Most server-side applications and many desktop applications contains data that is tied to a particular task that’s being executed. A common solution is to keep that kind of data in thread-local storage; to keep the data in variables bound to the executing thread. Convenient, but a practice based on a faulty assumption.
Bob Martin wrote an article about the problem of assuming that a unit-of-work has a one-to-one relationship with a thread.
ThreadLocal variables are a wonderfully convenient way to associate data with a given thread. Indeed, frameworks like Hibernate take advantage of this to hold session information. However, the practice depends upon assumption that a thread is equivalent to a unit-of-work. This is a faulty assumption.
ThreadLocal is the Java terminology, but the construct is common in multi-threaded environments. Bob remembers:
Thirteen years ago, while working on my first book, Jim Coplien and I were having a debate on the nature of threads and objects. He made a clarifying statement that has stuck with me since. He said: “An object is an abstraction of function. A thread is an abstraction of schedule.”
Mapping the data of a unit-of-work to one thread is a standard pattern today; a pattern that is found in many popular frameworks, and even though that approach works most of the time, there are situations where the faulty abstraction leaks.
It is not uncommon for a task to have different priorities for different parts of the task, and there are no rules that a task must be single-threaded. Bob exemplifies by saying that a unit-of-work may very well need responsive communication with an external service while performing a relatively long-time computation based upon the incoming data; a problem commonly solved by using two threads. He asks:
Where are the unit-of-work related variables? They can’t be kept in a ThreadLocal since each part of the task runs in a separate thread. They can’t be kept in static variables since there is more than one thread. The answer is that they have to be passed around between the threads as function arguments on the stack, and recorded in the data structured placed on the queue.
TapsaKoo encountered a similar situation a while ago. When trying to work in a domain-driven way in WinForms, he described his problems finding a place to keep session-specific data.
If the application has only one form open at a time, I could save the session-object into the CallContext. What if the application has multiple forms open at a time and each of them wants to have a separate instance of my session-class? CallContext is out of the question. So are all thread-specific alternatives. What is left? Nothing? I’m not the first person pondering this issue. A solution probably exists but I can’t find it. Do I really have to inject the session-object into every object instance that might need it? Or should I refactor a lot behavior from domain-classes into services and inject the session-object into them. I don’t like this approach because I want my classes to be more than data containers.
When reading Uncle Bobs post, TapsaKoo agreed that there are no easy solutions to the problem:
It doesn’t matter if you have 1 or 10 threads. The problem is always the same. UnitOfWork or SessionState should have a place that does not depend on threads. It’s a dangerous assumption that UnitOfWork is directly related to one single thread. That assumption seriously limits your other architectural choices.
Bob concludes that something seems to be missing:
So, though convenient, ThreadLocal variables confuse the issue of separating function from schedule. They tempt us to couple function and schedule together. This is unfortunate since the correspondence of function and schedule is weak and accidental.
What we’d really like is to be able to create UnitOfWorkLocal variables.
Community comments
Right theorical assumption, too much pain in practice
by Julien Delfosse,
Re: Right theorical assumption, too much pain in practice
by Alex Miller,
Re: Right theorical assumption, too much pain in practice
by Steven Devijver,
Re: Right theorical assumption, too much pain in practice
by Cameron Purdy,
Sharing context between threads.
by Peter Lawrey,
Java EE 5 solution: TransactionSynchronizationRegistry
by Patrick Linskey,
fluid variables
by Steven Shaw,
Right theorical assumption, too much pain in practice
by Julien Delfosse,
Your message is awaiting moderation. Thank you for participating in the discussion.
While I agree with this point theorically, I don't think multi threads unit of works can be applied practically (in most applications).
If you take transaction management, for example, it is clear that the transaction life must be bound to your unit of work. Problem is, in application servers, JTA transactions (and associated resources as JDBC connections obtained through JNDI lookup) are bound to a particular thread. Same goes for caller information and other useful, infrastructure provided data.
If you were to implement a multi thread unit of work, you'd have to abandon those low level features. It simply seems too much pain for me. I prefer using a logical unit of work (in the use case sense) that spans several technical units of work (in the ORM/tx sense of the term) that can be distributed amongst many thread. Mechanisms as workflows, JMS asynchronism, sync points, ... exist to support this way of programming, and you still can use all the features provided by your application server.
Re: Right theorical assumption, too much pain in practice
by Alex Miller,
Your message is awaiting moderation. Thank you for participating in the discussion.
JTA is particularly insidious in this regard (binding work to the current thread). Actually, that seems common in most transactional systems. I even think that's ok as long as you can strip the info off the thread at some point and move it to a different thread.
When I was at MetaMatrix, we used the Atomikos transaction manager and we were able to move transactions between threads (and include sub-transactions on different VMs). But it certainly wasn't pretty and didn't seem to be the way people normally used it.
Re: Right theorical assumption, too much pain in practice
by Steven Devijver,
Your message is awaiting moderation. Thank you for participating in the discussion.
Indeed, sharing connections and transactions between threads is unusual and unpractical. Also, the transaction schemantics of JTA and Spring - notably, the propagation levels - have clear behavior when it comes to delimiting the association of "unit-of-work" objects with the threads.
Sharing context between threads.
by Peter Lawrey,
Your message is awaiting moderation. Thank you for participating in the discussion.
There are two approaches I have taken to share context between threads. Both required a pool of threads to share a ThreadGroup.
1) Create a ThreadGroupLocal. This holds a value/values for each ThreadGroup shared between threads in that group.
2) Create a custom ThreadGroup which holds a thread safe Map of values.
Where you have something which is single threaded I associate a single thread pool and a cached (variable sized) pool with the thread group. Single threaded processing requires adding a task the first pool and any which isn't single threaded is added to the second pool.
This is the cut down version. You add sub-tasks to either the single pool or the multi threaded pool. The single threaded pool can kick off any number of tasks concurrently and then collect the results with the Future.get() method. Multi-threaded tasks can add sub-task which must run in the same threaded to the single pool.
This structure is used to run a large number of unrelated tasks at once. (Although the single threaded portions have to wait for each other)
Java EE 5 solution: TransactionSynchronizationRegistry
by Patrick Linskey,
Your message is awaiting moderation. Thank you for participating in the discussion.
In Java EE 5, this disparity has been addressed with the new TransactionSynchronizationRegistry interface. In particular, take a look at the putResource(Object,Object) and getResource(Object) methods.
-Patrick
--
Patrick Linskey
bea.com
Re: Right theorical assumption, too much pain in practice
by Cameron Purdy,
Your message is awaiting moderation. Thank you for participating in the discussion.
One can suspend() the transaction on one thread and resume() it on another.
Peace,
Cameron Purdy
Oracle Coherence: The Java Data Grid
fluid variables
by Steven Shaw,
Your message is awaiting moderation. Thank you for participating in the discussion.
Probably what you are missing is the idea of fluid variables (called special variables in Common Lisp).