This reverts commit f6ca7e4e.

The default collation of utf8 in MySQL is the `utf8_general_ci`, and
this should not be changed. This is because, the better collation in the
all locales is not exists, optimal collation in own application is not
known other than themselves.

The `utf8_unicode_ci` is known as Japanese killer in Japan, there are
serious impacts in search of Japanese.

MySQL implements the `utf8_unicode_ci` according to the Unicode
Collation Algorithm (UCA) described at http://www.unicode.org/reports/tr10/,
but the `utf8_unicode_ci` have only partial support for the UCA, only
primary level key comparison implemented (also known as L1 (Base
characters) comparison).

Because L1 (Base characters) comparison does not distinguish between the
presence or absence of the accent, if distinction of the accent is
important there is a serious impact (e.g. Japanese).

Example:

```
> SHOW CREATE TABLE `dicts`\G
*************************** 1. row ***************************
       Table: dicts
Create Table: CREATE TABLE `dicts` (
  `word` varchar(255) COLLATE utf8_unicode_ci DEFAULT NULL,
  `meaning` varchar(255) COLLATE utf8_unicode_ci DEFAULT NULL
) ENGINE=InnoDB DEFAULT CHARSET=utf8 COLLATE=utf8_unicode_ci
1 row in set (0.00 sec)

> INSERT INTO `dicts` VALUES ('ハハ', 'mother'), ('パパ', 'father');
Query OK, 2 rows affected (0.00 sec)

> SELECT * FROM `dicts` WHERE `word` = 'ハハ';
+--------+---------+
| word   | meaning |
+--------+---------+
| ハハ   | mother  |
| パパ   | father  |
+--------+---------+
2 rows in set (0.00 sec)

> CREATE UNIQUE INDEX `unique_index_word` ON `dicts`(`word`);
ERROR 1062 (23000): Duplicate entry 'ハハ' for key 'unique_index_word'
```

We should omit the collation entirely rather than providing a default.
Then the choice is the responsibility of the server and MySQL distribution.

These methods are defined in inheritance.rb

* `abstract_class?`
* `descends_from_active_record?`
* `compute_type`

we should be using a countdown latch instead of rolling our own
busy-loop.

- Addresses issue #22092
- Works on Postgres and MySQL
- Uses advisory locks because of two important properties:
  1. The can be obtained outside of the context of a transaction
  2. They are automatically released when the session ends, so if a
  migration process crashed for whatever reason the lock is not left
  open perpetually
- Adds get_advisory_lock and release_advisory_lock methods to database
  adapters
- Attempting to run a migration while another one is in process will
  raise a ConcurrentMigrationError instead of attempting to run in
  parallel with undefined behavior. This could be rescued and
  the migration could exit cleanly instead. Perhaps as a configuration
  option?

Technical Notes
==============

The Migrator uses generate_migrator_advisory_lock_key to build the key
for the lock. In order to be compatible across multiple adapters there
are some constraints on this key.
- Postgres limits us to 64 bit signed integers
- MySQL advisory locks are server-wide so we have to scope to the
  database
- To fulfil these requirements we use a Migrator salt (a randomly
  chosen signed integer with max length of 31 bits) that identifies
  the Rails migration process as the owner of the lock. We multiply
  this salt with a CRC32 unsigned integer hash of the database name to
  get a signed 64 bit integer that can also be converted to a string
  to act as a lock key in MySQL databases.
- It is important for subsequent versions of the Migrator to use the
  same salt, otherwise different versions of the Migrator will not see
  each other's locks.

This assumes only one query was ever executed, but it appears to
sometimes be loading schema information. We can just look at the array
of queries, rather than the "first" one that was run

Before this commit, if
`ActiveRecord::Base.belongs_to_required_by_default` is set to `true`,
then creating a record through `has_and_belongs_to_many` fails with the
cryptic error message `Left side must exist`. This is because
`inverse_of` isn't working properly in this case, presumably since we're
doing trickery with anonymous classes in the middle.

