Here's a patch that implements the Phantom Command ID idea that has been
discussed.

I didn't do anything about the system columns. We need to before
applying the patch, because the HeapTupleHeaderGetCmin/max functions
don't work properly if called outside the inserting/deleting
transaction. In fact, SELECT cmin,cmax FROM foo will cause assertion
failures if there's rows with phantom cids in the table.

I used the last free bit in t_infomask. At first I thought it wouldn't
be necessary, because you could detect that a row has a phantom command
id if both xmin and xmax are part of the current top-level transaction
(TransactionIdIsCurrentTransactionId). But that doesn't work because we
don't consider aborted subtransactions as current. Using the infomask
bit seems more robust anyway.

--
Heikki Linnakangas
EnterpriseDB http://www.enterprisedb.com



---------------------------(end of broadcast)---------------------------
TIP 5: don't forget to increase your free space map settings