Manual Core Data Migrations – Lessons Learned

So, I’ll try to keep this all brief.  In short, as everyone, including Core Data Ninja Master Marcus Zarra, has said:  Avoid heavy Core Data Migrations!

The tools are buggy, the crash bugs difficult to interpret.  I spent 4-5 hours already just sort of shooting at a black box.   So, I hope to share with you the things I learned and what caught me up.  If this is all too brief or out of context in places, it’s because I’m not a tutorial writer, and keep my blog posts quite often so that I can refer back to them.  So please feel free to write me if you need clarification.

I’ll also give the disclaimer that I’m not the most ‘canonical’ of programmers.  So, these are my findings.  I don’t know if they follow best practices, but I don’t think there are any best practices when it comes to this stuff…. otherwise I probably wouldn’t have made this lengthy blog post!

Scenario

Migration of old data store using an old Core Data model to a new data store with a brand new Core Data model

This documents my personal experience trying to migrate a Data Model I want to retire (i.e. completely separate Core Data Model file) to a brand new data model.

I wanted to move away from my old Objective-C pipeline which involves NSManagedObject subclasses using mogenerator that all have an Objective-C style Namespace and codegen, to a pure Swift version which uses the codegen feature of the Latest Xcode Core Data editor.

I kept these models separate because in a few releases when I know everyone has upgraded, I can eventually just remove the legacy model from the bundle and all the associated classes.

What I learned / experienced:

  • How to set up incremental migrations using a handy little helper from http://www.objc.io
  • How to migrate old data to new using Mapping Models
  • How to write and use custom NSEntityMappingPolicy objects
  • How to use the Mapping Model editor to call functions on these custom policies
  • All the various Pitfalls, Gotchas, and things that Apple really needs to improve on.

Useful Pre-requisite Reading:

How to set up incremental migrations

This is really what you want to do.  It makes your life easy in that you only need to make mapping models for each change to your data model.  There is a lot of code floating around the internet that helps you make an incremental migration manager.  I took one from objc.io and then modified it so it could support my use case where I am migrating a completely different store to a new one, and that new one should have its own URL on disk that may not follow from the original store.

Code for that migration manager can be found in a gist here.  Don’t just copy-paste:  You have to copy-paste into the correct files!  Also, it’s in objective-c so you’ll want to make a bridging header… It’s really not hard.

What is different in my version of MHWMigrationManager than from on objc.io:

  • It has a hack fix for a dubious migration error that has minimal debug help: ‘Mismatch between mapping and source/destination models’
  • It allows you to specify via the delegate a “Final Destination URL” once the Migration has successfully completed.  This can be useful as in my case where I have 2 different Stores and Models and want them to stay that way.

Set up your Core Data Stack

Basically, you start with Apple’s boilerplate code for Core Data and a persistent container, then you modify in a way that I’ve done here.  These 2 gists I’ve provided are all just starting points.  We haven’t even got to the annoying stuff yet.

If you create a brand new data model like I have done, be sure that it uses a different name than your old one.  You can’t rename the old one because you want the new one to have the same nice name.  Leave everything about the old one in place.

Pro Tip:  If you didn’t do this but already did the work of creating this new model version, if you create a new model either via duplication or copy-paste of entities, keep in mind that delete rules are not copied over, NOR are the inverse relationships.  So you’ll have to go through and set these by hand again.

Now Create Mapping Models

You can now proceed to create a mapping model in the ways you’ve probably read about on the internet already.   Pro Tips:

  • Only write your mapping model, once your Core Data Object Model is complete.  If you make a modification to your Core Object Model (.xcdatamodeld), any mappings that reference it will no longer be found by the Migration Manager.
  • Due to the nature of the migration process, you won’t have access to any subclass of NSManagedObject, and so this can make setting up NEW relationships difficult.  It is therefore a good idea to keep ‘foreign keys’ on your models with relationships.  (i.e. you extend your Post model to have an Attachment.  Attachment should have an attribute postId, or something identifiable)
  • If you use a subclass of NSEntityMigrationPolicy and you specify it in the Mapping Model editor, make sure that you also provide the Module name.  So, instead of LegacyToModernPolicy, I would put SongbookSimple.LegacyToModernPolicy, because my app’s module name is SongbookSimple.

Mapping Models are useful in that they allow a lot of customization, and you can sort of elect what you want to be done “automagically” and what you’d like to have more control over.  You get more control over these by using custom subclasses of NSEntityMigrationPolicy.