Rather than following this rabbit hole to try and get `inverse_of` to
work in a case that we know is not publicly supported, we can just turn
off this validation to match the behavior of 4.2 and earlier.

This is part of a refactoring to make it easier to allow `order` to use
sanitize like just about everything else on relation. The deleted test
doesn't give any reasoning as to why passing `nil` to `order` needs to
be supported, and it's rather nonsensical. I can almost see allowing an
empty string being passed (though I'm tempted to just disallow it...)

Look at `TestFixtures.set_fixture_class`. As documented, it
accepts a mapping of fixture identifiers (string or symbol) to Classes
(the model classes that implement the named fixture).

Look now at the initialization of `TestFixtures.fixture_class_names`.
It defines a Hash, which will return a string by default (where the
string is the estimated class name of the given fixture identifier).

Now look at TestFixtures.load_fixtures. It calls `FixtureSet.create_fixtures`,
passing in the mapping of `fixture_class_names`.

Following this on to `FixtureSet.create_fixtures`, this instantiates a
`FixtureSet::ClassCache`, passing in the map of class names.

`ClassCache`, in turn, calls `insert_class` for each value in the cache.
(Recall that `set_fixture_class` puts Class objects in there, while the
default proc for the mapping puts String objects.)

Look finally at `insert_class`. If the value is present, it checks to
see if the value is a subclass of `AR::Base`. Fair enough...but wait!
What if the value is a String? You get an exception, because a String
instance cannot be compared with a Class.

Judging from the implementation, it seems like the expected behavior
here is for `fixture_class_names` to have no default proc. Look-ups are
supposed to happen via `ClassCache`, with `fixture_class_names` existing
solely as a repository for explicitly-registered class mappings.

That is what this change does.

This reverts commit 52439460.

This fixes an issue with the build where tests would fail on mysql and
postgresql due to different ordering.

This reverts commit 6dc6a0b1, reversing
changes made to ec94f00b.

This pull request broke the build.

The first one is quite straightforward. We want to give the proper error
message in the case where a top level constant exists, but we're looking
for a nested one. We just need to port over the change to use
`subclass.name` into these changes.

The second set of failures, which are only present in the mysql adapter
tests, are stranger to me. The failure occurs because we were
previously comparing `subclass.name == self.name` instead of `subclass
== self`. However, I don't think that we need to support creating
anonymous classes which share a table with a class that uses STI,
overrides `name` to return the same name as athe class that we have no
other relationship with, when not assigned to a constant so it could
never be used anyway...

The commits around why that exist give no context, and I think they're
just poorly written tests (WTF does `test_schema` mean anyway, and why
does calling `.first` on some anonymous class test it?). We'll just
disable STI on that class.

We should call `scope.order!` and set `scope.reordering_value` to `true` if :reordering values are specified

