View of /Trellis/STM-Observer.txt

Parent Directory

Revision Log

The new algorithm lands - version bump to 0.6a1dev.
Too many changes to list!

=================================================
Software Transactional Memory (STM) And Observers
=================================================

The Trellis is built on a simple Python STM (Software Transactional Memory)
and "Observer Pattern" implementation.  This document specifies how that
implementation works, and tests it.

You should read this document if you plan to implement your own Trellis cell
types or other custom data structures, or if you just want to know how things
work "under the hood".


STM History
===========

An STM's job is to manage "atomic" changes to objects: i.e., multiple changes
that must happen as a unit, or else be rolled back.

To do this, the STM must have a **history**: a record of the actions taken
within a given "atomic" change set.  The PEAK implementation of an STM history
is found in the ``peak.events.stm`` module::

    >>> from peak.events.stm import STMHistory

    >>> hist = STMHistory()


A history object's ``atomically()`` method invokes a function as an atomic
operation, and its ``active`` attribute indicates whether it is currently
performing an atomic operation::

    >>> def is_active():
    ...     print "Active?", hist.active


    >>> is_active()
    Active? False

    >>> hist.atomically(is_active)
    Active? True

    >>> is_active()
    Active? False

Nested calls to ``atomically()`` simply execute the function, without affecting
the history::

    >>> def nested_operation():
    ...     hist.atomically(is_active)
    ...     is_active()

    >>> hist.atomically(nested_operation)
    Active? True
    Active? True


Commit/Abort Notices
--------------------

In order to get notification of commits or aborts, you can register Python
"context managers" with the history, using the ``manage()`` method.  The
``manage()`` method takes a context manager and calls its ``__enter__()``
method immediately, then calls its ``__exit__()`` method when the current
atomic operation is completed.  This allows things like locks or other
resources to be acquired as the operation progresses, and then get
automatically released when the operation is completed::

    >>> class DemoManager(object):
    ...     def __init__(self, num):
    ...         self.num = num
    ...     def __enter__(self):
    ...         print "Manager", self.num, "entering"
    ...     def __exit__(self, typ, val, tb):
    ...         print "Manager", self.num, "exiting", typ, val, tb

    >>> hist.manage(DemoManager(1))
    Traceback (most recent call last):
      ...
    AssertionError: Can't manage without active history

    >>> hist.atomically(hist.manage, DemoManager(2))
    Manager 2 entering
    Manager 2 exiting None None None

The same context manager can be passed to ``manage`` repeatedly, but its
enter/exit methods will only be called once::

    >>> def multi_manage():
    ...     mgr = DemoManager(3)
    ...     hist.manage(mgr)
    ...     hist.manage(mgr)

    >>> hist.atomically(multi_manage)
    Manager 3 entering
    Manager 3 exiting None None None

And if multiple context managers are registered, their ``__exit__`` methods
are called in the opposite order from their ``__enter__`` methods::

    >>> def multi_manage():
    ...     hist.manage(DemoManager(4))
    ...     hist.manage(DemoManager(5))

    >>> hist.atomically(multi_manage)
    Manager 4 entering
    Manager 5 entering
    Manager 5 exiting None None None
    Manager 4 exiting None None None

The ``__exit__()`` method is normally called with three ``None`` values, unless
an exception occurs during the operation.  In that case, the ``sys.exc_info()``
of the exception is passed in to the manager(s), before the exception is
re-raised::

    >>> def do_error():
    ...     hist.manage(DemoManager(6))
    ...     raise TypeError("Testing!")

    >>> try:
    ...     hist.atomically(do_error)
    ... except TypeError:
    ...     print "caught exception"
    Manager 6 entering
    Manager 6 exiting <type '...TypeError'> Testing! <traceback object...>
    caught exception

The ``__exit__()`` method should not raise an error.  Also note that, unlike
normal Python context managers, the return value of ``__exit__()`` is ignored.
(In other words, STM context managers cannot cause an exception to be ignored.)

If an ``__exit__()`` method *does* raise an error, however, subsequent context
managers will be passed the type, value, and traceback of the failing manager,
and the exception will be reraised::

    >>> class ErrorManager(DemoManager):
    ...     def __exit__(self, typ, val, tb):
    ...         super(ErrorManager, self).__exit__(typ, val, tb)
    ...         raise RuntimeError("Haha!")

