repaint method and design question

[conrad]

I have two classes:
one for the GUI the other for various scanning methods
on a file system.

I'm experiencing an issue where, while scanning,
the main pane, if minimized and brought back up,
is blank. So I'm thinking I need to constantly
repaint while scanning?

In terms of design: should I be passing around
GUI objects(instance variables) from my GUI class
to my file system scanning class?

Or should I handle it differently?

 

[Stefan Ram]

In terms of design: should I be passing around
GUI objects(instance variables) from my GUI class
to my file system scanning class?

If you use standard GUI components they should handle
it all for you.

You supply a model, and according to the observer pattern
your UI delegate will be notified of changes to the model
and will update the view.

When you design custom components you implement (overwrite)
»paintComponent«. But also in this case, it will be called
by Java, whenever necessary (for example, when an obscured part
of your component becomes visible). Maybe you want to study

http://google.to/search?q=java+custom-painting+tutorial

or

http://google.to/search?q=java+observer+ui-delegate+model

 

[Roedy Green]

I'm experiencing an issue where, while scanning,
the main pane, if minimized and brought back up,
is blank. So I'm thinking I need to constantly
repaint while scanning?

I don't know what you meant by "scanning" but if it is as all CPU
intensive or long lasting, you should do it on a separate thread, not
the EDT. See http://mindprod.com/jgloss/swingthreads.html
for details.

 

[John B. Matthews]

I don't know what you meant by "scanning" but if it is as all CPU
intensive or long lasting, you should do it on a separate thread, not
the EDT. See http://mindprod.com/jgloss/swingthreads.html
for details.

Roedy's right. Here's the outline of a simple model that may guide your
re-factoring. It uses a javax.swing.Timer to notify the GUI of interim
results (not compiled):

class Scanner extends Observable implements ActionListener, Runnable {

private int interval = 1000 / 4; // 4 Hz
private Timer timer;
private Results results;

public Scanner() {
super();
timer = new Timer(interval, this);
}

public void run() {
timer.start()
// lengthy code that generates results
timer.stop();
setChanged();
notifyObservers(results);
}

public void actionPerformed(ActionEvent e) {
// interim progress
setChanged();
notifyObservers(results);
}
}

public class ScanView extends JPanel implements Observer {
...
public void update(Observable o, Object arg) {
Results results = (Results) arg;
// fill in GUI components with results
this.repaint();
}
}
....
Scanner scanner = new Scanner();
new Thread((Runnable) scanner).start();

 

[John B. Matthews]

Please correct me if I'm wrong, but I see nothing in the code you posted
that ensures that the observer notification is done on the EDT. The
documentation specifically says that there are basically no guarantees as
to how the observer's update() method is called, never mind that it's
always called on the EDT (in fact, the docs also say that subclasses of
the Observable class may "deliver notifications on separate threads").

I appreciate your critical analysis! This is an outline of an approach
I've used (successfully) several times for simulations in which a view
displays the interim results calculated by an asynchronously running
model. I'm sure I've overlooked a critical caveat or two

My proposed Scanner implements Runnable, so it's a separate thread. I
recall reading the "separate threads" warning, but I understood the
warning to refer to notification order not being preserved across
threads. Timer(int, ActionListener) ensures that notification is done on
the EDT, but I'm not sure that is required. More importantly, the
Observer is running on the EDT, as it extends an AWT Component.

Since the notification happens in response to the timer event, and since
that event happens on the EDT, as long as the base Observable class always
executes notifications on the same thread on which notifyObservers() is
called, you're safe. But the docs don't restrict the Observable class in
this way. I personally wouldn't depend on the current implementation to
never change, unless there was some specific Java documentation that
promises that the Observable class will never change in this respect.

The javax.swing.Timer I proposed runs on a separate thread; only it's
action handler executes on the EDT. One _could_ put drawing code in the
handler, but sending a notification instead keeps the model and view
separate. I thinka java.util.Timer would do as well.

As long as I'm replying, I'll also point out that your sample code
conveniently skips demonstrating any of the synchronization code that
would be required in order to successfully respond to the timer while the
other thread is busily working on the time-consuming operation. This code
could in fact be included, yes...but it would not be an insignificant
change. In fact, it's arguably "the hard part" of the whole sample.

Yes, a caveat is warranted: The data structure that the model is
updating must be consistent for a given instant of simulated time, or
the view must receive a consistent copy/snapshot. In the OP's case, it
wouldn't do to have the scanner updating the same TableModel used by the
view!

Personally, I'd skip the timer altogether, and simply update the UI at
convenient intervals throughout processing based simply on progress (say,
every N files, where N is selected based on how long each file would take
to process). Then just use invokeAndWait() or invokeLater() directly from
the processing thread to call code that would update the GUI. This allows
the processing thread to package up the interesting data synchronously
without any actual explicit synchronization overhead, and send it off to
the GUI via the EDT, taking advantage of the EDT's own built-in
synchronization mechanisms.

This approach makes better sense than mine for the OP's file scanning
operation. Of course, the scanner thread still has to decide when it has
useful, self-consistent, interim results. My approach may be more suited
to watching a numerical simulation evolve, where there's no _a_priori_
way to know a preferred update point, other than "periodically".

