==================
Persistent Modules
==================

Document Overview
-----------------

This document seeks to capture technical information about persistent modules
to guide and document their design.

Goals
-----

These goals largely come from Zope 3.  It would be worth while considering
other applications.

- Persistent modules are used to support management of software using the
  ZODB.

- Software can be updated using network clients, such as web browsers and
  file-synchonozation tools.

- Application-server clusters can be updated transactionally without requiring
  server restarts.

- Persistent modules leverage a familiar model, modules, for managing Python
  software.

- Persistent modules can be synchronized to a file-system using the Zope
  file-system synchronization framework.  Persistent modules are synchronized
  for purposes including:

  o Use of traditional tools such as editors and code-analysis tools

  o Revision control

  Ideally, the file-system representation would consist of a Python source
  file.

Use cases
---------

- Create classes and functions that implement Zope 3 components.

  o Utility, Adapter, View, and service classes and factories.

  o Content components, which are typically persistent and/or
    pickleable.

- Define interfaces, including schema

- Import classes, functions, and interfaces from other modules.

- Import classes, functions, and interfaces from other persistent objects. For
  example, an adapter registration object might have a direct reference to a
  persistent-module-defined class.

- Change module source

  - Changes are reflected in module state

  - Changes are reflected in objects imported into other modules.

- Synchronize modules with a file-system representation.

Edge cases
----------

  ???

Fundamental dilema
------------------

Python modules were not designed to change at run time.  The source of a
Python module normally doesn't change while a Python program is running.
There is a crude reload tool that allows modules to be manually reloaded to
handle source changes.

Python modules contain mutable state.  A module has a dictionary that may be
mutated by application code. It may contain mutable data that is modified at
run time.  This is typeically used to implement global registries.

When a module is reloaded, it is reexecuted with a dictionary that includes
the results of the previous execution.

Programs using the ZODB may be said to have logical lifetimes that exceed the
lifetimes of individual processes. In addition, the program might exist as
multiple individual processes with overlapping run-times.

The lifetime of a persistent program is long enough that it is likely that
module source code will change during the life time of the program.

Issues
------

Should the state of a module be represented soley by the module source?

Consider the possibilities:

A. Module state is represented soley by it's source.

- This would be a departure from the behavior of standard Python modules.
  Standard Python modules retain a module dictionary that is not overwritten
  by reloads.  Python modules may be mutated from outside and may contain
  mutable data structures that are modified at run time.

  OTOH, a regular module's state is not persistent or shared accross
  processes.

  For standard Python modules, one could view the module source as an
  expression of the initial state of the module. (This isn't quite right
  either, since some modules are written in such a way that they anticipate
  module reloads.)

- Deleting variables from a module's source that have been imported by other
  modules or objects will cause the imported values to become disconnected
  from the module's source.  Even if the variables are added back later, the
  previously-imported values will be disconnected.

  It is tempting to introduce a data structure to record imports make from a
  module.  For example, suppose module M1 imports X from M2.  It's tempting to
  record that fact in M2, so that we disallow M2 to be removed or to be
  changed in such a way that M2 no-longer defines X.  Unfortunately, that
  would introduce state that isn't captured by my M2's source.

- Persistent modules could only be used for software. You wouldn't be able to
  use them to store mutable data, such as registries or counters, that are
  updated outside of the execution of the module source.

B. Module state isn't represented soley by it's source.

 - It would become possible to allow mutable data, such as registries in
   persistent modules.

 - It could be very difficult to see what a module's state is.  If a module
   contained mutable data, you'd need some way to get to that data so you
   could inspect and manipulate it.

 - When a module is synchronized to the file system, you'd need to syncronize
   it's source and you'd also need to synchronize it's contents in some
   way. Synchronization of the contents could be done using an XML pickle, but
   management of the data using file-system-based tools would be cumbersome.

   You'd end up with data duplicated between the two representations.  It
   would be cumbersome to manage the duplicated data in a consistent way.

C. Module state is represented soley by it's source, but allow additional meta
   data.

   This is the same as option A, except we support meta-data management.  The
   meta data could include dependency information. We'd keep track of external
   usage (import) of module variables to influence whether deletion of the
   module or defined variables is allowed, or whether to issue warnings when
   variables are deleted.

   Note that the management of the meta data need not be the responsibility of
   the module. This could be done via some application-defined facility, in
   which case, the module facility would need to provide an api for
   implimenting hooks for managing this information.

Special cases
-------------

This section contains examples that may introduce challenges for persistent
modules or that might motivate or highlight issues described above,

- Persistent classes

  Persistent classes include data that are not represented by the class
  sources.  A class caches slot definitions inherited from base classes.  This
  is information that is only indirectly represented by it's source.
  Similarly, a class manages a collection of it's subclasses.  This allows a
  class to invalidate cached slots in subclasses when a new slot definition is
  assigned (via a setattr).  The cached slots and collection of subclasses is
  not part of a persistent class' state.  It isn't saved in the database, but
  is recomputed when the class is loaded into memory or when it's subclasses
  are loaded into memory.

  Consider two persistent modules, M1, which defines class C1, and M2, which
  defines class C2.  C2 subclasses C1.  C1 defines a __getitem__ slot, which
  is inherited and cached by C2.

