Why do some webservers complain about memory leaks they create?

Question

The title might be a bit strong, but let me explain how I understand what happens. I guess this happened with Tomcat (and the message cited comes from Tomcat), but I'm not sure anymore.

TL;DR At the bottom there's a summary why I'm claiming that it is the web servers' fault.

I might be wrong (but without the possibility of being wrong there would be no reason to ask):

• An application uses a library
• the library uses a ThreadLocal
• the ThreadLocal refers to an object from the library
• each object refers to its ClassLoader

The webserver

• pools its worker threads for efficiency
• lends an arbitrary thread to an application
• does nothing special (w.r.t. the thread pool) when an application stops or redeploys

If I understand it correctly, after a redeploy the old "dirty" threads continue to be reused. Their ThreadLocals refer to the old classes which refer to their ClassLoader which refer to the whole old class hierarchy. So a lot of stuff stays in the PermGen space which over time leads to an OutOfMemoryError. Is this right so far?

---

I'm assuming two things:

• the redeploy frequency is a few time per hour
• the thread creation overhead is a fraction of a millisecond

So a complete thread pool renewal upon each redeploy costs a fraction of a millisecond a few times per hour, i.e., there's a time overhead of 0.0001 * 12/3600 * 100% i.e. 0.000033%.

But instead of accepting this tiny overhead, there are countless problems. Is my calculation wrong or what am I overlooking?

---

As a warning we get the message

The web application ... created a ThreadLocal with key of type ... and a value of type ... but failed to remove it when the web application was stopped.

which should be better stated as

The web server ... uses a thread pool but failed to renew it after stopping (or redeploying) an application.

Or am I wrong? The time overhead is negligible even when all threads get recreated from time to time. But clearing their ThreadLocals before they are provided to the applications would suffice and be even faster.

Summary

There are some real problems (recently this one) and the user can do nothing about it. The library writers sometimes can and sometimes can not. IMHO the web servers could solve it pretty easily. The thing happens and has a cause. So I'm blaming the only one party which could do anything about it.

Proposal for what the web server should exactly do

The title of this question is more provocative than correct, but it has its point. And so does the answer by raphw. This linked question has another open bounty.

I think the web servers could solve it as follows:

• ensure that each thread gets reused (or killed) sometime
• store a LastCleanupTimestamp in a ThreadLocal (for new threads it's the creation time)
• when re-using a thread, check if the cleanup timestamp is below some threshold (e.g., now minus some delta, e.g., 1 hour)
• if so, clean all ThreadLocals and set a new LastCleanupTimestamp

This would assure that no such leak exists longer than delta plus the duration of the longest request plus the thread turnaround time. The cost would compose as follows:

• checking a single ThreadLocal (i.e., some nanoseconds) per request
• cleaning all ThreadLocals reflectively (i.e., some more nanoseconds once each delta per thread)
• the cost from removing the data possibly useful for the application which stored them. This can't break an application as no application can assume to see a thread containing the thread locals it has set (since it can't even assume to see the thread itself anymore), but it may cost time needed to recreate the data (e.g., a cached DateFormat instance if someone still uses such a terrible thing).

It could be switched off by simply setting the thresold, if no app has been undeployed or redeployed recently.

Answer 1

TL;DR It's not web servers that create memory leaks. It's you.

Let me first state the problem more explicitly: ThreadLocal variables often refer to an instance of a Class that was loaded by a ClassLoader that was meant to be exclusively used by a container's application. When this application gets undeployed, the ThreadLocal reference gets orphaned. Since each instance keeps a reference to its Class and since each Class keeps a reference to its ClassLoader and since each ClassLoader keeps a reference to all classes it ever loaded, the entire class tree of the undeployed application cannot get garbage collected and the JVM instance suffers a memory leak.

Looking at this problem, you can optimize for either:

• Allow as many requests per second as possible even throughout a redeploy (thus keep response time short and reuse threads from a thread pool)
• Make sure that threads stay clean by discarding threads once they were used when a redeploy occurred (thus patch forgotten manual cleaning)

Most developers of web applications would argue that the first is more important since the second can be achieved by writing good code. And what would happen when a redeploy would happen concurrently to long lasting requests? You cannot shut down the old thread pool since this would interrupt running requests. (There is no globally defined maximum for how long a request cycle can take.) In the end, you would need a quite complex protocol for that and that would bring its own problems.