Just my two cents.

Worth much more! What, with compound interest and the price of copper,
etc.

 

[John B. Matthews]

[...]
My proposed Scanner implements Runnable, so it's a separate thread. I
recall reading the "separate threads" warning, but I understood the
warning to refer to notification order not being preserved across
threads.

Well, IMHO the docs are ambiguous. To me, the important part is that they
go out of their way to say that one should not rely on the notification
being done on a particular thread. The warning is in the context of
dealing with sub-classes of Observable, and so there _could_ be an
implication that as long as the sub-class doesn't change the default
implementation, the notification will be done on the same thread on which
notifyObservers() is called.

Yes, the subclass that implements Observable should behave correctly,
but I read the paragraph about notification order as a warning to
Observers that an Observable may change the order and thread.

But that's an assumption, and as we all know, assumptions are not nearly
as valuable as things printed in black and white.
Agreed!


Now this is exactly what I'm talking about. The Swing Timer class
does ensure that the Action event itself happens on the EDT.
However, there's nothing about Timer that guarantees that when you
call notifyObservers(), the notification itself happens on the same
thread on which that method was called.

For example, one implementation could create a copy of the observer
list and then ship that off to yet another thread to process the
list asynchronously.

I don't even know for sure that the _current_ implementation keeps
the notification on the same thread as which notifyObservers() was
called (though I do agree it's likely, I haven't bothered to check
it myself, and the fact is it could be implementation-dependent),
but it does seem to me that the docs don't promise that even if it's
true now, it always will be true.

I don't think it matters as long as the Observer respects the
order/thread warning above. In particular, my proposed Observer's
update(Observable, Object) method ends with repaint(). The hidden
assumption is that repaint() must be invoked from a JComponent, which
uses the RepaintManager to run from the EDT:

<http://java.sun.com/products/jfc/tsc/articles/painting/index.html#paint_
process>

It depends on what you mean by "is running on the EDT". It's true
that your Observer-implementing class is a Swing (not just AWT)
component. But that doesn't guarantee that code in that class is
run on the EDT. In fact, that's the whole issue: it _is_ a Swing
component and so when you execute Swing-dependent code in the class,
you have to go out of your way to ensure that happens on the EDT.

Good point! I should have said JComponent.

If that's what you meant, then yes...I agree. But if your statement was
intended to mean that because you've sub-classed JFrame, you're assured
that calls to your sub-class are always on the EDT, that would be wrong.

Yes, for JFrame. I proposed JPanel, which is a JComponent, but your
point is well taken.

See above. I agree that the Timer Action event is raised on the EDT,
but that doesn't necessarily guarantee that the Observer's update()
method is called on the EDT. There's a missing link between your
actionPerformed() method being called, and the Observer's update()
method being called that involves ambiguously defined behavior on
the part of the Java run-time.

As long as the Observer's update() repaints a lightweight container, I
don't think it matters if the notification doesn't come from the EDT.

Well, more specifically, since you know that the Timer Action event
will be raised on the EDT, and you also know that the scanning work
is done on a thread other than the EDT, you are assured that you
need some kind of synchronization mechanism. Even if this is as
simple as having the scanning thread making immutable copies of its
progress at regular intervals and assigned references to the data to
a volatile variable, _some_ sort of effort needs to be made to keep
things synchronized.

Agreed, although there are stochastic models where no synchronization at
all gives satisfactory results.

Agreed...to the extent that it might be difficult to determine what a
"preferred update point" is, a timer-based approach might work
better. I'm not sure that that applies to a numerical simulation,
but that's only because of the broadness of that genre of
algorithms. Some numerical simulations lend themselves just fine to
predictable update points, others might not.

Note, however, that the more likely that an algorithm is specifically
in need of a timer-based approach, the more complex the
synchronization mechanism will have to be (i.e. just copying
immutable results to a volatile variable won't do, since if you
could do that, it'd imply that you do in fact have predictable
update points ).

Don't similar synchronization problems attend to using the invoke*
utilities?

 

[John B. Matthews]

I see nothing in the Javadocs for JComponent#repaint() that indicates
that the method is thread safe.

Yes, the SDN article is referenced in Component, and perhaps elsewhere.

But you showed us

which explains that repaint() is thread safe.

Only for JComponents: "JComponent.repaint() registers an asynchronous
repaint request to the component's RepaintManager, which uses
invokeLater() to queue a Runnable to later process the request on the
event dispatching thread."

This worries me, as I had thought the Javadocs to be normative.

A linked article is less dispositive, I suppose, but not entirely
unreliable

How do variable values communicate across threads in that case?

Or by "stochastic" do you mean that it doesn't matter, because the
results are intended to be random?

Yes, the results evolve in apparent random order: many kinds of
fractals; diffusion limited aggregation, which has a fractal character;
any iteration over a set of lattice points in which you don't care that
some points are from iteration n-1 and some from iteration n.

I really do wonder how one handles the memory-model concerns in that
"stochastic models" scenario. Would you be so gracious as to explain?

My point was only that sometimes it doesn't matter if the model is
changing under you. When it matters, you have to synchronize access, of
course. Sorry if I seemed to be arguing to the contrary.

