// append an ascii-version of UDFunction into cachewa.qryText_
void UDFunction::generateCacheKey(CacheWA& cwa) const
{
NARoutine *routine = NULL;
NARoutine *action = NULL;
cwa += " nam:";
cwa += functionName_.getExternalName().data();
if (cwa.getPhase() >= CmpMain::BIND &&
getRoutineDesc() &&
(routine=getRoutineDesc()->getNARoutine()) != NULL)
{
char redefTime[40];
convertInt64ToAscii(routine->getRedefTime(), redefTime);
cwa += " redef:";
cwa += redefTime;
if ((routine->getSchemaLabelFileName()) &&
(str_len(routine->getSchemaLabelFileName()) > 0))
{
char schRedefTime[40];
convertInt64ToAscii(routine->getSchemaRedefTime(), schRedefTime);
cwa += " schredef:";
cwa += schRedefTime;
}
}
if (getRoutineDesc() &&
getRoutineDesc()->getActionNameAsGiven().length() != 0)
{
cwa += " actnam:";
cwa += getRoutineDesc()->getActionNameAsGiven();
if (cwa.getPhase() >= CmpMain::BIND &&
getRoutineDesc() &&
(action=getRoutineDesc()->getActionNARoutine()) != NULL)
{
char redefTime[40];
convertInt64ToAscii(action->getRedefTime(), redefTime);
cwa += " actredef:";
cwa += redefTime;
}
}
cwa += "(";
Lng32 arity = (Lng32) getArity();
for (Lng32 i = 0; i < arity; i++) {
if (i > 0) {
cwa += ", ";
}
child(i)->generateCacheKey(cwa);
}
if (getRoutineDesc()->getLocale() != 0 )
{
cwa += ", LOCALE: ";
char dFmt[20];
str_itoa(getRoutineDesc()->getLocale(), dFmt);
cwa += dFmt;
}
cwa += ")";
}
// is this entire expression cacheable after this phase?
NABoolean Scan::isCacheableExpr(CacheWA& cwa)
{
if (cwa.getPhase() >= CmpMain::BIND) {
// save scan's TableDesc
cwa.incNofScans(tabId_);
// native hbase access is not cacheable for now.
if ((getTableDesc()->getNATable()->isHbaseRowTable()) ||
(getTableDesc()->getNATable()->isHbaseCellTable()))
return FALSE;
if (stream_) { // pub-sub streams are not cacheable
return FALSE;
}
// mpalias SELECT is not cacheable unless explicitly requested
if (getTableDesc()->getNATable()->isAnMPTableWithAnsiName() &&
CmpCommon::getDefault(QUERY_CACHE_MPALIAS) == DF_OFF) {
return FALSE;
}
cwa.setConditionallyCacheable();
if (CmpCommon::getDefaultLong(MVQR_REWRITE_LEVEL) >= 1 &&
QRDescGenerator::hasRewriteEnabledMVs(getTableDesc())) {
cwa.setRewriteEnabledMV();
}
return RelExpr::isCacheableExpr(cwa);
}
return FALSE;
}
// Only alow UDFunctions to be cached after BIND time.
NABoolean UDFunction::isCacheableExpr(CacheWA& cwa)
{
if (cwa.getPhase() >= CmpMain::BIND)
return TRUE;
else
return FALSE;
}
// is any literal in this expr safely coercible to its target type?
NABoolean ItemExpr::isSafelyCoercible(CacheWA &cwa) const
{
if (cwa.getPhase() >= CmpMain::BIND) {
Int32 arity = getArity();
for (Int32 x = 0; x < arity; x++) {
if (!child(x)->isSafelyCoercible(cwa)) {
return FALSE;
}
}
if (arity == 2) {
// we have to disallow caching of the following types of exprs:
// expr + 123456789012345678901234567890
// expr || 'overlylongstringthatwouldoverflow'
ItemExpr *left = child(0), *right = child(1);
if (left->getOperatorType() == ITM_CONSTANT) {
if (right->getOperatorType() == ITM_CONSTANT) {
// "10 + 1" should be safely coercible
return TRUE;
} else {
return ((ConstValue*)left)->canBeSafelyCoercedTo
(right->getValueId().getType());
}
}
else if (right->getOperatorType() == ITM_CONSTANT) {
return ((ConstValue*)right)->canBeSafelyCoercedTo
(left->getValueId().getType());
}
// else both are nonliterals; fall thru
}
// else nondyadic expr; fall thru
return TRUE;
}
return FALSE;
}
// we want BiLogic to be cacheable
NABoolean BiLogic::isCacheableExpr(CacheWA& cwa)
{
if (cwa.getPhase() >= CmpMain::BIND && getArity() == 2) {
if (getOperatorType() == ITM_AND) {
ItemExpr *leftC=child(0), *rightC=child(1);
// we want to descend to both left & right
// so cwa.usedKeys & cwa.keyCols get updated
NABoolean leftOK = leftC->isCacheableExpr(cwa);
NABoolean rightOK = rightC->isCacheableExpr(cwa);
// return FALSE if both left & right are not cacheable.
if (!leftOK && !rightOK)
return FALSE;
// allow "select * from t where k1=1 and k2=2" as potentially cacheable
return TRUE;
// the definitive check whether a query's predicate covers
// all referenced tables' key columns can be (and is) done
// only in RelRoot::isCacheableExpr() when cwa.usedKeys &
// cwa.keyCols are fully defined. Otherwise, queries like
// "select * from t1 join t2 on a=x and a=7 where x=7"
// may be incorrectly rejected as non-cacheable if t1 has
// primary key(a) and t2 has primary key(x).
