Core Data only supports a small subset of ObjectiveC’s data. That’s fine – for most apps, the data you need to save can be broken down into smaller pieces of primitive data (that Apple supports).

Real properties and simulated properties

We often need to mix “supported” and “unsupported” data in a single class. There’s a couple of ways you can implement this – the most correct is to extend your CoreData class with a category that simulated a property using Associated Objects. The quicker way is to do the same but instead of an Ass. obj, simply write a short “setter” and “getter” (or for read-only: only a “getter”) in the category:

@interface MyCoreDataClass
@property(nonatomic) String storedFilename;
@end

@interface MyCoreDataClass(NonStoredAdditions)
/** simulated, derived readonly property that CoreData doesn't know about */
-(BOOL) fileExists;
@end

@implementation MyCoreDataClass(NonStoredAdditions)
/** simulated, derived readonly property that CoreData doesn't know about */
-(BOOL) fileExists
{
    ... complex code here to check for existence of the file
    ... save the result into an Associated Object so we don't re-calc every time
    ... MUST invalidate this code each time .storedFilename changes!

    ... Hmm. How?
}
@end

Either way, you need to react to “the other variable changing”. The lame way of doing this is that every time you update the CoreData-backed fields, you also manually update the front-end fields – and vice-versa. It’s … lots of code, and easy to get wrong.

But that’s error-prone, and I hate writing boilerplate (it’s bound to cause a bug sooner or later).

ObjectiveC’s listen-to-other-property: KVO (with shortcut)

The correct ObjectiveC response is to use KVO, which is fiendishly over-complicated – but has a couple of shortcuts for the common-cases:

-(id) init
{
    self = [super init];
    if( self )
    {
	[self addObserver:self forKeyPath:@"storedFilename" options:NSKeyValueObservingOptionNew context:self];
    }
    return self;
}

- (void)dealloc
{
    // REQUIRED: or your app will crash horribly at runtime, at random times:
    [self removeObserver:self forKeyPath:@"storedFilename" context:self];
	
    [super dealloc];
}


- (void)observeValueForKeyPath:(NSString *)keyPath ofObject:(id)object change:(NSDictionary *)change context:(void *)context
{
    if([keyPath isEqualToString:@"storedFilename"])
    {
        // react here by re-calculating the value of your derived property, and
        // ...re-store it into the ass. object
    }
}

Sadly, those shortcuts require you to put code in -init and -dealloc (otherwise you get random crashes, and some angry log messages). Argh! It’s a vicious circle of “CoreData doesn’t support X, so we do Y, but Y requires Z … which CoreData *also* doesn’t support”.

CoreData *does* support KVO listeners … just needs a tiny tweak

The correct way for Core Data is to use the “near equivalent” magic post-init methods, and the magic pre-dealloc method:

-(void)awakeFromFetch // called when pre-saved object is re-loaded from database
{
	[self addObserverForFileChanges];
}

-(void)awakeFromInsert // called when object is FIRST created
{
	[self addObserverForFileChanges];
}

-(void) addObserverForFileChanges
{
	[self addObserver:self forKeyPath:@"storedFilename" options:NSKeyValueObservingOptionNew context:self];
}
...
- (void)didTurnIntoFault
{
    [super didTurnIntoFault];
    [self removeObserver:self forKeyPath:@"storedFilename" context:self];
}

For more info:

• - (void)didTurnIntoFault

Core Data is Apple’s answer to “Wow! It’s difficult to store objects in an SQL database!”. Extended, over time, to do a lot more than just that – but that’s the core.

If you know what you’re doing – and you avoid the pit-traps along the way – it can be very good indeed. But we frequently see code written by iOS professionals that betrays a misunderstanding of what Apple intended, and misses out on some of the best features of Core Data.

Over time, we’ve coalesced some of our practices into re-useable code and techniques. All the code in this article is already up on GitHub, and I’ll be maintaining it periodically with improvements from our own projects.

(this post will be followed by a few others, or else extended over time)

Alternatives

NB … if you’re not interested in alternative approaches, skip ahead to the next section.

…re-write / extend Core Data?