Oh, and I hope I am always gracious!

 

[John B. Matthews]

Define "correctly". The point of the statements in the documentation
is that the Observable class is not under any restriction regarding
how it deals with notification. That's my point. There's a whole
range of what can be considered "correct", including any variety of
implementations that don't allow you assume that the Observer's
update() method is called on the EDT.

I think we agree on this for subclasses.

Yes, it may. And IMHO this can include the base Observable class.

I must demur: "The default implementation provided in the Observable
class will notify Observers in the order in which they registered
interest..." tells me that order is preserved if my subclass doesn't
change it. If I do change it, Observers must accept that.

The phrases "...this notification mechanism is has nothing to do with
threads..." and "Each observer has its update method called..." tell me
that my subclass will execute the update() method of a registered
Observer even if they're in different threads, and Observers have to
accept what that implies.

[...]

I don't even know for sure that the _current_ implementation keeps
the notification on the same thread as which notifyObservers()
was called (though I do agree it's likely, I haven't bothered to
check it myself, and the fact is it could be
implementation-dependent), but it does seem to me that the docs
don't promise that even if it's true now, it always will be true.

I don't think it matters as long as the Observer respects the
order/thread warning above. In particular, my proposed Observer's
update(Observable, Object) method ends with repaint(). The hidden
assumption is that repaint() must be invoked from a JComponent,
which uses the RepaintManager to run from the EDT:

<http://java.sun.com/products/jfc/tsc/articles/painting/index.html#p
aint_ process>

Unfortunately, the documentation is not clear as to whether repaint()
can be called from a thread other than the EDT. Just because it
queues a Runnable to the EDT, that doesn't necessarily mean that it
doesn't itself need to be also called from the EDT.

Well, the article claims "JComponent.repaint() registers an asynchronous
repaint request to the component's RepaintManager, which uses
invokeLater() to queue a Runnable," which is precisely the approach you
proposed earlier.

More significantly, even if repaint() is legal to be called from any
thread, not just the EDT, in presenting a code example we should be
VERY explicit that we are relying on that specific behavior. Other
interactions with a component (AWT or Swing) could easily still be
required to be done on the EDT, even if the repaint() method need not
be (and to be clear: I remain unconvinced that the documentation
promises that to be true).

A naïve reader of the code example might incorrectly extrapolate from
the lack of special handling for repaint() that no special handling
is needed generally.

I agree. The scheme relies on repainting JComponents, and then only if
the SDN article is dispositive. It's a hidden assumption that I should
document.

That's not what I am talking about. Whether you're talking Swing or
EDT, you need to make sure you are making calls to the component on
the EDT.

I agree. The Observer's update() method should defer drawing, but it can
get interim results and update component data models, e.g. a
BufferedImage or a TableModel. It then calls repaint() to schedule the
screen update on the EDT.

Again, for what I'm talking about, whether you are using JFrame or JPanel
doesn't matter. You still have the same "must be on EDT" requirement.


Well, I haven't seen any documentation that promises that. It could be
true, but I have no reason at the moment to believe that. But again, the
repaint() method is a very specific method; there is no general promise
that when interacting with an AWT or Swing component from the Observer's
update() method, it doesn't matter if the notification doesn't come from
the EDT. In many other scenarios, it definitely would matter.

Even if we grant that repaint() has special rules (and again, I'm not
convinced we should), the more general rule still needs to be followed for
other kinds of calls to the component. Presumably, if you're updating the
GUI, you will in fact find such calls, and they will need to be executed
on the EDT.

I welcome your skepticism, but the documentation for Component.repaint()
makes repeated reference to the SDN article, and the article mentions
calling paint() from the EDT as step one for both AWT and Swing
components.

[...]

Note, however, that the more likely that an algorithm is
specifically in need of a timer-based approach, the more complex
the synchronization mechanism will have to be (i.e. just copying
immutable results to a volatile variable won't do, since if you
could do that, it'd imply that you do in fact have predictable
update points ).

Don't similar synchronization problems attend to using the invoke*
utilities?

I don't really know what you mean here. By design, invokeLater() and
invokeAndWait() address synchronization issues related specifically
to those methods.

I see the run() method getting called from the EDT, but with no special
synchronization, other than being added to the end of the queue.

The code being invoked does need to deal with synchronization,
possibly. But when I write code like this, typically the only data
being used in the invoked code is the bare minimum data required to
update the GUI. In particular, the extraction of the relevant
information from my model has already been done, on the same thread
that's processing the model, and copied to some specific place used
by the invoked code (sometimes this is as simple as a "final" local
variable used in an anonymous class).

In this paradigm, because only one thread is ever actually using the
model itself (the processing thread for the model), no
synchronization issues exist with respect to the model, and the
invoke*() methods automatically deal with synchronization issues
that may exist with respect to the component itself.

Is the model also running on the EDT?

As I said, I'm not sure I understand your question. But my first
impression of what you're asking leads me to answer "no, similar
problems do not exist".

I appreciate your patience helping me vet the approach. It's immensely
helpful in understanding it's limitations.

