SQLAlchemy, Serializable transactions isolation an

Question

PostgreSQL and SQL defines a Serializable transaction isolation level. If you isolate transactions to this level, conflicting concurrent transactions abort and need retrying.

I am familiar with the concept of transaction retries from Plone / Zope world where the entire HTTP request can be replayed in the case there is a transaction conflict. How similar functionality could be achieved with SQLAlchemy (and potentially with zope.sqlalchemy)? I tried to read the documentation of zope.sqlalchemy and Zope transaction manager, but this is not obvious the me.

Specially I want something like this:

  # Try to do the stuff, if it fails because of transaction conflict do again until retry count is exceeded
  with transaction.manager(retries=3):
        do_stuff()

  # If we couldn't get the transaction through even after 3 attempts, fail with a horrible exception

Answer 1

So, after poking around two weeks and getting no off-the-shelf solution I came up with my own.

Here is a ConflictResolver class which provides managed_transaction function decorator. You can use the decorator to mark functions to be retryable. I.e. if there is an database conflict error when running the function, the function is run again, now with more hopes the db transaction which caused the conflict error would have finished.

The source code is here: https://bitbucket.org/miohtama/cryptoassets/src/529c50d74972ff90fe5b61dfbfc1428189cc248f/cryptoassets/core/tests/test_conflictresolver.py?at=master

The unit tests to cover it are here: https://bitbucket.org/miohtama/cryptoassets/src/529c50d74972ff90fe5b61dfbfc1428189cc248f/cryptoassets/core/tests/test_conflictresolver.py?at=master

Python 3.4+ only.

"""Serialized SQL transaction conflict resolution as a function decorator."""

import warnings
import logging
from collections import Counter

from sqlalchemy.orm.exc import ConcurrentModificationError
from sqlalchemy.exc import OperationalError


UNSUPPORTED_DATABASE = "Seems like we might know how to support serializable transactions for this database. We don't know or it is untested. Thus, the reliability of the service may suffer. See transaction documentation for the details."

#: Tuples of (Exception class, test function). Behavior copied from _retryable_errors definitions copied from zope.sqlalchemy
DATABASE_COFLICT_ERRORS = []

try:
    import psycopg2.extensions
except ImportError:
    pass
else:
    DATABASE_COFLICT_ERRORS.append((psycopg2.extensions.TransactionRollbackError, None))

# ORA-08177: can't serialize access for this transaction
try:
    import cx_Oracle
except ImportError:
    pass
else:
    DATABASE_COFLICT_ERRORS.append((cx_Oracle.DatabaseError, lambda e: e.args[0].code == 8177))

if not DATABASE_COFLICT_ERRORS:
    # TODO: Do this when cryptoassets app engine is configured
    warnings.warn(UNSUPPORTED_DATABASE, UserWarning, stacklevel=2)

#: XXX: We need to confirm is this the right way for MySQL, SQLIte?
DATABASE_COFLICT_ERRORS.append((ConcurrentModificationError, None))


logger = logging.getLogger(__name__)


class CannotResolveDatabaseConflict(Exception):
    """The managed_transaction decorator has given up trying to resolve the conflict.

    We have exceeded the threshold for database conflicts. Probably long-running transactions or overload are blocking our rows in the database, so that this transaction would never succeed in error free manner. Thus, we need to tell our service user that unfortunately this time you cannot do your thing.
    """


class ConflictResolver:

    def __init__(self, session_factory, retries):
        """

        :param session_factory: `callback()` which will give us a new SQLAlchemy session object for each transaction and retry

        :param retries: The number of attempst we try to re-run the transaction in the case of transaction conflict.
        """
        self.retries = retries

        self.session_factory = session_factory

        # Simple beancounting diagnostics how well we are doing
        self.stats = Counter(success=0, retries=0, errors=0, unresolved=0)

    @classmethod
    def is_retryable_exception(self, e):
        """Does the exception look like a database conflict error?

        Check for database driver specific cases.

        :param e: Python Exception instance
        """

        if not isinstance(e, OperationalError):
            # Not an SQLAlchemy exception
            return False

        # The exception SQLAlchemy wrapped
        orig = e.orig

        for err, func in DATABASE_COFLICT_ERRORS:
            # EXception type matches, now compare its values
            if isinstance(orig, err):
                if func:
                    return func(e)
                else:
                    return True

        return False

    def managed_transaction(self, func):
        """SQL Seralized transaction isolation-level conflict resolution.

        When SQL transaction isolation level is its highest level (Serializable), the SQL database itself cannot alone resolve conflicting concurrenct transactions. Thus, the SQL driver raises an exception to signal this condition.

        ``managed_transaction`` decorator will retry to run everyhing inside the function

        Usage::

            # Create new session for SQLAlchemy engine
            def create_session():
                Session = sessionmaker()
                Session.configure(bind=engine)
                return Session()

            conflict_resolver = ConflictResolver(create_session, retries=3)

            # Create a decorated function which can try to re-run itself in the case of conflict
            @conflict_resolver.managed_transaction
            def myfunc(session):

                # Both threads modify the same wallet simultaneously
                w = session.query(BitcoinWallet).get(1)
                w.balance += 1

            # Execute the conflict sensitive code inside a managed transaction
            myfunc()

        The rules:

        - You must not swallow all exceptions within ``managed_transactions``. Example how to handle exceptions::

            # Create a decorated function which can try to re-run itself in the case of conflict
            @conflict_resolver.managed_transaction
            def myfunc(session):

                try:
                    my_code()
                except Exception as e:
                    if ConflictResolver.is_retryable_exception(e):
                        # This must be passed to the function decorator, so it can attempt retry
                        raise
                    # Otherwise the exception is all yours

        - Use read-only database sessions if you know you do not need to modify the database and you need weaker transaction guarantees e.g. for displaying the total balance.

        - Never do external actions, like sending emails, inside ``managed_transaction``. If the database transaction is replayed, the code is run twice and you end up sending the same email twice.

        - Managed transaction section should be as small and fast as possible

        - Avoid long-running transactions by splitting up big transaction to smaller worker batches

        This implementation heavily draws inspiration from the following sources

        - http://stackoverflow.com/q/27351433/315168

        - https://gist.github.com/khayrov/6291557
        """

        def decorated_func():

            # Read attemps from app configuration
            attempts = self.retries

            while attempts >= 0:

                session = self.session_factory()
                try:
                    result = func(session)
                    session.commit()
                    self.stats["success"] += 1
                    return result

                except Exception as e:
                    if self.is_retryable_exception(e):
                        session.close()
                        self.stats["retries"] += 1
                        attempts -= 1
                        if attempts < 0:
                            self.stats["unresolved"] += 1
                            raise CannotResolveDatabaseConflict("Could not replay the transaction {} even after {} attempts".format(func, self.retries)) from e
                        continue
                    else:
                        session.rollback()
                        self.stats["errors"] += 1
                        # All other exceptions should fall through
                        raise

        return decorated_func

Answer 2

Postgres and Oracle conflict errors are marked as retryable by zope.sqlalchemy. Set your isolation level in the engine configuration and the transaction retry logic in pyramid_tm or Zope will work.