Before we start, I should point out that if you want “a better Core Data” there’s several things out there. I’ve tried a few, but none of them have ever made it onto our launched apps. This is *even though* I’ve been quietly impressed by some of them.

Why not? Well … here’s the issues that have affected us (and may affect you too)

1. Core Data – in fact, any Persistence Layer – is massively complex, it deliberately breaks the Objective-C/Cocoa platform, and it’s extremely dangerous to mess with. Some of the tech we’ve seen tries too hard to “fix” CoreData, and in the process has unexpected side-effects.
   • …when you’re dealing with USER DATA, that’s a terrifying prospect
   • Examples include: the Objective-C standard method “description” *corrupts yourd data if you alter it in any way”
   • Examples include: the Objective-C standard method “copy” *is hardcoded* to be unusable on CD objects – you get a nice error message at runtime from Apple if you try to use it, and your app crashes
   • …and other similar terrors. Unless you are 120% confident of your “CoreData++” library, don’t risk it.
2. Core Data is a single standard that Apple beats developers over the head with. There’s only one version, it’s very difficult to customize.
   • So: if you stick to standard CoreData method calls and behaviour, most developers can immediately understand your code without having to learn your proprietary setup
   • …which saves a lot of time and money on projects with multiple teams / companies involved
3. For some projects / clients / partners, the “license” on code is a serious issue, and code that you or I would happily use gets “forbidden” by the legal teams.
   • For some clients, anything we use has to be written by us, or have a massively open license (no GPL, no LGPL – not even MIT/BSD, in some cases)
   • For a very narrow set of clients, anything we use has to be reviewed *by us* (or by the client) line by line, and we have to warranty it. i.e. the code needs to be extremely short and simple!
4. Any code we write on top of this – can it be re-used in a “plain” CoreData project?
   • Often, you want to copy/paste snippets of code from one old project to a new one (assuming of course you – or same client – own the original code).
   • …if you’ve extended CoreData with new classes and methods, you might find this is very hard to do. (this has happened before on some iOS projects, with other libs we’ve used)

So, instead, we focus on making as few changes to CoreData as possible – and generally just sticking to Apple’s own concepts.

Making life easy, with minimal changes

We’re going to do three broad things:

1. Clean up Apple’s code, and enable some Apple-provided features you probably didn’t know existed
2. Very slightly extend Apple’s code (mostly: we’re going to make a start at adding some Blocks, since Apple still hasn’t!)
3. Offer some good practices in how you write YOUR code

1. Replace “100 lines of code” with a simple, easy, encapsulated class

One new class: CoreDataStack.m

One of the best parts of Core Data is that the high-level architecture is neatly split into 4 distinct layers. Collectively, Apple terms these the Core Data “stack” (although that term never appears in code – only in docs). Unfortunately, since all those layers are needed before you can do anything, it requires 100 lines of code just to read a single item from Core Data.

Apple’s image + text showing the layers (may require login to Apple Dev network)

1. Layer 1: file (on disk, or flash memory – Apple has NOT YET IMPLEMENTED a network version, or a remote DB one, although devs have been asking for this for many years. Unlikely to happen anytime soon)
2. Layer 2: Persistent Object Store (the DB part)
3. Layer 3: Persistence Store Co-ordinator (high-level management stuff)
4. Layer 4: NSManagedObjectContext (this is what 99% of your app code talks to)

Apple has largely ignored this problem; their “solution” is that when you create a new project in Xcode, they dump 100 lines (well, not quite 100, but close) of crud into your AppDelegate class.

This the wrong place to put that code – by Apple’s own rules, it should not go there. If your application is multi-threaded (90% of apps are) then it also MUST NOT go there (that’s almost guaranteed to cause data-loss bugs in the long run).

So, the first thing we did was to create a class – CoreDataStack – that encapsulates Apple’s 100 lines of boilerplate code:

GitHub link to CoreDataStack.m

You can drag/drop this into your projects (take the .h file too, of course) – it has NO DEPENDENCIES, except for CoreData.Framework (which is needed in all Core Data projects, by definition)

Using CoreDataStack

Instead of Apple’s large number of method calls, and large number of properties, in real-world projects you need just one property and just one method.