 

[John B. Matthews]

I finally figured that out, after spending some quality time with the
Javadocs.

It doesn't help that some of that outrigger documentation has been
superseded,
but the docs don't reflect the new knowledge.


But without synchronization, the display thread might only see
iteration zero forever.

In practice I haven't seen this. The model, an Observable, executes the
view's update(Observable Object) method, and the view either asks the
model for data or receives it as a parameter. For some models, the data
may be no more than a reference to an array of primitive types, which
the model continues to update even as the view accesses it to compose,
for example, a BufferedImage.

But don't model changes by very necessity have to be communicated
somewhere, eventually? Absent synchronization, either via the
'synchronized' keyword or via any of the several other mechanisms
that Java provides, there is no guarantee in the memory model that
those changes would ever escape to another thread that needs them.

I can see this for more complex, non-thread-safe data structures. I
imagine a Vector, which is synchronized, or one of the
Collections.synchronized[Collection|List|Map|etc.] would do as well.

And if another thread doesn't need them, then it's not a
multithreaded scenario.

I'm not sure failing to communicate model state actually contributes
to the stochasticism.

Yes, the model is intended to be stochastic, not factitious

 

[John B. Matthews]

[...]

but I read the paragraph about notification order as a warning
to Observers that an Observable may change the order and thread.

Yes, it may. And IMHO this can include the base Observable class.

I must demur: "The default implementation provided in the
Observable class will notify Observers in the order in which they
registered interest..." tells me that order is preserved if my
subclass doesn't change it. If I do change it, Observers must
accept that.

Nothing I wrote has anything to do with the _order_ in which
observers are notified.

Ah, I misunderstood. I thought you were saying that the base Observable
class might change the order, when the documentation says the opposite.
I understand that subclasses may do so.

What you wrote supports my point. Yet, you seem to be including it
with the intent to refute my point.

I can only conclude that either you misunderstand my point, or you
misunderstand the implications of "will execute the update() method
of a registered Observer even if they're in different threads".

Yes, I misunderstood your point.

On a slight tangent: it's not really clear from the above quote that
you understand an important point about threading. Specifically, an
_object_ is not on a specific thread. The _execution_ of some code
is on a specific thread. A given class can have code that is
executed on arbitrarily many threads. There is no connection
between a given object and a given thread, except whatever
connection is part of the actual explicit implementation.

If you're confused about this, that could explain some of the
disconnect here.

Of course, I don't know what I don't know, but I think I understand
Lew helpfully pointed out my ambiguous antecedent, so let me restate:

Starting a new thread for a Runnable [e.g. new Thread((Runnable)
scanner).start()] causes the Runnable's run() method to begin executing
on a new thread. Invoking notifyObservers() causes the registered
Observer's update() method to execute on that same thread, even though
other methods in the Observable may be executing on another thread. In
particular, an Observable that extends JPanel (or some other top-level
container) may have (at most) one method executing on the EDT.

I think this is the genesis of your welcomed criticism of my approach.

[...]

Unfortunately, the documentation is not clear as to whether
repaint() can be called from a thread other than the EDT. Just
because it queues a Runnable to the EDT, that doesn't necessarily
mean that it doesn't itself need to be also called from the EDT.

Well, the article claims "JComponent.repaint() registers an
asynchronous repaint request to the component's RepaintManager,
which uses invokeLater() to queue a Runnable," which is precisely
the approach you proposed earlier.

No, it's not.

To be clear: Lew has helped by pointing out _some_ documentation
(such as it is) that explains that the repaint() method can be
successfully called from a thread other than the EDT thread. So
we've got that point out of the way.

But, the approach I proposed has nothing to do with whether the
_painting_ of the control is done on the EDT or not (and in fact, I
take as granted that it is, whatever else the implementation may be,
since that's a requirement of AWT and Swing).

So, the fact that when you call repaint(), that ultimately queues a
Runnable to the event queue that will eventually do the actual
repainting of the controls has absolutely nothing to do with how
data is synchronized between the processing thread and the GUI (EDT)
thread.

I agree.

Well, more importantly, you gloss over the real work: actually taking the
model data and using it to update the GUI. The repaint() method itself
doesn't do anything except actually redraw the GUI. For that to be a
useful operation, you have to first _change_ the GUI so that when it
redraws, the state of the GUI is actually different, reflecting the
current state of processing.

So yes, while the lack of a guarantee that the update() method is called
on the EDT doesn't hurt us with respect to calling repaint() (sort
of...see * below), it certainly would if you actually meant to _do_
anything with the GUI that presents the current state of processing.

(*) Note my comments in my reply to Lew regarding the thread safety of
calling repaint(). The docs do explain that you can call repaint() from a
thread other than the EDT. But they also say that any method that's
thread-safe will specifically say so, and repaint() doesn't say so. This
means that while it's legal to call repaint() from a non-EDT thread, you
are still required to address synchronization yourself, since repaint()
isn't a thread-safe method.

In other words, you're really only safe calling repaint() in one respect.
There's another, arguably more important respect, in which calling it from
another thread without synchronization is _not_ safe.

Permit me to defer comment to that part of the thread.

[...]

