Thanks for your feedback and sorry for the late reply.
Hi Andy,
Sorry I'm not much help here.
Were you able to figure this out?
I've been working on this for a while. It would not work per
current PG server code. The biggest issue is when user
calls 'xa_end', he/she isn't intent to persist the 2pc
transaction, but the PG code would do that during the 'prepare
transaction' statement. In this case WE SHOULD NOT recover
it again during crash recovery. I'm hacking the pg server code
not do it until a user called 'xa_prepare'.
The internal way to do this is: during xa_end, I run the 'prepare
transaction xxx' statement and record a new xlog to rollback it
2pc, this is done at the end of the crash recovery stage. and
If the user runs the 'xa_prepare', I would add another xlog to
let the above thing not happen at the end of the crash recovery
stage.
(I hacked both JDBC driver and server code for the above stuff)
Besides this, I also find another issue that
connection 1:
xa_start('foo', ..);
insert into t select 1;
xa_end('foo', ..);
connection 2:
xa_start('foo', ..);
select * from t; ==> The above insert from connection 1 is visible for
current connection, but in our JDBC it doesn't.
I am also hacking this case as well, just record the transaction xid
which the 'foo' belongs to and add a new rule to the MVCC to let
the tuple be visible as well.
still tuned and testing now.