If autoloading of a table fails, don't register it in a metadata
instance. It seems that the original behaviour was accidentally
changed in f6198d9abf453182f4b111e0579a7a4ef1614e79, restore it.

Closes issue #2988

is currently being supported in addition to nose, and will likely
be preferred to nose going forward.   The nose plugin system used
by SQLAlchemy has been split out so that it works under pytest as
well.  There are no plans to drop support for nose at the moment
and we hope that the test suite itself can continue to remain as
agnostic of testing platform as possible.  See the file
README.unittests.rst for updated information on running tests
with pytest.

The test plugin system has also been enhanced to support running
tests against mutiple database URLs at once, by specifying the ``--db``
and/or ``--dburi`` flags multiple times.  This does not run the entire test
suite for each database, but instead allows test cases that are specific
to certain backends make use of that backend as the test is run.
When using pytest as the test runner, the system will also run
specific test suites multiple times, once for each database, particularly
those tests within the "dialect suite".   The plan is that the enhanced
system will also be used by Alembic, and allow Alembic to run
migration operation tests against multiple backends in one run, including
third-party backends not included within Alembic itself.
Third party dialects and extensions are also encouraged to standardize
on SQLAlchemy's test suite as a basis; see the file README.dialects.rst
for background on building out from SQLAlchemy's test platform.

concurrent ability to return connections from the pool means that the
"first_connect" event is now no longer synchronized either, thus leading
to dialect mis-configurations under even minimal concurrency situations.

SQLite allows column types that aren't technically understood in sqlite
by using 'data affinity', which is an algorithm for converting column
types in to some sort of useful type that can be stored and retrieved
from the db. Unfortunatly, this breaks reflection since we (previously)
expected a sqlite db to reflect column types that we permit in the
`ischema_names` for that dialect.

This patch changes the logic for 'unknown' column types during
reflection to instead run through SQLite's data affinity algorithm, and
assigns appropriate types from that.

It also expands the matching for column type to include column types
with spaces (strongly discouraged but allowed by sqlite) and also
completely empty column types (in which case the NullType is assigned,
which sqlite will treat as a Blob - or rather, Blob is treated as
NullType). These changes mean that SQLite will never raise an error for
an unknown type during reflection - there will always be some 'useful'
type returned, which follows the spirit of SQLite (accomodation before
sanity!).

applied to :class:`.Constraint` and :class:`.Index` objects.  Based
on a recipe in the wiki, the new feature uses schema-events to set up
names as various schema objects are associated with each other.  The
events then expose a configuration system through a new argument
:paramref:`.MetaData.naming_convention`.  This system allows production
of both simple and custom naming schemes for constraints and indexes
on a per-:class:`.MetaData` basis.  [ticket:2923]

commit 7e65e52c086652de3dd3303c723f98f09af54db8
Author: Mike Bayer <mike_mp@zzzcomputing.com>
Date:   Sat Feb 1 15:09:04 2014 -0500

    - first pass at new naming approach

reflection now updates the PKC in place.
- support the use case of the empty PrimaryKeyConstraint in order to specify
constraint options; the columns marked as primary_key=True will now be gathered
into the columns collection, rather than being ignored. [ticket:2910]
- add validation such that column specification should only take place
in the PrimaryKeyConstraint directly, or by using primary_key=True flags;
if both are present, they have to match exactly, otherwise the condition is
assumed to be ambiguous, and a warning is emitted; the old behavior of
using the PKC columns only is maintained.

type such as "charset" and "collation".  While MySQL wants all character-
based CAST calls to use the CHAR type, we now create a real CHAR
object at CAST time and copy over all the parameters it has, so that
an expression like ``cast(x, mysql.TEXT(charset='utf8'))`` will
render ``CAST(t.col AS CHAR CHARACTER SET utf8)``.

- Added new "unicode returns" detection to the MySQL dialect and
to the default dialect system overall, such that any dialect
can add extra "tests" to the on-first-connect "does this DBAPI
return unicode directly?" detection. In this case, we are
adding a check specifically against the "utf8" encoding with
an explicit "utf8_bin" collation type (after checking that
this collation is available) to test for some buggy unicode
behavior observed with MySQLdb version 1.2.3.  While MySQLdb
has resolved this issue as of 1.2.4, the check here should
guard against regressions.  The change also allows the "unicode"
checks to log in the engine logs, which was not previously
the case. [ticket:2906]

_reset_agent, so that it's local to the various begin_impl(),
rollback_impl(), etc.  this allows setting/resetting of the flag
to be symmetric.
- don't set _reset_agent if it's not None, don't unset it if it isn't
our own transaction.
- make sure we clean it out in close().
- basically, we're dealing here with pools using "threadlocal" that have a
counter, other various mismatches that the tests bring up
- test for recover() now has to invalidate() the previous connection,
because closing it actually rolls it back (e.g. this test was relying
on the broken behavior).