That's not what I am talking about. Whether you're talking Swing or
EDT, you need to make sure you are making calls to the component on
the EDT.

I agree. The Observer's update() method should defer drawing,

It does, by virtue of the repaint() method deferring actual drawing.

but it can
get interim results and update component data models, e.g. a
BufferedImage or a TableModel. It then calls repaint() to schedule the
screen update on the EDT.

But those "component data models" need to be synchronized. If you are not
executing your code on the EDT, then they may be being used by code
running on the EDT at the same time you are modifying them. And that's on
top of the fact that Swing does not automatically handle cross-thread
updates to data models for GUI components.

For example (from
http://java.sun.com/javase/6/docs/api/javax/swing/package-summary.html#threadi
ng):

...if a TableModel is attached to a JTable, the TableModel
should only be modified on the event dispatching thread.
If you modify the model on a separate thread you run the
risk of exceptions and possible display corruption

Conversely, if you _are_ executing your code on the EDT, then the data
model from which you're getting your actual information (that is, the
processing thread's model data structure) is what needs to be synchronized.

The bottom line is that you have a processing thread that's generating
data, but you can only display data on the EDT, which is by definition
going to be a different thread from the processing thread. You _must_
address synchronization of those threads any time you want data to flow
from one thread to the other. And the only way to present the results of
the processing (interim results or otherwise) to the user is to have data
flow from one thread to the other.

On top of all this, the Observable class doesn't give us any assurance
that Observer implementers are going to be called on the same thread on
which Observable.notifyObservers() is called. So your Observer
implementer cannot make the assumption that it's update() method is called
on the EDT, even if you call notifyObservers() on the EDT.

And finally (to conclude this particular logical progression), since your
Observer implementation cannot make the assumption that its update()
method is called on the EDT, it must itself explicitly ensure that any
interaction it has with Swing is done on the EDT, except possibly for
things that are explicitly documented as not needing that (in this case,
that would be the repaint() method, but not any of the other things you
might do in order to update the GUI).

These point are well taken, and thank you for pointing out the
TableModel caveat. I've been modifying a BufferedImage's
WriteableRaster, so any problems may have been more in-apparent than
impossible.

[...]

Even if we grant that repaint() has special rules (and again, I'm not
convinced we should), the more general rule still needs to be followed
for
other kinds of calls to the component. Presumably, if you're updating
the
GUI, you will in fact find such calls, and they will need to be executed
on the EDT.

I welcome your skepticism, but the documentation for Component.repaint()
makes repeated reference to the SDN article, and the article mentions
calling paint() from the EDT as step one for both AWT and Swing
components.

As I noted above, I concede the point that repaint() itself may be called
from a thread other than the EDT. However, please note:

-- that doesn't help with any of the other methods that you'll need to
call in order to update the GUI, and
-- since repaint() isn't documented as being thread-safe, you still
have a synchronization issue even with that method

[...]
Note, however, that the more likely that an algorithm is
specifically in need of a timer-based approach, the more complex
the synchronization mechanism will have to be (i.e. just copying
immutable results to a volatile variable won't do, since if you
could do that, it'd imply that you do in fact have predictable
update points ).

Don't similar synchronization problems attend to using the invoke*
utilities?

I don't really know what you mean here. By design, invokeLater() and
invokeAndWait() address synchronization issues related specifically
to those methods.

I see the run() method getting called from the EDT, but with no special
synchronization, other than being added to the end of the queue.

You see which "run() method" getting called from the EDT? How _isn't_
making sure some code is executed on the same thread as some other code
not obviously synchronizing those two pieces of code?

Is the model also running on the EDT?

No, not the model I'm referring to. As I said, "only one thread is ever
actually using the model itself (the processing thread for the model)".
Since by definition, the processing thread is not the EDT, obviously the
model is not "running on the EDT" (inasmuch as the model could be
"running" on any thread...I wouldn't describe it that way, but rather as
simply the _processing thread_ is what's "running", and it modifies the
model as it runs; that's a bit of a semantic nit-pick though).

Yes, I confounded the model as a class with the model's run() method.

On reflection, I see the clear contract the invokeLater() and
invokeAndWait() methods offer with respect to the EDT, and my approach
may have too many limitations.

 

[John B. Matthews]

The determination of the thread in which an action occurs depends on the
caller, not the called. It really doesn't make sense to say that an object
is
"in" a thread - only actions occur in a thread.

Between that and the dangling antecedent for "they", I'm not certain I
understand the sentence, but I think you are saying that the thread in which
an Observer's update() will occur is the thread in which the observed object
makes the call. That would be the case.

Yes. Thank you for pointing out the ambiguity. I tried to make myself
clearer in my response to Pete.

 

[John B. Matthews]

[...]

On a slight tangent: it's not really clear from the above quote
that you understand an important point about threading.
Specifically, an _object_ is not on a specific thread. The
_execution_ of some code is on a specific thread. A given class
can have code that is executed on arbitrarily many threads.
There is no connection between a given object and a given thread,
except whatever connection is part of the actual explicit
implementation.

If you're confused about this, that could explain some of the
disconnect here.

