Use pandasSQL transactions in sql test suite to avoid engine deadlocks

[WillAyd]

When trying to integrate the engine fixtures into these tests I have been running into a lot of deadlocking issues. From some research it looks like SQLAlchemy strongly discourages running doing Connectionless Execution

Note the comment:

Modern SQLAlchemy usage, especially the ORM, places a heavy stress on working within the context of a transaction at all times; the “implicit execution” concept makes the job of associating statement execution with a particular transaction much more difficult.

We probably need an overhaul of the sql module to more natively account for this, but I think it makes sense to unlock our test suite first so that getting better coverage via fixtures in #55074 would help us more confidently refactor

Here's an MRE to reproduce this deadlock locally:

import sqlalchemy as sa

from pandas.io.sql import pandasSQL_builder

engine = sa.create_engine("mysql+pymysql://root@localhost:3306/pandas")

pandasSQL = pandasSQL_builder(engine)
with pandasSQL.run_transaction() as trans:
    stmt = sa.text("CREATE TABLE test_trans (A INT, B TEXT)")
    trans.execute(stmt)

with pandasSQL.run_transaction() as trans:
    stmt = sa.text("INSERT INTO test_trans (A,B) VALUES (1, 'blah')")
    trans.execute(stmt)

result = pandasSQL.read_query("SELECT * FROM test_trans")

# Assume fixture tries to clean up all tables after function exit
with engine.begin() as conn:
    stmt = sa.text("DROP TABLE test_trans")
    conn.execute(stmt)

If you wrap the read_query call in a transaction there is no deadlock

(tip: if you ran the above locally you will probably want to kill the thread so it doesn't lock up your database. run show engine innodb status; inside the db, look for the hanging thread id and run KILL <thread_id>)

[mroeschke]

Thanks @WillAyd