}
else if (getOperatorType() == ITM_OR) {
return TRUE;
// ORs can be cacheable but are not necessarily
// parameterizable, see Bilogic::normalizeForCache below.
}
}
return FALSE;
}
// are RelExpr's kids cacheable after this phase?
NABoolean RelExpr::cacheableKids(CacheWA& cwa)
{
switch (cwa.getPhase()) {
case CmpMain::PARSE:
case CmpMain::BIND: {
Int32 arity = getArity();
if (arity <= 0) { // we have no kids
if (cwa.isConditionallyCacheable()) {
// we're conditionally cacheable and have no kids
setCacheableNode(cwa.getPhase());
return TRUE; // so, we're cachable
}
else {
return FALSE; // MAYBECACHEABLE is not cacheable at this phase
// don't mark this node non-cacheable because this
// RelExpr may be cacheable after the next phase.
}
}
// cacheability of child(ren) determine our cacheability
for (Int32 x = 0; x < arity; x++) {
if (!child(x) || // cases like "insert into t default values"
// return 1 from getArity() even if child(0) is NULL; so
// guard against this potential mxcmp crash and consider
// these cases non-cacheable during the PARSE stage.
child(x)->isNonCacheable()) {
// the 1st noncacheable child makes us noncacheable
setNonCacheable();
return FALSE;
}
else if (!child(x)->isCacheableExpr(cwa)) {
// noncacheable child
return FALSE;
// don't mark this node non-cacheable because this
// RelExpr may be cacheable after the next phase.
}
else { // cacheable child
continue; // look at next child
}
}
// all children are cacheable, so we're cacheable too
setCacheableNode(cwa.getPhase());
return TRUE;
}
default:
return FALSE;
}
}
// append an ascii-version of IsolatedScalarUDF into cachewa.qryText_
void IsolatedScalarUDF::generateCacheKey(CacheWA &cwa) const
{
NARoutine *routine = NULL;
NARoutine *action = NULL;
RelExpr::generateCacheKey(cwa);
cwa += " UDFname:";
cwa += getRoutineName().getQualifiedNameAsAnsiString().data();
if (cwa.getPhase() >= CmpMain::BIND &&
getRoutineDesc() &&
(routine=getRoutineDesc()->getNARoutine()) != NULL)
{
char redefTime[40];
convertInt64ToAscii(routine->getRedefTime(), redefTime);
cwa += " redef:";
cwa += redefTime;
}
if (getRoutineDesc() != NULL && getRoutineDesc()->isUUDFRoutine())
{
cwa += " action:";
cwa += getRoutineDesc()->getActionNameAsGiven();
if (cwa.getPhase() >= CmpMain::BIND &&
getRoutineDesc() &&
(action=getRoutineDesc()->getActionNARoutine()) != NULL)
{
char redefTime[40];
convertInt64ToAscii(action->getRedefTime(), redefTime);
cwa += " actredef:";
cwa += redefTime;
}
}
const ItemExpr *paramExpr = (getProcAllParamsTree() == NULL) ?
getProcInputParamsVids().rebuildExprTree(ITM_ITEM_LIST) :
getProcAllParamsTree();
if (paramExpr)
{
cwa += " arg:(";
paramExpr->generateCacheKey(cwa);
cwa += ")";
}
}
// is any literal in this expr safely coercible to its target type?
NABoolean ValueIdProxy::isSafelyCoercible(CacheWA& cwa) const
{
if (cwa.getPhase() >= CmpMain::BIND) {
if (!derivedFrom_.getItemExpr()->isSafelyCoercible(cwa))
return FALSE;
else
return TRUE;
}
return FALSE;
}
// does this ItemExpr (dis)qualify query to be cacheable after this phase?
NABoolean ConstValue::isCacheableExpr(CacheWA& cwa)
{
if (cwa.getPhase() <= CmpMain::PARSE && hasUnknownCharSet()) {
// string literal with unknown character set cannot be considered
// cacheable by pre-binder stages because the information needed
// to determine their character set is not known until bind-time.
// This fixes genesis case 10-041215-6141, soln 10-041215-2826.
return FALSE;
}
return ItemExpr::isCacheableExpr(cwa);
}
// we want BiRelat to be cacheable
NABoolean BiRelat::isCacheableExpr(CacheWA& cwa)
{
if (cwa.getPhase() >= CmpMain::BIND && getArity() == 2) {
if (getOperatorType() == ITM_EQUAL) {
ItemExpr *leftC=child(0), *rightC=child(1);
OperatorTypeEnum leftO = leftC->getOperatorType();
OperatorTypeEnum rightO = rightC->getOperatorType();
BaseColumn *base;
if (leftO == ITM_BASECOLUMN) {
base = (BaseColumn*)leftC;
if (base->isKeyColumnValue(*rightC)) {
cwa.addToUsedKeys(base);
}
return TRUE;
}
else if (rightO == ITM_BASECOLUMN) {
base = (BaseColumn*)rightC;
if (base->isKeyColumnValue(*leftC)) {
cwa.addToUsedKeys(base);
}
return TRUE;
}
else if (leftO == ITM_ITEM_LIST && rightO == ITM_ITEM_LIST &&
((ItemList*)leftC)->isListOfCacheableSelPred
(cwa, (ItemList*)rightC)) {
return TRUE;
}
else {
// we want all other equality comparisons to be cacheable, eg,
// retail_div_cd||acct_type_cd||substring(acct_id,1,8)='20V43085193'
return TRUE;
}
}
else {
return TRUE;
// other binary comparison predicates can be cacheable, but are
// not necessarily parameterizable, see BiRelat::normalizeForCache
}
}
return FALSE;
}
// does this query's selection predicate list qualify query
// to be cacheable after this phase?