Fixes #21886

This solves the following issue:

```
$ bin/test
Using sqlite3
/Users/senny/Projects/rails/activerecord/test/cases/adapters/mysql2/sp_test.rb:16:in `<class:Mysql2StoredProcedureTest>': undefined method `version' for #<ActiveRecord::ConnectionAdapters::SQLite3Adapter:0x007f8bab4b5b70> (NoMethodError)
	from /Users/senny/Projects/rails/activerecord/test/cases/adapters/mysql2/sp_test.rb:5:in `<top (required)>'
	from /Users/senny/Projects/rails/activesupport/lib/active_support/dependencies.rb:302:in `require'
	from /Users/senny/Projects/rails/activesupport/lib/active_support/dependencies.rb:302:in `block in require'
	from /Users/senny/Projects/rails/activesupport/lib/active_support/dependencies.rb:268:in `load_dependency'
	from /Users/senny/Projects/rails/activesupport/lib/active_support/dependencies.rb:302:in `require'
	from /Users/senny/Projects/rails/railties/lib/rails/test_unit/test_requirer.rb:11:in `block in require_files'
	from /Users/senny/Projects/rails/railties/lib/rails/test_unit/test_requirer.rb:10:in `each'
	from /Users/senny/Projects/rails/railties/lib/rails/test_unit/test_requirer.rb:10:in `require_files'
	from /Users/senny/Projects/rails/railties/lib/rails/test_unit/minitest_plugin.rb:69:in `plugin_rails_init'
	from /Users/senny/.rbenv/versions/2.2.3/lib/ruby/gems/2.2.0/gems/minitest-5.3.3/lib/minitest.rb:73:in `block in init_plugins'
	from /Users/senny/.rbenv/versions/2.2.3/lib/ruby/gems/2.2.0/gems/minitest-5.3.3/lib/minitest.rb:71:in `each'
	from /Users/senny/.rbenv/versions/2.2.3/lib/ruby/gems/2.2.0/gems/minitest-5.3.3/lib/minitest.rb:71:in `init_plugins'
	from /Users/senny/.rbenv/versions/2.2.3/lib/ruby/gems/2.2.0/gems/minitest-5.3.3/lib/minitest.rb:122:in `run'
	from bin/test:19:in `<main>'
```

This reverts commit 60c97012, reversing
changes made to 6a25202d.

Reason: Broken build

This triggers the JoinDependency work to reflect on the associations
and trigger an error as follows:

  ActiveRecord::ConfigurationError: Association named 'account' was
  not found on Company; perhaps you misspelled it?

Fix Company.of_first_firm joins association name

Should be `Company.joins(:accounts)` not `Company.joins(:account)`.

Do the same for Client.of_first_firm

* When tried to use `Company#accounts` test/models/company.rb I got:

```
ActiveRecord::StatementInvalid: SQLite3::SQLException: no such column:
accounts.company_id: SELECT COUNT(*) AS count_all, "companies"."firm_id"
AS companies_firm_id FROM "companies" INNER JOIN "accounts" ON
"accounts"."company_id" = "companies"."id" GROUP BY "companies"."firm_id"
```

* The refactor on Calculations class was just to simplify the code

Closes #21922

Let `Book(id, author_id)`, `Photo(id, book_id, author_id)` and `Author(id)`

Running `Book.group(:author_id).joins(:photos).count` will produce:

* Rails 4.2 - conflicts `author_id` in both projection and group by:
```sql
SELECT COUNT(*) AS count_all, author_id AS author_id
  FROM "books" INNER JOIN "photos" ON "photos"."book_id" = "books"."id"
 GROUP BY author_id
```

* Master (9d02a25d) - conflicts `author_id` only in projection:
```sql
SELECT COUNT(*) AS count_all, author_id AS author_id
  FROM "books" INNER JOIN "photos" ON "photos"."book_id" = "books"."id"
 GROUP BY "books"."author_id"
```

* With this fix:
```sql
SELECT COUNT(*) AS count_all, "books"."author_id" AS books_author_id
  FROM "books" INNER JOIN "photos" ON "photos"."book_id" = "books"."id"
 GROUP BY "books"."author_id"
```

This issue was resolved by #21687 already. But re-add args by #18856.
`#tables` extra args was only using by `#table_exists?`. This is for
internal API. This commit will remove these extra args again.

* If the create task fails for a reason other than the database already
  existing, processing should end. This is indicated by a non-zero exit
  status.
* Since the backtrace is already printed to screen, we forgo printing it
  again by using an explicit call to `exit`.
* ⚠ This modifies the behavior of the db:create task slightly in
  that the stack trace is no longer printed by default. If the `--trace`
  option is used, it will print the trace _after_ the error message.

I messed up the merge conflict, and accidentally removed a schema query
that needed to be ignored.

Column names inserted via `group` have to be qualified with table name.

Prior to this commit, Rails makes no differentiation between whether a
query uses bind parameters, and whether or not we cache that query as a
prepared statement. This leads to the cache populating extremely fast in
some cases, with the statements never being reused.