(NB: this assumes you’ve already used Apple’s GUI Editor to create a CoreDataModel file with one or more Entities and Attributes)

To start using Core Data:

/** one line complete init */
CoreDataStack* cdStack = [CoreDataStack coreDataStackWithModelName:@"MyModelName"];

/** one property you need for all Core Data method calls */
cdStack.managedObjectContext;

…where MyModelName is the filename of the Model file (the thing you click on to get the GUI interface for editing Entities and Attributes)

That’s it! Really! Everything else can be inferred (so we do it automatically, in code).

For compatibility, all the other items you *might, theoretically* want to access (e.g. the PersistenceStoreCoordinator, etc) are presented as properties on CoreDataStack.

Why is this a separate class, and yet not a Singleton?

This is actually very important: you DO NOT WANT CoreDataStack to be a Singleton! (although many people try to write their apps that way – c.f. Apple’s default template and its abuse of AppDelegate.m!)

Why?

Because Core Data was designed to have *multiple simultaneous instances* of NSManagedObjectContext, NSPersistentStoreCoordinator, etc. If you limit yourself to one copy of each – as per Apple’s template – you disable many of the features of CoreData.

Worse, some of those features are *required* if you’re going to use CoreData correctly on a real-world project.

The nearest you can get to making this a singleton is … have one instance per xcdatamodel file (an earlier version of the code would even cache this for you, although we since removed that because the performance impact was invisible – not worth the complexity).

More on this (multiple CoreDataStack instances) in the next section…

Tips and tricks: Improving your own code

Use multiple CoreDataStack instances

Here’s a little secret: Core Data was designed for you to access MULTIPLE “models” at once, within a single app … but Apple’s default template for iPhone projects makes that impossible (unless you replace it).

Also, remember the most important rule of Core Data: Apple’s source code is NOT THREAD SAFE and WILL CORRUPT YOUR DATA at random intervals if your app is multithreaded in the “wrong” way [click for more info].

…but here’s another, bigger, secret: if you’re clever, using multiple models at once, you can *avoid* the need for thread-safe code – both your code and Apple’s code. This is the only excuse for Apple’s code being unsafe: you can avoid the need for safety.

So, with CoreDataStack, to use a second model in your app, all you do is:

/** one line complete init */
CoreDataStack* cdOtherStack = [CoreDataStack coreDataStackWithModelName:@"MyOtherModelName"];

/** one property you need for all Core Data method calls */
cdOtherStack.managedObjectContext;

So long as you only have one thread reading and writing to each CoreDataStack instance, you will avoid all the bugs caused by Apple’s unsafe code. In most multi-threaded apps, it’s easy to split your data into multiple separate models, and have each model locked to a particular thread.

Multiple models … what?

Say you’re writing an Email client, that has the following classes for Core Data:

• Email.m : Subject, To, From, Body, isReadYet, DateReceived
• Person.m : EmailAddress, Name

Your app does the following:

• Shows a list of emails, that you can click on to read
• Shows a list of people – tap a person’s name to send them an email, or tap an “Add” button to add a new person
• Downloads emails in the background automatically
• Syncs your contacts to an addressbook server in the background

You now have a multi-threading problem – two sets of background threads accessing Core Data. This *will corrupt your data*. Apple has never tried to make CoreData thread-safe.

Now you have two choices:

• OPTION 1: learn how to write the bizarre code that lets CoreData function with multiple threads. Pray that none of your colleagues dares to edit any of your code – because if they do, there’s a high chance they’ll break it without realizing
• OPTION 2: Instead of having ONE model, have TWO models. One model contains just “Email” and the other contains just “Person”. Each background thread is associated 1:1 with a separate CoreDataStack instance – and suddenly everything is Thread-Safe. Nothing to worry about!

NB: this works because of CoreData’s fundamental design: Apple’s code is *not thread-safe for a single NSManagedObjectContext, but it IS thread-safe for multiple separate NSManagedObjectContext’s in memory at once*

Experts only: Single CoreData model, multiple threads

Further, if you know what you’re doing with multi-threading, there are many situations where you need to have two copies in memory of the SAME object-model. You’ll be manually cross-synching (there lie dragons).