In general, for property or relationship mappings, the auto-generated functionality can get you pretty far.  Otherwise, when you have specified a custom Policy, you can make method calls on that policy.

When to use custom policy methods

If you have a somewhat strange migration for a specific entity, you can just remove all Attribute mappings in for an Entity Mapping, then in your subclass, override the method below.  The method signature is misleading because you are actually asked to create the equivalent version of the old object in the new object.  So you’re manually specifying how one object “becomes” the new object, and in my case where a Song started to store its data in a separate SongData object, you have to create this SongData object and find a way to link the two (i.e. songFilename) because you create the actual relationships later.  EDIT:  What I’ve learned however is that ideally this createDestinationInstances(…) method should really be about 1:1 mapping.  If your migration ends up taking one object and splitting it into two, you should write to Policy Subclasses and have 2 Entity Mappings:  i.e OldSong -> Song and OldSong -> SongData, then in the 2nd pass of the migration where the relationships are re-established, that’s where you can associate them again.

    override func createDestinationInstances(forSource sInstance: NSManagedObject, in mapping: NSEntityMapping, manager: NSMigrationManager) throws {
        
        let dInstance = NSEntityDescription.insertNewObject(forEntityName: mapping.destinationEntityName!, into: manager.destinationContext)
        
        let text = sInstance.value(forKey: "text") as! String?
        
        dInstance.setValue(sInstance.value(forKey: "lastModified"), forKey: #keyPath(Song.createdAt))
        dInstance.setValue(sInstance.value(forKey: "lastModified"), forKey: #keyPath(Song.updatedAt))
        dInstance.setValue("txt",                                   forKey: #keyPath(Song.fileExtension))
        dInstance.setValue(sInstance.value(forKey: "filename"), forKey: #keyPath(Song.filename))
        dInstance.setValue(sInstance.value(forKey: "firstLetterUppercase"), forKey: #keyPath(Song.firstLetterUppercase))
        dInstance.setValue(sInstance.value(forKey: "name"), forKey: #keyPath(Song.title))
        dInstance.setValue(sInstance.value(forKey: "oldFilename"), forKey: #keyPath(Song.oldFilename))
        dInstance.setValue(self.songDataLengthOfText(text), forKey: #keyPath(Song.songDataLength))
        dInstance.setValue(text, forKey: #keyPath(Song.text))
        dInstance.setValue(SongDataType.plainText.rawValue, forKey: #keyPath(Song.typeRaw))
        dInstance.setValue(nil, forKey: #keyPath(Song.userInfoData))
        
        // NOTE: DON'T DO THIS.  Set up a OldSong -> SongData entity mapping!
        // Then fetch this instance later via its songFilename
        //let songData = NSEntityDescription.insertNewObject(forEntityName: SongData.entity().name!, into: manager.destinationContext)
        //songData.setValue(sInstance.value(forKey: "filename"), forKey: #keyPath(SongData.songFilename))
        //let structure = SongDataStructure()
        //structure.set(text: text)
        //songData.setValue(structure.serialize(), forKey: #keyPath(SongData.nsdata))
        
        // NOTE: DON'T DO THIS.  Set up a OldSong -> SongViewPreferences entity mapping!
        //let viewPreferences = NSEntityDescription.insertNewObject(forEntityName: SongViewPreferences.entity().name!, into: manager.destinationContext)
        //viewPreferences.setValue(sInstance.value(forKey: "filename"), forKey: #keyPath(SongViewPreferences.songFilename))
        
        manager.associate(sourceInstance: sInstance, withDestinationInstance: dInstance, for: mapping)
        return
    }

I will say this:  this method above is used to create instances.  You don’t associate them yet!  That’s why you can see I set songFilename on these 2 new data types (in my new model), so that I can create their relationship later…  this is for the second pass in the migration, where relationship mappings are created (more on that).

Quick Reference for Mapping Editor FUNCTION arguments:

NSMigrationManagerKey: $manager
NSMigrationSourceObjectKey: $source
NSMigrationDestinationObjectKey: $destination
NSMigrationEntityMappingKey: $entityMapping
NSMigrationPropertyMappingKey: $propertyMapping
NSMigrationEntityPolicyKey: $entityPolicy

If I don’t need such a detailed mapping of an object, I can also use custom methods to set a property in the editor for any attribute.  Pro Tip: Make sure you change the the “Source Fetch” on a Relationship mapping to “Use Custom Value Expression” if you do this.

