void PlanEnumerator::tagMemo(size_t id) {
QLOG() << "Tagging memoID " << id << endl;
NodeAssignment* assign = _memo[id];
verify(NULL != assign);
if (NULL != assign->pred) {
PredicateAssignment* pa = assign->pred.get();
verify(NULL == pa->expr->getTag());
verify(pa->indexToAssign < pa->first.size());
pa->expr->setTag(new IndexTag(pa->first[pa->indexToAssign]));
}
else if (NULL != assign->orAssignment) {
OrAssignment* oa = assign->orAssignment.get();
for (size_t i = 0; i < oa->subnodes.size(); ++i) {
tagMemo(oa->subnodes[i]);
}
}
else if (NULL != assign->newAnd) {
AndAssignment* aa = assign->newAnd.get();
if (AndAssignment::MANDATORY == aa->state) {
verify(aa->counter < aa->mandatory.size());
const OneIndexAssignment& assign = aa->mandatory[aa->counter];
for (size_t i = 0; i < assign.preds.size(); ++i) {
MatchExpression* pred = assign.preds[i];
verify(NULL == pred->getTag());
pred->setTag(new IndexTag(assign.index, assign.positions[i]));
}
}
else if (AndAssignment::PRED_CHOICES == aa->state) {
verify(aa->counter < aa->predChoices.size());
const OneIndexAssignment& assign = aa->predChoices[aa->counter];
for (size_t i = 0; i < assign.preds.size(); ++i) {
MatchExpression* pred = assign.preds[i];
verify(NULL == pred->getTag());
pred->setTag(new IndexTag(assign.index, assign.positions[i]));
}
}
else {
verify(AndAssignment::SUBNODES == aa->state);
verify(aa->counter < aa->subnodes.size());
tagMemo(aa->subnodes[aa->counter]);
}
}
else {
verify(0);
}
}
// static
void QueryPlannerIXSelect::stripUnneededAssignments(MatchExpression* node,
const std::vector<IndexEntry>& indices) {
if (MatchExpression::AND == node->matchType()) {
for (size_t i = 0; i < node->numChildren(); i++) {
MatchExpression* child = node->getChild(i);
if (MatchExpression::EQ != child->matchType()) {
continue;
}
if (!child->getTag()) {
continue;
}
// We found a EQ child of an AND which is tagged.
RelevantTag* rt = static_cast<RelevantTag*>(child->getTag());
// Look through all of the indices for which this predicate can be answered with
// the leading field of the index.
for (std::vector<size_t>::const_iterator i = rt->first.begin();
i != rt->first.end(); ++i) {
size_t index = *i;
if (indices[index].unique && 1 == indices[index].keyPattern.nFields()) {
// Found an EQ predicate which can use a single-field unique index.
// Clear assignments from the entire tree, and add back a single assignment
// for 'child' to the unique index.
clearAssignments(node);
RelevantTag* newRt = static_cast<RelevantTag*>(child->getTag());
newRt->first.push_back(index);
// Tag state has been reset in the entire subtree at 'root'; nothing
// else for us to do.
return;
}
}
}
}
for (size_t i = 0; i < node->numChildren(); i++) {
stripUnneededAssignments(node->getChild(i), indices);
}
}
void tagForSort(MatchExpression* tree) {
if (!Indexability::nodeCanUseIndexOnOwnField(tree)) {
size_t myTagValue = IndexTag::kNoIndex;
for (size_t i = 0; i < tree->numChildren(); ++i) {
MatchExpression* child = tree->getChild(i);
tagForSort(child);
IndexTag* childTag = static_cast<IndexTag*>(child->getTag());
if (NULL != childTag) {
myTagValue = std::min(myTagValue, childTag->index);
}
}
if (myTagValue != IndexTag::kNoIndex) {
tree->setTag(new IndexTag(myTagValue));
}
}
}
/**
* Traverse the subtree rooted at 'node' to remove invalid RelevantTag assignments to text index
* 'idx', which has prefix paths 'prefixPaths'.
*/
static void stripInvalidAssignmentsToTextIndex(MatchExpression* node,
size_t idx,
const unordered_set<StringData, StringData::Hasher>& prefixPaths) {
// If we're here, there are prefixPaths and node is either:
// 1. a text pred which we can't use as we have nothing over its prefix, or
// 2. a non-text pred which we can't use as we don't have a text pred AND-related.
if (Indexability::nodeCanUseIndexOnOwnField(node)) {
removeIndexRelevantTag(node, idx);
return;
}
// Do not traverse tree beyond negation node.
if (node->matchType() == MatchExpression::NOT
|| node->matchType() == MatchExpression::NOR) {
return;
}
// For anything to use a text index with prefixes, we require that:
// 1. The text pred exists in an AND,
// 2. The non-text preds that use the text index's prefixes are also in that AND.
if (node->matchType() != MatchExpression::AND) {
// It's an OR or some kind of array operator.
for (size_t i = 0; i < node->numChildren(); ++i) {
stripInvalidAssignmentsToTextIndex(node->getChild(i), idx, prefixPaths);
}
return;
}
// If we're here, we're an AND. Determine whether the children satisfy the index prefix for
// the text index.
invariant(node->matchType() == MatchExpression::AND);
bool hasText = false;
// The AND must have an EQ predicate for each prefix path. When we encounter a child with a
// tag we remove it from childrenPrefixPaths. All children exist if this set is empty at
// the end.
unordered_set<StringData, StringData::Hasher> childrenPrefixPaths = prefixPaths;
for (size_t i = 0; i < node->numChildren(); ++i) {
MatchExpression* child = node->getChild(i);
RelevantTag* tag = static_cast<RelevantTag*>(child->getTag());
if (NULL == tag) {
// 'child' could be a logical operator. Maybe there are some assignments hiding
// inside.
stripInvalidAssignmentsToTextIndex(child, idx, prefixPaths);
continue;
}
bool inFirst = tag->first.end() != std::find(tag->first.begin(),
tag->first.end(),
idx);
bool inNotFirst = tag->notFirst.end() != std::find(tag->notFirst.begin(),
tag->notFirst.end(),
idx);
if (inFirst || inNotFirst) {
// Great! 'child' was assigned to our index.
if (child->matchType() == MatchExpression::TEXT) {
hasText = true;
}
else {
childrenPrefixPaths.erase(child->path());
// One fewer prefix we're looking for, possibly. Note that we could have a
// suffix assignment on the index and wind up here. In this case the erase
// above won't do anything since a suffix isn't a prefix.
}
}
else {
// Recurse on the children to ensure that they're not hiding any assignments
// to idx.
stripInvalidAssignmentsToTextIndex(child, idx, prefixPaths);
}
}
// Our prereqs for using the text index were not satisfied so we remove the assignments from
// all children of the AND.
if (!hasText || !childrenPrefixPaths.empty()) {
for (size_t i = 0; i < node->numChildren(); ++i) {
stripInvalidAssignmentsToTextIndex(node->getChild(i), idx, prefixPaths);
}
}
}
static void stripInvalidAssignmentsTo2dsphereIndex(MatchExpression* node, size_t idx) {
if (Indexability::nodeCanUseIndexOnOwnField(node)) {
removeIndexRelevantTag(node, idx);
return;
}
const MatchExpression::MatchType nodeType = node->matchType();
// Don't bother peeking inside of negations.
if (MatchExpression::NOT == nodeType || MatchExpression::NOR == nodeType) {
return;
}
if (MatchExpression::AND != nodeType) {
// It's an OR or some kind of array operator.
for (size_t i = 0; i < node->numChildren(); ++i) {
stripInvalidAssignmentsTo2dsphereIndex(node->getChild(i), idx);
}
return;
}
bool hasGeoField = false;
for (size_t i = 0; i < node->numChildren(); ++i) {
MatchExpression* child = node->getChild(i);
RelevantTag* tag = static_cast<RelevantTag*>(child->getTag());
if (NULL == tag) {
// 'child' could be a logical operator. Maybe there are some assignments hiding
// inside.