But to be thread-safe, you have to guarantee that the entire stack – NSPersistenceCoordinator etc – is separate for each thread. The hard way to do this is to manually manage it. The easy way is to init a separate CoreDataStack instance per thread – and this will automatically be thread-safe, because each CoreDataStack is coded to NOT share any data or references between instances.

Good practices – applies to all CoreData projects

Don’t hard-code your class-names

Apple’s example source code for using CoreData is technically correct, but practically poor. It encourages a bad habit that – for most projects – is unnecessary and the cause of many bugs over time.

When you need to reference a CD object, Core Data is written so that you don’t neet to have access to the Class of that object. In theory, you can load a Class (from CoreData’s database, saved by a different app) that doesn’t exist in your project. We’re getting into some weird and freaky stuff here.

Just in case – even though 90% of coders will never use that feature – Apple tells you to use NSString to instantiate your CoreData objects. This is bad for most projects:

1. Apple’s refactor tools in Xcode are 10 years behind everyone else’s – they don’t support CoreData, and they ignore those strings – if you refactor a CoreData class, Xcode will break your project
2. It’s very very easy to make a typo when writing that string. Xcode 3 would auto-complete the name for you, but Xcode 4 removed this feature

So, each time you need to create a new object in the CoreData database/store, instead of this:

Email* newEmail = [NSEntityDescription insertNewObjectForEntityForName:@"Email" inManagedObjectContext:cdStack.managedObjectContext];

do this:

Email* newEmail = [NSEntityDescription insertNewObjectForEntityForName:NSStringFromClass([Email class]) inManagedObjectContext:cdStack.managedObjectContext];

…using Xcode’s autocomplete, this is the same or fewer keystrokes, even though it results in more text. More importantly, the compiler will now double-check that class for you, and refuse to build if you typo it. You’re also safe(r) with refactoring.

Similarly, when you do a Fetch with CoreData, instead of passing the string-name of the class, use the same NSStringFromClass call as above, for the same benefits.

Upgrading the CoreDataStack

Save Errors SHOULD NEVER BE IGNORED

Apple doesn’t like Exception Handlers and Assertions. Personally, I think 30 years of computer industry have proved them wrong there, but I’m willing to accept it.

Except for when a SAVE of CoreData fails; in this particular case, it is totally unacceptable for your app to silently ignore the error. Sadly, Apple’s default setup encourages you to do this.

How many times have you seen this in an app:

[self.managedObjectContext save:nil];

?

Sadly, I see it all the time. Because the alternative requires a minimum of 5 lines of code, including a double-pointer (that a lot of junior programmers and/or people who’ve never used C/C++ seem to feel uncomfortable with).

As a bonus, Apple’s code has some nasty behaviour with that save method. They have documented this (so I guess it’s a “feature” instead of a bug – but see what you think):

If managedObjectContex is nil, then “a save that fails” will ALWAYS seem to return “there was no error”.

This is one of the most painful things I’ve had to debug on CoreData projects. Many times.

So, we’re going to fix both of those. CoreDataStack.m has an optional method:

-(void) saveOrFail:(void(^)(NSError* errorOrNil)) blockFailedToSave

1. Because it’s a block, it’s *just as easy to log the error* as it is to ignore it (instead of requiring you to write error-creating + checkign code). Further, because it’s a block, it’s easy to nest attempted saves, logs, and failures.
2. If your NSManagedObjectContex is nil … instead of doing what Apple does (silent failure; pretends that the save has succeeded) … we explicitly FAIL (c.f. the source code for the saveOrFail method).

PS

Until 2011, this wasn’t even documented, but … not only is CoreData “not thread-safe”, but if you merely allocate/init a CoreData context from a different thread you will cause it to quietly self-destruct later on – after it’s taken your precious data. It’s easy to understand once you know – but this is not an obvious side-effect, and I’ve seen it catch a few people.

e.g. one of the patterns coders previously used to handle multi-threading – because Apple hadn’t documented this requirement – involved Thread A creating the context, adding notification listeners, and then passing the context to the thread that would use it.