NABoolean ItemList::isListOfCacheableSelPred
(CacheWA& cwa, ItemList *other) const
{
Int32 arity = getArity();
NABoolean result = FALSE;
if (cwa.getPhase() >= CmpMain::BIND &&
other && arity == other->getArity()) {
// assume this is an AND list, so, we need only one
// cacheable conjunct to consider the list cacheable.
for (Int32 x = 0; x < arity; x++) {
ItemExpr *leftC = child(x), *rightC = other->child(x);
OperatorTypeEnum leftO = leftC->getOperatorType();
OperatorTypeEnum rightO = rightC->getOperatorType();
BaseColumn *base;
if (leftO == ITM_BASECOLUMN) {
base = (BaseColumn*)leftC;
if (base->isKeyColumnValue(*rightC)) {
cwa.addToUsedKeys(base);
}
result = TRUE;
continue;
}
else if (rightO == ITM_BASECOLUMN) {
base = (BaseColumn*)rightC;
if (base->isKeyColumnValue(*leftC)) {
cwa.addToUsedKeys(base);
}
result = TRUE;
continue;
}
else if (leftO == ITM_ITEM_LIST && rightO == ITM_ITEM_LIST &&
((ItemList*)leftC)->isListOfCacheableSelPred
(cwa, (ItemList*)rightC)) {
result = TRUE;
continue;
}
}
}
return result;
}
// helper to change literals of a cacheable query into input parameters
void BiRelat::parameterizeMe(CacheWA& cwa, BindWA& bindWA, ExprValueId& child,
BaseColumn *base, ConstValue *val)
{
NABoolean parameterizableMVQRplan = cwa.hasRewriteEnabledMV() &&
CmpCommon::getDefault(MVQR_PARAMETERIZE_EQ_PRED) == DF_ON;
NABoolean parameterizablePlan = parameterizableMVQRplan ||
!cwa.hasRewriteEnabledMV();
if (val->isNull()) {
// val is null; keep null as is.
// this is part of a fix to genesis case: 10-010618-3484.
}
else if (base->isKeyColumnValue(*val) // keyCol = constant
&& cwa.isParameterizable(base) && parameterizablePlan) {
cwa.replaceWithNewOrOldConstParam(val, base->getNAColumn()->getType(),
child, bindWA);
if (parameterizableMVQRplan)
cwa.setParameterizedPred(TRUE);
}
else if (!base->canSafelyCoerce(*val)) {
// term is not parameterizable, keep it as is.
}
else { // nonKeyCol = constant
if (parameterizablePlan) {
computeAndAddSelParamIfPossible(cwa, bindWA, child, base, val);
if (parameterizableMVQRplan)
cwa.setParameterizedPred(TRUE);
}
}
}
// is any literal in this expr safely coercible to its target type?
NABoolean UDFunction::isSafelyCoercible(CacheWA& cwa) const
{
if (cwa.getPhase() >= CmpMain::BIND) {
Int32 arity = getArity();
for (Int32 x = 0; x < arity; x++) {
if (!child(x)->isSafelyCoercible(cwa)) {
return FALSE;
}
}
return TRUE;
}
return FALSE;
}
// is any literal in this expr safely coercible to its target type?
NABoolean Cast::isSafelyCoercible(CacheWA &cwa) const
{
if (cwa.getPhase() >= CmpMain::BIND) {
ItemExpr *opd = child(0);
if (!opd->isSafelyCoercible(cwa)) {
return FALSE;
}
if (opd->getOperatorType() == ITM_CONSTANT) {
return ((ConstValue*)opd)->canBeSafelyCoercedTo(*type_);
}
return TRUE;
}
return FALSE;
}
// is this entire expression cacheable after this phase?
NABoolean RelRoot::isCacheableExpr(CacheWA& cwa)
{
//queries prefixed by display are not cacheable e.g. display select * from ...
if(getDisplayTree())
return FALSE;
// Parallel extract producer queries are not cacheable
if (numExtractStreams_ > 0)
return FALSE;
// descend to scans early to get cwa.numberOfScans_
if (!RelExpr::isCacheableExpr(cwa)) {
return FALSE;
}
if (cwa.getPhase() == CmpMain::PARSE) {
if (compExprTree_ || compExpr_.entries() > 0) {
// insert-returning is not cacheable after parse
return FALSE;
}
}
else if (cwa.getPhase() >= CmpMain::BIND) {
// make sure select list is cacheable
if (compExprTree_) {
if (!compExprTree_->isCacheableExpr(cwa)) {
return FALSE;
}
}
else if (!compExpr_.isCacheableExpr(cwa)) {
return FALSE;
}
}
if (isAnalyzeOnly())
return FALSE;
return TRUE;
}
void computeAndAddSelParamIfPossible(
CacheWA& cwa, BindWA& bindWA, ExprValueId& child,
BaseColumn *base, ConstValue *val)
{
ColStatsSharedPtr cStatsPtr = (base->getTableDesc()->tableColStats()).
