When will completionBlock be called for dependencies in NSOperation

问题

From the docs:

The completion block you provide is executed when the value returned by the isFinished method changes to YES. Thus, this block is executed by the operation object after the operation’s primary task is finished or cancelled.

I'm using RestKit/AFNetworking, if that matters.

I have multiple dependencies in my NSOperation in a OperationQueue. I use the completion block to set some variables (appending the results to an array) that my child requires.

(task1,...,taskN) -> taskA

taskA addDependency: task1-taskN

Will taskA receive incomplete data since the child can execute before the completion block is fired?

Reference

Do NSOperations and their completionBlocks run concurrently?

I did a simple test by adding a sleep in my completion block and I had a different result. The completion block runs in the main thread. While all the completion block are sleeping, the child task ran.

回答1:

As I discuss below under "a few observations", you have no assurances that this final dependent operation will not start before your other sundry AFNetworking completion blocks have finished. It strikes me that if this final operation really needs to wait for these completion blocks to finish, then you have a couple of alternatives:

1. Use semaphores within each of the n the completion blocks to signal when they're done and have the completion operation wait for n signals; or

2. Don't queue this final operation up front, but rather have your completion blocks for the individual uploads keep track of how many pending uploads are still incomplete, and when it falls to zero, then initiate the final "post" operation.

3. As you pointed out in your comments, you could wrap your invocation of the AFNetworking operation and its completion handler in your own operation, at which point you can then use the standard addDependency mechanism.

4. You could abandon the addDependency approach (which adds an observer on the isFinished key of the operation upon which this operation is dependent, and once all those dependencies are resolved, performs the isReady KVN; the problem being that this can theoretically happen before your completion block is done) and replace it with your own isReady logic. For example, imagine you had a post operation which you could add your own key dependencies and remove them manually in your completion block, rather than having them removed automatically upon isFinished. Thus, you custom operation

   @interface PostOperation ()
   @property (nonatomic, getter = isReady) BOOL ready;
   @property (nonatomic, strong) NSMutableArray *keys;
   @end
   
   @implementation PostOperation
   
   @synthesize ready = _ready;
   
   - (void)addKeyDependency:(id)key {
       if (!self.keys)
           self.keys = [NSMutableArray arrayWithObject:key];
       else
           [self.keys addObject:key];
   
       self.ready = NO;
   }
   
   - (void)removeKeyDependency:(id)key {
       [self.keys removeObject:key];
   
       if ([self.keys count] == 0)
           self.ready = YES;
   }
   
   - (void)setReady:(BOOL)ready {
       if (ready != _ready) {
           [self willChangeValueForKey:@"isReady"];
           _ready = ready;
           [self didChangeValueForKey:@"isReady"];
       }
   }
   
   - (void)addDependency:(NSOperation *)operation{
       NSAssert(FALSE, @"You should not use addDependency with this custom operation");
   }
   

   Then, your app code could do something like, using addKeyDependency rather than addDependency, and explicitly either removeKeyDependency or cancel in the completion blocks:

   PostOperation *postOperation = [[PostOperation alloc] init];
   
   for (NSInteger i = 0; i < numberOfImages; i++) {
       NSURL *url = ...
       NSURLRequest *request = [NSURLRequest requestWithURL:url];
       NSString *key = [url absoluteString]; // or you could use whatever unique value you want
   
       AFHTTPRequestOperation *operation = [[AFHTTPRequestOperation alloc] initWithRequest:request];
       [operation setCompletionBlockWithSuccess:^(AFHTTPRequestOperation *operation, id responseObject) {
           // update your model or do whatever
   
           // now inform the post operation that this operation is done
   
           [postOperation removeKeyDependency:key];
       } failure:^(AFHTTPRequestOperation *operation, NSError *error) {
           // handle the error any way you want
   
           // perhaps you want to cancel the postOperation; you'd either cancel it or remove the dependency
   
           [postOperation cancel];
       }];
       [postOperation addKeyDependency:key];
       [queue addOperation:operation];
   }
   
   [queue addOperation:postOperation];
   

   This is using AFHTTPRequestOperation, and you'd obviously replace all of this logic with the appropriate AFNetworking operation for your upload, but hopefully it illustrates the idea.

