About a month ago, I was preparing my OracleOne talk and
tripped on a slide. I was trying to explain the new delivery process and how long the support of each version will last.
It wasn’t that clear at all.
So, I asked my felow Java Champions
about it. It triggered a discussion about the fact that it is indeed quite misunderstood.
We got together. Or to be honest, Martijn led the way and wrote an article to clarify the situation.
We then made multiple suggestions and corrections. Representatives of the main OpenJDK providers were also
After years missing it, I finally went to JCrete last year for the first time. I went with my
family and we had a great time. This year, I went back again but alone.
For those who never went, it is hard to describe the event. The boring way would be to say that it is a Java unconference.
But it is a really bad way to describe it.
Here is my current explanation.
Close your eyes. Imagine you can gather all the almighty experts in your field and put them in one place. Then imagine
this place has multiple beaches, great (cheap!) food and good wine. That everyone is friendly and wants to help each other without judging.
And that you will talk will all these experts all day, in formal classrooms and on the beach. You will gather a tremendous amount of knowledge.
This is JCrete.
This was my experience last year. This was my experience this year. It almost feels surreal.
I still don’t know how it happened. Is it Crete? Is it Heinz and Kirk’s magical touch?
Is it the Java community that is awesome? Probably all of the above sincerely.
Then, if you have never heard of JCrete let me explain a bit more what it is about.
First, it’s an unconference with a limited number of participants. Those participants are selected for their accomplishments.
It means that you are well-known in the Java community for something. Of course, a lot of Java Champions are
attending. To prevent too much inbreeding, a little amount of entropy coming from random senior lead developers is added.
However, the rule is that anyone attending should be able to propose a subject and talk about it. You can see it as a conference
with only speakers talking to one another.
Then, a typical day looks like that:
Go jogging with some of the attendees (chatting about interesting stuff)
Get breakfast with the attendees (chatting about interesting stuff)
Gather in the main room
Morning briefing and subject proposition for the day
Attend sessions with subjects you care about. Learn
Lunch (chatting about interesting stuff)
Go to some beach (chatting about interesting stuff)
Hack a bit. When stuck, there are high chances that one of the lead committers on the thing you are stuck on is attending too
Get unstuck by the lead committer (this year it was Matt Raible)
Go dinner to some restaurant or at an evening event (chatting about interesting stuff, around a bottle of wine)
Go to bed happy
Last year, this article when out of it. This year, some cool stuff that I will
talk about when it’s ready.
Here is the long awaited 3.5 version. It contains many bug fixes and some improvement. We allowed ourselves to possibly break
the compatibility with older versions for the greater good. So please read these notes thoroughly.
Java 5 is no longer supported. I dearly hope this won’t harm anyone
Java 9 is supported
TestNG support is added. Have a look at EasyMockListener
Class Mocking now works correctly for cross bundle mocking
verify() now checks for unexpected calls in case an AssertionError was swallowed during the test. It is in general
what you want but you can use verifyRecording() to bring back the old behavior
Default matcher for an array argument is now aryEq instead of eq. This should as well make sense for everyone and should
allow you to remove tons of aryEq all over your code. If you don’t like it, you can specify eq explicitly for the array
isNull and notNull with generic class parameter (#93)
Return a meaningful error when null on a primitive (#92)
Create opportunity to disable SingleThread checks (#88)
A long time ago (year 2000), the mock objects were
invented. It is now one of the most important parts of unit testing.
For those you don’t know, the idea of a mock object is to simulate a dependency to easily test a class. Quick example.
I have a class Pricer.
Being a nice human being, I want to test my calculations but I don’t want to use a real PriceFeed. The real implementation
has to an actual MQSeries queue that received prices from Reuters. It’s not something you want to do during your unit tests
(if at all).
So you will instead do a mock object, an object that will behave as you see fit for your test but that isn’t a real PriceFeed. It just
Not so long ago, Uncle Bob did a blog post about mock objects. He classifies
them into five different types or levels. Levels because each type is wiser than the previous one. He calls them “Test Doubles”.
I’ve decided to show you how to code them. Using EasyMock (of course), Mockito and by hand.
Type of mocks
From the most basic to the most advanced type.
A class that you pass into something when you don’t care how it’s used. e.g. As part of a test, when you must pass an
argument, but you know the argument will never be used.
In this EasyMock world, it is called a nice mock (Authorized mock = niceMock(Authorized.class)).
In the Mockito world, it’s just a mock (Authorized mock = mock(Authorized.class)).
In general, if a dummy is used, you will want it to throw an exception to tell you something is
wrong. So, with Mockito, you will in general get a NullPointerException (or not) if you do something like
if(mock.authorize(user, password)) // NPE
With EasyMock, you will generally prefer to use a normal mock (Authorized mock = mock(Authorized.class)) that
will make sure nothing unintended is called.
A class that returns a valid answer but always the same one.
In the EasyMock language, this is any mock with an expectation recorded (expect(mock.authorize(anyString(), anyString()).andStubReturn(true)).
In the Mockito language, this is a mock with behavior set on a method (when(mock.authorize(any(), any()).thenReturn(true).