getColStatsPtrForColumn(base->getValueId());
if (cStatsPtr == NULL )
return;
HistogramSharedPtr hist = cStatsPtr->getHistogram();
if ( hist == NULL )
return;
CostScalar sel;
NABoolean canComputeSelectivity = hist -> computeSelectivityForEquality(
val, cStatsPtr->getRowcount(), cStatsPtr->getTotalUec(),
sel);
if ( canComputeSelectivity == TRUE ) {
const NAType * newType = base->getNAColumn()->getType();
// for char datatypes, assign the caseinsensitive attribute from
// the const node.
if (newType->getTypeQualifier() == NA_CHARACTER_TYPE)
{
newType =
base->getNAColumn()->getType()->newCopy(cwa.wHeap());
((CharType*)newType)->setCaseinsensitive(((CharType*)val->getType())->isCaseinsensitive());
}
cwa.replaceWithNewOrOldSelParam(val, newType, Selectivity(sel),
child, bindWA);
}
}
// is this entire expression cacheable after this phase?
NABoolean Join::isCacheableExpr(CacheWA& cwa)
{
if (cwa.getPhase() >= CmpMain::BIND) {
// must first descend to scans to get cwa.numberOfScans_
if (!RelExpr::isCacheableExpr(cwa)) {
return FALSE;
}
if (isCacheableNode(cwa.getPhase())) {
cwa.setConditionallyCacheable();
}
// if we allow joins of views to be cached, query caching cannot
// distinguish between (see note at bottom of cachewa.h)
// select avg(f.a) from v f, v s group by f.b;
// select avg(s.a) from v f, v s group by f.b;
// select avg(t.a) from v f, t group by f.b;
// assuming v is "create view v from select * from t". We avoid
// false cache hits by detecting the possible occurrence of such
// view joins here and later using cwa.isViewJoin_ to include
// their query texts into their cache keys.
//
// A view is repsented by a renamed table with isView() returnning
// TRUE.
RelExpr *c0 = child(0);
RelExpr *c1 = child(1);
if ((c0->getOperatorType() == REL_RENAME_TABLE &&
((RenameTable *)c0)->isView() == TRUE)
||
(c1->getOperatorType() == REL_RENAME_TABLE &&
((RenameTable *)c1)->isView() == TRUE)) {
cwa.foundViewJoin();
}
// check its join predicate
ItemExpr *pred = joinPredTree_ ? joinPredTree_ :
joinPred_.rebuildExprTree();
if (pred) {
cwa.setHasPredicate();
// is join predicate cacheable?
if (pred->hasNoLiterals(cwa)) {
// predicate with no literals is cacheable
}
else {
cwa.setPredHasNoLit(FALSE);
if (!pred->isCacheableExpr(cwa)) {
// a non-cacheable predicate renders Join non-cacheable.
setNonCacheable();
return FALSE;
}
}
}
return TRUE; // join may be cacheable
}
return FALSE;
}
// is this entire expression cacheable after this phase?
NABoolean RelExpr::isCacheableExpr(CacheWA& cwa)
{
switch (cwa.getPhase()) {
case CmpMain::PARSE:
case CmpMain::BIND: {
// does query have too many ExprNodes?
if (cwa.inc_N_check_still_cacheable() == FALSE) {
// yes. query with too many ExprNodes is not cacheable.
return FALSE;
}
if (isNonCacheable()) { // this node is not cacheable
return FALSE; // so the entire expression is not cacheable
// don't mark this node non-cacheable because this
// RelExpr may be cacheable after the next phase.
}
if (isCacheableNode(cwa.getPhase())) {
// must be an INSERT, UPDATE, DELETE, or SELECT node;
// so, mark this expression as conditionally cacheable.
cwa.setConditionallyCacheable();
}
// must descend to scans to get cwa.numberOfScans_
if (!cacheableKids(cwa)) {
return FALSE;
}
// this node is either cacheable or maybecacheable
// check its selection predicate
ItemExpr *pred = selPredTree() ? selPredTree() :
getSelectionPred().rebuildExprTree();
if (pred) {
cwa.setHasPredicate();
// is selection predicate cacheable?
if (pred->hasNoLiterals(cwa)) {
// predicate with no literals is cacheable
}
else {
cwa.setPredHasNoLit(FALSE);
if (!pred->isCacheableExpr(cwa)) {
// a non-cacheable selection predicate
// renders entire RelExpr non-cacheable.
setNonCacheable();
return FALSE;
}
}
}
return TRUE; // RelExpr may be cacheable
}
default: { const NABoolean notYetImplemented = FALSE;
CMPASSERT(notYetImplemented);
return FALSE;
}
}
}
// append an ascii-version of Insert into cachewa.qryText_
void Insert::generateCacheKey(CacheWA &cwa) const
{
GenericUpdate::generateCacheKey(cwa);
if (insertColTree_) {
cwa += " insCol:";
insertColTree_->generateCacheKey(cwa);
}
// order by clause is important
ItemExpr *orderBy = orderByTree_ ? orderByTree_ :
reqdOrder_.rebuildExprTree();
if (orderBy) {
cwa += " order:";
orderBy->generateCacheKey(cwa);
}
const NATable *tbl;
if (cwa.getPhase() >= CmpMain::BIND &&
getTableDesc() && (tbl=getTableDesc()->getNATable()) != NULL) {
// If PARTITION clause has been used we must reflect that in the key.