stripInvalidAssignmentsTo2dsphereIndex(child, idx);
continue;
}
bool inFirst = tag->first.end() != std::find(tag->first.begin(),
tag->first.end(),
idx);
bool inNotFirst = tag->notFirst.end() != std::find(tag->notFirst.begin(),
tag->notFirst.end(),
idx);
// If there is an index assignment...
if (inFirst || inNotFirst) {
// And it's a geo predicate...
if (MatchExpression::GEO == child->matchType() ||
MatchExpression::GEO_NEAR == child->matchType()) {
hasGeoField = true;
}
}
else {
// Recurse on the children to ensure that they're not hiding any assignments
// to idx.
stripInvalidAssignmentsTo2dsphereIndex(child, idx);
}
}
// If there isn't a geo predicate our results aren't a subset of what's in the geo index, so
// if we use the index we'll miss results.
if (!hasGeoField) {
for (size_t i = 0; i < node->numChildren(); ++i) {
stripInvalidAssignmentsTo2dsphereIndex(node->getChild(i), idx);
}
}
}
// static
Status QueryPlanner::plan(const CanonicalQuery& query,
const QueryPlannerParams& params,
std::vector<QuerySolution*>* out) {
LOG(5) << "Beginning planning..." << endl
<< "=============================" << endl
<< "Options = " << optionString(params.options) << endl
<< "Canonical query:" << endl
<< query.toString() << "=============================" << endl;
for (size_t i = 0; i < params.indices.size(); ++i) {
LOG(5) << "Index " << i << " is " << params.indices[i].toString() << endl;
}
bool canTableScan = !(params.options & QueryPlannerParams::NO_TABLE_SCAN);
// If the query requests a tailable cursor, the only solution is a collscan + filter with
// tailable set on the collscan. TODO: This is a policy departure. Previously I think you
// could ask for a tailable cursor and it just tried to give you one. Now, we fail if we
// can't provide one. Is this what we want?
if (query.getParsed().isTailable()) {
if (!QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR) && canTableScan) {
QuerySolution* soln = buildCollscanSoln(query, true, params);
if (NULL != soln) {
out->push_back(soln);
}
}
return Status::OK();
}
// The hint or sort can be $natural: 1. If this happens, output a collscan. If both
// a $natural hint and a $natural sort are specified, then the direction of the collscan
// is determined by the sign of the sort (not the sign of the hint).
if (!query.getParsed().getHint().isEmpty() || !query.getParsed().getSort().isEmpty()) {
BSONObj hintObj = query.getParsed().getHint();
BSONObj sortObj = query.getParsed().getSort();
BSONElement naturalHint = hintObj.getFieldDotted("$natural");
BSONElement naturalSort = sortObj.getFieldDotted("$natural");
// A hint overrides a $natural sort. This means that we don't force a table
// scan if there is a $natural sort with a non-$natural hint.
if (!naturalHint.eoo() || (!naturalSort.eoo() && hintObj.isEmpty())) {
LOG(5) << "Forcing a table scan due to hinted $natural\n";
// min/max are incompatible with $natural.
if (canTableScan && query.getParsed().getMin().isEmpty() &&
query.getParsed().getMax().isEmpty()) {
QuerySolution* soln = buildCollscanSoln(query, false, params);
if (NULL != soln) {
out->push_back(soln);
}
}
return Status::OK();
}
}
// Figure out what fields we care about.
unordered_set<string> fields;
QueryPlannerIXSelect::getFields(query.root(), "", &fields);
for (unordered_set<string>::const_iterator it = fields.begin(); it != fields.end(); ++it) {
LOG(5) << "Predicate over field '" << *it << "'" << endl;
}
// Filter our indices so we only look at indices that are over our predicates.
vector<IndexEntry> relevantIndices;
// Hints require us to only consider the hinted index.
// If index filters in the query settings were used to override
// the allowed indices for planning, we should not use the hinted index
// requested in the query.
BSONObj hintIndex;
if (!params.indexFiltersApplied) {
hintIndex = query.getParsed().getHint();
}
// Snapshot is a form of a hint. If snapshot is set, try to use _id index to make a real
// plan. If that fails, just scan the _id index.
if (query.getParsed().isSnapshot()) {
// Find the ID index in indexKeyPatterns. It's our hint.
for (size_t i = 0; i < params.indices.size(); ++i) {
if (isIdIndex(params.indices[i].keyPattern)) {
hintIndex = params.indices[i].keyPattern;
break;
}
}
}
size_t hintIndexNumber = numeric_limits<size_t>::max();
if (hintIndex.isEmpty()) {
QueryPlannerIXSelect::findRelevantIndices(fields, params.indices, &relevantIndices);
} else {
// Sigh. If the hint is specified it might be using the index name.
BSONElement firstHintElt = hintIndex.firstElement();
if (str::equals("$hint", firstHintElt.fieldName()) && String == firstHintElt.type()) {
string hintName = firstHintElt.String();
for (size_t i = 0; i < params.indices.size(); ++i) {
if (params.indices[i].name == hintName) {
LOG(5) << "Hint by name specified, restricting indices to "
<< params.indices[i].keyPattern.toString() << endl;
relevantIndices.clear();
relevantIndices.push_back(params.indices[i]);
hintIndexNumber = i;
hintIndex = params.indices[i].keyPattern;
break;
}
}
} else {
for (size_t i = 0; i < params.indices.size(); ++i) {
if (0 == params.indices[i].keyPattern.woCompare(hintIndex)) {
relevantIndices.clear();
relevantIndices.push_back(params.indices[i]);
LOG(5) << "Hint specified, restricting indices to " << hintIndex.toString()
<< endl;
hintIndexNumber = i;
break;
}
}
}
if (hintIndexNumber == numeric_limits<size_t>::max()) {
return Status(ErrorCodes::BadValue, "bad hint");
}
}
// Deal with the .min() and .max() query options. If either exist we can only use an index
// that matches the object inside.
if (!query.getParsed().getMin().isEmpty() || !query.getParsed().getMax().isEmpty()) {
BSONObj minObj = query.getParsed().getMin();
BSONObj maxObj = query.getParsed().getMax();
// The unfinished siblings of these objects may not be proper index keys because they
// may be empty objects or have field names. When an index is picked to use for the
// min/max query, these "finished" objects will always be valid index keys for the
// index's key pattern.
BSONObj finishedMinObj;
BSONObj finishedMaxObj;
// This is the index into params.indices[...] that we use.
size_t idxNo = numeric_limits<size_t>::max();
// If there's an index hinted we need to be able to use it.
if (!hintIndex.isEmpty()) {
if (!minObj.isEmpty() && !indexCompatibleMaxMin(minObj, hintIndex)) {
LOG(5) << "Minobj doesn't work with hint";
return Status(ErrorCodes::BadValue, "hint provided does not work with min query");
}
if (!maxObj.isEmpty() && !indexCompatibleMaxMin(maxObj, hintIndex)) {
LOG(5) << "Maxobj doesn't work with hint";
return Status(ErrorCodes::BadValue, "hint provided does not work with max query");
}
const BSONObj& kp = params.indices[hintIndexNumber].keyPattern;
finishedMinObj = finishMinObj(kp, minObj, maxObj);
finishedMaxObj = finishMaxObj(kp, minObj, maxObj);
// The min must be less than the max for the hinted index ordering.
if (0 <= finishedMinObj.woCompare(finishedMaxObj, kp, false)) {
LOG(5) << "Minobj/Maxobj don't work with hint";
return Status(ErrorCodes::BadValue,
"hint provided does not work with min/max query");
}
idxNo = hintIndexNumber;
} else {
// No hinted index, look for one that is compatible (has same field names and
// ordering thereof).
for (size_t i = 0; i < params.indices.size(); ++i) {
const BSONObj& kp = params.indices[i].keyPattern;
BSONObj toUse = minObj.isEmpty() ? maxObj : minObj;
if (indexCompatibleMaxMin(toUse, kp)) {
// In order to be fully compatible, the min has to be less than the max
// according to the index key pattern ordering. The first step in verifying
// this is "finish" the min and max by replacing empty objects and stripping
// field names.
finishedMinObj = finishMinObj(kp, minObj, maxObj);
finishedMaxObj = finishMaxObj(kp, minObj, maxObj);
// Now we have the final min and max. This index is only relevant for
// the min/max query if min < max.
if (0 >= finishedMinObj.woCompare(finishedMaxObj, kp, false)) {
// Found a relevant index.
idxNo = i;
break;
}
// This index is not relevant; move on to the next.
