primary_keys(table) needs to query various metadata tables in Postgres to
determine the primary key for the table. Previously, it did so using a
complex common table expression against pg_constraint and pg_attribute.

This patch simplifies the query by joining pg_index against pg_attribute
instead of going through pg_constraint. This avoids an expensive unnest,
window function query, and common table expression.

EXPLAINing these queries in Postgres against a database with a single
table with a composite primary key shows a 66% reduction in the plan and
execute latencies. This is significant during application startup time,
especially against very large schemas, where these queries would be even
slower and more numerous.

Closes #27949

Column options are passed as an hash args then used as `options` hash in
`add_column_options!`. Converting args to attributes is inconvinient for
using options as an hash.

Follow up to #26266.

The default type of `primary_key` and `references` were changed to
`bigint` since #26266. But legacy migration and sqlite3 adapter should
keep its previous behavior.

`ColumnDumper#migration_keys` was extracted to customize keys for
standardized column arguments widths. But the feature was removed in
df84e986. The internal method is no longer used for that.

The PR #27384 changed migration compatibility behaviour.

```ruby
class CreateMasterData < ActiveRecord::Migration[5.0]
  def change
    create_table :master_data, id: :integer do |t|
      t.string :name
    end
  end
end

```

Previously this migration created non-autoincremental primary key
expected. But after the PR, the primary key changed to autoincremental,
it is unexpected.

This change restores the behaviour of the compatibility layer.

The PR #27384 changed integer-like primary key to be autoincrement
unless an explicit default. This means that integer-like primary key is
restored as autoincrement unless dumping the default nil explicitly.
We should dump integer-like primary key with default nil correctly.

MySQL generated columns: https://dev.mysql.com/doc/refman/5.7/en/create-table-generated-columns.html
MariaDB virtual columns: https://mariadb.com/kb/en/mariadb/virtual-computed-columns/

Declare virtual columns with `t.virtual name, type: …, as: "expression"`.
Pass `stored: true` to persist the generated value (false by default).

Example:

    create_table :generated_columns do |t|
      t.string  :name
      t.virtual :upper_name, type: :string,  as: "UPPER(name)"
      t.virtual :name_length, type: :integer, as: "LENGTH(name)", stored: true
      t.index :name_length  # May be indexed, too!
    end

Closes #22589

column_definitions() needs to fetch the collation for every column, if
present. Previously, it did so using a correlated subquery - a subquery
that references results from the outer scope.

This patch updates the query to remove the subquery in favor of a
simpler and more efficient JOIN clause.

Running the two queries through EXPLAIN against Postgres additionally
shows that the original form with a correlated subquery requires a
Nested Loop Left Join, while the new form with a simple JOIN can use a
more efficient Merge Left Join.

this probably fixes #25840

One empty string allocation per column per row.

These methods are obviously for internal use.

These internal initialize methods are no longer used internally.

This reverts commit 39c77eb1, reversing
changes made to 9f6f51be.

* Don't force PKs on tables that have explicitly opted out
* All integer-like PKs are autoincrement unless they have an explicit
  default

Remove `initialize_schema_migrations_table` and `initialize_internal_metadata_table` internal public methods

These internal methods accidentally appeared in the doc, and so almost
useless. It is enough to create these internal tables directly, and
indeed do so in several places.

https://github.com/rails/rails/blob/v5.0.1/activerecord/lib/active_record/schema.rb#L55
https://github.com/rails/rails/blob/v5.0.1/activerecord/lib/active_record/railties/databases.rake#L6
https://github.com/rails/rails/blob/v5.0.1/activerecord/lib/active_record/tasks/database_tasks.rb#L230

https://www.sqlite.org/foreignkeys.html

Raise `ActiveRecord::InvalidForeignKey` when a record cannot be inserted
or updated because it references a non-existent record for SQLite3
adapter.

If a process is forked more than once, the pool was grabbing the oldest
spec, not the most recent spec. This wasn't noticed before because most
folks are lilely forking the process only once.

If you're forking the process multiple times however the wrong spec name
will be returned and an incorrect connection will be used for the
process.