if (tbl->isPartitionNameSpecified()) {
cwa += " partition:";
cwa += tbl->getClusteringIndex()->getFileSetName().getQualifiedNameAsString().data();
}
// If PARTITION range has been used we must reflect that in the key.
else if (tbl->isPartitionRangeSpecified()) {
cwa += " partition:";
char str[100];
sprintf(str, " from %d to %d",
tbl->getExtendedQualName().getPartnClause().getBeginPartitionNumber() ,
tbl->getExtendedQualName().getPartnClause().getEndPartitionNumber());
cwa += str;
}
}
if (isUpsert())
{
cwa += " upsert:";
}
}
// is any literal in this expr safely coercible to its target type?
NABoolean Assign::isSafelyCoercible(CacheWA &cwa) const
{
if (cwa.getPhase() >= CmpMain::BIND) {
// we have to disallow caching of the following types of updates:
// update t set col = overlylongliteral
ItemExpr *src = getSource().getItemExpr();
if (src->getOperatorType() == ITM_CONSTANT) {
// source is a literal; make sure this update is cacheable only if
// literal can be safely coerced into its target type.
return ((ConstValue*)src)->canBeSafelyCoercedTo(getTarget().getType());
}
else { // source is not a literal
// we need to descend into this expr to verify that no
// errors can occur during backpatching. For example,
// update t set i = i + 123456789012345678901234567890
// should not be cacheable if i is a smallint.
// reject "update t set c=(subquery)" as noncacheable
// as part of a fix to CR 10-020108-8401.
return src->isSafelyCoercible(cwa) && src->isCacheableExpr(cwa);
}
}
return FALSE;
}
// is this entire expression cacheable after this phase?
NABoolean GenericUpdate::isCacheableExpr(CacheWA& cwa)
{
// descend to scans early to get cwa.numberOfScans_
if (!RelExpr::isCacheableExpr(cwa)) {
return FALSE;
}
// Make "{update|delete} ... where current of cursor" non-cacheable
// so that stale cache will not lead to timestamp mismatch error at
// runtime. AQR attempts to handle this error, but only after the
// referenced cursor is closed due to transaction rollback. This is
// Solution 10-100425-9755.
if (currOfCursorName()) {
return FALSE;
}
if (cwa.getPhase() >= CmpMain::BIND) {
// make sure any literals in the assignment clause can be safely
// cast and assigned to their target types at plan-backpatch-time
ItemExpr *newExpr = newRecExprTree_ ? newRecExprTree_ :
newRecExpr_.rebuildExprTree(ITM_ITEM_LIST);
if (newExpr && !newExpr->isSafelyCoercible(cwa)) {
return FALSE;
}
// reject as non-cacheable queries such as
// prepare s from select * from (update t042qT8 set b=7
// set on rollback c=12345678901234567890 where a=2) as t;
ItemExpr *setOnRollback;
if (newRecBeforeExpr_.entries() > 0 &&
(setOnRollback=newRecBeforeExpr_.rebuildExprTree(ITM_ITEM_LIST))
&& !setOnRollback->isSafelyCoercible(cwa)) {
return FALSE;
}
// make sure any executor predicate is cacheable
ItemExpr *execPred = executorPredTree_ ? executorPredTree_ :
executorPred_.rebuildExprTree();
if (execPred) {
cwa.setHasPredicate();
if (execPred->hasNoLiterals(cwa)) {
// predicate with no literals is cacheable
}
else {
cwa.setPredHasNoLit(FALSE);
return execPred->isCacheableExpr(cwa);
}
}
// at this time, not cacheable if subquery is specified in
// UPDATE SET clause.
// This could be enabled later.
if (subqInUpdateAssign()) {
return FALSE;
}
}
else {
if ((getTableName().isPartitionNameSpecified()) ||
(getTableName().isLocationNameSpecified()) ||
(getTableName().isPartitionRangeSpecified()))
return FALSE; // If PartnClause is used no cache hit before bind stage.
}
return TRUE; // may be cacheable
}
// does this entire ItemExpr qualify query to be cacheable after this phase?
NABoolean ItemExpr::isCacheableExpr(CacheWA& cwa)
{
switch (cwa.getPhase()) {
default: { const NABoolean notYetImplemented = FALSE;
CMPASSERT(notYetImplemented);
break;
}
case CmpMain::PARSE:
// ItemExpr::isCacheableExpr is used by Tuple::isCacheable to
// determine whether a tuple-insert or a tuplelist-insert is
// cacheable after parse. It traverses a tuple's values list
// looking for non-cacheable ItemExprs such as HostVar,
// DynamicParam, Subquery, etc.