This works most of the time – I know, because I used this pattern back in 2009 – and I’d learnt it from someone else who’d been using it for a while themselves. But it also fails some of the time, unpredictably, with strange crashes deep in Apple’s code. Now we know better, of course.

My take-home: multi-threaded CoreData is to be avoided unless you *really* know what you’re doing with CoreData. Given that Apple has been very slow (years slow!) to document this aspect of CoreData, I don’t recommend it, unless you’re willing to spend a lot of time learning the “community wisdom” on the topic.

Our recent iPad app – London Unfurled – draws enormous images (90,000 pixels wide) very fast, with zooming and panning. We built it using a custom OpenGL rendering loop – we did NOT use CATiledLayer.

When I gave a couple of talks on this recently, several people wondered if “CATiledLayer would do it all for you?”. Ah … no. Try the app. You’ll quickly see that it’s much faster than CATL, even at CATL’s fastest.

It seems that CATL is a little slower, and uses more memory, than many of us expected. But it’s still very useful – e.g. we’re using it in a new game project right now. So … what’s going on? When should you use CATL, and when should you avoid it? What works, and what doesn’t?

Why all the confusion?

Three problems:

1. Until iOS4, CATL routinely crashed people’s apps, with little or no explanation. It was a documented feature in Apple’s rendering code (Apple has now changed this to be more programmer-friendly – no crashes), quite easy to workaround once you knew it. But … it created a lot of FUD and frustration. A lot of experienced iOS programmers have learnt to live without CATL, and haven’t used it in a shipping app
2. There’s very little documentation for CATL. A total of 5 sentences for the class, and 5 sentences covering configuration. Even the most Frequently Asked Question about CATL is missing from the docs (“how do I get rid of the fade/flash effect?”).
3. CATL appears to fill an obvious hole in Apple’s libraries: Apple hasn’t (yet) provided classes to handle lazy rendering and “automatic redraw” when zooming. CATL doesn’t necessarily fill those holes – but with the lack of docs, lots of programmers jump to some (reasonable) conclusions here

What do iOS programmers try to use CATiledLayer for?

So, let’s have a look at some common use cases, and some common misconceptions:

1. Drawing images that are “too big” to place in a single UIImageView (e.g. anything over 2096×2096 pixels)
2. Reducing the memory footprint of “huge” images, so that an inifinitely large image can be rendered
3. Rendering “huge” images very quickly, using automatic caching
4. Adding “infinite zooming” to your application (e.g. as used in the Google Maps app)
5. Making CGPath objects zoom correctly when you zoom in a UIScrollView (without CATL, they go blurry when you zoom)

What is CATiledLayer actually good for?

CATiledLayer:

• GOOD INTEGRATION with UIScrollView, automatically “Just Works” (almost; see below)
• When you zoom in a UIScrollView, CATL will AUTOMATICALLY RE-RENDER ITS CONTENTS AT HIGH RESOLUTION
• When tiles are cached, panning and zooming is FAST and SMOOTH
• When some tiles are NOT cached, panning is JUDDERY, but zooming is SMOOTH
• Uses MORE MEMORY than manually managing memory; on iOS, if you’re rendering truly huge images, and you want them to run fast … then you need more fine-grained control over the exact amount of RAM you’re using from second to second
• Uses a SIMPLISTIC, OFTEN POOR caching algorithm: e.g. for any content that is not “UIImage / CGImageRef”, it will REDUCE rendering speed
• Even for images, if an image is more than one tile in size, CATL will ADD A FLICKERING EFFECT TO THE RENDERING (like the tiles in Google Maps “flicking” into existence). This CANNOT BE REMOVED.
• Additionally, by default, it will ANIMATE THE FLICKERING to make it less offensive – but more obvious. This CAN BE REMOVED.

So, in summary, when to use CATL?