}
}
}
if (idxNo == numeric_limits<size_t>::max()) {
LOG(5) << "Can't find relevant index to use for max/min query";
// Can't find an index to use, bail out.
return Status(ErrorCodes::BadValue, "unable to find relevant index for max/min query");
}
LOG(5) << "Max/min query using index " << params.indices[idxNo].toString() << endl;
// Make our scan and output.
QuerySolutionNode* solnRoot = QueryPlannerAccess::makeIndexScan(
params.indices[idxNo], query, params, finishedMinObj, finishedMaxObj);
QuerySolution* soln = QueryPlannerAnalysis::analyzeDataAccess(query, params, solnRoot);
if (NULL != soln) {
out->push_back(soln);
}
return Status::OK();
}
for (size_t i = 0; i < relevantIndices.size(); ++i) {
LOG(2) << "Relevant index " << i << " is " << relevantIndices[i].toString() << endl;
}
// Figure out how useful each index is to each predicate.
QueryPlannerIXSelect::rateIndices(query.root(), "", relevantIndices);
QueryPlannerIXSelect::stripInvalidAssignments(query.root(), relevantIndices);
// Unless we have GEO_NEAR, TEXT, or a projection, we may be able to apply an optimization
// in which we strip unnecessary index assignments.
//
// Disallowed with projection because assignment to a non-unique index can allow the plan
// to be covered.
//
// TEXT and GEO_NEAR are special because they require the use of a text/geo index in order
// to be evaluated correctly. Stripping these "mandatory assignments" is therefore invalid.
if (query.getParsed().getProj().isEmpty() &&
!QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR) &&
!QueryPlannerCommon::hasNode(query.root(), MatchExpression::TEXT)) {
QueryPlannerIXSelect::stripUnneededAssignments(query.root(), relevantIndices);
}
// query.root() is now annotated with RelevantTag(s).
LOG(5) << "Rated tree:" << endl
<< query.root()->toString();
// If there is a GEO_NEAR it must have an index it can use directly.
MatchExpression* gnNode = NULL;
if (QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR, &gnNode)) {
// No index for GEO_NEAR? No query.
RelevantTag* tag = static_cast<RelevantTag*>(gnNode->getTag());
if (0 == tag->first.size() && 0 == tag->notFirst.size()) {
LOG(5) << "Unable to find index for $geoNear query." << endl;
// Don't leave tags on query tree.
query.root()->resetTag();
return Status(ErrorCodes::BadValue, "unable to find index for $geoNear query");
}
LOG(5) << "Rated tree after geonear processing:" << query.root()->toString();
}
// Likewise, if there is a TEXT it must have an index it can use directly.
MatchExpression* textNode = NULL;
if (QueryPlannerCommon::hasNode(query.root(), MatchExpression::TEXT, &textNode)) {
RelevantTag* tag = static_cast<RelevantTag*>(textNode->getTag());
// Exactly one text index required for TEXT. We need to check this explicitly because
// the text stage can't be built if no text index exists or there is an ambiguity as to
// which one to use.
size_t textIndexCount = 0;
for (size_t i = 0; i < params.indices.size(); i++) {
if (INDEX_TEXT == params.indices[i].type) {
textIndexCount++;
}
}
if (textIndexCount != 1) {
// Don't leave tags on query tree.
query.root()->resetTag();
return Status(ErrorCodes::BadValue, "need exactly one text index for $text query");
}
// Error if the text node is tagged with zero indices.
if (0 == tag->first.size() && 0 == tag->notFirst.size()) {
// Don't leave tags on query tree.
query.root()->resetTag();
return Status(ErrorCodes::BadValue,
"failed to use text index to satisfy $text query (if text index is "
"compound, are equality predicates given for all prefix fields?)");
}
// At this point, we know that there is only one text index and that the TEXT node is
// assigned to it.
invariant(1 == tag->first.size() + tag->notFirst.size());
LOG(5) << "Rated tree after text processing:" << query.root()->toString();
}
// If we have any relevant indices, we try to create indexed plans.
if (0 < relevantIndices.size()) {
// The enumerator spits out trees tagged with IndexTag(s).
PlanEnumeratorParams enumParams;
enumParams.intersect = params.options & QueryPlannerParams::INDEX_INTERSECTION;
enumParams.root = query.root();
enumParams.indices = &relevantIndices;
PlanEnumerator isp(enumParams);
isp.init();
MatchExpression* rawTree;
while (isp.getNext(&rawTree) && (out->size() < params.maxIndexedSolutions)) {
LOG(5) << "About to build solntree from tagged tree:" << endl
<< rawTree->toString();
// The tagged tree produced by the plan enumerator is not guaranteed
// to be canonically sorted. In order to be compatible with the cached
// data, sort the tagged tree according to CanonicalQuery ordering.
std::unique_ptr<MatchExpression> clone(rawTree->shallowClone());
CanonicalQuery::sortTree(clone.get());
PlanCacheIndexTree* cacheData;
Status indexTreeStatus =
cacheDataFromTaggedTree(clone.get(), relevantIndices, &cacheData);
if (!indexTreeStatus.isOK()) {
LOG(5) << "Query is not cachable: " << indexTreeStatus.reason() << endl;
}
unique_ptr<PlanCacheIndexTree> autoData(cacheData);
// This can fail if enumeration makes a mistake.
QuerySolutionNode* solnRoot = QueryPlannerAccess::buildIndexedDataAccess(
query, rawTree, false, relevantIndices, params);
if (NULL == solnRoot) {
continue;
}
QuerySolution* soln = QueryPlannerAnalysis::analyzeDataAccess(query, params, solnRoot);
if (NULL != soln) {
LOG(5) << "Planner: adding solution:" << endl
<< soln->toString();
if (indexTreeStatus.isOK()) {
SolutionCacheData* scd = new SolutionCacheData();
scd->tree.reset(autoData.release());
soln->cacheData.reset(scd);
}
out->push_back(soln);
}
}
}
// Don't leave tags on query tree.
query.root()->resetTag();
LOG(5) << "Planner: outputted " << out->size() << " indexed solutions.\n";
// Produce legible error message for failed OR planning with a TEXT child.
// TODO: support collection scan for non-TEXT children of OR.
if (out->size() == 0 && textNode != NULL && MatchExpression::OR == query.root()->matchType()) {
MatchExpression* root = query.root();
for (size_t i = 0; i < root->numChildren(); ++i) {
if (textNode == root->getChild(i)) {
return Status(ErrorCodes::BadValue,
"Failed to produce a solution for TEXT under OR - "
"other non-TEXT clauses under OR have to be indexed as well.");
}
}
}
// An index was hinted. If there are any solutions, they use the hinted index. If not, we
// scan the entire index to provide results and output that as our plan. This is the
// desired behavior when an index is hinted that is not relevant to the query.
if (!hintIndex.isEmpty()) {
if (0 == out->size()) {
QuerySolution* soln = buildWholeIXSoln(params.indices[hintIndexNumber], query, params);
verify(NULL != soln);
LOG(5) << "Planner: outputting soln that uses hinted index as scan." << endl;
out->push_back(soln);
}
return Status::OK();
}
// If a sort order is requested, there may be an index that provides it, even if that
// index is not over any predicates in the query.