// determine cacheability of tuple INSERT's itemexprs here
if (isNonCacheable() || !cwa.isConditionallyCacheable()) {
return FALSE;
}
else { // we're either cacheable or maybecacheable
Int32 arity = getArity();
if (arity <= 0) { // we have no kids & we're conditionally cacheable
return TRUE; // we're cacheable
}
// cacheability of child(ren) determine our cacheability
for (Int32 x = 0; x < arity; x++) {
if (!child(x) || child(x)->isNonCacheable()) {
// the 1st noncacheable child makes us noncacheable
setNonCacheable();
return FALSE;
}
else if (!child(x)->isCacheableExpr(cwa)) {
// noncacheable child
return FALSE;
}
else { // cacheable child
continue; // look at next child
}
}
// all children are cacheable, so we're cacheable too
setCacheableNode(cwa.getPhase());
return TRUE;
}
break;
case CmpMain::BIND:
// does query have too many ExprNodes?
if (cwa.inc_N_check_still_cacheable() == FALSE) {
// yes. query with too many ExprNodes is not cacheable.
return FALSE;
}
// ItemExpr::isCacheableExpr is used in "after BIND" cases to
// visit all operands of an ItemExpr looking for noncacheable
// expressions. For example, it discovers that
// SELECT (SELECT a FROM t2 WHERE MAX(o.c)>1) FROM t1 o;
// is noncacheable only after it visits the outer query's select list.
if (isNonCacheable()) {
return FALSE;
}
else if (isCacheableNode(cwa.getPhase())) {
return TRUE;
}
else { // we're either cacheable or maybecacheable
// Assume this ItemExpr is an operand of a function or a list. In
// this case, a single noncacheable child renders us noncacheable.
Int32 arity = getArity();
for (Int32 x = 0; x < arity; x++) {
if (!child(x) || !child(x)->isCacheableExpr(cwa)) {
// noncacheable child
return FALSE;
}
else { // cacheable child
continue; // look at next child
}
}
// all children are cacheable, so we're cacheable too
setCacheableNode(cwa.getPhase());
return TRUE;
}
break;
}
return FALSE;
}
// does this entire ItemExpr qualify query to be cacheable after this phase?
NABoolean DynamicParam::isCacheableExpr(CacheWA& cwa)
{
// a dynamic parameter is not cacheable after parse
return (cwa.getPhase() <= CmpMain::PARSE) ? FALSE : TRUE;
}
// append an ascii-version of GenericUpdate into cachewa.qryText_
void GenericUpdate::generateCacheKey(CacheWA& cwa) const
// NB: This comment applies to all generateCacheKey methods.
// generateCacheKey is used to generate a string representation s of the
// "parameterized" query. Since this string s is used by QCache::lookUp
// to determine if a query is in the cache, it is essential that:
// (1) two different queries have different string representations
// (2) two queries that differ only in their query literals should
// have the same string representations
// One possible implementation of generateCacheKey is to use the query's
// original query text. But, original query text does not satisfy (2).
// To get (2), we call generateCacheKey() from RelRoot::normalizeForCache
// which, by definition, replaced query literals with constant parameters.
// However, generateCacheKey must also satisfy (1). generateCacheKey must
// generate two different strings for two logically different queries.
//
// To satisfy requirements (1) and (2), generateCacheKey and
// normalizeForCache must be in sync -- every user-specified expr that
// generateCacheKey emits into cwa.qryText_ must be examined by
// normalizeForCache for possible replacement of any literal there into
// a constant parameter.
//
// In order for the literal-into-constantparameter replacement to be safe,
// isCacheableExpr must visit all user-specified exprs to make sure that
// only constants that can be safely cast into the query's target types
// are considered cacheable. For example, given this update query
// update t set a = 'xyz' where pk = 1;
// isCacheableeExpr, normalizeForCache, and generateCacheKey must cooperate
// so that:
// 1) isCacheableExpr rejects the query as noncacheble if 'xyz' cannot be
// safely cast into a's target type, eg, 'xyz' may be too long if a's
// type is char(1).
// 2) normalizeForCache must visit and replace both 'xyz' and 1 with
// appropriate constant parameters.
// 3) generateCacheKey must emit some string representation of the
// parameterized query, eg, "update t set a = % where pk = %".
// generateCacheKey can emit more stuff, eg, internally specified
// begin/end-key predicates, but it must emit a string representation
// of all user-specified parts of the query.
{
// append to cwa.qryText_ GenericUpdate's "essential" data members
RelExpr::generateCacheKey(cwa);
// An extension of the fix to 10-010618-3505, 10-010619-3515:
// for "after bind" Insert/Update/Delete queries, include table's
// RedefTime into cwa.qryText_ to make sure we get a cache hit only on
// query that reference table(s) that have not changed since the query's
// addition to the cache. The queries that reference altered table(s)
// will never be hit again and will eventually age out of the cache.
// This is not strictly necessary, but it speeds up the processing
// of insert/update/delete queries on altered tables.
const NATable *tbl;
if (cwa.getPhase() >= CmpMain::BIND &&
getTableDesc() && (tbl=getTableDesc()->getNATable()) != NULL) {
char redefTime[40];
convertInt64ToAscii(tbl->getRedefTime(), redefTime);
cwa += " redef:";
cwa += redefTime;
}
ItemExpr *newExpr = newRecExprTree_ ? newRecExprTree_ :
newRecExpr_.rebuildExprTree(ITM_ITEM_LIST);
if (newExpr) {
cwa += " newRecExpr:";
newExpr->generateCacheKey(cwa);
}
// make sure cache key can distinguish these 2 queries:
// prepare s from select * from (update t042qT8 set b=7 where a=2) as t;
// prepare s from select * from (update t042qT8 set b=7 set on rollback c=2
// where a=2) as t;
ItemExpr *setOnRollback;
if (newRecBeforeExpr_.entries() > 0 &&
(setOnRollback=newRecBeforeExpr_.rebuildExprTree(ITM_ITEM_LIST))) {
cwa += " setOnRollback:";
setOnRollback->generateCacheKey(cwa);
}
ItemExpr *execPred = executorPredTree_ ? executorPredTree_ :
executorPred_.rebuildExprTree();
if (execPred) {
cwa += " execPred:";
execPred->generateCacheKey(cwa);
}
// MVs --
// The NOLOG parameter is essential.
