[potatolicious]

I work on a fairly complex Core Data implementation in a heavily multi-threaded app.

It requires a lot of discipline, and we rely on a lot of learning-through-pain since the documentation beyond the most basic implementation is mostly non-existent.

Core Data is pretty simple if you do all of your work on the main thread, but then your'e also blocking your main execution loop for operations that can take a long time. A large save will easily reach into >200ms land on an iOS device, and your app will feel like crap.

Thread safety in NSManagedObjectContexts is somewhat documented, but has a lot of quirkiness to it. There is some general knowledge around what is and isn't thread-safe, but almost nothing surrounding how to build a multithreaded app and not be murdered by Core Data. Much of this is trial and error and observing what open source projects are doing.

Once you settle into a decent threading scheme that won't kill you (probably with multiple NSManagedObjectContexts) you have to worry about communicating changes across multiple contexts - then you delve into parent-child contexts, where the document gets even more bare. Then you get into propagation delays when crossing context boundaries.

Keeping your code clean through all of this is also a challenge. There are also runtime issues - our implementation in Core Data experiences unbounded growth in memory usage on iOS5, but not in iOS6, forcing us to manually mitigate memory use (which Core Data is supposed to take care of), creating even more thread-safety concerns.

All in all, it's a pain, but it's also the only realistic object store/graph on the platform.

[gurkendoktor]

> A large save will easily reach into >200ms land on an iOS device, and your app will feel like crap.

Can you give an order of magnitude at this would happen? I usually only have 50-200 entities in my database in total, saving maybe half of them at the same time. Yet libraries like RestKit insist on a two-context setup to avoid the saving delay you mention. I'm not sure if I should fight the libraries and go back to a single-threaded setup (that I understand) or just go with the parent/child context flow.

[delackner]

Fetch and save operations get pretty slow with even just 10,000 entities (if you touch all those entities). It gets a little better if you pre-fetch all the entities rather than one at a time, but the save is still very slow. Breaking that out into a backround thread is feasible, but as the previous commenter mentions, then you risk momentary divergence between the UI and the backend.

[seivan]

Index properly. Save in a queue.

[delackner]

This is fetching every entity by its indexed key. My understanding of saving in a background queue is that during that 1-2 seconds, the ui will still be using the pre-modification data set, so very briefly the ui may appear incorrect, no? Not a huge deal, but if the ui will be wrong, then for our app that would still force us to pause the ui until the save finishes.

[jlmendezbonini]

>Much of this is trial and error and observing what open source projects are doing.

Could you name some of those open source projects? Thanks.