// static
bool IDHackRunner::canUseProjection(const CanonicalQuery& query) {
const ParsedProjection* proj = query.getProj();
// No projection is OK - ID Hack will fetch entire document.
if (!proj) {
return true;
}
// If there is a projection, it has to be a covered projection on
// the _id field only.
if (proj->requiresDocument()) {
return false;
}
const std::vector<std::string>& requiredFields = proj->getRequiredFields();
if (1U != requiredFields.size()) {
return false;
}
if ("_id" != requiredFields[0]) {
return false;
}
// Can use this projection with ID Hack.
return true;
}
// 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();
}
std::unique_ptr<QuerySolution> QueryPlannerAnalysis::analyzeDataAccess(
const CanonicalQuery& query,
const QueryPlannerParams& params,
std::unique_ptr<QuerySolutionNode> solnRoot) {
auto soln = std::make_unique<QuerySolution>();
soln->filterData = query.getQueryObj();
soln->indexFilterApplied = params.indexFiltersApplied;
solnRoot->computeProperties();
analyzeGeo(params, solnRoot.get());
// solnRoot finds all our results. Let's see what transformations we must perform to the
// data.
// If we're answering a query on a sharded system, we need to drop documents that aren't
// logically part of our shard.
if (params.options & QueryPlannerParams::INCLUDE_SHARD_FILTER) {
if (!solnRoot->fetched()) {
// See if we need to fetch information for our shard key.
// NOTE: Solution nodes only list ordinary, non-transformed index keys for now
bool fetch = false;
BSONObjIterator it(params.shardKey);
while (it.more()) {
BSONElement nextEl = it.next();
if (!solnRoot->hasField(nextEl.fieldName())) {
fetch = true;
break;
}
}
if (fetch) {
FetchNode* fetchNode = new FetchNode();
fetchNode->children.push_back(solnRoot.release());
solnRoot.reset(fetchNode);
}
}
ShardingFilterNode* sfn = new ShardingFilterNode();
sfn->children.push_back(solnRoot.release());
solnRoot.reset(sfn);
}
bool hasSortStage = false;
solnRoot.reset(analyzeSort(query, params, solnRoot.release(), &hasSortStage));
// This can happen if we need to create a blocking sort stage and we're not allowed to.
if (!solnRoot) {
return nullptr;
}
// A solution can be blocking if it has a blocking sort stage or
// a hashed AND stage.
bool hasAndHashStage = hasNode(solnRoot.get(), STAGE_AND_HASH);
soln->hasBlockingStage = hasSortStage || hasAndHashStage;
const QueryRequest& qr = query.getQueryRequest();
if (qr.getSkip()) {
auto skip = std::make_unique<SkipNode>();
skip->skip = *qr.getSkip();
skip->children.push_back(solnRoot.release());
solnRoot = std::move(skip);
}
// Project the results.
if (query.getProj()) {
solnRoot = analyzeProjection(query, std::move(solnRoot), hasSortStage);
// If we don't have a covered project, and we're not allowed to put an uncovered one in,
// bail out.
if (solnRoot->fetched() && params.options & QueryPlannerParams::NO_UNCOVERED_PROJECTIONS)
return nullptr;
} else {
// If there's no projection, we must fetch, as the user wants the entire doc.
if (!solnRoot->fetched() && !(params.options & QueryPlannerParams::IS_COUNT)) {
FetchNode* fetch = new FetchNode();
fetch->children.push_back(solnRoot.release());
solnRoot.reset(fetch);
}
}
// When there is both a blocking sort and a limit, the limit will
// be enforced by the blocking sort.
// Otherwise, we need to limit the results in the case of a hard limit
// (ie. limit in raw query is negative)
if (!hasSortStage) {
// We don't have a sort stage. This means that, if there is a limit, we will have
// to enforce it ourselves since it's not handled inside SORT.
if (qr.getLimit()) {
LimitNode* limit = new LimitNode();
limit->limit = *qr.getLimit();
limit->children.push_back(solnRoot.release());
solnRoot.reset(limit);
} else if (qr.getNToReturn() && !qr.wantMore()) {
// We have a "legacy limit", i.e. a negative ntoreturn value from an OP_QUERY style
// find.
LimitNode* limit = new LimitNode();
limit->limit = *qr.getNToReturn();
limit->children.push_back(solnRoot.release());
solnRoot.reset(limit);
}
}
soln->root = std::move(solnRoot);
return soln;
}