if (isNoLogOperation()) {
cwa += " NOLOG";
}
// "current of cursor/hostvar" is essential
if (currOfCursorName_) {
currOfCursorName_->generateCacheKey(cwa);
}
// not sure if the following are essential, but better to be safe &
// slightly inefficient than to deliver a false hit (ie, wrong plan)
cwa += mtsStatement_ ? "m1" : "m0";
cwa += noFlow_ ? "n1" : "n0";
cwa += noRollback_ ? "o1" : "o0";
cwa += noCheck_ ? "nc" : "dc";
// not sure if the following are essential, but we don't know how
// to quickly & cheaply include them into our cachekey:
// updatedTableName_, tabId_, updateToSelectMap_, indexDesc_,
// newRecExprArray_, usedColumns_, newRecBeforeExpr_,
// newRecBeforeExprArray_, usedBeforeColumns_, potentialOutputs_
// indexNumberArray_, scanIndexDesc_, rowsAffected_, stoi_,
// oldToNewMap_
// The following data members are not "essential" to generateCacheKey
// (at least "after bind") because they are either covered by other
// data members (eg, beginKeyPred and endKeyPred_ are covered by the
// selection pred in RelExpr) or they are not yet defined until later
// (eg, after the optimize phase):
// indexNewRecExprArrays_, beginKeyPred_, endKeyPred_,
// pathKeys_, partKeys_, indexBeginKeyPredArray_,
// indexEndKeyPredArray_, checkConstraints_
}
// append an ascii-version of RelRoot into cachewa.qryText_
void RelRoot::generateCacheKey(CacheWA &cwa) const
{
RelExpr::generateCacheKey(cwa);
ItemExpr *cExpr = compExprTree_ ? compExprTree_ :
compExpr_.rebuildExprTree();
if (cExpr) {
// append any select list into cache key
cwa += " cExpr:";
cExpr->generateCacheKey(cwa);
// reflect any "[first n]"
cwa += ((RelRoot*)this)->needFirstSortedRows() ? " 1stN" : " ";
// Should the select_list aliases be a part of the cache key?
// Their not affecting the compiled plan argues for their exclusion.
// Their affecting sqlci's expected output argues for their inclusion.
RETDesc *rDesc = getRETDesc();
CollIndex degree, x;
if (rDesc && (degree=rDesc->getDegree()) > 0) {
cwa += " sla:";
for (x = 0; x < degree; x++){
cwa += rDesc->getColRefNameObj(x).getColName().data();
cwa += " ";
}
// fix 0-061115-0532 (query cache didn't handle select with embedded
// update correctly). New/Old corr. names are recorded for embedded
// updates here for exact match. This is important because otherwise
// a reuse of a query returning the old/new version of values for
// a query requesting new/old version is totally possible and
// unacceptable.
//
// Sample embedded update queries
// select * from (update tab1 set x = x + 1 where x > 1 return new.*) y;
// select * from (update tab1 set x = x + 1 where x > 1 return new.x, old.y) y;
//
if ( cwa.isUpdate() && isTrueRoot() == FALSE ) {
cwa += " corrNamTok:";
cwa += rDesc->getBindWA()->getCorrNameTokens();
}
}
}
// order by clause is important
ItemExpr *orderBy = orderByTree_ ? orderByTree_ :
reqdOrder_.rebuildExprTree();
if (orderBy) {
cwa += " order:";
orderBy->generateCacheKey(cwa);
}
// statement-level access type & lock mode are important for multiuser
// applications. both are reflected in the stmt-level and/or context-wide
// TransMode. So, we mimic RelRoot::codeGen logic here: "copy the current
// context-wide TransMode, then overlay with this stmt's 'FOR xxx ACCESS'
// setting, if any".
