On Fri, May 28, 2004 at 04:45:28PM -0400, Tom Lane wrote:
> However, I just remembered why we rejected this idea to start with :-(.
> If we do it this way then when the overall xact commits, we no longer
> have state that tells which particular tuples are good or not. We would
> have to trawl for tuples written by aborted subtransactions and mark
> them dead before committing, else other transactions would think they
> were good.
>
> What this says is that we still need persistent pg_subtrans status.
> I'm not sure if we can use CIDs as subtrans IDs this way and still have
> a reasonably efficient storage representation for the global pg_subtrans
> table.
I'm not sure if I understand your last assertion. We can in no way use
the CID as subtrans ID in pg_subtrans, of course (it starts from 0 at
each main transaction start).
So pg_subtrans remains the same, and we assign a new Xid to each
subtransaction. Each tuple gets Xmin/Xmax according to the Xid of the
current subtransaction. Within the transaction tree we don't use the
Xid to check for visibility, but Cmin/Cmax and the abort bitmap.
When the Xmin/xmax does not belong to our transaction tree, we use
pg_subtrans and pg_clog.
--
Alvaro Herrera (<alvherre[a]dcc.uchile.cl>)
"We are who we choose to be", sang the goldfinch
when the sun is high (Sandman)
---------------------------(end of broadcast)---------------------------
TIP 9: the planner will ignore your desire to choose an index scan if your
joining column's datatypes do not match
