Savepoints

Savepoints provide a way to save to disk intermediate work done during a transaction allowing:

  • partial transaction (subtransaction) rollback (abort)
  • state of saved objects to be freed, freeing on-line memory for other uses

Savepoints make it possible to write atomic subroutines that don’t make top-level transaction commitments.

Applications

To demonstrate how savepoints work with transactions, we’ve provided a sample data manager implementation that provides savepoint support. The primary purpose of this data manager is to provide code that can be read to understand how savepoints work. The secondary purpose is to provide support for demonstrating the correct operation of savepoint support within the transaction system. This data manager is very simple. It provides flat storage of named immutable values, like strings and numbers.

>>> import transaction
>>> from transaction.tests import savepointsample
>>> dm = savepointsample.SampleSavepointDataManager()
>>> dm['name'] = 'bob'

As with other data managers, we can commit changes:

>>> transaction.commit()
>>> dm['name']
'bob'

and abort changes:

>>> dm['name'] = 'sally'
>>> dm['name']
'sally'
>>> transaction.abort()
>>> dm['name']
'bob'

Now, let’s look at an application that manages funds for people. It allows deposits and debits to be entered for multiple people. It accepts a sequence of entries and generates a sequence of status messages. For each entry, it applies the change and then validates the user’s account. If the user’s account is invalid, we roll back the change for that entry. The success or failure of an entry is indicated in the output status. First we’ll initialize some accounts:

>>> dm['bob-balance'] = 0.0
>>> dm['bob-credit'] = 0.0
>>> dm['sally-balance'] = 0.0
>>> dm['sally-credit'] = 100.0
>>> transaction.commit()

Now, we’ll define a validation function to validate an account:

>>> def validate_account(name):
...     if dm[name+'-balance'] + dm[name+'-credit'] < 0:
...         raise ValueError('Overdrawn', name)

And a function to apply entries. If the function fails in some unexpected way, it rolls back all of its changes and prints the error:

>>> def apply_entries(entries):
...     savepoint = transaction.savepoint()
...     try:
...         for name, amount in entries:
...             entry_savepoint = transaction.savepoint()
...             try:
...                 dm[name+'-balance'] += amount
...                 validate_account(name)
...             except ValueError as error:
...                 entry_savepoint.rollback()
...                 print("%s %s" % ('Error', str(error)))
...             else:
...                 print("%s %s" % ('Updated', name))
...     except Exception as error:
...         savepoint.rollback()
...         print("%s" % ('Unexpected exception'))

Now let’s try applying some entries:

>>> apply_entries([
...     ('bob',   10.0),
...     ('sally', 10.0),
...     ('bob',   20.0),
...     ('sally', 10.0),
...     ('bob',   -100.0),
...     ('sally', -100.0),
...     ])
Updated bob
Updated sally
Updated bob
Updated sally
Error ('Overdrawn', 'bob')
Updated sally

>>> dm['bob-balance']
30.0

>>> dm['sally-balance']
-80.0

If we provide entries that cause an unexpected error:

>>> apply_entries([
...     ('bob',   10.0),
...     ('sally', 10.0),
...     ('bob',   '20.0'),
...     ('sally', 10.0),
...     ])
Updated bob
Updated sally
Unexpected exception

Because the apply_entries used a savepoint for the entire function, it was able to rollback the partial changes without rolling back changes made in the previous call to apply_entries:

>>> dm['bob-balance']
30.0

>>> dm['sally-balance']
-80.0

If we now abort the outer transactions, the earlier changes will go away:

>>> transaction.abort()

>>> dm['bob-balance']
0.0

>>> dm['sally-balance']
0.0

Savepoint invalidation

A savepoint can be used any number of times:

>>> dm['bob-balance'] = 100.0
>>> dm['bob-balance']
100.0
>>> savepoint = transaction.savepoint()

>>> dm['bob-balance'] = 200.0
>>> dm['bob-balance']
200.0
>>> savepoint.rollback()
>>> dm['bob-balance']
100.0

>>> savepoint.rollback()  # redundant, but should be harmless
>>> dm['bob-balance']
100.0

>>> dm['bob-balance'] = 300.0
>>> dm['bob-balance']
300.0
>>> savepoint.rollback()
>>> dm['bob-balance']
100.0

However, using a savepoint invalidates any savepoints that come after it:

>>> dm['bob-balance'] = 200.0
>>> dm['bob-balance']
200.0
>>> savepoint1 = transaction.savepoint()

>>> dm['bob-balance'] = 300.0
>>> dm['bob-balance']
300.0
>>> savepoint2 = transaction.savepoint()

>>> savepoint.rollback()
>>> dm['bob-balance']
100.0

>>> savepoint2.rollback() 
Traceback (most recent call last):
...
InvalidSavepointRollbackError: invalidated by a later savepoint

>>> savepoint1.rollback() 
Traceback (most recent call last):
...
InvalidSavepointRollbackError: invalidated by a later savepoint

>>> transaction.abort()

Databases without savepoint support

Normally it’s an error to use savepoints with databases that don’t support savepoints:

>>> dm_no_sp = savepointsample.SampleDataManager()
>>> dm_no_sp['name'] = 'bob'
>>> transaction.commit()
>>> dm_no_sp['name'] = 'sally'
>>> transaction.savepoint() 
Traceback (most recent call last):
...
TypeError: ('Savepoints unsupported', {'name': 'bob'})

>>> transaction.abort()

However, a flag can be passed to the transaction savepoint method to indicate that databases without savepoint support should be tolerated until a savepoint is rolled back. This allows transactions to proceed if there are no reasons to roll back:

>>> dm_no_sp['name'] = 'sally'
>>> savepoint = transaction.savepoint(1)
>>> dm_no_sp['name'] = 'sue'
>>> transaction.commit()
>>> dm_no_sp['name']
'sue'

>>> dm_no_sp['name'] = 'sam'
>>> savepoint = transaction.savepoint(1)
>>> savepoint.rollback() 
Traceback (most recent call last):
...
TypeError: ('Savepoints unsupported', {'name': 'sam'})

Failures

If a failure occurs when creating or rolling back a savepoint, the transaction state will be uncertain and the transaction will become uncommitable. From that point on, most transaction operations, including commit, will fail until the transaction is aborted.

In the previous example, we got an error when we tried to rollback the savepoint. If we try to commit the transaction, the commit will fail:

>>> transaction.commit() 
Traceback (most recent call last):
...
TransactionFailedError: An operation previously failed, with traceback:
...
TypeError: ('Savepoints unsupported', {'name': 'sam'})

We have to abort it to make any progress:

>>> transaction.abort()

Similarly, in our earlier example, where we tried to take a savepoint with a data manager that didn’t support savepoints:

>>> dm_no_sp['name'] = 'sally'
>>> dm['name'] = 'sally'
>>> savepoint = transaction.savepoint() 
Traceback (most recent call last):
...
TypeError: ('Savepoints unsupported', {'name': 'sue'})

>>> transaction.commit() 
Traceback (most recent call last):
...
TransactionFailedError: An operation previously failed, with traceback:
...
TypeError: ('Savepoints unsupported', {'name': 'sue'})


>>> transaction.abort()

After clearing the transaction with an abort, we can get on with new transactions:

>>> dm_no_sp['name'] = 'sally'
>>> dm['name'] = 'sally'
>>> transaction.commit()
>>> dm_no_sp['name']
'sally'
>>> dm['name']
'sally'