This bug occurs when an attribute of an ActiveRecord model is an
ActiveRecord::Type::Integer type or a ActiveRecord::Type::Decimal type (or any
other type that includes the ActiveRecord::Type::Numeric module. When the value
of the attribute is negative and is set to the same negative value, it is marked
as changed.

Take the following example of a Person model with the integer attribute age:

    class Person < ActiveRecord::Base
      # age          :integer(4)
    end

The following will produce the error:

    person = Person.new(age: -1)
    person.age = -1
    person.changes
    => { "age" => [-1, -1] }
    person.age_changed?
    => true

The problematic line is here:

    module ActiveRecord
      module Type
        module Numeric
          ...

          def non_numeric_string?(value)
            # 'wibble'.to_i will give zero, we want to make sure
            # that we aren't marking int zero to string zero as
            # changed.
            value.to_s !~ /\A\d+\.?\d*\z/
          end
        end
      end
    end

The regex match doesn't accept numbers with a leading '-'.

`ActiveRecord::Base#[]` has overhead that was introduced in 4.2. The
`foo["id"]` working with PKs other than ID isn't really a case that we
want to support publicly, but deprecating was painful enough that we
avoid it. `_read_attribute` was introduced as the faster alternative for
use internally. By using that, we can save a lot of overhead. We also
save some overhead by reading the attribute one fewer times in
`stale_state`.

Fixes #18151

If there is a method defined such as `find_and_do_stuff(id)`, which then
gets called on an association, we will perform statement caching and the
parent ID will not change on subsequent calls.

Fixes #18117

Calling `changed_attributes` will ultimately check if every mutable
attribute has changed in place. Since this gets called whenever an
attribute is assigned, it's extremely slow. Instead, we can avoid this
calculation until we actually need it.

Fixes #18029

PG will warn without it, but mysql2 errors out.

Changes `rails g model Post user:references` from

    def change
      create_table :posts do |t|
        t.references :user, index: true
      end

      add_foreign_key :posts, :users
    end

to

    def change
      create_table :posts do |t|
        t.references :user, index: true, foreign_key: true
      end
    end

Changes `rails g migration add_user_to_posts user:references` from

    def change
      add_reference :posts, :users, index: true
      add_foreign_key :posts, :users
    end

to

    def change
      add_reference :posts, :users, index: true, foreign_key: true
    end

This has the same comments as 9af90ffa00ba35bdee888e3e1ab775ba0bdbe72c,
however it affects the `add_reference` method, and `t.references` in the
context of a `change_table` block.

There is a lot of duplication of code between creating and updating
tables. We should re-evaluate the structure of this code from a high
level so changes like this don't need to be made in two places. (Note to
self)

While we still aren't accepting PRs that only make changes like this,
it's fine when we're actively working on a method if it makes our lives
easier.

Rather than having to do:

    create_table :posts do |t|
      t.references :user
    end

    add_foreign_key :posts, :users

You can instead do:

    create_table :posts do |t|
      t.references :user, foreign_key: true
    end

Similar to the `index` option, you can also pass a hash. This will be
passed as the options to `add_foreign_key`. e.g.:

    create_table :posts do |t|
      t.references :user, foreign_key: { primary_key: :other_id }
    end

is equivalent to

    create_table :posts do |t|
      t.references :user
    end

    add_foreign_key :posts, :users, primary_key: :other_id

While we aren't taking PRs with these kinds of changes just yet, they
are fine if we're actively working on the method and it makes things
easier.

If the test is interrupted in a way that the teardown block fails to
run, the tests will fail to run until the table is removed manually
without this option.

PG doesn't register it's types using the `int(4)` format that others do.
As such, if we alias `int8` to the other integer types, the range
information is lost. This is fixed by simply registering it separately.

The other option (which I specifically chose to avoid) is to pass the
information of the original type that was being aliased as an argument.
I'd rather avoid that, since an alias should truly be treated the same.
If we need different behavior for a different type, we should explicitly
register it with that, and not have a conditional based on aliasing.