//
if (!query.getParsed().getSort().isEmpty() &&
!QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR) &&
!QueryPlannerCommon::hasNode(query.root(), MatchExpression::TEXT)) {
// See if we have a sort provided from an index already.
// This is implied by the presence of a non-blocking solution.
bool usingIndexToSort = false;
for (size_t i = 0; i < out->size(); ++i) {
QuerySolution* soln = (*out)[i];
if (!soln->hasBlockingStage) {
usingIndexToSort = true;
break;
}
}
if (!usingIndexToSort) {
for (size_t i = 0; i < params.indices.size(); ++i) {
const IndexEntry& index = params.indices[i];
// Only regular (non-plugin) indexes can be used to provide a sort, and only
// non-sparse indexes can be used to provide a sort.
//
// TODO: Sparse indexes can't normally provide a sort, because non-indexed
// documents could potentially be missing from the result set. However, if the
// query predicate can be used to guarantee that all documents to be returned
// are indexed, then the index should be able to provide the sort.
//
// For example:
// - Sparse index {a: 1, b: 1} should be able to provide a sort for
// find({b: 1}).sort({a: 1}). SERVER-13908.
// - Index {a: 1, b: "2dsphere"} (which is "geo-sparse", if
// 2dsphereIndexVersion=2) should be able to provide a sort for
// find({b: GEO}).sort({a:1}). SERVER-10801.
if (index.type != INDEX_BTREE) {
continue;
}
if (index.sparse) {
continue;
}
// Partial indexes can only be used to provide a sort only if the query predicate is
// compatible.
if (index.filterExpr && !expression::isSubsetOf(query.root(), index.filterExpr)) {
continue;
}
const BSONObj kp = QueryPlannerAnalysis::getSortPattern(index.keyPattern);
if (providesSort(query, kp)) {
LOG(5) << "Planner: outputting soln that uses index to provide sort." << endl;
QuerySolution* soln = buildWholeIXSoln(params.indices[i], query, params);
if (NULL != soln) {
PlanCacheIndexTree* indexTree = new PlanCacheIndexTree();
indexTree->setIndexEntry(params.indices[i]);
SolutionCacheData* scd = new SolutionCacheData();
scd->tree.reset(indexTree);
scd->solnType = SolutionCacheData::WHOLE_IXSCAN_SOLN;
scd->wholeIXSolnDir = 1;
soln->cacheData.reset(scd);
out->push_back(soln);
break;
}
}
if (providesSort(query, QueryPlannerCommon::reverseSortObj(kp))) {
LOG(5) << "Planner: outputting soln that uses (reverse) index "
<< "to provide sort." << endl;
QuerySolution* soln = buildWholeIXSoln(params.indices[i], query, params, -1);
if (NULL != soln) {
PlanCacheIndexTree* indexTree = new PlanCacheIndexTree();
indexTree->setIndexEntry(params.indices[i]);
SolutionCacheData* scd = new SolutionCacheData();
scd->tree.reset(indexTree);
scd->solnType = SolutionCacheData::WHOLE_IXSCAN_SOLN;
scd->wholeIXSolnDir = -1;
soln->cacheData.reset(scd);
out->push_back(soln);
break;
}
}
}
}
}
// geoNear and text queries *require* an index.
// Also, if a hint is specified it indicates that we MUST use it.
bool possibleToCollscan =
!QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR) &&
!QueryPlannerCommon::hasNode(query.root(), MatchExpression::TEXT) && hintIndex.isEmpty();
// The caller can explicitly ask for a collscan.
bool collscanRequested = (params.options & QueryPlannerParams::INCLUDE_COLLSCAN);
// No indexed plans? We must provide a collscan if possible or else we can't run the query.
bool collscanNeeded = (0 == out->size() && canTableScan);
if (possibleToCollscan && (collscanRequested || collscanNeeded)) {
QuerySolution* collscan = buildCollscanSoln(query, false, params);
if (NULL != collscan) {
SolutionCacheData* scd = new SolutionCacheData();
scd->solnType = SolutionCacheData::COLLSCAN_SOLN;
collscan->cacheData.reset(scd);
out->push_back(collscan);
LOG(5) << "Planner: outputting a collscan:" << endl
<< collscan->toString();
}
}
return Status::OK();
}
// static
void QueryPlanner::plan(const CanonicalQuery& query,
const QueryPlannerParams& params,
vector<QuerySolution*>* out) {
QLOG() << "=============================\n"
<< "Beginning planning, options = " << optionString(params.options) << endl
<< "Canonical query:\n" << query.toString() << endl
<< "============================="
<< endl;
// The shortcut formerly known as IDHACK. See if it's a simple _id query. If so we might
// just make an ixscan over the _id index and bypass the rest of planning entirely.
if (!query.getParsed().isExplain() && !query.getParsed().showDiskLoc()
&& isSimpleIdQuery(query.getParsed().getFilter())
&& !query.getParsed().hasOption(QueryOption_CursorTailable)) {
// See if we can find an _id index.
for (size_t i = 0; i < params.indices.size(); ++i) {
if (isIdIndex(params.indices[i].keyPattern)) {
const IndexEntry& index = params.indices[i];
QLOG() << "IDHACK using index " << index.toString() << endl;
// If so, we make a simple scan to find the doc.
IndexScanNode* isn = new IndexScanNode();
isn->indexKeyPattern = index.keyPattern;
isn->indexIsMultiKey = index.multikey;
isn->direction = 1;
isn->bounds.isSimpleRange = true;
BSONObj key = getKeyFromQuery(index.keyPattern, query.getParsed().getFilter());
isn->bounds.startKey = isn->bounds.endKey = key;
isn->bounds.endKeyInclusive = true;
isn->computeProperties();
QuerySolution* soln = QueryPlannerAnalysis::analyzeDataAccess(query, params, isn);
if (NULL != soln) {
out->push_back(soln);
QLOG() << "IDHACK solution is:\n" << (*out)[0]->toString() << endl;
// And that's it.
return;
}
}
}
}
for (size_t i = 0; i < params.indices.size(); ++i) {
QLOG() << "idx " << i << " is " << params.indices[i].toString() << endl;
}
bool canTableScan = !(params.options & QueryPlannerParams::NO_TABLE_SCAN);
// If the query requests a tailable cursor, the only solution is a collscan + filter with
// tailable set on the collscan. TODO: This is a policy departure. Previously I think you
// could ask for a tailable cursor and it just tried to give you one. Now, we fail if we
// can't provide one. Is this what we want?
if (query.getParsed().hasOption(QueryOption_CursorTailable)) {
if (!QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR)
&& canTableScan) {
QuerySolution* soln = buildCollscanSoln(query, true, params);
if (NULL != soln) {
out->push_back(soln);
}
}
return;
}
// The hint can be $natural: 1. If this happens, output a collscan. It's a weird way of
// saying "table scan for two, please."
if (!query.getParsed().getHint().isEmpty()) {
BSONElement natural = query.getParsed().getHint().getFieldDotted("$natural");
if (!natural.eoo()) {
QLOG() << "forcing a table scan due to hinted $natural\n";
if (canTableScan) {
QuerySolution* soln = buildCollscanSoln(query, false, params);
if (NULL != soln) {
out->push_back(soln);
}
}
return;
}
}
// NOR and NOT we can't handle well with indices. If we see them here, they weren't
// rewritten to remove the negation. Just output a collscan for those.
if (QueryPlannerCommon::hasNode(query.root(), MatchExpression::NOT)
|| QueryPlannerCommon::hasNode(query.root(), MatchExpression::NOR)) {
// If there's a near predicate, we can't handle this.