    >>> def manage_with_error():
    ...     hist.manage(DemoManager(7))
    ...     hist.manage(ErrorManager("error"))
    ...     hist.manage(DemoManager(8))

    >>> try:
    ...     hist.atomically(manage_with_error)
    ... except RuntimeError:
    ...     print "caught exception"
    Manager 7 entering
    Manager error entering
    Manager 8 entering
    Manager 8 exiting None None None
    Manager error exiting None None None
    Manager 7 exiting <type '...RuntimeError'> Haha! <traceback object...>
    caught exception

In other words, all context managers are guaranteed to have their ``__exit__``
methods called, even if one fails.  The exception that comes out of the
``atomically()`` call will be the most recently-raised exception.

Last, but not least, history objects have a ``in_cleanup`` attribute that
indicates whether they are currently in the process of comitting or aborting
an operation.  This can be useful if context managers might call code that
needs to behave differently during a commit/abort than during an atomic
operation::

    >>> hist.in_cleanup
    False

    >>> class CleanupManager:
    ...     def __enter__(self):
    ...         print "on entry:", hist.in_cleanup
    ...     def __exit__(self, typ, val, tb):
    ...         print "on exit:", hist.in_cleanup

    >>> hist.atomically(hist.manage, CleanupManager())
    on entry: False
    on exit: True


Undo, Rollback and Save Points
------------------------------

While you can use context managers to implement some forms of commit/rollback,
it's easier for most things to use a history object's "undo" log.  The log
records "undo actions": functions (and optional positional arguments) that can
be used to undo whatever operations have been done so far.  For example::

    >>> def undoing(msg):
    ...     print "undoing", msg

    >>> def with_undo():
    ...     hist.on_undo(undoing, "op 1")
    ...     hist.on_undo(undoing, "op 2")

    >>> hist.atomically(with_undo)  # success, nothing gets undone

Nothing happened here, because the operation completed successfully.  But if
an error occurs, the undo functions are called in reverse order::

    >>> def with_undo():
    ...     hist.on_undo(undoing, "op 1")
    ...     hist.on_undo(undoing, "op 2")
    ...     raise TypeError("foo")

    >>> try:
    ...     hist.atomically(with_undo)
    ... except TypeError:
    ...     print "caught exception"
    undoing op 2
    undoing op 1
    caught exception

But this does NOT happen if the error occurs in a manager's ``__exit__()``
method::

    >>> def with_undo():
    ...     hist.manage(ErrorManager("error"))
    ...     hist.on_undo(undoing, "op 1")
    ...     hist.on_undo(undoing, "op 2")

    >>> try:
    ...     hist.atomically(with_undo)
    ... except RuntimeError:
    ...     print "caught exception"
    Manager error entering
    Manager error exiting None None None
    caught exception

(That's because managers may be used to control locking or other context setups
that need to still be in place for the undo's to be safe.)

Note, by the way, that undo functions must NEVER raise errors, under any
circumstances.  If they do, any undo functions that have *not* been called
yet, will never *be* called.

In addition to the automatic rollback on error, you can also record savepoints
within an atomic operation, and then rollback to that savepoint at any time
later::

    >>> def with_savepoint():
    ...     hist.on_undo(undoing, "op 1")
    ...     sp = hist.savepoint()
    ...     hist.on_undo(undoing, "op 2")
    ...     hist.on_undo(undoing, "op 3")
    ...     hist.rollback_to(sp)

    >>> hist.atomically(with_savepoint)
    undoing op 3
    undoing op 2


Commit Callbacks
----------------

If you want to defer an action until the overal operation is ready to commit,
you can use the ``on_commit(func, *args)`` method::

    >>> def do_something(msg):
    ...     print "committing", msg

    >>> def do_with_commit():
    ...     hist.on_commit(do_something, "hello!")
    ...     print "registered commit operation"

    >>> hist.atomically(do_with_commit)
    registered commit operation
    committing hello!