In particular, the two problematic cases are `where(foo: [1, 2, 3])` and
`where("foo = ?", 1)`. In both cases we'll end up quoting the values
rather than using a bind param, causing a cache entry for every value
ever used in that query.

It was noted that we can probably eventually change `where("foo = ?",
1)` to use a bind param, which would resolve that case. Additionally, on
PG we can change our generated query to be `WHERE foo = ANY($1)`, and
pass an array for the bind param. I hope to accomplish both in the
future.

For SQLite and MySQL, we still end up preparing the statements anyway,
we just don't cache it. The statement will be cleaned up after it is
executed. On postgres, we skip the prepare step entirely, as an API is
provided to execute with bind params without preparing the statement.

I'm not 100% happy on the way this ended up being structured. I was
hoping to use a decorator on the visitor, rather than mixing a module
into the object, but the way Arel has it's visitor pattern set up makes
it very difficult to extend without inheritance. I'd like to remove the
duplication from the various places that are extending it, but that'll
require a larger restructuring of that initialization logic. I'm going
to take another look at the structure of it soon.

This changes the signature of one of the adapter's internals, and will
require downstream changes from third party adapters. I'm not too
worried about this, as worst case they can simply add the parameter and
always ignore it, and just keep their previous behavior.

Fixes #21992.

With `unscope!` called last, it undoes `where` constraints of the same
value when the `where` is chained after the `unscope`. This is what a
`rewhere` does. This is undesirable behavior.

The included tests demonstrate both the `unscope(...).where(...)`
behavior as well as the direct use of `rewhere(...)`.

This is in reference to #21955.

This commit follow up of 4d8f62dc.
The difference from 4d8f62dc are below:

* Change `WhereClauseFactory` to accept `Arel::Nodes::Node`
* Change test cases of `relation_test.rb`

[#20473]

The default timestamp used for AR is `updated_at` in nanoseconds! (:nsec) This causes issues on any machine that runs an OS that supports nanoseconds timestamps, i.e. not-OS X, where the cache_key of the record persisted in the database (milliseconds precision) is out-of-sync with the cache_key in the ruby VM.

This commit adds:

A test that shows the issue, it can be found in the separate file `cache_key_test.rb`, because
- model couldn't be defined inline
- transactional testing needed to be turned off to get it to pass the MySQL tests
This seemed cleaner than putting it in an existing testcase file.

It adds :usec as a dateformat that calculates datetime in microseconds

It sets precision of cache_key to :usec instead of :nsec, as no db supports nsec precision on timestamps

When passing an instance of `ActiveRecord::Base` to `#update`, it would
internally call `#find`, resulting in a misleading deprecation warning.

This change gives this deprecated use of `#update` its own, meaningful
warning.

Currently `tinyblob` is dumped to `t.binary "tiny_blob", limit: 255`.
But `t.binary ... limit: 255` is generating SQL to `varchar(255)`.
It is incorrect. This commit fixes this problem.

Closes #21563.

The `name` argument of `add_references` was both used to generate the
column name `<name>_id` and as the target table for the foreign key
`name.pluralize`.

It's primary purpose is to define the column name. In cases where the
`to_table` of the foreign key is different than the column name we
should be able to specify it individually.

Fractional Seconds supported version is 5.6.4 or higher.

http://dev.mysql.com/doc/refman/5.6/en/fractional-seconds.html

`pk_and_sequence_for` is implemented for PG and MySQL adapters (not
implemented for Sqlite3 adapter). But MySQL adapters are not using
`pk_and_sequence_for` already.

Any gems or libraries which do work with serialization or YAML will
ultimately need to compare these objects (albeit indirectly) to ensure
correctness. These will likely never get used internally (as they're
slow), but we should still expose them for others.

In this commit, find_by doesn't cache arguments
so that find_by with association subquery works correctly.

Fixes #20817