Commit 33d3d6f6 authored by Tim Peters's avatar Tim Peters

Checkpointing progress made with Jim so far today.

New interface IResourceManager intends to combine the
best aspects of IDataManager and IDataManagerOriginal,
and will become "the" interface.
New interface ISavePoint captures (along with models in
collaborations.txt) what should be a clean design for a
savepoint/rollback facility that actually works.
parent 51e6ab41
......@@ -18,6 +18,103 @@ $Id$
import zope.interface
class IResourceManager(zope.interface.Interface):
"""Objects that manage resources transactionally.
These objects may manage data for other objects, or they may manage
non-object storages, such as relational databases.
IDataManagerOriginal is the interface currently provided by ZODB
database connections, but the intent is to move to the newer
IDataManager.
"""
# Two-phase commit protocol. These methods are called by the
# ITransaction object associated with the transaction being
# committed.
def tpc_begin(transaction):
"""Begin two-phase commit, to save data changes.
An implementation should do as much work as possible without
making changes permanent. Changes should be made permanent
when tpc_finish is called (or aborted if tpc_abort is called).
The work can be divided between tpc_begin() and tpc_vote(), and
the intent is that tpc_vote() be as fast as possible (to minimize
the period of uncertainty).
transaction is the ITransaction instance associated with the
transaction being committed.
"""
def tpc_vote(transaction):
"""Verify that a resource manager can commit the transaction.
This is the last chance for a resource manager to vote 'no'. A
resource 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.
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.
transaction is the ITransaction instance associated with the
transaction being committed.
All changes made by the current transaction are aborted. Note
that this includes all changes stored in any savepoints that may
be associated with the current transaction.
tpc_abort() can be called at any time, either in or out of the
two-phase commit.
This should never fail.
"""
# The savepoint/rollback API.
def savepoint(transaction):
"""Save partial transaction changes.
There are two purposes:
1) To allow discarding partial changes without discarding all
dhanges.
2) To checkpoint changes to disk that would otherwise live in
memory for the duration of the transaction.
Returns an object implementing ISavePoint2 that can be used
to discard changes made since the savepoint was captured.
An implementation that doesn't support savepoints should implement
this method by returning a savepoint object that raises an
exception when its rollback method is called. The savepoint method
shouldn't raise an error. This way, transactions that create
savepoints can proceed as long as an attempt is never made to roll
back a savepoint.
"""
def discard(transaction):
"""Discard changes within the transaction since the last savepoint.
That means changes made since the last savepoint if one exists, or
since the start of the transaction.
"""
class IDataManagerOriginal(zope.interface.Interface):
"""Objects that manage transactional storage.
......@@ -297,6 +394,35 @@ class ITransaction(zope.interface.Interface):
# XXX is this this allowed to cause an exception here, during
# the two-phase commit, or can it toss data silently?
class ISavePoint(zope.interface.Interface):
"""ISavePoint objects represent partial transaction changes.
Sequences of savepoint objects are associated with transactions,
and with IResourceManagers.
"""
def rollback():
"""Discard changes made after this savepoint.
This includes discarding (call the discard method on) all
subsequent savepoints.
"""
def discard():
"""Discard changes saved by this savepoint.
That means changes made since the immediately preceding
savepoint if one exists, or since the start of the transaction,
until this savepoint.
Once a savepoint has been discarded, it's an error to attempt
to rollback or discard it again.
"""
next_savepoint = zope.interface.Attribute(
"""The next savepoint (later in time), or None if self is the
most recent savepoint.""")
class IRollback(zope.interface.Interface):
def rollback():
......
##############################################################################
#
# Copyright (c) 2004 Zope Corporation and Contributors.
# All Rights Reserved.
#
# This software is subject to the provisions of the Zope Public License,
# Version 2.1 (ZPL). A copy of the ZPL should accompany this distribution.
# THIS SOFTWARE IS PROVIDED "AS IS" AND ANY AND ALL EXPRESS OR IMPLIED
# WARRANTIES ARE DISCLAIMED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
# WARRANTIES OF TITLE, MERCHANTABILITY, AGAINST INFRINGEMENT, AND FITNESS
# FOR A PARTICULAR PURPOSE.
#
##############################################################################
"""Sample objects for use in tests
$Id$
"""
class ResourceManager(object):
"""Sample resource manager.
This class provides a trivial resource-manager implementation and doc
strings to illustrate the protocol and to provide a tool for writing
tests.
Our sample resource manager has state that is updated through an inc
method and through transaction operations.