// TODO: Should canonicalized query detect this?
if (QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR)) {
warning() << "Can't handle NOT/NOR with GEO_NEAR";
return;
}
QLOG() << "NOT/NOR in plan, just outtping a collscan\n";
if (canTableScan) {
QuerySolution* soln = buildCollscanSoln(query, false, params);
if (NULL != soln) {
out->push_back(soln);
}
}
return;
}
// Figure out what fields we care about.
unordered_set<string> fields;
QueryPlannerIXSelect::getFields(query.root(), "", &fields);
for (unordered_set<string>::const_iterator it = fields.begin(); it != fields.end(); ++it) {
QLOG() << "predicate over field " << *it << endl;
}
// Filter our indices so we only look at indices that are over our predicates.
vector<IndexEntry> relevantIndices;
// Hints require us to only consider the hinted index.
BSONObj hintIndex = query.getParsed().getHint();
// Snapshot is a form of a hint. If snapshot is set, try to use _id index to make a real
// plan. If that fails, just scan the _id index.
if (query.getParsed().isSnapshot()) {
// Find the ID index in indexKeyPatterns. It's our hint.
for (size_t i = 0; i < params.indices.size(); ++i) {
if (isIdIndex(params.indices[i].keyPattern)) {
hintIndex = params.indices[i].keyPattern;
break;
}
}
}
size_t hintIndexNumber = numeric_limits<size_t>::max();
if (!hintIndex.isEmpty()) {
// Sigh. If the hint is specified it might be using the index name.
BSONElement firstHintElt = hintIndex.firstElement();
if (str::equals("$hint", firstHintElt.fieldName()) && String == firstHintElt.type()) {
string hintName = firstHintElt.String();
for (size_t i = 0; i < params.indices.size(); ++i) {
if (params.indices[i].name == hintName) {
QLOG() << "hint by name specified, restricting indices to "
<< params.indices[i].keyPattern.toString() << endl;
relevantIndices.clear();
relevantIndices.push_back(params.indices[i]);
hintIndexNumber = i;
hintIndex = params.indices[i].keyPattern;
break;
}
}
}
else {
for (size_t i = 0; i < params.indices.size(); ++i) {
if (0 == params.indices[i].keyPattern.woCompare(hintIndex)) {
relevantIndices.clear();
relevantIndices.push_back(params.indices[i]);
QLOG() << "hint specified, restricting indices to " << hintIndex.toString()
<< endl;
hintIndexNumber = i;
break;
}
}
}
if (hintIndexNumber == numeric_limits<size_t>::max()) {
// This is supposed to be an error.
warning() << "Can't find hint for " << hintIndex.toString();
return;
}
}
else {
QLOG() << "Finding relevant indices\n";
QueryPlannerIXSelect::findRelevantIndices(fields, params.indices, &relevantIndices);
}
for (size_t i = 0; i < relevantIndices.size(); ++i) {
QLOG() << "relevant idx " << i << " is " << relevantIndices[i].toString() << endl;
}
// Figure out how useful each index is to each predicate.
// query.root() is now annotated with RelevantTag(s).
QueryPlannerIXSelect::rateIndices(query.root(), "", relevantIndices);
QLOG() << "rated tree" << endl;
QLOG() << query.root()->toString() << endl;
// If there is a GEO_NEAR it must have an index it can use directly.
// XXX: move into data access?
MatchExpression* gnNode = NULL;
if (QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR, &gnNode)) {
// No index for GEO_NEAR? No query.
RelevantTag* tag = static_cast<RelevantTag*>(gnNode->getTag());
if (0 == tag->first.size() && 0 == tag->notFirst.size()) {
return;
}
GeoNearMatchExpression* gnme = static_cast<GeoNearMatchExpression*>(gnNode);
vector<size_t> newFirst;
// 2d + GEO_NEAR is annoying. Because 2d's GEO_NEAR isn't streaming we have to embed
// the full query tree inside it as a matcher.
for (size_t i = 0; i < tag->first.size(); ++i) {
// GEO_NEAR has a non-2d index it can use. We can deal w/that in normal planning.
if (!is2DIndex(relevantIndices[tag->first[i]].keyPattern)) {
newFirst.push_back(i);
continue;
}
// If we're here, GEO_NEAR has a 2d index. We create a 2dgeonear plan with the
// entire tree as a filter, if possible.
GeoNear2DNode* solnRoot = new GeoNear2DNode();
solnRoot->nq = gnme->getData();
if (MatchExpression::GEO_NEAR != query.root()->matchType()) {
// root is an AND, clone and delete the GEO_NEAR child.
MatchExpression* filterTree = query.root()->shallowClone();
verify(MatchExpression::AND == filterTree->matchType());
bool foundChild = false;
for (size_t i = 0; i < filterTree->numChildren(); ++i) {
if (MatchExpression::GEO_NEAR == filterTree->getChild(i)->matchType()) {
foundChild = true;
filterTree->getChildVector()->erase(filterTree->getChildVector()->begin() + i);
break;
}
}
verify(foundChild);
solnRoot->filter.reset(filterTree);
}
solnRoot->numWanted = query.getParsed().getNumToReturn();
if (0 == solnRoot->numWanted) {
solnRoot->numWanted = 100;
}
solnRoot->indexKeyPattern = relevantIndices[tag->first[i]].keyPattern;
// Remove the 2d index. 2d can only be the first field, and we know there is
// only one GEO_NEAR, so we don't care if anyone else was assigned it; it'll
// only be first for gnNode.
tag->first.erase(tag->first.begin() + i);
QuerySolution* soln = QueryPlannerAnalysis::analyzeDataAccess(query, params, solnRoot);
if (NULL != soln) {
out->push_back(soln);
}
}
// Continue planning w/non-2d indices tagged for this pred.
tag->first.swap(newFirst);
if (0 == tag->first.size() && 0 == tag->notFirst.size()) {
return;
}
}
// Likewise, if there is a TEXT it must have an index it can use directly.
MatchExpression* textNode;
if (QueryPlannerCommon::hasNode(query.root(), MatchExpression::TEXT, &textNode)) {
RelevantTag* tag = static_cast<RelevantTag*>(textNode->getTag());
if (0 == tag->first.size() && 0 == tag->notFirst.size()) {
return;
}
}
// If we have any relevant indices, we try to create indexed plans.
if (0 < relevantIndices.size()) {
// The enumerator spits out trees tagged with IndexTag(s).
PlanEnumerator isp(query.root(), &relevantIndices);
isp.init();
MatchExpression* rawTree;
while (isp.getNext(&rawTree)) {
QLOG() << "about to build solntree from tagged tree:\n" << rawTree->toString()
<< endl;
// This can fail if enumeration makes a mistake.
QuerySolutionNode* solnRoot =
QueryPlannerAccess::buildIndexedDataAccess(query, rawTree, false, relevantIndices);
if (NULL == solnRoot) { continue; }
QuerySolution* soln = QueryPlannerAnalysis::analyzeDataAccess(query, params, solnRoot);
if (NULL != soln) {
QLOG() << "Planner: adding solution:\n" << soln->toString() << endl;
out->push_back(soln);
}
}
}
QLOG() << "Planner: outputted " << out->size() << " indexed solutions.\n";
// An index was hinted. If there are any solutions, they use the hinted index. If not, we
// scan the entire index to provide results and output that as our plan. This is the
// desired behavior when an index is hinted that is not relevant to the query.
if (!hintIndex.isEmpty() && (0 == out->size())) {
QuerySolution* soln = buildWholeIXSoln(params.indices[hintIndexNumber], query, params);
if (NULL != soln) {
QLOG() << "Planner: outputting soln that uses hinted index as scan." << endl;
out->push_back(soln);
}
return;
}
// If a sort order is requested, there may be an index that provides it, even if that
// index is not over any predicates in the query.