Commit actions are called in the order they are registered (before any
managers) and their registrations are subject to undo::

    >>> def do_with_commit_and_undo():
    ...     hist.manage(DemoManager('test'))
    ...     hist.on_commit(do_something, 1)
    ...     sp = hist.savepoint()
    ...     hist.on_commit(do_something, 2)
    ...     hist.rollback_to(sp)
    ...     hist.on_commit(do_something, 3)

    >>> hist.atomically(do_with_commit_and_undo)
    Manager test entering
    committing 1
    committing 3
    Manager test exiting None None None

Commit actions are also allowed to record undo actions, in case there is
an error later in the commit process::

    >>> def do_with_commit_and_undo():
    ...     def f1():
    ...         print "f1 running"
    ...         hist.on_undo(f3)
    ...     def f2():
    ...         print "f2 running"
    ...         raise AssertionError
    ...     def f3():
    ...         print "f3 running"
    ...     hist.on_commit(f1)
    ...     hist.on_commit(f2)

    >>> try:
    ...     hist.atomically(do_with_commit_and_undo)
    ... except AssertionError:
    ...     print "caught exception"
    f1 running
    f2 running
    f3 running
    caught exception

And of course, commit actions are not run if an error occurs before they have
a chance to run:

    >>> def do_no_commit():
    ...     hist.on_commit(do_something, "should not happen")
    ...     raise AssertionError
    >>> try:
    ...     hist.atomically(do_no_commit)
    ... except AssertionError:
    ...     print "caught exception"
    caught exception

    
Logged Setattr
--------------

As a convenience, you can use the ``setattr()`` method of history objects to
automatically log an undo function to restore the old value of the attribute
being set::

    >>> class SomeObject:
    ...     pass

    >>> s1 = SomeObject()
    >>> s1.foo = 'bar'

    >>> def setattr_normally():
    ...     hist.change_attr(s1, 'foo', "baz")

    >>> hist.atomically(setattr_normally)
    >>> s1.foo
    'baz'

    >>> def setattr_rollback():
    ...     hist.change_attr(s1, 'foo', "spam")
    ...     raise TypeError

    >>> hist.atomically(setattr_rollback)
    Traceback (most recent call last):
      ...
    TypeError
    
    >>> s1.foo      # result is unchanged after rollback
    'baz'

As you can see, this saves you the work of recording the undo operation
manually.


The Observer Framework
======================

Above and beyond the bare STM system, the Trellis also needs a strong
"observer" system that manages subjects, listeners, and the links between
them::

    >>> from peak.events import stm


Subjects, Listeners, and Links
------------------------------

The ``AbstractSubject`` and ``AbstractListener`` base classes are used to
implement the observer pattern.  They must be subclassed to be used.  If you
will be using a lot of instances, you can conserve memory by using ``__slots__``
in your subclass to store the needed attributes, as shown::

    >>> class Listener(stm.AbstractListener):
    ...     __slots__ = '__weakref__', 'layer', 'next_subject'

    >>> class Subject(stm.AbstractSubject):
    ...     __slots__ = 'layer', 'next_listener'

Listeners and Subjects can be many-to-many related.  That is, each listener may
listen to zero or more subjects, and each subject can have zero or more
listeners.  These relationships are created using ``Link`` objects::

    >>> l1 = Listener()
    >>> l2 = Listener()
    >>> s1 = Subject()
    >>> s2 = Subject()

    >>> link11 = stm.Link(s1, l1)
    >>> link21 = stm.Link(s2, l1)
    >>> link12 = stm.Link(s1, l2)
    >>> link22 = stm.Link(s2, l2)

    >>> list(l1.iter_subjects())==list(l2.iter_subjects())==[s2, s1]
    True
    >>> list(s1.iter_listeners())==list(s2.iter_listeners())==[l2, l1]
    True

The relationship between a subject and its listeners is "weak"; that is, if
a listener goes out of existence, all of the relevant links are unlinked from
the relevant subjects' listener chains::

    >>> del l1
    >>> list(s1.iter_listeners())==list(s2.iter_listeners())==[l2]
    True

You can also explicitly break a link by calling its ``.unlink()`` method::

    >>> link12.unlink()
    >>> list(s1.iter_listeners())
    []
    >>> list(s2.iter_listeners())==[l2]
    True

Normally, however, you will not create or manage links explicitly, as they
are automatically managed by the controller.