TransMode tmode;
tmode.updateTransMode(CmpCommon::transMode());
StmtLevelAccessOptions &opts = ((RelRoot*)this)->accessOptions();
if (opts.accessType() != ACCESS_TYPE_NOT_SPECIFIED_) {
tmode.updateAccessModeFromIsolationLevel
(TransMode::ATtoIL(opts.accessType()));
tmode.setStmtLevelAccessOptions();
}
if (isTrueRoot()) {
// these are needed by Javier's qc stats virtual tbl interface
cwa.setIsoLvl(tmode.getIsolationLevel());
cwa.setAccessMode(tmode.getAccessMode());
cwa.setAutoCommit(tmode.getAutoCommit());
cwa.setFlags(tmode.getFlags());
cwa.setRollbackMode(tmode.getRollbackMode());
cwa.setAutoabortInterval(tmode.getAutoAbortIntervalInSeconds());
cwa.setMultiCommit(tmode.getMultiCommit());
}
// needed to distinguish these queries and avoid a false hit
// select * from (delete from t where a=2) as t;
// select * from (delete from t where a=2 for SKIP CONFLICT ACCESS) as t;
char mode[40];
str_itoa(tmode.getIsolationLevel(), mode); cwa += " il:"; cwa += mode;
str_itoa(tmode.getAccessMode(), mode); cwa += " am:"; cwa += mode;
// Solution: 10-060418-5903
str_itoa(cwa.getIsoLvlForUpdates(), mode); cwa += " ilu:"; cwa += mode;
str_itoa(tmode.getAutoCommit(), mode); cwa += " ac:"; cwa += mode;
str_itoa(tmode.getFlags(), mode); cwa += " fl:"; cwa += mode;
str_itoa(tmode.getRollbackMode(), mode); cwa += " rm:"; cwa += mode;
str_itoa(tmode.getAutoAbortIntervalInSeconds(), mode); cwa += " ai:"; cwa += mode;
str_itoa(tmode.getMultiCommit(), mode); cwa += " mc:"; cwa += mode;
if (opts.lockMode() != LOCK_MODE_NOT_SPECIFIED_) {
// need to distinguish these queries and avoid a false hit
// select * from t in share mode;
// select * from t in exclusive mode;
str_itoa(opts.lockMode(), mode); cwa += " lm:"; cwa += mode;
}
// updatableSelect_ is essential. Otherwise, queries like
// "select * from t" and "select * from t for update" can confuse
// query caching into a false hit, causing fullstack/test051 to fail.
if (updatableSelect_) {
cwa += " 4updt ";
}
// for update of col [,col]... clause is important
ItemExpr *updCol = updateColTree_ ? updateColTree_ :
updateCol_.rebuildExprTree();
if (updCol) {
cwa += " updCol:";
updCol->generateCacheKey(cwa);
}
// making the CQS part of the key is more efficient than calling
// CompilerEnv::changeEnv() in ControlDB::setRequiredShape()
if (reqdShape_) {
reqdShape_->unparse(cwa.reqdShape_);
}
}
// append an ascii-version of Scan into cachewa.qryText_
void Scan::generateCacheKey(CacheWA &cwa) const
{
RelExpr::generateCacheKey(cwa);
// Fix to 10-010618-3505, 10-010619-3515: include this Scan table's
// RedefTime into cwa.qryText_ to make sure we get a cache hit only on
// query that reference table(s) that have not changed since the query's
// addition to the cache. The queries that reference altered table(s)
// will never be hit again and will eventually age out of the cache.
const NATable *tbl;
if (cwa.getPhase() >= CmpMain::BIND &&
getTableDesc() && (tbl=getTableDesc()->getNATable()) != NULL) {
char redefTime[40];
convertInt64ToAscii(tbl->getRedefTime(), redefTime);
cwa += " redef:";
cwa += redefTime;
if (tbl->isHiveTable()) {
char lastModTime[40];
Int64 mTime = tbl->getClusteringIndex()->getHHDFSTableStats()->getModificationTS();
convertInt64ToAscii(mTime, lastModTime);
cwa += " lastMod:";
cwa += lastModTime;
cwa += " numFiles:";
char numFiles[20];
Int64 numberOfFiles = tbl->getClusteringIndex()->getHHDFSTableStats()->getNumFiles();
sprintf(numFiles, " %ld", numberOfFiles);
cwa += numFiles ;
}
// save pointer to this table. later, QueryCache::addEntry will use
// this pointer to get to this table's histograms's timestamp
cwa.addTable( (NATable*)tbl );
// If PARTITION clause has been used we must reflect that in the key.
if (tbl->isPartitionNameSpecified()) {
cwa += " partition:";
cwa += tbl->getClusteringIndex()->getFileSetName().getQualifiedNameAsString().data();
}
// If PARTITION range has been used we must reflect that in the key.
else if (tbl->isPartitionRangeSpecified()) {
cwa += " partition:";
char str[100];
sprintf(str, " from %d to %d",
tbl->getExtendedQualName().getPartnClause().getBeginPartitionNumber() ,
tbl->getExtendedQualName().getPartnClause().getEndPartitionNumber());
cwa += str;
}
}
// We must reflect userTableName_.location into cache key.
// Otherwise, two queries which differ only in location such as
// table table (table T058a, location $system.zsd12345.x1234500);
// table table (table T058a, location $data .zsd12345.x1234500);
// can confuse our query caching code to return a false hit and
// cause fullstack/test058 to fail.
cwa += userTableName_.getLocationName().data();
// Same with stream_ because queries like
// "select * from t" and "select * from stream(t)" can
// confuse query caching into a false hit causing test079 to fail.
if (stream_) {
cwa += " stream ";
}
// mark mpalias queries so they can be decached upon user request
if (getTableDesc()->getNATable()->isAnMPTableWithAnsiName()) {
cwa += AM_AN_MPALIAS_QUERY;
}
if (getHbaseAccessOptions())
{
cwa += " hbaseVersions: ";
char numVersions[20];
sprintf(numVersions, " %d", getHbaseAccessOptions()->getHbaseVersions());
cwa += numVersions ;
}
}
NABoolean Update::isCacheableExpr(CacheWA& cwa)
{
cwa.setIsUpdate(TRUE);
return GenericUpdate::isCacheableExpr(cwa);
}
