Example:

    create_table :users do |t|
      t.string :name
      t.index 'lower(name) varchar_pattern_ops'
    end

Fixes #19090.
Fixes #21765.
Fixes #21819.
Fixes #24359.
Signed-off-by: NJeremy Daer <jeremydaer@gmail.com>

There is no need to fetch all table indexes in remove_index if name is specified. If name is wrong, then StatementInvalid will be raised.
Signed-off-by: NJeremy Daer <jeremydaer@gmail.com>

Make sure we handle explicitly passed nil's to lock_version as well.
An explicitly passed nil value is now converted to 0 on LockingType,
so that we don't end up with ActiveRecord::StaleObjectError in update record
optimistic locking

Fixes #24695

https://github.com/rails/rails/commit/42dd2336b31a8d98776d039a2b9fd7f834156a78 changed INSERT INTO versions to run in 1 single query.
This breaks for sqlite versions < 3.7.11, which is especially the case on Ubuntu 12.04 LTS, that has SQLite version 3.7.9 as default.

So we check for support for multi insert, before performing single query inserts, else fallback to older version of running multiple queries.
[Vipul A M & Yasuo Honda]

When the query cache completes, if Active Record is still inside of a
transaction, it is because the transaction is meant to be left open
above this unit of work (such as transactional fixtures in tests). There
were several tests around the behavior of "tests" that were invalid, as
tests are not run through the executor. They have been changed to
reflect the new behavior, which is closer to what actually occurs in
Rails tests.

Fixes #23989
Fixes #24491
Close #24500

If the adapter supports indexes in create table, it generates SQL
in one query.

Fix isolated test failures due to referencing ActiveRecord::Tasks::DatabaseAlreadyExists before another test happened to make it available

* Switch to keyword args where we can without breaking compat.
* Use add_table_options! for :options, too.
* Some code polish.

Comments are specified in migrations, stored in database itself (in its schema),
and dumped into db/schema.rb file.

This allows to generate good documentation and explain columns and tables' purpose
to everyone from new developers to database administrators.

For PostgreSQL and MySQL only. SQLite does not support comments at the moment.

See docs for PostgreSQL: http://www.postgresql.org/docs/current/static/sql-comment.html
See docs for MySQL: http://dev.mysql.com/doc/refman/5.7/en/create-table.html

Follow up to #24542.

In MySQL and PostgreSQL, a time column value is saved as ignored the
date part of it. But in SQLite3, a time column value is saved as a string.

We should keep previous quoting behavior in sqlite3 adapter.

```
sqlite> CREATE TABLE "foos" ("id" INTEGER PRIMARY KEY AUTOINCREMENT NOT NULL, "start" time(0), "finish" time(4));
sqlite> INSERT INTO "foos" ("start", "finish") VALUES ('2000-01-01 12:30:00', '2000-01-01 12:30:00.999900');
sqlite> SELECT "foos".* FROM "foos";
1|2000-01-01 12:30:00|2000-01-01 12:30:00.999900
sqlite> SELECT  "foos".* FROM "foos" WHERE "foos"."start" = '2000-01-01 12:30:00' LIMIT 1;
1|2000-01-01 12:30:00|2000-01-01 12:30:00.999900
sqlite> SELECT  "foos".* FROM "foos" WHERE "foos"."start" = '12:30:00' LIMIT 1;
sqlite>
```

Context #24522.

TIME column on MariaDB doesn't ignore the date part of the string when
it coerces to time.

```
root@localhost [test] > CREATE TABLE `foos` (`id` int AUTO_INCREMENT PRIMARY KEY, `start` time(0), `finish` time(4)) ENGINE=InnoDB;
Query OK, 0 rows affected (0.02 sec)

root@localhost [test] > INSERT INTO `foos` (`start`, `finish`) VALUES ('2000-01-01 12:30:00', '2000-01-01 12:30:00.999900');
Query OK, 1 row affected, 2 warnings (0.00 sec)

Note (Code 1265): Data truncated for column 'start' at row 1
Note (Code 1265): Data truncated for column 'finish' at row 1
root@localhost [test] > SELECT `foos`.* FROM `foos`;
+----+----------+---------------+
| id | start    | finish        |
+----+----------+---------------+
|  1 | 12:30:00 | 12:30:00.9999 |
+----+----------+---------------+
1 row in set (0.00 sec)

root@localhost [test] > SELECT  `foos`.* FROM `foos` WHERE `foos`.`start` = '2000-01-01 12:30:00' LIMIT 1;
Empty set (0.00 sec)

root@localhost [test] > SELECT  `foos`.* FROM `foos` WHERE `foos`.`start` = '12:30:00' LIMIT 1;
+----+----------+---------------+
| id | start    | finish        |
+----+----------+---------------+
|  1 | 12:30:00 | 12:30:00.9999 |
+----+----------+---------------+
1 row in set (0.00 sec)
```

- Specify we want to run on latest stable ruby for mariadb

- change in runs of builds

Make mariadb? method publicly available

Previously we were assuming that the only valid types for encoding were
arrays and hashes. However, any JSON primitive is an accepted value by
both PG and MySQL.