• If the “raw” content of your layer CHANGES RARELY OR NEVER
• If your content can be rendered at MULTIPLE RESOLUTIONS (e.g. text, e.g. CAShapeLayer, e.g. CGPath)

If you need to change content, there’s workarounds that work nicely (see below) – but it’s only workable if the changes are small and/or relatively infrequent (i.e. no more than a couple / second).

Making use of CATiledLayer…

…CATL does NOT automatically re-draw!

Every time you make ANY change to the contents of a CATL, you must manually call:

[(CATiledLayer*) YOUR_OBJECT setNeedsDisplay];

…effectively, you are really calling the private internal method:

[(CATiledLayer*) YOUR_OBJECT invalidateCache]; // just guessing. This?
[(CATiledLayer*) YOUR_OBJECT reGenerateTiles]; // ...or maybe this?

…but that’s not entirely obvious.

…with UIScrollView

For “perfect” integration, the CATL would use the UISV’s zoom settings to automatically determine it’s own zoom settings.

Fortunately, this is relatively easy to implement manually. There’s some excellent walk-through info on this at Things That Were Not Immediately Obvious To Me (NB: if you also read the “Part 1″ of that page, don’t panic – it’s doing a LOT more work than you need to. Probably ignore it for now).

…with CAShapeLayer

If you place a CAShapeL inside a UIScrollView, and zoom in … it all goes blurry. .

However, if you place CATiledL inside a UIScrollView, and write your CAShapeL’s into the CATiledL, and zoom in … you get high-resolution. Yay!

But … performance drops through the floor. Because CATiledL isn’t doing any bitmap caching.

Fortunately, you can flip the “shouldRasterize” property of CAShapeL to “TRUE” just before drawing it to the CATiledL, and it wil output something that’s already a bitmap, so you get the best of both worlds.

NB: shouldRasterize can have unexpected consequences when using CA Animations – however, it’s a huge performance boost for some things, like drop shadows. So, if your rendering looks wrong when you use this trick, Google for the extensive tutorials and support questions on shouldRasterize…

Example code:

-(void)drawRect:(CGRect)rect
{	
	CGContextRef context = UIGraphicsGetCurrentContext();
	
	for( CAShapeLayer* s in self.myShapes )
	{

// NB: this "if" assumes that you manually created your CAShapeLayer
//    the correct "frame" such that it fits the embedded CGPath.
// By default, Apple doesn't do this for you - your CAShapeLayer will
//    have a frame of {{0,0} {0,0}}.
// You can omit the "if" altogether - it will reduce performance a little,
//    but the tile is going to be cached anyway, so it's not fatal
		if( CGRectIntersectsRect(rect, s.frame ) 
		{
			CGContextSaveGState(context);
				
			/** 
			 this effectively causes the CATiledLayer to cache a bitmap instead of complete CGPath's
				 
			 so ... rendering performance is noticeably faster, even with just 5-6 paths on screen!
			 */
			s.shouldRasterize = TRUE;
			
			/** 
			CGContextTranslateCTM(context, s.frame.origin.x, s.frame.origin.y );
			[s renderInContext:context];

			CGContextRestoreGState(context);
		}
	}
}

…with small changes to contents

If you redraw a CALayer, Apple waits till you’ve drawn the whole thing, then moves it to the screen. It appears all in one go, maybe with a tiny delay.

If you change even a TINY part of a CATL, even if you tell the CATL that “only this small piece has changed”, then CATL updates the screen lots of times, once for each tile that is generated. This is slower than doing it all in once, and creates an artificial “flicker” on the screen.

(NB: to do “only a small change”, use [setNeedsDisplayInRect:] – I’ve tested, and this appears to only de-cache the tiles in that rect. It is noticeably faster than [setNeedsDisplay], when the rect is small)

There is no way around this – CATL is just a little weak in its core algorithm. I believe Apple is introducing an artificial delay so that it can execute on a background thread without “using too much CPU time”.