FUNCTION($entityPolicy, "filenameForLibraryWith:" , $manager)

This is syntax in the editor.  It’s saying “to get the value for this destination attribute, call a method on the (custom) policy called

func filename(forLibraryWith manager: NSMigrationManager) -> String

Remember you can use the arguments listed about in the Quick Reference section for knowing what kind of arguments you can pass into these methods!

You can also use these custom methods when performing relationship mappings.  Now, I personally haven’t had to override the method:

func createRelationships(forDestination dInstance: NSManagedObject, in mapping: NSEntityMapping, manager: NSMigrationManager) throws

I’ve got by with the Relationship mappings in the Mapping Model editor and custom methods.

Do not reference subclasses of NSManagedObject during migration!

It would seem that this is how Core Data ensures that the migration can work reliably.  It sucks because you need to remember:

  • awakeToInsert is not called, because no instances of your subclass are used in migration, so whatever you would set in awakeToInsert, you should do here too.
  • The relationship accessors aren’t present either.  So be safe and specify the relationship and its inverse.

I created a handy helper method to find NSManagedObject instances:

class LegacyToModernPolicy: NSEntityMigrationPolicy {
    
    static func find(entityName: String,
                     in context: NSManagedObjectContext,
                     sortDescriptors: [NSSortDescriptor],
                     with predicate: NSPredicate? = nil,
                     limit: Int? = nil) throws -> [NSManagedObject] {
        
        let fetchRequest: NSFetchRequest = NSFetchRequest(entityName: entityName)
        fetchRequest.predicate = predicate
        fetchRequest.sortDescriptors = sortDescriptors
        if let limit = limit {
            fetchRequest.fetchLimit = limit
        }
        
        do {
            let results = try context.fetch(fetchRequest)
            return results
        } catch {
            log.error("Error fetching: \(error.localizedDescription)")
            throw error
        }
    }
  }

Pitfalls, Gotchas, Other Tips, Things I’ve already said

Here’s a list of things I might have mentioned already, but in this modern age where attention spans are low, and we need headings to actually read something, here is a list of things that I discovered in this long journey over 2 days:

    1. If you created your new Data Model and it started with the same name as your legacy one (i.e. SongbookSimple), it’s better to just create a new Core Data Model file with a new name (e.g. SongbookSimpleDataModel), then rebuild your data model by hand. This in practice means copy-pasting all the entities from the one to the other. Note! The Delete rules are not copied over, NOR are the inverse relationships. So you’ll have to go in and for each Entity, connect these up again.
    2. If you get an error during Migration such as:
      Couldn’t create mapping policy for class named (LegacyToModernArtistPolicy)
      It’s probably because you have Swift module-related errors. See Sandeep’s comment in this Question:
      So LegacyToModernArtistPolicy becomes SongbookSimple.LegacyToModernArtistPolicy and it works
    3. If you change your current model to accommodate the migration process, AND you’ve already run the app, you’ve altered the data model in a bad way. You don’t know if this is a typical state to be in, so better play it safe. You’ve got to remove the app from the simulator, re-install an old version so to populate an old data store, then run your new version in development.
    4. If you are doing custom fetch requests so to satisfy creating relationships (in your NSEntityMigrationPolicy), it would seem that fetch requests are returning NSManagedObject objects and not the subclasses. Nor can they be casted. For example, you might have seen a crash error:Could not cast value of type 'NSManagedObject_Library_' (0x6100000504d0) to 'SongbookSimple.Library' (0x101679180).This is why the helper above can be useful.  I fixed the bug I mention here with:
          func libraryForManager(_ manager: NSMigrationManager) -> NSManagedObject {
              
              do {
                  var library: NSManagedObject? = try LegacyToModernPolicy.find(entityName: Library.entity().name!,
                                                          in: manager.destinationContext,
                                                          sortDescriptors: [NSSortDescriptor(key: "filename", ascending: true)],
                                                          with: nil,
                                                          limit: 1).first
                  if library == nil {
                      let dInstance = NSEntityDescription.insertNewObject(forEntityName: Library.entity().name!, into: manager.destinationContext)
                      
                      // awakeFromInsert is not called, so I have to do the things I did there, here:
                      dInstance.setValue(Library.libraryFilename, forKey: #keyPath(Library.filename))
                      dInstance.setValue(NSDate(timeIntervalSince1970: 0), forKey: #keyPath(Library.updatedAt))
                      library = dInstance
                  }
                  
                  return library!
                  