---

Original answer:

A few observations:

1. As I think you concluded, when your operation completes, it (a) initiates its completion block; (b) makes the queue available for other operations (either operations that had not yet started because of maxConcurrentOperationCount, or because of dependencies between the operations). I do not believe that you have any assurances that the completion block will be done before that next operation commences.

   Empirically, it looks like the dependent operation does not actually trigger until after the completion blocks are done, but (a) I don't see that documented anywhere and (b) this is moot because if you're using AFNetworking's own setCompletionBlockWithSuccess, it ends up dispatching the block asynchronously to the main queue (or the defined successCallbackQueue), thereby thwarting any (undocumented) assurances of synchrony.

2. Furthermore, you say that the completion block runs in the main thread. If you're talking about the built in NSOperation completion block, you have no such assurances. In fact, the setCompletionBlock documentation says:

   The exact execution context for your completion block is not guaranteed but is typically a secondary thread. Therefore, you should not use this block to do any work that requires a very specific execution context. Instead, you should shunt that work to your application’s main thread or to the specific thread that is capable of doing it. For example, if you have a custom thread for coordinating the completion of the operation, you could use the completion block to ping that thread.

   But if you're talking about one of AFNetworking's custom completion blocks, e.g. those that you might set with AFHTTPRequestOperation's setCompletionBlockWithSuccess, then, yes, it's true that those are generally dispatched back to the main queue. But AFNetworking does this using the standard completionBlock mechanism, so the above concerns still apply.

回答2:

It matters if your NSOperation is a subclass of AFHTTPRequestOperation. AFHTTPRequestOperation uses the NSOperation's property completionBlock for its own purpose in method setCompletionBlockWithSuccess:failure. In that case, don't set the property completionBlock yourself!

It seems, AFHTTPRequestOperation's success and failure handler will run on the main thread.

Otherwise, the execution context of NSOperation's completion block is "undefined". That means, the completion block can execute on any thread/queue. In fact it executes on some private queue.

IMO, this is the preferred approach, unless the execution context shall be explicitly specified by the call-site. Executing completion handlers on threads or queues which instances are accessible (the main thread for example) can easily cause dead locks by an unwary developer.

---

Edit:

If you want to start a dependent operation after the completion block of the parent operation has been finished, you can solve that by making the completion block content itself a NSBlockOperation (a new parent) and add this operation as a dependency to the children operation and start it in a queue. You may realize, that this quickly becomes unwieldy, though.

Another approach would require an utility class or class library which is especially suited to solve asynchronous problems in a more concise and easy way. ReactiveCocoa would be capable to solve such (an easy) problem. However, it's unduly complex and it actually has a "learning curve" - and a steep one. I wouldn't recommend it, unless you agree to spend a few weeks in learning it and have a lot other asynchronous use cases and even much more complex ones.

A simpler approach would utilize "Promises" which are pretty common in JavaScript, Python, Scala and a few other languages.

Now, please read carefully, the (easy) solution is actually below:

"Promises" (sometimes called Futures or Deferred) represent the eventual result of an asynchronous task. Your fetch request is such asynchronous task. But instead specifying a completion handler, the asynchronous method/task returns a Promise:

-(Promise*) fetchThingsWithURL:(NSURL*)url;

You obtain the result - or the error - with registering a success handler block or a failure handler block like so:

Promise* thingsPromise = [self fetchThingsWithURL:url];
thingsPromise.then(successHandlerBlock, failureHandlerBlock);

or, the blocks inlined:

thingsPromise.then(^id(id things){
   // do something with things
   return <result of success handler>
}, ^id(NSError* error){
   // Ohps, error occurred
   return <result of failure handler>
});

And shorter:

[self fetchThingsWithURL:url]
.then(^id(id result){
     return [self.parser parseAsync:result];
}, nil);

Here, parseAsync: is an asynchronous method which returns a Promise. (Yes, a Promise).

---

You might wonder how to get the result from the parser?

