core", "in-ns");
		Var refer = var("clojure.core", "refer");
		in_ns.invoke(USER);
		refer.invoke(CLOJURE);
		maybeLoadResourceScript("user.clj");


		// start socket servers
		Var require = var("clojure.core", "require");
		Symbol SERVER = Symbol.intern("clojure.core.server");
		require.invoke(SERVER);
		Var start_servers = var("clojure.core.server", "start-servers");
		start_servers.invoke(System.getProperties());
	}
	catch(Exception e) {
		throw Util.sneakyThrow(e);
	}
	finally {
		Var.popThreadBindings();
	}
}

...

Among other things, doInit function requires clojure.core.server namespace, which is the missing namespace in this scenario.

Maybe invoking RT.init() before requiring clojure.main in the Polylith's ClassLoader creation could solve the issue. We could add (invoke-in class-loader clojure.lang.RT/init []) after loading clojure.lang.RT as following:

 (defn create-class-loader [paths color-mode] 
   (try 
     (let [url-array (into-array URL (map path->url paths)) 
           ^URLClassLoader class-loader (url-classloader url-array ext-classloader)] 
       (.loadClass class-loader "clojure.lang.RT")
       (invoke-in class-loader clojure.lang.RT/init []) ; <- new addition
       (eval-in* class-loader '(require 'clojure.main)) 
       class-loader) 
     (catch Exception e 
       (println (str (color/error color-mode "Couldn't create classloader for paths: ") paths " " 
                     (color/error color-mode e)))))) 

I haven't had time to test this yet. I'll do that when I have some time.