So … instead, if you want the screen to update instantly with no flicker … you have to create an “overlay” layer above the CATL, and write your changes into that layer.

e.g. I recently made a game which showed all the countries of the world, and let you click a country to select it. When you select it, I wanted to change the fill-colour of the country. Implementation:

• UIScrollView (requires ONE AND ONLY ONE subview, or it doesn’t work correctly)
  • UIView (with embedded CALayer) “container”
    • UIView (with embedded CATiledLayer) “all countries of the world”
    • UIView (with embedded CALayer) “OVERLAY for temporary changes to the CATL
      • … if nothing selected, this layer is empty
      • … if a country is selected, I clone the country’s pixels, change the colours, and add them to this layer

This works very fast – when you select a country, there’s no flicker, and it appears to happen instantaneously.

Without this, the update is slow, and you can see the tiles getting rendered one by one.

What does CATL actually do?

Through trial and error, we can see that CATL works something like this…

Sublayers

CATL layers *cannot have* CALayer sublayers.

Try it. Everything goes horribly wrong. Your sublayers might render, if you’re lucky (I saw this happen approx 1 time in 100 – I think it was a bug).

You can force them to render, by adding them to your tiles. Don’t do this: it *destroys* the performance of CATL.

If you want sublayers … you need to create a separate “container” CALayer, add it to the CATL’s *superlayer* … and then add your “sublayers” to that new “container”.

However, generally, it’s better to make a container UIView, and add to the CALayer’s UIView’s superview. Why? Because then you can still add sublayers (UIView.layer is always there), but you can also add things like UIButton too…

Tiles

CATL – surprise, surprise! – works by keeping a list of Tile objects internally, and each time the rendering system asks it to render itself, it blits one or more tiles onto the screen to cover the dirty rectangle.

Unfortunately, Apple *does not* allow us access to the NSObject (or, possibly, the struct) that they use to represent individual “tiles”. They give us an approximation – when they are creating a new Tile, they send us a CGRect that is the exact frame of the tile to generate (i.e. an offset/origin, and a width/height).

All Tiles are the same size *in pixels*, but they end up different sizes when you start zooming (see below).

There is also some magic about the meaning of “size” when a view starts to zoom, more on that later.

You can change the tile-size (it’s a config option for CATL), but that resets the CATL, and deletes all existing tiles (it seems).

The process for “generating new tiles” used to be complex, but now it’s very very simple:

/*! Override "drawRect:".
With CATL, the "rect" object has a special definition: it is the exact *frame* of the Tile that you are
generating - whatever you render will be cached as "TILE-N" (educated guess: no-one knows how Apple implements this internally  ).

NOTE1: Apple hasn't documented this in the CATL class; you have to watch the WWDC videos, or use trial and error to discover it.
NOTE2: because "rect" is a *frame*, it captures both the tile size AND the tile offset. e.g. rect might be: {{512,256}, {256,256}}
*/
-(void)drawRect:(CGRect)rect
{	
	CGContextRef context = UIGraphicsGetCurrentContext();

	for( CALayer* l in self.myInternalLayers )
	{		
		if( CGRectIntersectsRect( rect, l.frame ) )
		{
				CGContextSaveGState(context);				
				CGContextTranslateCTM(context, l.frame.origin.x, l.frame.origin.y );
				[l renderInContext:context];
				CGContextRestoreGState(context);
		}
		else
			;//DEBUG: NSLog(@"ignored non-intersecting tile");
	}
}

Now: here there IS a surprise. By default, CATL stores the tiles in a very low performance way – it stores them as rendering commands. I think most people assume it stores them as blit’able bitmaps – that would make sense: it’s fast, it’s efficient.

IF you’re rendering pure image data in the CATL, then de facto a Tile is “only slightly more memory” than a blit. If you’re rendering anything else, you should write your own code to convert your render commands to a bitmap, and render the bitmap to CATL. If you do this, you can often get literally 5x increase in render speed.

So, for instance, if you draw CGPath objects to a CATL, they will be stored *multiple times over* in memory, increasing your memory usage, and reducing your rendering speed. Yes: used naively, CATL can make rendering slower, and take more memory.

Tile Caching

When CATL renders to screen, it only generates Tiles that it doesn’t already have inside its cache. When it hits some arbitrary limit, it deletes some existing tiles. Apple provides zero info on what the limit is, or which tiles get deleted first.