//
// XXX XXX: Can we do this even if the index is sparse? Might we miss things?
if (!query.getParsed().getSort().isEmpty()
&& !QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR)
&& !QueryPlannerCommon::hasNode(query.root(), MatchExpression::TEXT)) {
// See if we have a sort provided from an index already.
bool usingIndexToSort = false;
for (size_t i = 0; i < out->size(); ++i) {
QuerySolution* soln = (*out)[i];
if (!soln->hasSortStage) {
usingIndexToSort = true;
break;
}
}
if (!usingIndexToSort) {
for (size_t i = 0; i < params.indices.size(); ++i) {
const BSONObj& kp = params.indices[i].keyPattern;
if (providesSort(query, kp)) {
QLOG() << "Planner: outputting soln that uses index to provide sort."
<< endl;
QuerySolution* soln = buildWholeIXSoln(params.indices[i], query, params);
if (NULL != soln) {
out->push_back(soln);
break;
}
}
if (providesSort(query, QueryPlannerCommon::reverseSortObj(kp))) {
QLOG() << "Planner: outputting soln that uses (reverse) index "
<< "to provide sort." << endl;
QuerySolution* soln = buildWholeIXSoln(params.indices[i], query, params, -1);
if (NULL != soln) {
out->push_back(soln);
break;
}
}
}
}
}
// TODO: Do we always want to offer a collscan solution?
// XXX: currently disabling the always-use-a-collscan in order to find more planner bugs.
if ( !QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR)
&& !QueryPlannerCommon::hasNode(query.root(), MatchExpression::TEXT)
&& ((params.options & QueryPlannerParams::INCLUDE_COLLSCAN) || (0 == out->size() && canTableScan)))
{
QuerySolution* collscan = buildCollscanSoln(query, false, params);
if (NULL != collscan) {
out->push_back(collscan);
QLOG() << "Planner: outputting a collscan:\n";
QLOG() << collscan->toString() << endl;
}
}
}
// static
Status QueryPlanner::plan(const CanonicalQuery& query,
const QueryPlannerParams& params,
std::vector<QuerySolution*>* out) {
QLOG() << "=============================\n"
<< "Beginning planning, options = " << optionString(params.options) << endl
<< "Canonical query:\n" << query.toString() << endl
<< "============================="
<< endl;
for (size_t i = 0; i < params.indices.size(); ++i) {
QLOG() << "idx " << i << " is " << params.indices[i].toString() << endl;
}
bool canTableScan = !(params.options & QueryPlannerParams::NO_TABLE_SCAN);
// If the query requests a tailable cursor, the only solution is a collscan + filter with
// tailable set on the collscan. TODO: This is a policy departure. Previously I think you
// could ask for a tailable cursor and it just tried to give you one. Now, we fail if we
// can't provide one. Is this what we want?
if (query.getParsed().hasOption(QueryOption_CursorTailable)) {
if (!QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR)
&& canTableScan) {
QuerySolution* soln = buildCollscanSoln(query, true, params);
if (NULL != soln) {
out->push_back(soln);
}
}
return Status::OK();
}
// The hint can be $natural: 1. If this happens, output a collscan. It's a weird way of
// saying "table scan for two, please."
if (!query.getParsed().getHint().isEmpty()) {
BSONElement natural = query.getParsed().getHint().getFieldDotted("$natural");
if (!natural.eoo()) {
QLOG() << "forcing a table scan due to hinted $natural\n";
// min/max are incompatible with $natural.
if (canTableScan && query.getParsed().getMin().isEmpty()
&& query.getParsed().getMax().isEmpty()) {
QuerySolution* soln = buildCollscanSoln(query, false, params);
if (NULL != soln) {
out->push_back(soln);
}
}
return Status::OK();
}
}
// Figure out what fields we care about.
unordered_set<string> fields;
QueryPlannerIXSelect::getFields(query.root(), "", &fields);
for (unordered_set<string>::const_iterator it = fields.begin(); it != fields.end(); ++it) {
QLOG() << "predicate over field " << *it << endl;
}
// Filter our indices so we only look at indices that are over our predicates.
vector<IndexEntry> relevantIndices;
// Hints require us to only consider the hinted index.
BSONObj hintIndex = query.getParsed().getHint();
// Snapshot is a form of a hint. If snapshot is set, try to use _id index to make a real
// plan. If that fails, just scan the _id index.
if (query.getParsed().isSnapshot()) {
// Find the ID index in indexKeyPatterns. It's our hint.
for (size_t i = 0; i < params.indices.size(); ++i) {
if (isIdIndex(params.indices[i].keyPattern)) {
hintIndex = params.indices[i].keyPattern;
break;
}
}
}
size_t hintIndexNumber = numeric_limits<size_t>::max();
if (hintIndex.isEmpty()) {
QueryPlannerIXSelect::findRelevantIndices(fields, params.indices, &relevantIndices);
}
else {
// Sigh. If the hint is specified it might be using the index name.
BSONElement firstHintElt = hintIndex.firstElement();
if (str::equals("$hint", firstHintElt.fieldName()) && String == firstHintElt.type()) {
string hintName = firstHintElt.String();
for (size_t i = 0; i < params.indices.size(); ++i) {
if (params.indices[i].name == hintName) {
QLOG() << "hint by name specified, restricting indices to "
<< params.indices[i].keyPattern.toString() << endl;
relevantIndices.clear();
relevantIndices.push_back(params.indices[i]);
hintIndexNumber = i;
hintIndex = params.indices[i].keyPattern;
break;
}
}
}
else {
for (size_t i = 0; i < params.indices.size(); ++i) {
if (0 == params.indices[i].keyPattern.woCompare(hintIndex)) {
relevantIndices.clear();
relevantIndices.push_back(params.indices[i]);
QLOG() << "hint specified, restricting indices to " << hintIndex.toString()
<< endl;
hintIndexNumber = i;
break;
}
}
}
if (hintIndexNumber == numeric_limits<size_t>::max()) {
return Status(ErrorCodes::BadValue, "bad hint");
}
}
// Deal with the .min() and .max() query options. If either exist we can only use an index
// that matches the object inside.
if (!query.getParsed().getMin().isEmpty() || !query.getParsed().getMax().isEmpty()) {
BSONObj minObj = query.getParsed().getMin();
BSONObj maxObj = query.getParsed().getMax();
// This is the index into params.indices[...] that we use.
size_t idxNo = numeric_limits<size_t>::max();
// If there's an index hinted we need to be able to use it.
if (!hintIndex.isEmpty()) {
if (!minObj.isEmpty() && !indexCompatibleMaxMin(minObj, hintIndex)) {
QLOG() << "minobj doesnt work w hint";
return Status(ErrorCodes::BadValue,
"hint provided does not work with min query");
}
if (!maxObj.isEmpty() && !indexCompatibleMaxMin(maxObj, hintIndex)) {
QLOG() << "maxobj doesnt work w hint";
return Status(ErrorCodes::BadValue,
"hint provided does not work with max query");
}
idxNo = hintIndexNumber;
}
else {
// No hinted index, look for one that is compatible (has same field names and
// ordering thereof).
for (size_t i = 0; i < params.indices.size(); ++i) {
const BSONObj& kp = params.indices[i].keyPattern;
BSONObj toUse = minObj.isEmpty() ? maxObj : minObj;
if (indexCompatibleMaxMin(toUse, kp)) {
idxNo = i;
break;
}
}
}
if (idxNo == numeric_limits<size_t>::max()) {
QLOG() << "Can't find relevant index to use for max/min query";
// Can't find an index to use, bail out.
return Status(ErrorCodes::BadValue,
"unable to find relevant index for max/min query");
}
// maxObj can be empty; the index scan just goes until the end. minObj can't be empty
// though, so if it is, we make a minKey object.
if (minObj.isEmpty()) {
BSONObjBuilder bob;
bob.appendMinKey("");
minObj = bob.obj();
}
else {
// Must strip off the field names to make an index key.
minObj = stripFieldNames(minObj);
}
if (!maxObj.isEmpty()) {
// Must strip off the field names to make an index key.
maxObj = stripFieldNames(maxObj);
}
QLOG() << "max/min query using index " << params.indices[idxNo].toString() << endl;
// Make our scan and output.
