//---------------------------------------------------------------------------
// @function:
// CPhysicalDML::PosComputeRequired
//
// @doc:
// Compute required sort order based on the key information in the table
// descriptor:
// 1. If a table has no keys, no sort order is necessary.
//
// 2. If a table has keys, but they are not modified in the update, no sort
// order is necessary. This relies on the fact that Split always produces
// Delete tuples before Insert tuples, so we cannot have two versions of the
// same tuple on the same time. Consider for example tuple (A: 1, B: 2), where
// A is key and an update "set B=B+1". Since there cannot be any other tuple
// with A=1, and the tuple (1,2) is deleted before tuple (1,3) gets inserted,
// we don't need to enforce specific order of deletes and inserts.
//
// 3. If the update changes a key column, enforce order on the Action column
// to deliver Delete tuples before Insert tuples. This is done to avoid a
// conflict between a newly inserted tuple and an old tuple that is about to be
// deleted. Consider table with tuples (A: 1),(A: 2), where A is key, and
// update "set A=A+1". Split will generate tuples (1,"D"), (2,"I"), (2,"D"), (3,"I").
// If (2,"I") happens before (2,"D") we will have a violation of the key constraint.
// Therefore we need to enforce sort order on Action to get all old tuples
// tuples deleted before the new ones are inserted.
//
//---------------------------------------------------------------------------
COrderSpec *
CPhysicalDML::PosComputeRequired
(
IMemoryPool *pmp,
CTableDescriptor *ptabdesc
)
{
COrderSpec *pos = GPOS_NEW(pmp) COrderSpec(pmp);
const DrgPbs *pdrgpbsKeys = ptabdesc->PdrgpbsKeys();
if (1 < pdrgpbsKeys->UlLength() && CLogicalDML::EdmlUpdate == m_edmlop)
{
// if this is an update on the target table's keys, enforce order on
// the action column, see explanation in function's comment
const ULONG ulKeySets = pdrgpbsKeys->UlLength();
BOOL fNeedsSort = false;
for (ULONG ul = 0; ul < ulKeySets && !fNeedsSort; ul++)
{
CBitSet *pbs = (*pdrgpbsKeys)[ul];
if (!pbs->FDisjoint(m_pbsModified))
{
fNeedsSort = true;
break;
}
}
if (fNeedsSort)
{
IMDId *pmdid = m_pcrAction->Pmdtype()->PmdidCmp(IMDType::EcmptL);
pmdid->AddRef();
pos->Append(pmdid, m_pcrAction, COrderSpec::EntAuto);
}
}
else if (m_ptabdesc->FPartitioned())
{
COptimizerConfig *poconf = COptCtxt::PoctxtFromTLS()->Poconf();
BOOL fInsertSortOnParquet = FInsertSortOnParquet();
BOOL fInsertSortOnRows = FInsertSortOnRows(poconf);
if (fInsertSortOnParquet || fInsertSortOnRows)
{
GPOS_ASSERT(CLogicalDML::EdmlInsert == m_edmlop);
m_fInputSorted = true;
// if this is an INSERT over a partitioned Parquet or Row-oriented table,
// sort tuples by their table oid
IMDId *pmdid = m_pcrTableOid->Pmdtype()->PmdidCmp(IMDType::EcmptL);
pmdid->AddRef();
pos->Append(pmdid, m_pcrTableOid, COrderSpec::EntAuto);
}
}
return pos;
}
