Transactions in WSGI: Using a Transaction-Aware Data Manager Under repoze.tm

Overview

One of the purposes of creating repoze.tm (see http://repoze.org/viewcvs/repoze.tm/trunk/ for direct source view) was to allow for systems that participate in a WSGI pipeline to make use of the existing two-phase commit transaction management provided by the ZODB transaction package. But it occurs to me that there's no real "getting started" documentation about using the ZODB transaction machinery outside the context of Zope itself. So let's provide one.

Defining A Problem

Let's pretend we have an existing system that places data into a relational database when someone submits a form. The system has been running for a while, and our code handles the database commit and rollback for us explicitly; if the form processing succeeds, our code commits the database transaction. If it fails, our code rolls back the database transaction. Everything works fine.

Now our customer asks us if we can also place data into another separate relational database when the form is submitted as well as continuing to place data in the original database. We need to put data in both databases, and if we want to ensure that no records exist in one that don't exist in the other as a result of a form submission, we're going to need to do a pretty complicated commit and rollback dance in each place in our code which needs to write to both data stores. We can't just blindly commit one, then commit the other, because the second commit may fail and we'll be left with "orphan" data in the first, and we'll either need to clean it up manually or leave it there to trip over later.

A transaction manager helps us ensure that no data is committed to either database unless both participating data stores can commit. Once the transaction manager determines that both data stores are willing to commit, it will commit them both in very quick succession, so that there is only a minimal chance that the second data store will fail to commit. If it does, the system will raise an error that makes it impossible to begin another transaction until the system restarts, so the damage is minimized. In practice, this error almost never occurs unless the code that interfaces the database to the transaction manager has a bug.

Mocking Up A Data Manager

The piece of code you need to write in order to participate in ZODB transactions is called a data manager. It is typically a class. Here's the interface that you need to implement in the code for a data manager:

      class IDataManager(zope.interface.Interface):
          """Objects that manage transactional storage.

          These objects may manage data for other objects, or they
          may manage non-object storages, such as relational
          databases.  For example, a ZODB.Connection.

          Note that when some data is modified, that data's data
          manager should join a transaction so that data can be
          committed when the user commits the transaction.  """

          transaction_manager = zope.interface.Attribute(
              """The transaction manager (TM) used by this data
              manager.

              This is a public attribute, intended for read-only
              use.  The value is an instance of ITransactionManager,
              typically set by the data manager's constructor.  """
              )

          def abort(transaction):
              """Abort a transaction and forget all changes.

              Abort must be called outside of a two-phase commit.

              Abort is called by the transaction manager to abort transactions
              that are not yet in a two-phase commit.
              """

          # Two-phase commit protocol.  These methods are called by
          # the ITransaction object associated with the transaction
          # being committed.  The sequence of calls normally follows
          # this regular expression: tpc_begin commit tpc_vote
          # (tpc_finish | tpc_abort)

          def tpc_begin(transaction):

              """Begin commit of a transaction, starting the
              two-phase commit.

              transaction is the ITransaction instance associated with the
              transaction being committed.
              """

          def commit(transaction):

              """Commit modifications to registered objects.

              Save changes to be made persistent if the transaction
              commits (if tpc_finish is called later).  If tpc_abort
              is called later, changes must not persist.

              This includes conflict detection and handling.  If no
              conflicts or errors occur, the data manager should be
              prepared to make the changes persist when tpc_finish
              is called.  """

          def tpc_vote(transaction):
              """Verify that a data manager can commit the transaction.

              This is the last chance for a data manager to vote 'no'.  A
              data manager votes 'no' by raising an exception.

              transaction is the ITransaction instance associated with the
              transaction being committed.
              """

          def tpc_finish(transaction):

              """Indicate confirmation that the transaction is done.

              Make all changes to objects modified by this
              transaction persist.

              transaction is the ITransaction instance associated
              with the transaction being committed.

              This should never fail.  If this raises an exception,
              the database is not expected to maintain consistency;
              it's a serious error.  """

          def tpc_abort(transaction):

              """Abort a transaction.

              This is called by a transaction manager to end a
              two-phase commit on the data manager.  Abandon all
              changes to objects modified by this transaction.

              transaction is the ITransaction instance associated
              with the transaction being committed.