Of course, I don't know what I don't know, but I think I
understand Lew helpfully pointed out my ambiguous antecedent, so
let me restate:

Starting a new thread for a Runnable [e.g. new Thread((Runnable)
scanner).start()] causes the Runnable's run() method to begin
executing on a new thread. Invoking notifyObservers() causes the
registered Observer's update() method to execute on that same
thread, even though other methods in the Observable may be
executing on another thread.

Well, that depends on how you call notifyObservers(). The code you
posted does _not_ actually do this though. It calls
notifyObservers() from the actionPerformed() method, which is called
by the Timer class on the EDT, not the thread on which the
Observable's run() method was called.

Worse! A serious flaw in the outline I proposed is that update() runs on
the EDT when sent by the Timer's actionPerformed() method, but the final
notification executes update() on the thread on which the
Observable's run() method was called.

At _best_, this will cause the registered Observer's update() method
to be executed on the EDT, but in any case certainly not on the
thread on which the Observable's run() method was called. At worst,
the notifyObservers() method itself might actually use yet another
thread for the purpose of calling the Observer's update() method
(I'd be surprised in the Sun implementation of the base Observable
class did that, but you never really know and the docs go to some
length to warn that you cannot count on that happening).

I am somewhat reassured that "... the action event handlers for Timers
execute on ... the event-dispatching thread," and that the Observable's
"... notification mechanism has nothing to do with threads..."

I suppose that depends on your definition of "executing". Personally, I'd
say that any method still in the call stack is "executing", and in that
sense, any class at all can have any number of methods executing on the
EDT.

Well, "executing" in the sense that drawing methods should execute only
on a single thread, the EDT.

If by that you mean that, since your restatement of your previous
statement is flawed, then so too is your proposed approach to the original
question, then yes...I suppose this could be that "genesis".

Permit me to amend my proposed outline, clarifying the need to
synchronize shared data and to execute update() on the EDT:

/**
* A model capable of running on a separate thread.
* A javax.swing.Timer is used for interim updates so that the
* actionPerformed() method executes on the event-dispatching thread.
*/
class Model extends Observable implements ActionListener, Runnable {

private int interval = 1000 / 4; // 4 Hz
private Timer timer; // javax.swing.Timer required
private Results results; // a synchronized Collection
private boolean done;

public Model() {
super();
timer = new Timer(interval, this);
}

public void run() {
done = false;
timer.start()
// generate results
done = true;
}

public void actionPerformed(ActionEvent e) {
if (done) timer.stop;
// notify Observer(s) of progress
setChanged();
notifyObservers(results);
}
}

/**
* A view capable of processing interim results from a model.
*/
public class View extends JPanel implements Observer {
View(Model model) {
...
model.addObserver(this);
}

/**
* Act on notifications from the model.
* This method relies on notifications being sent from the
* actionPerformed() method of a javax.swing.Timer in order
* to execute on the event-dispatching thread.
*/
public void update(Observable o, Object arg) {
Results results = (Results) arg;
// process results for display
this.repaint();
}
}
....
Model model = new Model();
new Thread((Runnable) model).start();

[...] These point are well taken, and thank you for pointing out
the TableModel caveat. I've been modifying a BufferedImage's
WriteableRaster, so any problems may have been more in-apparent
than impossible.

I suspect that you were just lucky to have chosen an implementation
that is unlikely to have run into problems with WriteableRaster.
BufferedImage/WriteableRaster don't have the same EDT-related
limitations that JTable, etc. would have. But they aren't
thread-safe either.

In your case however, my impression is that you are simply updating
the BufferedImage periodically, and my guess is that your period is
much longer than it would take to display the image on-screen. In
that particular implementation, you pretty much guarantee that the
EDT is done with the BufferedImage by the time you get around to
trying to change it again, and you wouldn't normally cause the EDT
to try to access it again until after you've finished changing it.

You probably can, if you try, cause a subtle display bug by forcing a
redraw of the image at the exact moment that the processing thread
is updating the BufferedImage. But you'd have to have just the
right timing (this is typical of threading bugs...the vast majority
of the time, things work fine, but once in a blue moon you hit the
timing just right, and things fail...sometimes quite dramatically
). And even doing that, in this particular case the error would
probably be very subtle, depending on how exactly you're updating
the BufferedImage (if you're just adding new things to the image,
then at worst the image displayed would have a partial update...if
you clear the image and redraw it from scratch, then you might have
a more easily-noticed problem).

Yes. In previous tests, the shared data was simply a byte[][], so the
display was correct even in the worst case: a final update _not_
executing on the EDT, while paintComponent() was still executing _on_
the EDT.

[...]
On reflection, I see the clear contract the invokeLater() and
invokeAndWait() methods offer with respect to the EDT, and my approach
may have too many limitations.

I guess that depends on your definition of "limitations". But personally,
I'd say the contrary is true: your approach has too few limitations.
In particular, using invokeLater/AndWait() is somewhat more limited. But
with that limitation comes a clearer, simpler way to move data from one
thread to another.

Your approach is actually more flexible, but it carries with it a heavy
burden to fully address all of the threading issues, which include both
the requirement to ensure that all methods (with few exceptions, like
repaint()) for a AWT/Swing component are called on the EDT, and the
requirement that you synchronize access to shared data structures between
any code running on the EDT and on your processing thread.