You use a spy when you want to be sure that the authorize() method is called by the system.
In EasyMock, it means you are not stubbing anymore. You want to record a precise call (expect(mock.authorize(anyString(), anyString()).andReturn(true))
and then verify that the call actually occurred (verify(mock)).
In Mockito, you still stub the call and then verify the call occurred (verify(authorizer).authorize(any(), any())). Note that Mockito has its own
concept of a spy, which is different. A Mockito spy is a shell over an actual class that allows to verify calls to them. It is indeed useful
but it isn’t a mock. So don’t get lost in the semantic.
A true mock is a mock that knows how to verify itself. In fact, EasyMock and Mockito mocks are always true mocks. So their
implementations of a true mock is the same as for the spy.
A Fake has business behavior. You can drive a fake to behave in different ways by giving it different data. They are usually
used for integration testing to simulate other parts of your system.
I rarely use a mocking framework for them. It tends to make the code more complicated than coding by hand. Still, a mocking
framework can be used.
When jumping from one flavor to another, you should make sure you really need to. Because the more complicated your
mocking is, the more coupling you will have with the actual implementation. It makes the test more fragile. But you
still need to make sure everything is working as expected!
Modify to test
If your code isn’t easy to test, modify your code. Do whatever is needed. You will end up with a better design
anyway. A test should not be complicated. If it needs to, something is wrong.
Provide a testing framework
If you build something, you should provide a nice framework to test it. Spring has spring-test. You are
responsible for making what you do testable, mockable, etc.
Use as less mocks as possible
Usually, unit tests should use at worst 3 mocks. It you have more, you probably should split your code in
smaller parts. A lot of mocks makes the code unreadable.
Explain and document your tests
Tests are harder to understand than actual production code. When someone reads test code, he should understand
the purpose. So use a nice test name to explain what you wanted to do. Use javadoc. Use line comments to explain the flow.
I’m refactoring my tests a lot. A lot. Nice methods preventing copy & paste. Testing frameworks. Fixtures. Base test
classes. Everything to make it as pretty as my production code.
I felt on some code yesterday and had to think a bit about it before deciding that it wasn’t working as expected. And then went on to
wonder if I could make it work. I found it interesting so I thought I should tell you about it.
So. What’s going on here?
We have a list
A CompletableFuture is waiting for the list to be filled
We wait on the future until is has finished
If it takes too long, we timeout
It works. If the list is filled, the test will be successful, if the list is never filled, the get will timeout after
1 seconds (throwing a TimeoutException) and the test will fail.
The only problem is that if the test does fail, the future task itself will never finish. It is still stuck in the while
Because nothing can stop it. supplyAsync is submitting a task to the common pool. This task will run in a thread we are
not managing. The task won’t stop until the list isn’t empty anymore. That’s it.
A timeout of the CompletableFuture won’t change anything. It just means that we are not waiting on the get anymore. But
it has no power over the task itself.
What can we do?
Maybe we can interrupt it? Let’s try.
In the original loop, nothing could be interrupted. Now we introduce a sleep. It can be interrupted. However, it won’t.
The timeout on the get doesn’t trigger an interrupt on the thread running the task.
That doesn’t work either?
Can it be cancelled you say?
Here the code is a bit more complicated. We can’t access the future from the lambda directly. Because the lambda starts
before the assignment is made. So we use an AtomicReference, wait until its content isn’t null anymore and then wait
for cancellation… that never arrives.
Yes. The timeout on the get won’t trigger a cancellation. This is on purpose. There is nothing preventing you from waiting
a bit on the get, do something else, and come back to get again.
Enough of that! Tell me what works!
OK. OK. Calm down. I’ll show you but it’s not pretty.
So now we are cancelling the task ourselves. That works. The task correctly finishes. One funny thing to mention is that the
assert won’t fail. In fact, it’s the call to future.get() in assertThat that will throw a CancellationException and
make the test fail.
OK. We now have a pretty ugly and complicated solution that works.
Can we simplify?
You might have noticed that cancel() take a parameter named mayInterruptIfRunning. That sounds promising! We can get
an interruption! Let’s try.
No atomic reference anymore. But doesn’t work. The task doesn’t get interrupted. If I quote the javadoc for cancel():
mayInterruptIfRunning this value has no effect in this implementation because interrupts are not used to control processing.
Basically, that means CompletableFuture are not supposed to be interrupted. They are at a higher level of abstraction.
Here is the nicest solution I know about.
Yes, it’s just a simple flag. But it works nicely.
Still, you might ask me: “Why is it that complicated?”
In fact, I’m not totally sure. I think CompletableFuture are not meant to be used like this. They are supposed to
complete. Or to fail exceptionally. Normal Future are the ones that are supposed to loop like that.
But I’ll need more digging to be more conclusive. Right now, I only wanted to share my How to cancel a CompletableFuture?
I was at JCrete two weeks ago. For those who don’t know, it is an awesome Java unconference where everyone with
their family to talk about Java. It has been created by Heinz Kabutz a.k.a The Java Specialist. I was
really happy to see a bunch of heads I haven’t seen since moving back to Montreal.