This does involve a minor breaking change in the handling of `default`
in the schema dumper. This is easily worked around, as passing a
hash/array literal would have worked fine in previous versions of Rails.
However, because of this, I will not be backporting this to 4.2 or
earlier.

Fixes #24234

In 04ac5655 I assumed that we would
never want to pass the "table_name.column_name" form to where with a
symbol. However, in Ruby 2.2 and later, you can quote symbols using the
new hash syntax, so it's a semi-reasonable thing to do if we want to
support the dot notation (which I'd rather deprecate, but that would be
too painful of a migration).

Instead we've changed the definition of "this is a table name with a
dot" to when the value associated is a hash. It would make very little
sense to write `where("table_name.column_name": { foo: :bar })` in any
scenario (other than equality for a JSON column which we don't support
through `where` in this way).

Close #24514.

- The negative scenario test case was missing earlier.

We support microsecond datetime precision for MySQL 5.6.4+. MariaDB has
supported it since 5.3.0, but even 10.x versions return a compatible
version string like `5.5.5-10.1.8-MariaDB-log` which we parse as 5.5.5,
before MySQL supported microsecond precision.

Specialize our version check to account for MariaDB to fix.

* Fix undefined method `owners' for NullPreloader:Class

Fixing undefined method `owners' for
ActiveRecord::Associations::Preloader::NullPreloader:Class

* Use Ruby 1.9 hash format

Use Ruby 1.9 hash format

#24192

[Rafael Mendonça França + Ladislav Smola]

Because we define `QUOTED_TRUE` as `"1"` and `QUOTED_FALSE` as `"0"`.
And add test cases to ensure this commit does not break current
behavior even if the value of `attributes_before_type_cast` is false.

This reverts commit 7b82e1c7.

This would have removed the ability to reference a schema when using PG

When prepared statements are enabled, the statement cache caches the SQL
directly, including the bind parameters. If a similar query is run later
with prepared statements disabled, we need to use a separate cache
instead of trying to share the same one.

Fixes #24351

Dots have special meaning in most backends (e.g. everything except
SQLite3), as well as most methods that work with table or column names.
This isn't something that we ever explicitly supported, but there's at
least one case of somebody using this (see #24367), so we'll go through a deprecation
cycle as normal.

When a proc is given as a default value, the form builder ends up
displaying `Proc#to_s` when the default is used. That's because we
didn't handle the proc until type casting. This issue technically can
occur any time that a proc is the value before type casting, but in
reality the only place that will occur is when a proc default is
provided through the attributes API, so the best place to handle this
edge case is there.

I've opted to memoize instead of just moving the `Proc#call` up, as this
made me realize that it could potentially interact very poorly with
dirty checking.

The code here is a little redundant, but I don't want to rely on how
`value_before_type_cast` is implemented in the super class, even if it's
just an `attr_reader`.

Fixes #24249

Close #24306

Without clearing the caches afterward, removals done in migrations would
not be reflected in a separate task in the same process. That is, given
a table with a migration to remove a column, the schema cache would
still reflect that a table has that in something such as the
'db:seed' task:

`rake db:migrate db:seed`
(A common thing to do in a script for a project ala `bin/setup`)

vs

`rake db:migrate && rake db:seed`
(Two processes)

The first would not reflect that the column was removed.
The second would (cache reset).

The error is raised because user passed invalid version number to a public api of
`ActiveRecord`, so `ArgumentError` is more suitable.
And add a test case checking if an error is raised when unknown migration version
is passed, because these test cases are not implemented.

This reverts a334425c.

The main reason is that now the workflow is inconsistent when using
spring.

When using spring `RAILS_ENV` is always set, so only one database is
created.

This means that in development `bin/rake db:create` and `bundle exec
rake db:create` have different results.

It also breaks the `bin/setup` script since `bin/rake db:setup
db:test:prepare` will fail.

In this patch, all log-related stuff in `ActiveSupport::Dependencies`
is removed because the logging is no longer useful.

For keep the default SQL mode.

Follow up to #18228.

In MySQL and PostgreSQL, primary key is to be `NOT NULL` implicitly.
But in SQLite it must be specified `NOT NULL` explicitly.

I'm unsure how this passed CI in the pull request.

Before:

```ruby
create_table "big_numbers", force: :cascade do |t|
  t.integer "bigint_column", limit: 8
end
```

After:

```ruby
create_table "big_numbers", force: :cascade do |t|
  t.bigint "bigint_column"
end
```