This fixes the issue by reversing the list of spec names so we can grab
the most recent spec rather than the oldest spec.

because Struct.new returns a Class, we just can give it a name and use it directly without inheriting from it

(I personally prefer writing one string in one line no matter how long it is, though)

Need `table_name` to avoid all databases scan.
See https://dev.mysql.com/doc/refman/5.7/en/information-schema-optimization.html.

```
> EXPLAIN SELECT fk.referenced_table_name AS 'to_table', fk.referenced_column_name AS 'primary_key', fk.column_name AS 'column', fk.constraint_name AS 'name', rc.update_rule AS 'on_update', rc.delete_rule AS 'on_delete' FROM information_schema.key_column_usage fk JOIN information_schema.referential_constraints rc USING (constraint_schema, constraint_name) WHERE fk.referenced_column_name IS NOT NULL AND fk.table_schema = 'activerecord_unittest' AND fk.table_name = 'fk_test_has_pk' AND rc.table_name = 'fk_test_has_pk'\G
*************************** 1. row ***************************
           id: 1
  select_type: SIMPLE
        table: fk
   partitions: NULL
         type: ALL
possible_keys: NULL
          key: TABLE_SCHEMA,TABLE_NAME
      key_len: NULL
          ref: NULL
         rows: NULL
     filtered: NULL
        Extra: Using where; Open_full_table; Scanned 0 databases
*************************** 2. row ***************************
           id: 1
  select_type: SIMPLE
        table: rc
   partitions: NULL
         type: ALL
possible_keys: NULL
          key: TABLE_NAME
      key_len: NULL
          ref: NULL
         rows: NULL
     filtered: NULL
        Extra: Using where; Open_full_table; Scanned 1 database; Using join buffer (Block Nested Loop)
2 rows in set, 1 warning (0.00 sec)
```

Fixes #27579.

Revert "Merge pull request #21233 from mtsmfm/disable-referential-integrity-without-superuser-privileges"

This reverts commit eeac6151, reversing
changes made to 5c40239d.

Reason: Broke the isolated tests.
https://travis-ci.org/rails/rails/builds/188721346

Passing `name` to `tables` is already deprecated at #21601.
Passing `name` to `indexes` is also unused.

Related #22973, #24708.

`select_all`, `select_one`, `select_value`, and `select_values` method
signature is `(arel, name = nil, binds = [])`.
But `select_rows` is `(sql, name = nil, binds = [])`.

Per the regression commit below, the commit changes the behavior of
`#changed?`to consult the `#changed_in_place?` method on `Type::Value` classes.
Per this change, `PostgreSQL::OID::Hstore` needs to override this method
in order to compare the deserialized forms of the two arguments. In
Ruby, two hashes are considered equal even if their key order is
different. This commit helps to bring that behavior to `Hstore` values.

Fixes regression introduced by 8e633e50

Fixes #27502

In f1a0fa9e we moved backend specific timestamp behavior out of the type
and into the adapter. This was in line with our general attempt to
reduce the number of adapter specific type subclasses. However, on PG,
the array type performs all serialization, including database encoding
in its serialize method.

This means that we have converted the value into a string before
reaching the database, so no adapter specific logic can be applied (and
this also means that timestamp arrays were using the default `.to_s`
method on the given object, which likely meant timestamps were being
ignored in certain cases as well)

Ultimately I want to do a more in depth refactoring which separates
database serializer objects from the active model type objects, to give
us a less awkward API for introducing the attributes API onto Active
Model.

However, in the short term, we follow the solution we've applied
elsewhere for this. Move behavior off of the type and into the adapter,
and use a data object to allow the type to communicate information up
the stack.

Fixes #27514.

As pointed out by @matthewd this change makes ImmutableString aware
of MysqlString's existence whereas previously MysqlString was only
overriding public API.

cc @kamipo

This reverts commit e632c2fa, reversing
changes made to 334a7dcf.

The only difference between `Type::ImmutableString` and its subclasses
is the representation of the casted booleans. Prefer extracting
`casted_true`/`casted_false` and override these by subclasses.

Currently schema dumper does not dump array subtype `precision` and
`scale` options. This commit fixes the issue.

The visibility of all internal protected methods was changed to private
since 5b14129d.

Some methods were added to public API in
5b14129d and they should be not part of
the public API.

Currently uniqueness validator is coupled with building Arel ASTs.
This commit extracts `WhereClauseFactory#build_for_case_sensitive` for
decouple the building Arel ASTs.