One of the sessions I’ve led was about data visibility according to the JMM. I didn’t care about
lock and synchronization. Just data visibility.
If I write to this variable, will this other thread see the new value for sure?
I was lucky enough to gather with a handful of subject matter experts.
My goal was to give really simple examples of what works or not. The final code can be found on
But I will still describe it here. It’s based on the concept of “Will the thread ever get out of the loop?”. All the examples are almost identical.
A thread is started
It loops until a field is flipped to true
The main thread flips the field to true
If the field visibility is correct, when the main thread flips the field, the child thread will see the value and correctly exit the loop. If
it’s not, the thread might go in an infinite loop according to the JMM. I’m saying “might” because depending on the CPU architecture,
JVM version and the way the wind is blowing, it might see the value correctly on your machine. That doesn’t mean it will work ever after.
I recommend that you first guess the result before looking at the answer.
The first example is a normal field (no final or volatile) without any kind of synchronization.
Does it work: No
Data visibility to another thread is not guaranteed by a normal field. The child thread might never see stopNormalField ever value changed.
Now, let’s turn it to a volatile field.
Does it work: Yes
Volatile ensure data visibility. When a volatile field is changed, all other threads are seeing the value right away. Volatile tells the JVM
that the value shouldn’t be kept local to the thread.
So far so good.
What about a volatile array?
Does it work: No
Data visibility of array elements is not guaranteed. Elements of a volatile array are not volatile. Yes, it is non-intuitive.
Let’s now start with higher abstraction from java.util.concurrent (JUC). First, an atomic field.
Does it work: Yes
Of course it works. Data visibility of the content of an atomic is guaranteed. Note, however, that it doesn’t apply to the field referencing
the atomic field (stopAtomicField). It still works on our case because we assign (stopAtomicField = new AtomicBoolean()) before starting
the thread. The JVM makes sure a starting thread will see everything that happened-before.
Now. What about synchronization?
Does it work: Yes
Two threads synchronizing on the same mutex are seeing the same thing inside the synchronized section. All words in bold are really
For instance, not synchronizing on the same mutex means all bets are off.
Does it work: No
By the way, watch out for lambda and inner classes. They don’t have the same this. For a lambda, this is the class where the lambda is
defined. For an inner class, it is the inner class.
Does it work: Yes
Synchronization works. Fair enough. But can I use a lock? I was told locks are better than synchronization.
Does it work: Yes
Locking provides the same data visibility as synchronizing.
In fact, all JUCs are providing the necessary memory barriers to get the correct visibility. Below, countDown and await will make sure
the child thread sees the world correctly.
Does it work: Yes
As soon as you use JUC abstractions, things tend to magically work in fact. Which is a good thing even though everything magically working
is always a bit frightening.
OK. We will finish this post with the Don’t do this at home example. It means it is trickier to get right and you won’t need it for business
Does it work: Yes
Here we cause the synchronization of the normal field by using a volatile field. Writing to stopVolatileField causes a happens-before
relationship. Happens-before is JMM jargon. It has a script definition. That means what it means in English. That you are sure something
happened before. In our case, it means the value written to stopNormalField will be seen by the child thread after it reads stopVolatileField.
Thanks again to all the contributors to this session (you will find some of them on the
Objenesis went out last week partly because Google App Engine is now supporting Java 8 in
beta. And they are no security manager anymore preventing Objenesis to work.
The problem was that Objenesis was checking if it was running on GAE to pick an instantiating strategy. So it was still running in degraded
more on GAE Java 8. It is now fixed.
However, Spring uses an embedded version of Objenesis. So they need to upgrade. Luckily, they are super fast to do so. You can follow the
issue but it will be in Spring 4.3.10 and Spring Boot 1.5.5.
Meanwhile, you might be eager to test your shiny new app on this new Google App Engine platform and would be really happy to make it work.
I have a solution for you. It isn’t pretty but it works.
In Java, in case you don’t know, you just need to put your class file in front of another one in the classpath to “shadow” it. It means your
class will be used instead of the original implementation. In a war, everything in WEB-INF/classes goes before WEB-INF/lib. So you can
easily shadow a class.
That’s what I did. My SpringObjenesis.java
version shadows Spring one. It delegates to the real Objenesis implementation instead of the Spring embedded one. You only need to add the latest
Objenesis (2.6) as a dependency to your project.
An all new shiny Objenesis is out. The framework everybody uses without even knowing it.
It had fun playing with the brand new Google App Engine platform supporting Java 8. They dropped the terrible security manager they used to
have to know Objenesis is working perfectly on it. This was brought to my attention by GAE developers. They are now testing mainstream frameworks
on their platform to make sure it works. This should make GAE adoption easier.
Java 9 was pretty much already working but I got rid of all the “Illegal accesses”. Not that obvious since in Objenesis, this is how we roll.
Finally, a bit of cleanup in the serialization specification was done. So now a serializing instantiator should instantiate a class by calling
the no-arg constructor of the first non-serializable class in the hierarchy. And do nothing else. See the documentation