Yes, I am reluctant to abandon the approach completely; I like the
Observer/Observable pattern too well Thank you, too for guiding me to
the (latent) bug described above.

 

[John B. Matthews]

[...]

Starting a new thread for a Runnable [e.g. new Thread((Runnable)
scanner).start()] causes the Runnable's run() method to begin
executing on a new thread. Invoking notifyObservers() causes the
registered Observer's update() method to execute on that same
thread, even though other methods in the Observable may be
executing on another thread.

Well, that depends on how you call notifyObservers(). The code you
posted does _not_ actually do this though. It calls
notifyObservers() from the actionPerformed() method, which is called
by the Timer class on the EDT, not the thread on which the
Observable's run() method was called.

Please note, from the above: "...not the thread on which the Observable's
run() method was called"

Worse! A serious flaw in the outline I proposed is that update() runs on
the EDT when sent by the Timer's actionPerformed() method, but the final
notification executes update() on the thread on which the
Observable's run() method was called.

Ugh.

No, you are incorrect. Again, this seems to be a flaw in your
understanding about how threads work. It is _definitely_ not true that
"the final notificatione executes update() on the thread on which the
Observable's run() method was called". There is no way for Java to do
that, without you implementing it explicitly yourself.

I even reiterated the point here:

Note in particular the last phrase in the sentence, starting with "but in
any case...".

You've been extraordinarily patient, so I am reluctant to be
contradictory. In my original post, notifyObservers() is called each
time actionPerformed() is called by the Timer; and one final time in
run(), after the timer stops:

public void run() {
timer.start()
// lengthy code that generates results
timer.stop();
setChanged();
notifyObservers(results);
}

public void actionPerformed(ActionEvent e) {
// interim progress
setChanged();
notifyObservers(results);
}

The call stack shows a series of calls to update() via the EDT and a
final one in an anonymous thread on which the Observable's run() method
was called. It is definitely true. It was a bug. I fixed it in the
amended code below.

I don't see how that reassures you.

Observable states "that this notification mechanism has nothing to do
with threads..." The method notifyObservers() says, "Each observer has
its update method called..." Where does "yet another thread" come from?
Moreover, the warning specifically refers to subclasses of Observable
and the Observer's reliance on order. As long as my subclass of
Observable doesn't start another thread to issue notifications,
notifyObservers() will execute update() on the EDT, as arranged by Timer.

Again, you're not really being clear. Drawing methods may call other
drawing methods.

Either you are saying that the current instruction being executed can only
be in one method at a time (which is trivially true and uninteresting) or
you are saying that the EDT executes exactly one method per event queue
item (which is obviously false and could lead to some unfortunate mistakes
if it's believed).

I see your point that "executing" includes methods still in the call
stack. I understand that AWTEvents are dispatched from the EventQueue
sequentially and in the same order as they are put on the queue.

It's not clear why the statement you made is relevant to the question at
hand anyway, but inasmuch as you believe it to be, that could be at least
partially responsible for your errors.


The clarification is fine as far as it goes. But you still have a
synchronization bug and you are still relying on undocumented behavior.

I understand the need for synchronization; I specified a synchronized
Collection, but the comment was elided. I believe the behavior is
documented, as described above.

In particular:


The above method runs on thread A and modifies the data structure that
keeps the results of the processing.


The above method runs on the EDT. In the current implementation of the
Java run-time I'm testing on, it winds up calling this method...


...on the same thread (i.e. the EDT).

I see the same result using the amended code.

On the one hand, this is good for the purposes of applying whatever
changes have happened to whatever components are used to display the
results. On the other hand, you are relying on undocumented behavior, and
in addition you fail to synchronize access to the data structure that is
being modified on thread A.

I specified a synchronized Collection above, but this might be more
clear:

Results results = (Results) arg;
synchronized(results) {
// process results for display
}

The one positive aspect of the example is dependent on one of the negative
aspects, and fails to address the other negative aspects. I don't think
this is a good thing.

[...]

Your approach is actually more flexible, but it carries with it a
heavy burden to fully address all of the threading issues, which
include both the requirement to ensure that all methods (with few
exceptions, like repaint()) for a AWT/Swing component are called
on the EDT, and the requirement that you synchronize access to
shared data structures between any code running on the EDT and on
your processing thread.

Yes, I am reluctant to abandon the approach completely; I like the
Observer/Observable pattern too well Thank you, too for guiding me to
the (latent) bug described above.

I'm afraid I can't take any credit yet, since so far you seem to be
believing there's a bug where there isn't, and failing to understand the
bugs where they are.

No, the spurious, final notification was calling update() on a thread
other than the EDT. That was a bug waiting to happen. You helped me look
at this critically. Thank you.

I think we agree that the contract for Timer specifies that action event
handlers for Timers execute on the EDT.

There's nothing fundamentally unfixable by your approach, but at the very
least, I think the difficulty in even _explaining_ the bugs is strongly
illustrative of the challenges involved in getting that approach to work
correctly. It's definitely not an implementation that someone
inexperienced with multi-threaded code should be trying to attempt.