The cache is private, opaque, of “undocumented” size, with “undocumented” behaviour. This – if nothing else – makes CATL useless in many real-world cases. Opaque caches are evil.

Also, c.f. notes above on Tiles: this is a RENDER COMMAND cache, not a BITMAP cache. If you’re going to do render commands (i.e. 95% of rendering in a modern iOS app!), make sure you generate bitmaps on the fly, and render those to the CATL instead.

LOD (Levels of Detail)

CATL innately supports zooming: it has a special feature where you can insert it inside a UIScrollView, and it will intelligently redraw itself *at higher resolution* whenever the UIScrollView zooms in.

(NB: you can also achieve this behaviour manually, without a UIScrollView, by manually updating the low-level CALayer properties that tell a layer the size/area/zoom it should render with)

When you zoom (the complex process where you change the CALayer size, frame, contentScale, transform, etc), the CATL simply creates a whole NEW set of Tiles, and adds them to its cache. The old ones are NOT deleted (unless it runs out of space).

If you then zoom back out again, CATL will render very quickly, because it has the data cached. Probably.

i.e. the CATL cache looks something like this:

• TILE1 — ZOOM = 1:1 — SIZE = 256×256 — POSITION = 0,0
• TILE2 — ZOOM = 1:1 — SIZE = 256×256 — POSITION = 0,1
• TILE3 — ZOOM = 1:1 — SIZE = 256×256 — POSITION = 1,0
• TILE4 — ZOOM = 1:2 — SIZE = 256×256 — POSITION = 0,0
• TILE5 — ZOOM = 1:2 — SIZE = 256×256 — POSITION = 0,1
• TILE6 — ZOOM = 1:2 — SIZE = 256×256 — POSITION = 1,0

Rendering

NB: what follows comes from trial and error and educated guesses. I’d be very happy for someone from Apple to correct this with what’s *actually* happening – but I think this is close enough for us to understand how to use CATL.

The process that CATL uses goes something like this:

1. Look at the layer’s contentsScale (I think that’s the property it uses? Or … maybe it reads the affineTransform property instead?), and the layer’s “normal” bounds (widthxheight)
2. Use the “scale” info to decide “how many on-screen pixels” a standard tile would cover
   • e.g. if you’ve zoomed-in by a factor of 1.5, and you’re using the default Tile size of 256×256, then the “on-screen” size of a tile is now 384×384
3. WHILE the “on-screen pixel size” is greater than 2 x the default size, switch LOD level (i.e. use tiles that have smaller and smaller “layer.bounds” size)
4. Look at the CGRect that the windowing system has told the CATL to “drawRect:” in
   • if you’ve zoomed in, this will be much smaller area than the layer’s normal bounds
   • if you’ve zoomed out, this will be much bigger area than the layer’s normal bounds
5. Use that CGRect to calculate a list of tile-offsets that are needed to cover the VISIBLE area
   • e.g. for a VISIBLE frame of “{{100,0}, {500,256}}”, and 256×256 tiles, you’d need tiles at: {0,0}, {256,0}, {512,0}.
6. For each tile that is in the cache, render it immediately *BY REPLAYING RENDER COMMANDS*
   • c.f. above, this can be SLOW. So slow that you see tiles appearing one-by-one on screen
7. Display to screen
8. …
9. While that’s appearing on scren, in a background thread, do:
   1. For each tile that is NOT in the cache, call the user-written “drawRect:” method to generate the tiles.
   2. Each time a tile completes, schedule an update to the mainthread that will OVERWRITE the screen contents with the new tile
      • i.e. some time after the CATL was rendered on screen (typically “half a second, up to several seconds”), while your main app is now running some other piece of code, small pieces of the CATL start magically appearing
   3. With the default CATL implementation, each new tile is animated in, using a “flash/fade” animation that takes 0.25 seconds per tile
      • For most apps, most of the time, this is painfully slow. Most people subclass CATL, and override the “+(CSTimeInterval) fadeDuration” method to return “0.0″ instead.
      • c.f. above: even with a fadeDuration of 0.0, you can often STILL “see” the tiles appear on screen, because CATL uses an inefficient tile rendering / tile caching algorithm