The ThreadLocal induced leak can however be avoided by always writing:

myThreadLocal.set( ... );
try {
  // Do something here.
} finally {
  myThreadLocal.remove();
}

That way, your thread will always turn out clean. (On a side note, this is almost like creating global variables: It is almost always a terrible idea. There are some web frameworks like for example Wicket that make a lot of use of this. Web frameworks like this are terrible to use when you need to do things concurrently and get very unintuitive for others to use. There is a trend away from the typical Java one thread per request model such as demonstrated with Play and Netty. Do not get stuck with this anti-pattern. Do use ThreadLocal sparingly! It is almost always a sign of bad design.)

You should further be aware that memory leaks that are induced by ThreadLocal are not always detected. Memory leaks are detected by scanning the web server's worker thread pool for ThreadLocal variables. If a ThreadLocal variable was found the variable's Class reveals its ClassLoader. If this ClassLoader or one of its parents is that of the web application that just got undeployed, the web server can safely assume a memory leak.

However, imagine that you stored some large array of Strings in a ThreadLocal variable. How can the web server assume that this array belongs to your application? The String.class was of course loaded with the JVM's bootstrap ClassLoader instance and cannot be associated with a particular web application. By removing the array, the web server might break some other application that is running in the same container. By not removing it, the web server might leak a large amount of memory. (This time, it is not a ClassLoader and its Classes that are leaked. Depending on the size of the array, this leak might however even be worse.)

And it gets worse. This time, imagine that you stored an ArrayList in your ThreadLocal variable. The ArrayList is part of the Java standard library and therefore loaded with the system ClassLoader. Again, there is no way of telling that the instance belongs to a particular web application. However, this time your ClassLoader and all its Classes will leak as well as all instances of such classes that are stored in the thread local ArrayList. This time, the web server even cannot certainly determine that a memory leak occurred when it finds that the ClassLoader was not garbage collected since garbage collection can only be recommended to a JVM (via System#gc()) but not enforced.

Renewing the thread pool is not as cheap as you might assume.

A web application cannot just go and throw away all threads in a thread pool whenever an application is undeployed. What if you stored some values in those threads? When a web application recycles a thread, it should (I am not sure if all web servers do this) find all non-leaking thread local variables and reregister them in the replaced Thread. The numbers you stated about efficiency would therefore not longer hold.

At the same time, the web server need to implement some logic that manages the replacement of all thread pool's Threads what does neither work in favor of your proposed time calculation. (You might have to deal with long lasting requests - think of running an FTP server in a servlet container -- such that this thread pool transition logic might be active for quite a long time.)

Furthermore, ThreadLocal is not the only possibility of creating a memory leak in a servlet container.

Setting a shut down hook is another example. (And it is unfortunately a common one. Here, you should manually remove the shut down hook when your application is undeployed. This problem would not be solved by discarding threads.) Shut down hooks are furthermore instances of custom subclasses of Thread that were always loaded by an application's class loader.

In general, any application that keeps a reference to an object that was loaded by a child class loader might create a memory leak. (This is generally possible via Thread#getContextClassLoader().) In the end, it is the developer's resposibility to not cause memory leaks, even in Java applications where many developer's misinterpret the automatic garbage collection as there are no memory leaks. (Think of Jochua Bloch's famous stack implementation example.)

After this general statement, I want to comment on Tomcat's memory leak protection:

Tomcat does not promise you to detect all memory leaks but covers specific types of such leaks as they are listed in their wiki. What Tomcat actually does:

Each Thread in the JVM is examined, and the internal structures of the Thread and ThreadLocal classes are introspected to see if either the ThreadLocal instance or the value bound to it were loaded by the WebAppClassLoader of the application being stopped.

Some versions of Tomcat even try to compensate for the leak:

Tomcat 6.0.24 to 6.0.26 modify internal structures of the JDK (ThreadLocalMap) to remove the reference to the ThreadLocal instance, but this is unsafe (see #48895) so that it became optional and disabled by default from 6.0.27. Starting with Tomcat 7.0.6, the threads of the pool are renewed so that the leak is safely fixed.

However, you have to properly configure Tomcat to do so. The wiki entry on its memory leak protection even warns you how you can break other applications when TimerThreads are involved or how you might leak memory leaks when starting your own Threads or ThreadPoolExecutors or when using common dependencies for several web applications.

All the clean up work offered by Tomcat is a last resort! Its nothing you want to have in your production code.

Summarized: It is not Tomcat that creates a memory leak, it is your code. Some versions of Tomcat try to compensate for such leaks which are detectable if it is configured to do so. However, it is your responsibility to take care of memory leaks and you should see Tomcat's warnings as an invitation to fix your code rather than to reconfigure Tomcat to clean up your mess. If Tomcat detects memory leaks in your application, there might even be more. So take a heap and thread dump out of your application and find out where your code is leaking.