I'm no expert on multi-threading, but I sometimes get to track down bugs
in code written by people who are.

Note also that it's not that there's anything wrong with using the
Observer/Observable pattern per se. You could almost as easily apply that
pattern to the approach I've suggested. The biggest problem is the lack
of any assurance that the Observer's update() method will be called on the
same thread that the Observable used to call notifyObservers(), but that's
addressable with a (potentially redundant) call to invokeLater/AndWait().

I believe my interpretation is reasonable, but I remain open.

 

[John B. Matthews]

[...]
In my original post, notifyObservers() is called each time
actionPerformed() is called by the Timer; and one final time in
run(), after the timer stops:

Ah, I see. I failed to process the crucial word "final" (as in, "but the
final notification executes...") in your previous post. I thought you
were making the statement about _all_ notifications. (I'll blame it on
the fact that "final" wound up at the end of a line, and it's easy to miss
words at the end of a line ).

At least I was correct about there being "no way for Java to do that,
without implementing it explicitly yourself". You _did_ implement it
explicitly yourself.

Sorry for the confusion.

Not at all; I'd overlooked it for much longer! Plus, I always see
something new in the debugger and profiler

[...]
Observable states "that this notification mechanism has nothing to do
with threads..." The method notifyObservers() says, "Each observer has
its update method called..." Where does "yet another thread" come from?

It comes from either a change in the Java run-time implementation, or from
a sub-class that modifies the default behavior. The latter you already
understand. The former seems unlikely to me, but given that the
documentation doesn't provide a promise against the possibility, I'm
loathe to actually _depend_ on that.

Ah, I see the value in this distinction: the former may be less likely,
but it still deserves notice; the latter definitely requires a clear
prohibition against monkeying with the notification thread, as
established by Timer.

But even the possibility of a different sub-class of Observable is an
important possibility. Sample code almost never gets run as-is. People
start there and then modify it. There's no way to guarantee that the
Observable class is the default one, or a sub-class that doesn't modify
the default threading behavior.

This is an issue that documentation, perhaps even to a degree that seems
excessive, can address successfully. But it does need to be
well-documented in the sample.

I agree emphatically.

[...]
I understand the need for synchronization; I specified a
synchronized Collection [in] the comment[...]

You did have a comment, but since your code doesn't actually demonstrate
the initialization and use of the Results, I'm afraid that a naïve reader
may gloss over or completely ignore the advice. IMHO it should be much
more explicit. Also consider that many scenarios don't even lend
themselves to having results of processing being stored in a collection.

It's nice to have collection classes that internalize synchronization, but
a person without that option may find themselves lost without more
explicit guidance as to how to synchronize across the threads (or may not
even realize the importance of the comment describing the "Results" class
as a synchronized collection).

Yes, a single, easily-overlooked comment is insufficient. The shared
data may range from a single primitive type to an arbitrarily complex
data structure, and synchronization may be required for none, some or
all of the data. I've had excellent results where "none" was required,
but I want to experiment more with "some" and "all" for a time.

[...]

 

[John B. Matthews]

Any time you share data between threads, it is never the case that
the amount of synchronization required is "none". There are
circumstances where it might appear that the code is not broken for a
while, but it's about as safe as smoking cigarettes thinking you
don't risk cancer just because it hasn't happened yet.

No smoking!

Without some form of synchronization, there is absolutely no guarantee that
writes made by one thread will ever be visible for reads in another.

None.

Certainly a program must be correctly synchronized, but I'd naturally
like to use explicit synchronization no more than is required.

For example, elements of an int array are modified atomically and do not
need synchronization to ensure sequential consistency. While iterating,
a single writer and a single reader can race on at most one element; the
reader simply sees the previous value. Each full iteration is perceived
as atomic, although one element may be from iteration n-1. The Timer's
final notification ensures that the last complete writer iteration
happens before the start of the last reader iteration. The result is
correct with no explicit synchronization. In the absence of word
tearing, a similar case may be made for byte or short array. This is
what I meant by "none"; please correct me if I'm wrong.

In contrast, consider a model that reports interim progress as a
percentage of completion. If an Observer receives a new Number as the
actual parameter in update(), there is no shared data and no need for
synchronization at all.

I see the classes of java.util.concurrent.atomic extend the notion of
volatile, but they are new to me.

Of course, I already love the names.

 

[John B. Matthews]

Certainly a program must be correctly synchronized, but I'd naturally
like to use explicit synchronization no more than is required.

It is required *every* time threads share data. [...]
Atomicity is not the problem. Visibility is. [...]
Changes written by one thread may never be seen by another. [...]
The happens-before relationship will not exist without synchronization.

[...]

Ah, I think I'm back on the reservation. The memory model doesn't
_preclude_ visibility across threads, it just can't _guarantee_ it
without synchronization.

As long as the thread is the same.


Yes, they are a mere nearly four years old and introduced with a version now
in its End-of-Life processing.

Yes, I'm late to this corner of the party.

The important thing is that they implement explicit synchronization
internally.

Without synchronization, there is no happens-before established. Without
happens-before, changes may *never* propagate to a different thread.

Thanks!