Tests failing on Rails 5.2.0.rc1

[shioyama]

They were passing on beta, but some seem to be failing now on rc1. Need to investigate.

[shioyama]

If anyone is reading, the issues are all about the dirty plugin, as usual. Nothing else seems to be affected.

[shioyama]

For reference, I think the commit that broke things was rails/rails@a19e91f0fab

[shioyama]

Arg. This time Rails has really made a muck of things for Mobility, or any gem that wants to have pseudo-attributes on an AR model.

I don't quite know how to get around these changes without really digging deep into AR/AM internals, which seems like a bad idea. The basic problem is that Mobility tracks changes to localized attributes like title_en, but now with these changes AR/AM will not include these attributes in changes, so it looks like they haven't changed.

For now, dirty tracking in AR 5.2 is broken. I wish I could just hack out a fix but this one will take more time I'm afraid. I wonder if other gem maintainers may also be having an issue with this... I think it's a bad direction for Rails, but I confess that I don't know the internals well enough to make a detailed criticism.

[shioyama]

The key issue is that in ActiveModel, the mutation tracker is initialized with a set of attributes where the default falls through to a hash with an automatic default value:

https://github.com/rails/rails/blob/6a97a17f195a925959866edda2e951e20d7b1e76/activemodel/lib/active_model/dirty.rb#L252-L256

@attributes = AttributeSet.new(
  Hash.new { |h, attr|
    h[attr] = Attribute.with_cast_value(attr, _clone_attribute(attr), Type.default_value)
  }
)

For AR models, the attributes hash does not have a default value like this.

I think the best way to tackle this is probably to somehow change the mutation tracker on AR models to be aware of virtual translated attributes. I've tried actually modifying the attributes hash itself but this seems much more dangerous. Maybe we could subclass the mutation tracker and swap it out or something... no really good options here.

I hope we don't have to give up on the dirty plugin for AR models, but AR is certainly not making it easy to build this type of functionality.

[shioyama]

For now in 27585cc I've locked tests to 5.2.0.beta2, which passes. They fail in rc1.

[shioyama]

Wow good news: the main change that broke things in AR rc1 has been reverted, since it "makes assumptions that I do not want in the code base" according to @sgrif. I'm glad since I agree it did seem to make some quite sweeping assumptions.

That said, a smaller number of specs (I see six with postgres) are still failing, so will need to pinpoint why. In #166 I've changed the Gemfile to look at the master branch until the next release candidate is available.

[shioyama]

Fixed in #166, and I'll release 0.5.1 with the fix now.