When we create a sample resource manager:
>>> rm = ResourceManager()
It has two pieces state, state and delta, both initialized to 0:
>>> rm.state
0
>>> rm.delta
0
state is meant to model committed state, while delta represents
tentative changes within a transaction. We change the state by
calling inc:
>>> rm.inc()
which updates delta:
>>> rm.delta
1
but state isn't changed until we commit the transaction:
>>> rm.state
0
To commit the changes, we use 2-phase commit. We execute the first
stage by calling prepare. We need to pass a transation. Our
sample resource managers don't really use the transactions for much,
so we'll be lazy and use strings for transactions. The sample
resource manager updates the state when we call tpc_vote:
>>> t1 = '1'
>>> rm.tpc_begin(t1)
>>> rm.state, rm.delta
(0, 1)
>>> rm.tpc_vote(t1)
>>> rm.state, rm.delta
(1, 1)
Now if we call tpc_finish:
>>> rm.tpc_finish(t1)
Our changes are "permanent". The state reflects the changes and the
delta has been reset to 0.
>>> rm.state, rm.delta
(1, 0)
"""
def __init__(self):
self.state = 0
self.sp = 0
self.transaction = None
self.delta = 0
self.txn_state = None
def _check_state(self, *ok_states):
if self.txn_state not in ok_states:
raise ValueError("txn in state %r but expected one of %r" %
(self.txn_state, ok_states))
def _checkTransaction(self, transaction):
if (transaction is not self.transaction
and self.transaction is not None):
raise TypeError("Transaction missmatch",
transaction, self.transaction)
def inc(self, n=1):
self.delta += n
def tpc_begin(self, transaction):
"""Prepare to commit data.
>>> rm = ResourceManager()
>>> rm.inc()
>>> t1 = '1'
>>> rm.tpc_begin(t1)
>>> rm.tpc_vote(t1)
>>> rm.tpc_finish(t1)
>>> rm.state
1
>>> rm.inc()
>>> t2 = '2'
>>> rm.tpc_begin(t2)
>>> rm.tpc_vote(t2)
>>> rm.tpc_abort(t2)
>>> rm.state
1
It is an error to call tpc_begin more than once without completing
two-phase commit:
>>> rm.tpc_begin(t1)
>>> rm.tpc_begin(t1)
Traceback (most recent call last):
...
ValueError: txn in state 'tpc_begin' but expected one of (None,)
>>> rm.tpc_abort(t1)
If there was a preceeding savepoint, the transaction must match:
>>> rollback = rm.savepoint(t1)
>>> rm.tpc_begin(t2)
Traceback (most recent call last):
,,,
TypeError: ('Transaction missmatch', '2', '1')
>>> rm.tpc_begin(t1)
"""
self._checkTransaction(transaction)
self._check_state(None)
self.transaction = transaction
self.txn_state = 'tpc_begin'
def tpc_vote(self, 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.
"""
self._checkTransaction(transaction)
self._check_state('tpc_begin')
self.state += self.delta
self.txn_state = 'tpc_vote'
def tpc_finish(self, transaction):
"""Complete two-phase commit
>>> rm = ResourceManager()
>>> rm.state
0
>>> rm.inc()
We start two-phase commit by calling prepare:
>>> t1 = '1'
>>> rm.tpc_begin(t1)
>>> rm.tpc_vote(t1)
We complete it by calling tpc_finish:
>>> rm.tpc_finish(t1)
>>> rm.state
1
It is an error ro call tpc_finish without calling tpc_vote:
>>> rm.inc()
>>> t2 = '2'
>>> rm.tpc_begin(t2)
>>> rm.tpc_finish(t2)
Traceback (most recent call last):
...
ValueError: txn in state 'tpc_begin' but expected one of ('tpc_vote',)
>>> rm.tpc_abort(t2) # clean slate
>>> rm.tpc_begin(t2)
>>> rm.tpc_vote(t2)
>>> rm.tpc_finish(t2)
Of course, the transactions given to tpc_begin and tpc_finish must
be the same:
>>> rm.inc()
>>> t3 = '3'
>>> rm.tpc_begin(t3)
>>> rm.tpc_vote(t3)
>>> rm.tpc_finish(t2)
Traceback (most recent call last):
...
TypeError: ('Transaction missmatch', '2', '3')
"""
self._checkTransaction(transaction)
self._check_state('tpc_vote')
self.delta = 0
self.transaction = None
self.prepared = False
self.txn_state = None
def tpc_abort(self, transaction):
"""Abort a transaction
The abort method can be called before two-phase commit to
throw away work done in the transaction:
>>> rm = ResourceManager()
>>> rm.inc()
>>> rm.state, rm.delta
(0, 1)
>>> t1 = '1'
>>> rm.tpc_abort(t1)
>>> rm.state, rm.delta
(0, 0)
The abort method also throws away work done in savepoints:
>>> rm.inc()
>>> r = rm.savepoint(t1)
>>> rm.inc()
>>> r = rm.savepoint(t1)
>>> rm.state, rm.delta
(0, 2)
>>> rm.tpc_abort(t1)
>>> rm.state, rm.delta
(0, 0)
If savepoints are used, abort must be passed the same
transaction:
>>> rm.inc()
>>> r = rm.savepoint(t1)
>>> t2 = '2'
>>> rm.tpc_abort(t2)
Traceback (most recent call last):
...
