The fact that this only includes column names is an oversight.

There are some minor changes to the point type as I had forgotten that
this will affect the behavior of `t.point` in migrations and the schema
dumper so we need to handle those as well.

I'll say this again so I can convince myself to come up with a better
structure... TYPES SHOULD NOT CARE ABOUT SCHEMA DUMPING AND WE NEED TO
BETTER SEPARATE THESE.

This reverts commit 3a1f6fe7.

This commit takes the code in a direction that I am looking to avoid.
The predicate builder should be purely concerned with AST construction
as it matters to methods like `where`. Things like case sensitivity
should continue to be handled elsewhere.

I still think that this is something that should be handled in the pg
gem, but it's not going to end up happening there so we'll do it here
instead. Once we bump to pg 0.19 we can pass the encoding to the
`encode` method instead.

This issue occurs because C has no concept of encoding (or strings,
really). The bytes that we pass here when sending the value to the
database will always be interpreted as whatever encoding the connection
is currently configured to use. That means that roundtripping to the
database will lose no information

However, after assigning we round trip through our type system without
hitting the database. The only way that we can do the "correct" thin
here would be to actually give a reference to the connection to the
array type and have it check the current value of the connection's
encoding -- which I'm strongly opposed to. We could also pass in the
encoding when it's constructed, but since that can change independently
of the type I'm not a huge fan of that either.

This feels like a reasonable middle ground, where if we have an array of
strings we simply use the encoding of the string we're given.

Fixes #26326.

This was almost every case where we are overriding `respond_to?` in a
way that mirrors a parallel implementation of `method_missing`. There is
one remaining case in Active Model that should probably do the same
thing, but had a sufficiently strange implementation that I want to
investigate it separately.

Fixes #26333.

If a parent association was accessed in an `after_find` or
`after_initialize` callback, it would always end up loading the
association, and then immediately overwriting the association we just
loaded. If this occurred in a way that the parent's `current_scope` was
set to eager load the child, this would result in an infinite loop and
eventually overflow the stack.

For records that are created with `.new`, we have a mechanism to
perform an action before the callbacks are run. I've introduced the same
code path for records created with `instantiate`, and updated all code
which sets inverse instances on newly loaded associations to use this
block instead.

Fixes #26320.

Rails dropped Ruby 1.9 support, but this comment still true.

Follow up to #26301.

The performance difference between `Hash[]` and `Hash.dup` looks to have
been narrowed by @tenderlove via this commit --> https://github.com/ruby/ruby/commit/b3803cc49ad382e23291d75ce57ffb2b74bb9577#diff-eff9999082c8ce7d8ba1fc1d79f439cf.
Since this commit first appeared in Ruby 2.0.0, and since Rails now
requires a minimum Ruby version of 2.2.2, this performance boost should
be available for all users.

Relevant links:

- This behavior was originally added via https://github.com/rails/rails/commit/02174a3efc6fa8f2e5e6f114e4cf0d8a06305b6a
- The conversation on the Ruby issue tracker lives here --> https://bugs.ruby-lang.org/issues/7166

Previously, if the the association was previously loaded and then
the foreign key changed by itself, a #save call would trigger a
load of the new associated record during autosave. This is unnecessary
and the autosave code (in that case) didn't use the loaded record
anyways.

For simplicity.

For reduce instantiating `Type::Value`.

Introduced low level methods #set_value and #get_value for setting query attributes:

  relation.set_value(:where, {id: 1})
  relation.get_value(:includes)

Used those internally when working with relation's attributes
at the abstract level

`format_string` is used for standardized column types/arguments spaces.
Now the standardization was removed at df84e986.

Now the schema dumper by default doesn't align the types and arguments
in the ruby format anymore.

This is to close #26111

Developers need to be aware that `dependent: :destroy` on a scoped
`has_many` association would only destroy the associated objects in
that scope. Potentially leaving other associated objects outside that
scope untouched.

Fixes #26137.

Simply use its own methods because `CollectionProxy` inherits `Relation`.

