In passing, allow multi-word adapters to be referenced in a URL:
underscored_name must become hyphened-name.

If the supplied string doesn't contain a colon, it clearly cannot be a
database URL. They must have intended to do a key lookup, so even though
it failed, give the explanatory deprecation warning, and raise the
exception that lists the known configs.

Conveniently, this also simplifies our logical behaviour: if the string
matches a known configuration, or doesn't contain a colon (and is
therefore clearly not a URL), then we output a deprecation warning, and
behave exactly as we would if it were a symbol.

this lets arel know how to correctly quote the value

In our normal usage, it's rare for this to make a difference... but is
more technically correct.

As well as a spec that proves this is a good idea, let's also add a more
sane-looking one that just covers basic to_sql functionality. There
aren't many places where we actually use escape_bytea, but that's one
that won't be going away.

There is no reason for the PG adapter to have a default limit of 255 on :string
columns. See this snippet from the PG docs:

    Tip: There is no performance difference among these three types, apart
    from increased storage space when using the blank-padded type, and a
    few extra CPU cycles to check the length when storing into a
    length-constrained column. While character(n) has performance
    advantages in some other database systems, there is no such advantage
    in PostgreSQL; in fact character(n) is usually the slowest of the
    three because of its additional storage costs. In most situations text
    or character varying should be used instead.

The keys are quite sufficient; we shouldn't be throwing passwords
around.

This is all about the case where we have a `DATABASE_URL`, and we have a
`database.yml` present, but the latter doesn't contain the key we're
looking for.

If the key is a symbol, we'll always connect to `DATABASE_URL`, per the
new behaviour in 283a2ede.

If the key is a string, on the other hand, it should always be a URL:
the ability to specify a name not present in `database.yml` is new in
this version of Rails, and that ability does not stretch to the
deprecated use of a string in place of a symbol.

Uncovered by @guilleiguaran while investigating #14495 -- this actually
may be related to the original report, but we don't have enough info to
confirm.

That which was now relative is now absolute.

Restore the 4.0 behaviour for 'sqlite3:///', but deprecate it. We'll
change to the absolute-path interpretation in 4.2.

The current "correct" spellings for in-memory, relative, and absolute
URLs, respectively, are:

    sqlite3::memory:
    sqlite3:relative/path
    sqlite3:/full/path

Substantially reverses/defers fbb79b51.
Uncovered by @guilleiguaran while investigating #14495, though that
sounds like a different issue.

Expand the query used in #table_exists? to include materialized views in the
kinds of relations it searches.

* Clarify what the situation is and what to do.
* Advise loading schema using `rake db:setup` instead of migrating.
* Use a rescue in the initializer rather than extending the error
  message in-place.
* Preserve the original backtrace of other errors by using `raise`
  rather than raising again with `raise error`.

References 0ec45cd1

This is a follow-up fix to f7a6b115 and
06082f66

This patch registers custom domains in our OID-type_map.
They will behave exactly as the type specified by `pg_type.typbasetype`.

/cc @matthewd

I ran the whole test suite and compared the old to the new types.
Following is the list of types that did change with this patch:

```
DIFFERENT TYPE FOR mood: NEW: enum, BEFORE:
DIFFERENT TYPE FOR floatrange: NEW: floatrange, BEFORE: float
```

The `floatrange` is a custom type. The old type `float` was simply a coincidence
form the name `floatrange` and our type-guessing.

We have `connection_adapters/column.rb` so it's easier to remember
that the column in in a separate file.

fixes #10613

It is already called inside type_to_sql method.

This method is not being called anywhere in our code and a GitHub search
show it is not being used in any project.

* cache `o.name` value
* Avoid extra `concat` call
* Avoid extra `<<` call

These comments will likely be outdated with time and doesn't include
any information that can't be found in the adapters

Also improve some of the code conventions

Also:
  - updates tests by stubbing table_exists? method
  - adds entry for creating indexes in CREATE TABLE to changelog

This is important, because adding an index on a temporary table after
it has been created would commit the transaction

Conflicts:
	activerecord/CHANGELOG.md

The pk_an_sequence_for query previously joined against pg_class's oid
for rows in pg_depend, but pg_depend's objid may point to other system
tables, such as pg_attrdef.  If a row in one of those other tables
coincidentally has the same oid as an (unrelated) sequence, that
sequence name may be returned instead of the real one.

This ensures that only the pg_depend entries pointing to pg_class are
considered.

.. not a general timeout.

Now, if a thread checks out a connection then dies, we can immediately
recover that connection and re-use it.

This should alleviate the pool exhaustion discussed in #12867. More
importantly, it entirely avoids the potential issues of the reaper
attempting to check whether connections are still active: as long as the
owning thread is alive, the connection is its business alone.

As a no-op reap is now trivial (only entails checking a thread status
per connection), we can also perform one in-line any time we decide to
sleep for a connection.

It wasn't doing anything beyond clearing the statement cache.

mitigates #14323

[fixes #14361]
[related #13886]

You might want to branch it to include this only for 5.6, but
passing these values to < 5.6 doesn't cause issues either.