What Does Rails Do With Both :dependent => :destroy and cascade delete/nullify/restrict

Question

I'm trying to decide how best to set up (if at all) foreign key constraints for my rails application. I have a model Response that belongs_to a Prompt. I would like to use :dependent => :destroy to have destroy called on every Response that belongs to a deleted Prompt and I'm trying to decide what delete constraint I should place on my foreign key.

In short I want advice about how I can get best take advantage of both the destroy method on dependent objects and foreign key constraints to ensure cruft doesn't accumulate and reflect the logical structure of the data being stored. Several earlier questions such as Should I use ON DELETE CASCADE, :dependent => :destroy, or both? and Rails: delete cascade vs dependent destroy asked which was better but they don't really say much about how the two choices interact and in what order they are triggered or seemed vague on the point.

As I see it the considerations seem to break up into a few pieces:

1. Does :dependent => :destroy call destroy first on the dependent objects before removing the parent from the database so destroy will still be called on these objects even if I use cascade delete?

2. Does :dependent => :destroy remove the dependent objects from the database before (or in a transaction with) removing the parent from the database? In other words if I set cascade to nullify will the database end up wastefully nullifying the references on the child objects before they are deleted?

3. Are deletes issued as a result of the original destroy and chained :dependent => :destroy options wrapped in a transaction or will unfortunately timed crashes leave cruft in the database if I don't set cascade delete?

4. Finally will :dependent => :destroy ensure the parent object is deleted from the database if I use restrict as the foreign key on_delete option?

Unfaithful

Answer 1

With dependent: :destroy in a transaction rails first destroys all dependencies, and only then deletes the record itself.

There may be a race condition: if a dependent record was added just after rails read collection for destroying, but not deleted parent yet - it may be left over. Let's call these "race condition records" below.

1. yes, you can use dependent: :destroy and on delete cascade, this way some children (race condition ones) can be deleted without callbacks. If callbacks are mandatory - on delete restrict together with some locking and explicit children deletion may be better. This is somewhat like validates :some_field, uniqueness: true that is better to be backed by unique index, only database itself can ensure data consistency.

2. since parent is deleted last, on delete nullify will not get in the way (you'll get nullified race condition records)

3. there's transaction wrapping all deletes, only race condition records can be left over

4. on delete restrict over dependent: :destroy will trigger only for race condition records (and roll back whole transaction), but if there was no race condition - rails will happily delete everything.

Denier

Thanks so much. But I'm a little confused by why there should be any race condition records. I thought the idea was that if you use transactions then you can linearly order all interactions with the database as if they took place in a certain order with all the commands in the transaction executed consecutively. So if properly transacted why doesn't this prevent the new dependent records from being added? Or is refetching the id of the depended on obj avoided on creation for perf reasons? – Unfaithful

@PeterGerdes most DBs have default transaction isolation level at "read committed" (for performance reasons), so "non-repeatable read" is allowed. Also rails logic lives outside of db. Example: transaction A adds a dependent record R, but is not commited. Transaction B is the parent destroy transaction, in dependent records it will not read R because A is not commited. Then A finally commits, foreign key is not violated because parent record still exists (B is not commited and also might have not deleted it yet), so R is written. B can re-read items and have R in the set, but it's too late – Denier