Fixes #18144

[Sean Griffin & ysbaddaden]

Fixes #18122

1. Specify that you need to create the test databases, and that no special
   Rails command needs to be run to do that.
2. Although the underscore style of `rake test_mysql` works, make the
   documentation of running the tests in RUNNING_UNIT_TESTS.rdoc
   consistent with the "Contributing..." guide.
3. Promote "Testing Active Record" to not be a subsection of
   "Running a Single Test," since it doesn't make sense as a subsection
   of that.

Add nodoc to some constants [skip ci]

I find that `Specifies the starting point for the batch processing.`
does not give enough information for me to understand what this
parameter actually does.

Closes #17945

`db:test:prepare` still purges the database to always keep the test
database in a consistent state.

This patch introduces new problems with `db:schema:load`. Prior
to the introduction of foreign-keys, we could run this file against
a non-empty database. Since every `create_table` containted the
`force: true` option, this would recreate tables when loading the schema.

However with foreign-keys in place, `force: true` wont work anymore and
the task will crash.

/cc @schneems

Apparently PG does not validate against RFC 4122. The intent of the original
patch is just to protect against PG errors (which potentially breaks txns, etc)
because of bad user input, so we shouldn't try any harder than PG itself.

Closes #17931

In the case of serialized columns, we would expect the unserialized
value as input, not the serialized value. The original issue which made
this distinction, #14163, introduced a bug. If you passed serialized
input to the method, it would double serialize when it was sent to the
database. You would see the wrong input upon reloading, or get an error
if you had a specific type on the serialized column.

To put it another way, `update_column` is a special case of
`update_all`, which would take `['a']` and not `['a'].to_yaml`, but you
would not pass data from `params` to it.

Fixes #18037

Because we're only using the `connection` so passing the entire tracker
isn't unnecessary.

Eventually only the `connection` will be passed to `add_constraints`
with later refactoring but curretly that's not possible because of
`construct_tables` method.

https://github.com/rails/rails/commit/39542fba54328ca048fb75a5d5b37f8e1d4c1f37#commitcomment-8938379

Move microseconds formatting to `AbstractAdapter`.

The test added in 42418cfc wasn't
actually testing anything, since the bug was with TZ aware attributes
only.

PostgreSQL for example, allows infinity as a valid value for date time
columns. The PG type has explicit handling for that case. However, time
zone conversion will end up trampling that handling. Unfortunately, we
can't call super and then convert time zones.

However, if we get back nil from `.in_time_zone`, it's something we
didn't expect so we can let the superclass handle it.

Fixes #17971

The `RecursiveCallbackDeveloper` and `ImmutableMethodDeveloper` classes
are not used anymore in tests, and neither is the `@Cancelled` variable.

To be possible to use a custom column name to save/read the polymorphic
associated type in a has_many or has_one polymorphic association, now users
can use the option :foreign_type to inform in what column the associated object
type will be saved.

`freeze` will ultimately end up freezing the `AttributeSet`, which in
turn freezes its `@attributes` hash. However, we actually insert a
special object to lazily instantiate the values of the hash on demand.
When it does need to actually instantiate all of them for iteration (the
only case is `ActiveRecord::Base#attributes`, which calls
`AttributeSet#to_h`), it will set an instance variable as a performance
optimization

Since it's just an optimization for subsequent calls, and that method
being called at all is a very uncommon case, we can just leave the ivar
alone if we're frozen, as opposed to coming up with some overly
complicated mechanism for freezing which allows us to continue to modify
ourselves.

Fixes #17960

The user is able to pass PG string literals in 4.1, and have it
converted to an array. This is also possible in 4.2, but it would remain
in string form until saving and reloading, which breaks our
`attr = save.reload.attr` contract. I think we should deprecate this in
5.0, and only allow array input from user sources. However, this
currently constitutes a breaking change to public API that did not go
through a deprecation cycle.

The validation added in 5a3dc809 will
reject values for the `:on` option for after_commit and after_rollback
callbacks that are string values like `"create"`.

However, the error message says ":on conditions for after_commit and
after_rollback callbacks have to be one of create,destroy,update". That
looks like a string value *would* be valid.

This commit changes the error message to say ":on conditions for
after_commit and after_rollback callbacks have to be one of [:create,
:destroy, :update]", making it clearer that symbols are required.