QuerySolutionNode* solnRoot = QueryPlannerAccess::makeIndexScan(params.indices[idxNo],
query,
params,
minObj,
maxObj);
QuerySolution* soln = QueryPlannerAnalysis::analyzeDataAccess(query, params, solnRoot);
if (NULL != soln) {
out->push_back(soln);
}
return Status::OK();
}
for (size_t i = 0; i < relevantIndices.size(); ++i) {
QLOG() << "relevant idx " << i << " is " << relevantIndices[i].toString() << endl;
}
// Figure out how useful each index is to each predicate.
// query.root() is now annotated with RelevantTag(s).
QueryPlannerIXSelect::rateIndices(query.root(), "", relevantIndices);
QLOG() << "rated tree" << endl;
QLOG() << query.root()->toString() << endl;
// If there is a GEO_NEAR it must have an index it can use directly.
// XXX: move into data access?
MatchExpression* gnNode = NULL;
if (QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR, &gnNode)) {
// No index for GEO_NEAR? No query.
RelevantTag* tag = static_cast<RelevantTag*>(gnNode->getTag());
if (0 == tag->first.size() && 0 == tag->notFirst.size()) {
QLOG() << "unable to find index for $geoNear query" << endl;
return Status(ErrorCodes::BadValue, "unable to find index for $geoNear query");
}
GeoNearMatchExpression* gnme = static_cast<GeoNearMatchExpression*>(gnNode);
vector<size_t> newFirst;
// 2d + GEO_NEAR is annoying. Because 2d's GEO_NEAR isn't streaming we have to embed
// the full query tree inside it as a matcher.
for (size_t i = 0; i < tag->first.size(); ++i) {
// GEO_NEAR has a non-2d index it can use. We can deal w/that in normal planning.
if (!is2DIndex(relevantIndices[tag->first[i]].keyPattern)) {
newFirst.push_back(i);
continue;
}
// If we're here, GEO_NEAR has a 2d index. We create a 2dgeonear plan with the
// entire tree as a filter, if possible.
GeoNear2DNode* solnRoot = new GeoNear2DNode();
solnRoot->nq = gnme->getData();
if (NULL != query.getProj()) {
solnRoot->addPointMeta = query.getProj()->wantGeoNearPoint();
solnRoot->addDistMeta = query.getProj()->wantGeoNearDistance();
}
if (MatchExpression::GEO_NEAR != query.root()->matchType()) {
// root is an AND, clone and delete the GEO_NEAR child.
MatchExpression* filterTree = query.root()->shallowClone();
verify(MatchExpression::AND == filterTree->matchType());
bool foundChild = false;
for (size_t i = 0; i < filterTree->numChildren(); ++i) {
if (MatchExpression::GEO_NEAR == filterTree->getChild(i)->matchType()) {
foundChild = true;
filterTree->getChildVector()->erase(filterTree->getChildVector()->begin() + i);
break;
}
}
verify(foundChild);
solnRoot->filter.reset(filterTree);
}
solnRoot->numWanted = query.getParsed().getNumToReturn();
if (0 == solnRoot->numWanted) {
solnRoot->numWanted = 100;
}
solnRoot->indexKeyPattern = relevantIndices[tag->first[i]].keyPattern;
// Remove the 2d index. 2d can only be the first field, and we know there is
// only one GEO_NEAR, so we don't care if anyone else was assigned it; it'll
// only be first for gnNode.
tag->first.erase(tag->first.begin() + i);
QuerySolution* soln = QueryPlannerAnalysis::analyzeDataAccess(query, params, solnRoot);
if (NULL != soln) {
out->push_back(soln);
}
}
// Continue planning w/non-2d indices tagged for this pred.
tag->first.swap(newFirst);
if (0 == tag->first.size() && 0 == tag->notFirst.size()) {
return Status::OK();
}
}
// Likewise, if there is a TEXT it must have an index it can use directly.
MatchExpression* textNode;
if (QueryPlannerCommon::hasNode(query.root(), MatchExpression::TEXT, &textNode)) {
RelevantTag* tag = static_cast<RelevantTag*>(textNode->getTag());
if (0 == tag->first.size() && 0 == tag->notFirst.size()) {
return Status::OK();
}
}
// If we have any relevant indices, we try to create indexed plans.
if (0 < relevantIndices.size()) {
// The enumerator spits out trees tagged with IndexTag(s).
PlanEnumeratorParams enumParams;
enumParams.intersect = params.options & QueryPlannerParams::INDEX_INTERSECTION;
enumParams.root = query.root();
enumParams.indices = &relevantIndices;
PlanEnumerator isp(enumParams);
isp.init();
MatchExpression* rawTree;
// XXX: have limit on # of indexed solns we'll consider. We could have a perverse
// query and index that could make n^2 very unpleasant.
while (isp.getNext(&rawTree)) {
QLOG() << "about to build solntree from tagged tree:\n" << rawTree->toString()
<< endl;
// This can fail if enumeration makes a mistake.
QuerySolutionNode* solnRoot =
QueryPlannerAccess::buildIndexedDataAccess(query, rawTree, false, relevantIndices);
if (NULL == solnRoot) { continue; }
QuerySolution* soln = QueryPlannerAnalysis::analyzeDataAccess(query, params, solnRoot);
if (NULL != soln) {
QLOG() << "Planner: adding solution:\n" << soln->toString() << endl;
out->push_back(soln);
}
}
}
QLOG() << "Planner: outputted " << out->size() << " indexed solutions.\n";
// An index was hinted. If there are any solutions, they use the hinted index. If not, we
// scan the entire index to provide results and output that as our plan. This is the
// desired behavior when an index is hinted that is not relevant to the query.
if (!hintIndex.isEmpty()) {
if (0 == out->size()) {
QuerySolution* soln = buildWholeIXSoln(params.indices[hintIndexNumber], query, params);
verify(NULL != soln);
QLOG() << "Planner: outputting soln that uses hinted index as scan." << endl;
out->push_back(soln);
}
return Status::OK();
}
// If a sort order is requested, there may be an index that provides it, even if that
// index is not over any predicates in the query.
//
if (!query.getParsed().getSort().isEmpty()
&& !QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR)
&& !QueryPlannerCommon::hasNode(query.root(), MatchExpression::TEXT)) {
// See if we have a sort provided from an index already.
bool usingIndexToSort = false;
for (size_t i = 0; i < out->size(); ++i) {
QuerySolution* soln = (*out)[i];
if (!soln->hasSortStage) {
usingIndexToSort = true;
break;
}
}
if (!usingIndexToSort) {
for (size_t i = 0; i < params.indices.size(); ++i) {
const IndexEntry& index = params.indices[i];
if (index.sparse) {
continue;
}
const BSONObj kp = LiteParsedQuery::normalizeSortOrder(index.keyPattern);
if (providesSort(query, kp)) {
QLOG() << "Planner: outputting soln that uses index to provide sort."
<< endl;
QuerySolution* soln = buildWholeIXSoln(params.indices[i], query, params);
if (NULL != soln) {
out->push_back(soln);
break;
}
}
if (providesSort(query, QueryPlannerCommon::reverseSortObj(kp))) {
QLOG() << "Planner: outputting soln that uses (reverse) index "
<< "to provide sort." << endl;
QuerySolution* soln = buildWholeIXSoln(params.indices[i], query, params, -1);
if (NULL != soln) {
out->push_back(soln);
break;
}
}
}
}
}
// TODO: Do we always want to offer a collscan solution?