  Suppose we have a process, P1, which has M1 and M2 in memory.  C2 in P1 has
  a (cached) __getitem__ slot filled with the definition inherited from C1 in
  P1.  C1 in P1 has C2 in it's collection of subclasses. In P1, we modify M1,
  by editing and recompiling its source.  When we recompile M1's source, we'll
  update the state of C1 by calling it's __setstate__ method, passing the new
  class dictionary.  The __setstate__ method will, in turn, use setattr to
  assign the values from the new dictionary.  If we set a slot attribute, the
  __setattribute__ method in C1 will notify each of it's subclasses that the
  slot has changed.  Now, suppose that we've added a __len__ slot definition
  when we modified the source.  When we set the __len__ attribute in C1, C2
  will be notified that there is a new slot definition for __len__.

  Suppose we have a process P2, which also has M1 and M2 loaded into memory.
  As in P1, C2 in P2 caches the __getitem__ slot and C1 in P2 has C2 in P2 in
  it's collection of subclasses.  Now, when M1 in P1 is modified and the
  corresponding transaction is committed, an invalidation for M1 and all of
  the persistent objects it defines, including C1, is sent to all other
  processes. When P2 gets the invalidation for C1, it invalidates C1. It
  happens that persistent classes are not allowed to be ghosts.  When a
  persistent class is invalidated, it immediately reloads it's state, rather
  than converting itself into a ghost. When C2's state is reloaded in P2, we
  assign it's attributes from the new class dictionary. When we assign slots,
  we notify it's subclasses, including C2 in P2.

  Suppose we have a process P3, that only has M1 in memory.  In P3, M2 is not
  in memory, nor are any of it's subobjects.  In P3, C2 is not in the
  collection of subclasses of C1, because C2 is not in memory and the
  collection of subclasses is volatile data for C1.  When we modify C1 in P1
  and commit the transaction, the state of C1 in P3 will be updated, but the
  state of C2 is not affected in P3, because it's not in memory.

  Finally, consider a process, P4 that has M2, but not M1 in memory.  M2 is
  not a ghost, so C2 is in memory. Now, since C2 is in memory, C1 must be in
  memory, even though M1 is not in memory, because C2 has a reference to C1.
  Further, C1 cannot be a ghost, because persistent classes are not allowed to
  be ghosts. When we commit the transation in P1 that updates M1, an
  invalidation for C1 is sent to P4 and C1 is updated.  When C1 is updated,
  it's subclasses (in P4), including C2 are notified, so that their cached
  slot definitions are updated.

  When we modify M1, all copies in memory of C1 and C2 are updated properly,
  even though the data they cache is not cached persistently. This works, and
  only works, because persistent classes are never ghosts.  If a class could
  be a ghost, then invalidating it would have not effect and non-ghost
  dependent classes would not be updated.

- Persistent interfaces

  Like classes, Zope interfaces cache certain information.  An interface
  maintains a set of all of the interfaces that it extends.  In addition,
  interfaces maintain a collection of all of their sub-interfaces.  The
  collection of subinterfaces is used to notify sub=interfaces when an
  interface changes.

  (Interfaces are a special case of a more general class of objects, called
   "specifications", that include both interfaces and interface declareations.
   Similar caching is performed for other specifications and related data
   structures. To simplify the discussion, however, we'll limit ourselves to
   interfaces.)

  When designing persistent interfaces, we have alternative approaches to
  consider:

  A. We could take the same approach as that taken with persistent classes.
     We would not save cached data persistently.  We would compute it as
     objects are moved into memory.

     To take this approach, we'd need to also make persistent interfaces
     non-ghostifiable.  This is necessary to properly propigate object
     changes.

     One could argue that non-ghostifiability if classes is a necessary wart
     forced on us by details of Python classes that are beyond our control,
     and that we should avoid creating new kinds of objects that require
     non-ghostifiability.

  B. We could store the cached data persistently.  For example, we could store
     the set of extended interfaces and the set of subinterfaces in persistent
     dictionaries.

     A significant disadvantage of this approach is that persistent interfaces
     would accumulate state is that not refelcted in their source code,
     however, it's worth noting that, while the dependency and cache data
     cannot be derived from a single module source, it *can* be derived from
     the sources of all of the modules in the system.  We can implement
     persistent interface in such a way that execution of module code causes
     all dependcies among module-defined interfaces to be recomputed
     correctly.

     (This is, to me, Jim, an interesting case: state that can be computed
      during deserialization from other serialized state. This should not be
      surprising, as we are essentially talking about cached data used for
      optimization purposes.)

Proposals
---------

- A module's state must be reprersented, directly or indirectly, by it's
  source.  The state may also include information, such as caching data, that
  is derivable from it's source-represented state.

  It is unclear if or how we will enforce this.  Perhaps it will be just a
  guideline.  The module-synchronization adapters used in Zope will only
  synchronize the module source.  If a module defines state that is not
  represented by or derivable from it's source, then that data will be lost in
  synchronization.  Of course, applications that don't use the synchronization
  framework would be unaffected by this limitation. Alternatively, one could
  develop custom module-synchronization adapters that handled extra module
  data, however, development of such adapters will be outside the scope of the
  Zope project.

Notes
-----

- When we invalidate a persistent class, we need to delete all of the
  attributes defined by it's old dictionary that are not defined by the new
  class dictionary.

