[DB-SIG] DBAPI two phase commit implementation in psycopg2

[James Henstridge]

On Sat, Sep 25, 2010 at 12:04 AM, Daniele Varrazzo
<daniele.varrazzo at gmail.com> wrote:
> On Fri, Sep 24, 2010 at 9:48 AM, M.-A. Lemburg <mal at egenix.com> wrote:
>
>> I'd also look around to check how other tools that interoperate with
>> PG in two-phase commits handle this. XA is a widely used standard
>> in the industry, so I assume the problem must have popped up
>> elsewhere as well.
>
> MySQL uses a XA model, so the mapping is direct. Other high profile
> databases hide the commands they use under tons of APIs and I've not
> been able to find implementation details.
>
> I've found instead the JDBC implementation of the XA - PG mapper [1]:
> it is probably a good idea to use the same format to allow some form
> of interoperation between tools. They use str(format_id) + '_' +
> Base64(gtrid) + '_' + Base64(bqual) and on recover they refuse to work
> on anything that doesn't follow this model. I don't agree on the
> latter point because I think the driver should allow the user to
> leverage everything the database permits, so I will probably find a
> way to parse back a generic string into a XID. But I guess this is an
> implementation detail better discussed on the psycopg mailing lists.
>
> Thank you very much. Have a nice weekend.

When writing the TPC additions for the spec, I did take PostgreSQL
into account (I thought I'd have time to write the psycopg2
implementation back then).  The reasoning for using the three part
identifiers was that the XA-style identifiers were quite common and it
was easier to losslessly encode the three part identifiers as a string
than vice versa.

The spec should allow you to manage identifiers that don't match your
mapping though.  The user can only get references to transaction ID
objects via method calls on the connection.  So while transaction IDs
are required to provide tuple like behaviour, the adapter doesn't have
to use actual tuple objects.  A custom object type could easily be
used here to round trip the foreign IDs between tpc_recover() and
tpc_commit()/tpc_abort().

James.