              } catch {
                  fatalError("Not sure why this is failing!")
              }
          }
      
    5. You think you are far along because all of your NSEntityMigrationPolicy objects are having their custom methods called, nothing is failing, but at the end of the migration process you’re still getting errors. SpecificallyCocoa error 1570. This implies that data validations are failing.It could easily be that it’s because you made a false assumption that awakeFromInsert will get called on your new instances during the migration process. For example you added a non-optional attribute to your model who gets his default value set in awakeFromInsert. You should specifically make sure these are getting set either via the mapping editor and your custom policy (with a custom method), or try to handle the errors in performCustomValidationForEntityMapping:manager:error:I prefer graphic tools. So I just write some boring methods that will set default values and it’s all good.
    6. IF YOU CHANGE YOUR MODEL, THEN THE MAPPING FILE WILL BREAK AND NOT BE FOUND DURING MIGRATION.IF YOU CHANGE YOUR MODEL, THEN THE MAPPING FILE WILL BREAK AND NOT BE FOUND DURING MIGRATION.This means you have to rebuild the Mapping File.
    7. In all of the NSEntityMigrationPolicy subclass, make sure you are always dealing with NSManagedObject and not their subclasses!! If you don’t, your automatic mappings (that use custom methods on the policy subclass) will fail.
    8. You don’t HAVE to create mappings in the Mapping Model editor. Basically, the part that has attribute mappings is a way of NOT overriding the method
      func createDestinationInstances(forSource sInstance: NSManagedObject, in mapping: NSEntityMapping, manager: NSMigrationManager) throws

      and the Relationship Mappings editor is a way of not having to subclass

      func createRelationships(forDestination dInstance: NSManagedObject, in mapping: NSEntityMapping, manager: NSMigrationManager) throws

      ONE THING you DO need to understand is that during the migration process, you are dealing with NSManagedObject instances, and not instances of your (likely) subclasses. So you don’t get all the benefits of the accessor methods that would ensure object graph integrity. What does this mean? You should make sure that you connect both sides of a relationship via relationship mappings. This means, in your data model it makes sense to keep a field to a ‘primary key’, so that you can use one object to find another.

    9. YAY! The Migration manager says it migrated… except… I got an exception:
      'Mismatch between mapping and source/destination models'

      I have to say, Core Data documentation sucks. There’s just not enough of it, and no explanation of why it fails during migration. Totally crap. I’ve spent a whole day just trying to shoot into the dark, hoping I’ll figure out what’s going on with this giant black box.

      The fix for this occurs in my modified MHWMigrationManager gist, and it’s what sorted me out!

Still Missing from everything that I learned:

  • I don’t really know how one is supposed to work with the createRelationships(…) method.  I only do that aspect of migration via the Mapping Model Editor and custom methods on the Policy
  • I haven’t done anything with validation yet.  So I don’t know how that comes into play.

Wrapping Up

This was a super long post.  But hopefully it can be seen as a good reference guide or at least things to think about when trying to do heavyweight migrations.  Now that I understand how they’re done, they’re not AS scary anymore.  What was annoying was trying to figure this all out.

If you benefit from this, I would be happy if you let me know!

Advertisements

Don’t learn your Localization lessons the hard way

There’s a great word in German.  It’s called “Arscharbeit”.  It literally means “ass work”.  But what it really means is a task that basically just sucks.  Is brute force, requires very little thinking, and is just boring and tedious.

Today and tomorrow I have to do some “Arscharbeit” because we didn’t plan localization well enough and now want to re-organize a lot of the way it is done.

We chose a wishy-washy, not clearly defined key-naming system.  In the end it wasn’t quite clear what context the key was defining, so we just had a mash-up of weird key names.  The task now is to rename all of these, put certain keys into specific tables so that re-use will be easier.  And I know that in the end the entire app will still require a complete visual inspection because of values for keys gone missing. 

I won’t go into the details of the refactoring job because it got messy.  I will provide you with two links and just say “Don’t waste time and money.  Read these and do something about them AT THE BEGINNING OF YOUR PROJECT.”  Fixing later is annoying, boring, and expensive. 

http://www.objc.io/issue-9/string-localization.html

http://www.innovaptor.com/blog/2013/02/07/a-localization-workflow-that-works-well-in-practice.html