[self fetchThingsWithURL:url]
.then(^id(id result){
     return [self.parser parseAsync:result];
}, nil)
.then(^id(id parserResult){
    NSLog(@"Parser returned: %@", parserResult);
    return nil;  // result not used
}, nil);

This actually starts async task fetchThingsWithURL:. Then when finished successfully, it starts async task parseAsync:. Then when this finished successfully, it prints the result, otherwise it prints the error.

Invoking several asynchronous tasks sequentially, one after the other, is called "continuation" or "chaining".

Note that the whole statement above is asynchronous! That is, when you wrap the above statement into a method, and execute it, the method returns immediately.

---

You might wonder how to catch any errors, say fetchThingsWithURL: fails, or parseAsync::

[self fetchThingsWithURL:url]
.then(^id(id result){
     return [self.parser parseAsync:result];
}, nil)
.then(^id(id parserResult){
    NSLog(@"Parser returned: %@", parserResult);
    return nil;  // result not used
}, nil)
.then(/*succes handler ignored*/, ^id (NSError* error){
    // catch any error
    NSLog(@"ERROR: %@", error);
    return nil; // result not used
});

Handlers execute after the corresponding task has been finished (of course). If the task succeeds, the success handler will be called (if any). If the tasks fails, the error handler will be called (if any).

Handlers may return a Promise (or any other object). For example, if an asynchronous task finished successfully, its success handler will be invoked which starts another asynchronous task, which returns the promise. And when this is finished, yet another one can be started, and so force. That's "continuation" ;)

---

You can return anything from a handler:

Promise* finalResult = [self fetchThingsWithURL:url]
.then(^id(id result){
     return [self.parser parseAsync:result];
}, nil)
.then(^id(id parserResult){
    return @"OK";
}, ^id(NSError* error){
    return error;
});

Now, finalResult will either eventually become the value @"OK" or an NSError.

---

You can save the eventual results into an array:

array = @[
    [self task1],
    [self task2],
    [self task3]
];

and then continue when all tasks have been finished successfully:

[Promise all:array].then(^id(results){
    ...
}, ^id (NSError* error){
    ...
});

---

Setting a promise's value will be called: "resolving". You can resolve a promise only ONCE.

You may wrap any asynchronous method with a completion handler or completion delegates into a method which returns a promise:

- (Promise*) fetchUserWithURL:(NSURL*)url 
{
    Promise* promise = [Promise new];

    HTTPOperation* op = [[HTTPOperation alloc] initWithRequest:request 
        success:^(NSData* data){
            [promise fulfillWithValue:data];
        } 
        failure:^(NSError* error){
            [promise rejectWithReason:error];
        }];

    [op start];

    return promise;
}

Upon completion of the task, the promise can be "fulfilled" passing it the result value, or it can be "rejected" passing it the reason (error).

Depending on the actual implementation, a Promise can also be cancelled. Say, you hold a reference to a request operation:

self.fetchUserPromise = [self fetchUsersWithURL:url];

You can cancel the asynchronous task as follows:

- (void) viewWillDisappear:(BOOL)animate {
    [super viewWillDisappear:animate];
    [self.fetchUserPromise cancel];
    self.fetchUserPromise = nil;
}

---

In order to cancel the associated async task, register a failure handler in the wrapper:

- (Promise*) fetchUserWithURL:(NSURL*)url 
{
    Promise* promise = [Promise new];

    HTTPOperation* op = ... 
    [op start];

    promise.then(nil, ^id(NSError* error){
        if (promise.isCancelled) {
            [op cancel];
        }
        return nil; // result unused
    });

    return promise;
}

Note: you can register success or failure handlers, when, where and as many as you want.

---

So, you can do a lot with promises - and even more than in this brief introduction. If you read up to here, you might get an idea how to solve your actual problem. It's right there - and it's a few lines of code.

I admit, that this short introduction into promises was quite rough and it's also quite new to Objective-C developers, and may sound uncommon.

You can read a lot about promises in the JS community. There are one or three implementations in Objective-C. The actual implementation won't exceed a few hundred lines of code. It happens, that I'm the author of one of it:

RXPromise.

Take it with a grain of salt, I'm probably totally biased, and apparently all others ever dealt with Promises, too. ;)

来源：https://stackoverflow.com/questions/18745635/when-will-completionblock-be-called-for-dependencies-in-nsoperation