[ZODB-Dev] breaking out the transaction module from ZODB

[Jim Fulton]

On Nov 7, 2007, at 11:38 PM, Chris McDonough wrote:

> I've begun work on breaking out the transaction module so it can be  
> used independently from ZODB.

Great!

>
> Here's what I've done so far:
>
> - I've moved TransactionError and TransactionFailedError from  
> ZODB.POSException into transaction.interfaces, e.g.:
>
>   class TransactionError(Exception):
>       """An error occurred due to normal transaction processing."""

+1

>   class TransactionFailedError(Exception):
>       """Cannot perform an operation on a transaction that  
> previously failed.
>
>       An attempt was made to commit a transaction, or to join a  
> transaction,
>       but this transaction previously raised an exception during an  
> attempt
>       to commit it.  The transaction must be explicitly aborted,  
> either by
>       invoking abort() on the transaction, or begin() on its  
> transaction
>       manager.
>       """

Why not subclass TransactionError?

>
> - I've caused ZODB.POSException to add the POSError base class to  
> both TransactionError and TransactionFailedError
>   after importing them from transaction.interfaces, e.g.:
>
>   from transaction.interfaces import TransactionError
>   from transaction.interfaces import TransactionFailedError
>
>   # We want to be able to distribute the transaction module  
> independent
>   # from ZODB but we need to maintain backwards compatibility with  
> older
>   # ZODB releases, where TransactionError and TransactionFailedError
>   # were actually defined within ZODB.POSException, and inherited from
>   # POSError.  With this solution, if ZODB is present,  
> TransactionError
>   # and TransactionFailedError will have POSError as a base class.  If
>   # ZODB is not present, they won't.  Thanks to Ian Bicking for
>   # suggesting this solution; as ugly as it is, it does the job.
>
>   TransactionError.__bases__ += (POSError,)
>   TransactionFailedError.__bases__ += (POSError,)

Is this *really* necessary?  It's obviously a bit evil.  Let's  
explore alternatives to this:

1. Just don't do it.  I'd be a bit surprised if there was code  
actually catching POSError.

2,. If we really (really really) need to support catching POSError,  
then perhaps we should mobe POSError
      to the transaction package.


> - I've created a zc.zodbutils package that is essentially the code  
> that currently lives in
>   the ZODB.utils module; I've also moved the TimeStamp.c code that  
> currently lives
>   in 'persistent' into it.  A stub ZODB.utils module exists that  
> just does
>   "from zc.zodbutils import *", and in the persistent package's  
> __init__.py, I
>   do "from zc.zodbutils import TimeStamp" for backwards compatibility.

I'd rather not do this.  Let's be a bit more selective here.  The  
number of imports from ZODB are pretty limited. Many of them should  
move to transaction.  Some of them are just test utilities that can  
be duplicated.

I think the biggest challenge is WeakSet.  This could be broken out  
into a separate package, but I think it's not as general as its name  
implies and should probably just be moved to transaction.

> The intention is that the "transaction" distribution will depend  
> only on zc.zodbutils (as will of course the ZODB distro, along with  
> its other current dependencies plus the transaction distribution).

I think this is overly complicated.  Let's just move or copy a few  
things to transaction.


>   I'm wondering about version numbering and naming for the separate  
> packages.. I suspect we shouldn't try to marry the transaction  
> distribution version number to the ZODB distribution version number  
> because they really won't be tied together that way.

Agreed,

>   Maybe just start transaction at "1.0" or something.

Yup.

>   And I'm thinking that the transaction distribution should be  
> named just "transaction".

Yes, unless we decide to move the package.  I think transaction is a  
bit presumptuous. :)

There is a more important issue that also suggests that moving the  
package.  A potential danger with distutils is that different  
distributions can provide the same Python package, which is a recipe  
for extreme confusion at best.  Imagine someone installing  
transaction 1.0 and ZODB 3.8.

I'd be *very* tempted to start a "z" namespace package (as I wish I'd  
done a long time ago :) and put it there.  Granted that claiming "z"  
is also a bit presumptuous, but I think we'd have a reasonable clam  
to it. :)  Moving the package avoids accidentally installing 2  
transaction modules at the same time.

>   And the name "zc.zodbutils" is just a placeholder, suggestions  
> from interested parties would be helpful.

Let's not even do this.  See above,

> I haven't adjusted any imports in tests, nor have I repackaged the  
> transaction module using setuptools yet.  I wanted to get a sense  
> of whether folks thought what I've done so far is reasonable or if  
> you might have done it differently.

Thanks for working on this.  See my comments above.

Jim

--
Jim Fulton
Zope Corporation