Thanks to @thedarkone for pointing out that an instance of this object
is used in a shared context.

See also PR: #17610

`.reflections` public API changed to return a String instead of a Symbol
as keys.

see commit 1f314884 and 6259e4e2

[fixes #16928]
[fixes #17610]

For some reason changing `.find` to `.unscoped.find` in
https://github.com/rails/rails/commit/f766abd4cf3eb75469d3646cfb6d85e668c619f3
caused `scoping` to leak in some tests when run in isolation (looks like
a concurrency issue?). `relation_scoping_test.rb` is a case that failed.
From what I can tell it should not be possible, but changing to the
block form fixes it. There is a deeper issue that I can't seem to find.
/cc @senny

Also checked to make sure this does not affect foreign key constraints.
(It doesn't).

Fixes #12856
Closes #14088

We can't change the signature without a deprecation cycle.

on the joined assoiciation

Prior to this patch you'd end up with an error like:

```
ActiveRecord::RecordNotFound: Couldn't find <Model> with 'id'=<id> [WHERE (<default_scope condition>)]
```

I grepped the source code for code snippets wrapped in backticks in the comments
and replaced the backticks with plus signs so they are correctly displayed in
the Rails documentation.

[ci skip]

This makes the following changes:
  * warn if `:null` is not passed to `add_timestamps`
  * `timestamps` method docs link to `add_timestamps` docs
  * explain where additional options go
  * adjust examples to include `null: false` (to prevent deprecation warnings)

It doesn't make sense for the subclass to implement this method, and not
have it on the parent. We can also DRY up the implementation of
`#lookup` to be defined in terms of fetch, which will give us a single
point of entry

This reverts commit da99a2a2.

We added a comparison to "id", and call to `self.class.primary_key` a
*lot*. We also have performance hits from `&block` all over the place.
We skip the check in a new method, in order to avoid breaking the
behavior of `read_attribute`

We don't have the check the range when the value is coming from the DB,
so override type_cast_from_database to short-circuit the extra work.
The difference is huge but the absolute gain is quite small. That being
said this is a hotspot and it showed up on the radar when benchmarking
discourse.

This appears to be a performance hotspot, see #17655.

Performance improvement, as well as improved code clarity

This reverts commit 6f7910ae and 52c70d49.

Query params are type cased through the same method, so this approach doesn't work.

See comment on 6f7910ae

We don't have the check the range when the value is coming from the DB, so
override type_cast_from_database to short-circuit the extra work.

type_cast_from_database (small)  3437507.5 (±29.2%) i/s -   14223135 in   4.996973s
type_cast_from_database (large)  3158588.7 (±28.3%) i/s -   13265628 in   4.992121s
type_cast (small)                 481984.8 (±14.2%) i/s -    2352012 in   5.005694s
type_cast (large)                 477331.8 (±14.2%) i/s -    2332824 in   5.012365s

Comparison:
type_cast_from_database (small):  3437507.5 i/s
type_cast_from_database (large):  3158588.7 i/s - 1.09x slower
type_cast (small):                 481984.8 i/s - 7.13x slower
type_cast (large):                 477331.8 i/s - 7.20x slower

The difference is huge but the absolute gain is quite small. That being said
this is a hotspot and it showed up on the radar when benchmarking discourse.

Oh hey, we got to remove some code because of that!

Arel handles this for us automatically. Updated tests, as BindParam is
no longer a subclass of SqlLiteral. We should remove the second argument
to substitute_at entirely, as it's no longer used

If something causes the teardown block to not get run (errors,
interrupting test runs), we have to manually delete them, which is a
pain.

This caused a pretty major performance regression for 4.2, as this is a
hotspot for query construction. We're still slightly slower than 4.1,
but it's much less significant.

This class no longer has any private methods

`default_proc` makes a hash unmarshallable, and adds unneccessary
overhead. Since we control all access to the hash, let's just handle it
in that method. This has the side effect of improving performance on
initialization (but not neccessarily on access). We'll need to profile
further once the tests are passing.

In real usage, we give the builder a types hash with a default value of
`Type::Value.new`. This means we need to explicitly check for the key,
rather than the truthiness of the type to determine if it's a known but
uninitialized attribute

We don't know which attributes will or won't be used, and we don't want
to create massive bottlenecks at instantiation. Rather than doing *any*
iteration over types and values, we can lazily instantiate the object.

The lazy attribute hash should not fully implement hash, or subclass
hash at any point in the future. It is not meant to be a replacement,
but instead implement its own interface which happens to overlap.