              This should never fail.
              """

          def sortKey():

              """Return a key to use for ordering registered
              DataManagers.

              ZODB uses a global sort order to prevent deadlock when
              it commits transactions involving multiple resource
              managers.  The resource manager must define a
              sortKey() method that provides a global ordering for
              resource managers.  """
              # Alternate version:
              #"""Return a consistent sort key for this connection.
              # #This allows ordering multiple connections that use
              the same storage in #a consistent manner. This is
              unique for the lifetime of a connection, #which is
              good enough to avoid ZEO deadlocks.  #"""

Let's implement a mock data manager. Our mock data manager will write data to a file if the transaction commits. It will not write data to a file if the transaction aborts:

      class MockDataManager:

          transaction_manager = None

          def __init__(self, data, path):
              self.data = data
              self.path = path

          def abort(self, transaction):
              pass

          def tpc_begin(self, transaction):
              pass

          def commit(self, transaction):
              import tempfile
              self.tempfn = tempfile.mktemp()
              temp = open(self.tempfn, 'wb')
              temp.write(self.data)
              temp.flush()
              temp.close()

          def tpc_vote(self, transaction):
              import os
              if not os.path.exists(self.tempfn):
                  raise ValueError('%s doesnt exist' % self.tempfn)
              if os.path.exists(self.path):
                  raise ValueError('file already exists')

          def tpc_finish(self, transaction):
              import os
              os.rename(self.tempfn, self.path)

          def tpc_abort(self, transaction):
              import os
              try:
                  os.remove(self.tempfn)
              except OSError:
                  pass

We can create a datamanager and join it into the currently running transaction:

      dm = MockDataManager('heres the data',  '/tmp/file')
      import transaction
      t = transaction.get()
      t.join(dm)

When the transaction commits, a file will be placed in /tmp/file containing heres the data. If the transaction aborts, no file will be created.

If more than one data manager is joined to the transaction, all of them must be willing to commit or the entire transaction is aborted and none of them commit. If you can imagine creating two of the mock data managers we've made within application code, if one has a problem during "tpc_vote", neither will actually write a file to the ultimate location.

Integrating With repoze.tm

The repoze.tm transaction management machinery has an implicit policy. When it is in the WSGI pipeline, a transaction is started when the middleware is invoked. Thus, in your application code, calling "import transaction; transaction.get()" will return the transaction object created by the repoze.tm middleware. You needn't call t.commit() or t.abort() within your application code. You only need to call t.join, to register your data manager with the transaction. repoze.tm will abort the transaction if an exception is raised by your application code or lower middleware before it returns a WSGI response. If your application or lower middleware raises an exception, the transaction is aborted.

When a repoze.tm is in the WSGI pipeline, a boolean key is present in the environment (repoze.tm.active). A utility function named isActive can be imported from the repoze.tm package and passed the WSGI environment to check for activation:

      from repoze.tm import isActive
      tm_active = isActive(wsgi_environment)

If an application needs to perform an action after a transaction ends, the "after_end" registry may be used to register a callback. The after_end.register function accepts a callback (accepting no arguments) and a transaction instance:

      from repoze.tm import after_end
      import transaction
      t = transaction.get() # the current transaction
      def func():
          pass # close a connection, etc
      after_end.register(func, t)

after_end callbacks should only be registered when the transaction manager is active, or a memory leak will result (registration cleanup happens only on transaction commit or abort, which is managed by repoze.tm while in the pipeline).

Further Documentation

Many database adapters written for Zope (e.g. for Postgres, MySQL, etc) use this transaction manager, so it should be possible to take a look in these places to see how to implement a more real-world transaction-aware database connector that uses this module in non-Zope applications:

      http://svn.zope.org/ZODB/branches/3.7/src/transaction/

      http://svn.zope.org/ZODB/branches/3.8/src/transaction/

      http://mysql-python.sourceforge.net/ (ZMySQLDA)

      http://www.initd.org/svn/psycopg/psycopg2/trunk/ (ZPsycoPGDA)

A demonstration file that matches the mock data manager I outlined above is also available .