TypeError: ('Transaction missmatch', '2', '1')
>>> rm.tpc_abort(t1)
The abort method is also used to abort a two-phase commit:
>>> rm.inc()
>>> rm.state, rm.delta
(0, 1)
>>> rm.tpc_begin(t1)
>>> rm.state, rm.delta
(0, 1)
>>> rm.tpc_vote(t1)
>>> rm.state, rm.delta
(1, 1)
>>> rm.tpc_abort(t1)
>>> rm.state, rm.delta
(0, 0)
Of course, the transactions passed to prepare and abort must
match:
>>> rm.tpc_begin(t1)
>>> rm.tpc_abort(t2)
Traceback (most recent call last):
...
TypeError: ('Transaction missmatch', '2', '1')
>>> rm.tpc_abort(t1)
This should never fail.
"""
self._checkTransaction(transaction)
if self.transaction is not None:
self.transaction = None
if self.txn_state == 'tpc_vote':
self.state -= self.delta
self.txn_state = None
self.delta = 0
def savepoint(self, transaction):
"""Provide the ability to rollback transaction state
Savepoints provide a way to:
- Save partial transaction work. For some resource managers, this
could allow resources to be used more efficiently.
- Provide the ability to revert state to a point in a
transaction without aborting the entire transaction. In
other words, savepoints support partial aborts.
Savepoints don't use two-phase commit. If there are errors in
setting or rolling back to savepoints, the application should
abort the containing transaction. This is *not* the
responsibility of the resource manager.
Savepoints are always associated with a transaction. Any work
done in a savepoint's transaction is tentative until the
transaction is committed using two-phase commit.
>>> rm = ResourceManager()
>>> rm.inc()
>>> t1 = '1'
>>> r = rm.savepoint(t1)
>>> rm.state, rm.delta
(0, 1)
>>> rm.inc()
>>> rm.state, rm.delta
(0, 2)
>>> r.rollback()
>>> rm.state, rm.delta
(0, 1)
>>> rm.tpc_begin(t1)
>>> rm.tpc_vote(t1)
>>> rm.tpc_finish(t1)
>>> rm.state, rm.delta
(1, 0)
Savepoints must have the same transaction:
>>> r1 = rm.savepoint(t1)
>>> rm.state, rm.delta
(1, 0)
>>> rm.inc()
>>> rm.state, rm.delta
(1, 1)
>>> t2 = '2'
>>> r2 = rm.savepoint(t2)
Traceback (most recent call last):
...
TypeError: ('Transaction missmatch', '2', '1')
>>> r2 = rm.savepoint(t1)
>>> rm.inc()
>>> rm.state, rm.delta
(1, 2)
If we rollback to an earlier savepoint, we discard all work
done later:
>>> r1.rollback()
>>> rm.state, rm.delta
(1, 0)
and we can no longer rollback to the later savepoint:
>>> r2.rollback()
Traceback (most recent call last):
...
TypeError: ('Attempt to roll back to invalid save point', 3, 2)
We can roll back to a savepoint as often as we like:
>>> r1.rollback()
>>> r1.rollback()
>>> r1.rollback()
>>> rm.state, rm.delta
(1, 0)
>>> rm.inc()
>>> rm.inc()
>>> rm.inc()
>>> rm.state, rm.delta
(1, 3)
>>> r1.rollback()
>>> rm.state, rm.delta
(1, 0)
But we can't rollback to a savepoint after it has been
committed:
>>> rm.tpc_begin(t1)
>>> rm.tpc_vote(t1)
>>> rm.tpc_finish(t1)
>>> r1.rollback()
Traceback (most recent call last):
...
TypeError: Attempt to rollback stale rollback
"""
if self.txn_state is not None:
raise TypeError("Can't get savepoint during two-phase commit")
self._checkTransaction(transaction)
self.transaction = transaction
self.sp += 1
return SavePoint(self)
def discard(self, transaction):
pass
class SavePoint(object):
def __init__(self, rm):
self.rm = rm
self.sp = rm.sp
self.delta = rm.delta
self.transaction = rm.transaction
def rollback(self):
if self.transaction is not self.rm.transaction:
raise TypeError("Attempt to rollback stale rollback")
if self.rm.sp < self.sp:
raise TypeError("Attempt to roll back to invalid save point",
self.sp, self.rm.sp)
self.rm.sp = self.sp
self.rm.delta = self.delta
def discard(self):
pass
def test_suite():
from doctest import DocTestSuite
return DocTestSuite()
if __name__ == '__main__':
unittest.main()
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment