On Tue, Sep 13, 2022 at 11:04 AM Laurenz Albe <laurenz.a...@cybertec.at> wrote: > > On Tue, 2022-09-13 at 11:47 +0300, Nikita Malakhov wrote: > > On Tue, Sep 13, 2022 at 11:06 AM Laurenz Albe <laurenz.a...@cybertec.at> > > wrote: > > > Shouldn't such tuples be considered dead right away, even if the inserting > > > transaction is still active? That would allow cleaning them up even > > > before > > > the transaction is done. > > > > > > There is this code in HeapTupleSatisfiesVacuumHorizon: > > > > > > else if > > > (TransactionIdIsCurrentTransactionId(HeapTupleHeaderGetRawXmin(tuple))) > > > { > > > [...] > > > /* inserted and then deleted by same xact */ > > > if > > > (TransactionIdIsCurrentTransactionId(HeapTupleHeaderGetUpdateXid(tuple))) > > > return HEAPTUPLE_DELETE_IN_PROGRESS; > > > > > > Why HEAPTUPLE_DELETE_IN_PROGRESS and not HEAPTUPLE_DEAD? > > > > Please correct me if I'm wrong, despite tuples being inserted and deleted > > by the same > > transaction - they are visible inside the transaction and usable by it, so > > considering them > > dead and cleaning up during execution is a bad idea until the transaction > > is ended. > > But once they are deleted or updated, even the transaction that created them > cannot > see them any more, right?
Forgive me if this is not related but if there is a savepoint between the insertion and deletion, wouldn't it be possible for the transaction to recover the deleted tuples? Best regards Pantelis Theodosiou