// XXX: currently disabling the always-use-a-collscan in order to find more planner bugs.
if ( !QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR)
&& !QueryPlannerCommon::hasNode(query.root(), MatchExpression::TEXT)
&& hintIndex.isEmpty()
&& ((params.options & QueryPlannerParams::INCLUDE_COLLSCAN) || (0 == out->size() && canTableScan)))
{
QuerySolution* collscan = buildCollscanSoln(query, false, params);
if (NULL != collscan) {
out->push_back(collscan);
QLOG() << "Planner: outputting a collscan:\n";
QLOG() << collscan->toString() << endl;
}
}
return Status::OK();
}
bool PlanEnumerator::prepMemo(MatchExpression* node) {
if (Indexability::nodeCanUseIndexOnOwnField(node)) {
// We only get here if our parent is an OR, an array operator, or we're the root.
// If we have no index tag there are no indices we can use.
if (NULL == node->getTag()) { return false; }
RelevantTag* rt = static_cast<RelevantTag*>(node->getTag());
// In order to definitely use an index it must be prefixed with our field.
// We don't consider notFirst indices here because we must be AND-related to a node
// that uses the first spot in that index, and we currently do not know that
// unless we're in an AND node.
if (0 == rt->first.size()) { return false; }
// We know we can use an index, so grab a memo spot.
size_t myMemoID;
NodeAssignment* assign;
allocateAssignment(node, &assign, &myMemoID);
assign->pred.reset(new PredicateAssignment());
assign->pred->expr = node;
assign->pred->first.swap(rt->first);
return true;
}
else if (MatchExpression::OR == node->matchType()) {
// For an OR to be indexed, all its children must be indexed.
for (size_t i = 0; i < node->numChildren(); ++i) {
if (!prepMemo(node->getChild(i))) {
return false;
}
}
// If we're here we're fully indexed and can be in the memo.
size_t myMemoID;
NodeAssignment* assign;
allocateAssignment(node, &assign, &myMemoID);
OrAssignment* orAssignment = new OrAssignment();
for (size_t i = 0; i < node->numChildren(); ++i) {
orAssignment->subnodes.push_back(_nodeToId[node->getChild(i)]);
}
assign->orAssignment.reset(orAssignment);
return true;
}
else if (MatchExpression::AND == node->matchType() || Indexability::arrayUsesIndexOnChildren(node)) {
// map from idx id to children that have a pred over it.
unordered_map<IndexID, vector<MatchExpression*> > idxToFirst;
unordered_map<IndexID, vector<MatchExpression*> > idxToNotFirst;
vector<MemoID> subnodes;
for (size_t i = 0; i < node->numChildren(); ++i) {
MatchExpression* child = node->getChild(i);
if (Indexability::nodeCanUseIndexOnOwnField(child)) {
RelevantTag* rt = static_cast<RelevantTag*>(child->getTag());
for (size_t j = 0; j < rt->first.size(); ++j) {
idxToFirst[rt->first[j]].push_back(child);
}
for (size_t j = 0 ; j< rt->notFirst.size(); ++j) {
idxToNotFirst[rt->notFirst[j]].push_back(child);
}
}
else {
if (prepMemo(child)) {
verify(_nodeToId.end() != _nodeToId.find(child));
size_t childID = _nodeToId[child];
subnodes.push_back(childID);
}
}
}
if (idxToFirst.empty() && (subnodes.size() == 0)) { return false; }
AndAssignment* newAndAssignment = new AndAssignment();
newAndAssignment->subnodes.swap(subnodes);
// At this point we know how many indices the AND's predicate children are over.
newAndAssignment->predChoices.resize(idxToFirst.size());
// This iterates through the predChoices.
size_t predChoicesIdx = 0;
// For each FIRST, we assign nodes to it.
for (unordered_map<IndexID, vector<MatchExpression*> >::iterator it = idxToFirst.begin(); it != idxToFirst.end(); ++it) {
OneIndexAssignment* assign = &newAndAssignment->predChoices[predChoicesIdx];
++predChoicesIdx;
// Fill out the OneIndexAssignment with the preds that are over the first field.
assign->index = it->first;
// We can swap because we're never touching idxToFirst again after this loop over it.
assign->preds.swap(it->second);
// If it's a multikey index, we can't intersect the bounds, so we only want one pred.
if ((*_indices)[it->first].multikey) {
// XXX: pick a better pred than the first one that happens to wander in.
// XXX: see and3.js, indexq.js, arrayfind7.js
QLOG() << "Index " << (*_indices)[it->first].keyPattern.toString()
<< " is multikey but has >1 pred possible, should be smarter"
<< " here and pick the best one"
<< endl;
assign->preds.resize(1);
}
assign->positions.resize(assign->preds.size(), 0);
//
// Compound analysis here and below.
//
// Don't compound on multikey indices. (XXX: not whole story...)
if ((*_indices)[it->first].multikey) { continue; }
// Grab the expressions that are notFirst for the index whose assignments we're filling out.
unordered_map<size_t, vector<MatchExpression*> >::const_iterator compoundIt = idxToNotFirst.find(it->first);
if (compoundIt == idxToNotFirst.end()) { continue; }
const vector<MatchExpression*>& tryCompound = compoundIt->second;
// Walk over the key pattern trying to find
BSONObjIterator kpIt((*_indices)[it->first].keyPattern);
// Skip the first elt as it's already assigned.
kpIt.next();
size_t posInIdx = 0;
while (kpIt.more()) {
BSONElement keyElt = kpIt.next();
++posInIdx;
bool fieldAssigned = false;
for (size_t j = 0; j < tryCompound.size(); ++j) {
MatchExpression* maybe = tryCompound[j];
// Sigh we grab the full path from the relevant tag.
RelevantTag* rt = static_cast<RelevantTag*>(maybe->getTag());
if (keyElt.fieldName() == rt->path) {
assign->preds.push_back(maybe);
assign->positions.push_back(posInIdx);
fieldAssigned = true;
}
}
// If we have (a,b,c) and we can't assign something to 'b' don't try
// to assign something to 'c'.
if (!fieldAssigned) { break; }
}
}
// Some predicates *require* an index. We stuff these in 'mandatory' inside of the
// AndAssignment.
//
// TODO: We can compute this "on the fly" above somehow, but it's clearer to see what's
// going on when we do this as a separate step.
//
// TODO: Consider annotating mandatory indices in the planner as part of the available
// index tagging.
// Note we're not incrementing 'i' in the loop. We may erase the i-th element.
for (size_t i = 0; i < newAndAssignment->predChoices.size();) {
const OneIndexAssignment& oie = newAndAssignment->predChoices[i];
bool hasPredThatRequiresIndex = false;
for (size_t j = 0; j < oie.preds.size(); ++j) {
MatchExpression* expr = oie.preds[j];
if (MatchExpression::GEO_NEAR == expr->matchType()) {
hasPredThatRequiresIndex = true;
break;
}
if (MatchExpression::TEXT == expr->matchType()) {
hasPredThatRequiresIndex = true;
break;
}
}
if (hasPredThatRequiresIndex) {
newAndAssignment->mandatory.push_back(oie);
newAndAssignment->predChoices.erase(newAndAssignment->predChoices.begin() + i);
}
else {
++i;
}
}
newAndAssignment->resetEnumeration();
size_t myMemoID;
NodeAssignment* assign;
allocateAssignment(node, &assign, &myMemoID);
// Takes ownership.
assign->newAnd.reset(newAndAssignment);
return true;
}
// Don't know what the node is at this point.
return false;
}