`length` is delegated to `records` (`load_target`) by
`ActiveRecord::Delegation`.

https://github.com/rails/rails/blob/v5.0.0/activerecord/lib/active_record/relation/delegation.rb#L38

This reverts #23067. #23067 is for propagating `pk` value from
`sql_for_insert` to `exec_insert` (avoiding extra query for pg adapter).
Now `exec_insert` includes `sql_for_insert` since #26002 therefore this
propagating is no longer needed.

Strict mode controls how MySQL handles invalid or missing values in
data-change statements such as INSERT or UPDATE. If strict mode is not
in effect, MySQL inserts adjusted values for invalid or missing values
and produces warnings.

```ruby
  def test_mysql_not_null_defaults_non_strict
    using_strict(false) do
      with_mysql_not_null_table do |klass|
        record = klass.new
        assert_nil record.non_null_integer
        assert_nil record.non_null_string
        assert_nil record.non_null_text
        assert_nil record.non_null_blob

        record.save!
        record.reload

        assert_equal 0,  record.non_null_integer
        assert_equal "", record.non_null_string
        assert_equal "", record.non_null_text
        assert_equal "", record.non_null_blob
      end
    end
  end
```

It is inconsistent with other types that only text/blob defaults treated
as an empty string. This commit fixes the inconsistency.

invalid column type
    /activerecord/lib/active_record/connection_adapters
    /abstract/schema_definitions.rb:306
    type = type.to_sym

Changed to the following to handle nil case:
    type = type.to_sym if type

Added regression test for this case:
  /activerecord/test/cases/migration_test.rb:554
  if current_adapter?(:SQLite3Adapter)
    def test_allows_sqlite3_rollback_on_invalid_column_type
      Person.connection.create_table :something, force: true do |t|
        t.column :number, :integer
        t.column :name, :string
        t.column :foo, :bar
      end
      assert Person.connection.column_exists?(:something, :foo)
      assert_nothing_raised { Person.connection.remove_column :something, :foo, :bar }
      assert !Person.connection.column_exists?(:something, :foo)
      assert Person.connection.column_exists?(:something, :name)
      assert Person.connection.column_exists?(:something, :number)
    ensure
      Person.connection.drop_table :something, if_exists: true
    end
  end

Currently `CollectionProxy` inherits `Relation` therefore we can use
its own methods rather than delegating to collection association.

Currently `CollectionProxy` inherits `Relation` and `Relation` includes
`QueryMethods`. This method is completely duplicated.

https://github.com/rails/rails/blob/v5.0.0/activerecord/lib/active_record/relation/query_methods.rb#L271-L275

Because `sql_for_insert` is only called in `use_insert_returning?` is
true since #26002.

Because `sanitize_conditions` protected method is only used in one place.

Otherwise CollectionProxy's bang methdos cannot respect dirty target.

`#second`, `#third`, etc finder methods was added in 03855e79.
But the signature of these methods is inconsistent with the original
finder methods. And also the signature of `#first` and `#last` methods
is different from the original. This commit fixes the inconsistency.

The current behaviour of checking if there is a LEFT OUTER JOIN arel
node to detect if we are doing eager_loading is wrong. This problem
wasn't frequent before as only some pretty specific cases would add
a LEFT OUTER JOIN arel node. However, the recent new feature
left_outer_joins also add this node and made this problem happen
frequently.

Since in the perform_calculation function, we don't have access to
eager_loading information, I had to extract the logic for the distinct
out to the calculate method.

As I was in the file for left_outer_join tests, I fixed a few that had
bugs and I replaced some that were really weak with something that
will catch more issues.

In relation tests, the first test I changed would have failed if it
had validated the hash returned by count instead of just checking how
many pairs were in it. This is because this merge of join currently
transforms the join node into an outer join node, which then made
count do a distinct. So before this change, the return was
{1=>1, 4=>1, 5=>1}.

Currently uniqueness validator is coupled with building Arel ASTs.
This commit extracts `PredicateBuilder::CaseSensitiveHandler` for
decouple the building Arel ASTs.

If handled as an associated predicate even though a table has the
column, will generate invalid SQL by valid column name treated as a
table name.

Style/SpaceBeforeBlockBraces
Style/SpaceInsideBlockBraces
Style/SpaceInsideHashLiteralBraces

Fix all violations in the repository.