// static
Status QueryPlanner::planFromCache(const CanonicalQuery& query,
const QueryPlannerParams& params,
CachedSolution* cachedSoln,
QuerySolution** out) {
// Create a copy of the expression tree. We use cachedSoln to annotate this with indices.
MatchExpression* clone = query.root()->shallowClone();
// XXX: Use data in cachedSoln to tag 'clone' with the indices used. The tags use an index
// ID which is an index into some vector of IndexEntry(s). How do we maintain this across
// calls to plan? Do we want to store in the soln the keypatterns of the indices and just
// map those to an index into params.indices? Might be easiest thing to do, and certainly
// most intelligible for debugging.
// Use the cached index assignments to build solnRoot. Takes ownership of clone.
QuerySolutionNode* solnRoot =
QueryPlannerAccess::buildIndexedDataAccess(query, clone, false, params.indices);
// XXX: are the NULL cases an error/when does this happen / can this happen?
if (NULL != solnRoot) {
QuerySolution* soln = QueryPlannerAnalysis::analyzeDataAccess(query, params, solnRoot);
if (NULL != soln) {
QLOG() << "Planner: adding cached solution:\n" << soln->toString() << endl;
*out = soln;
}
}
// XXX: if any NULLs return error status?
return Status::OK();
}
void run() {
for (int i = 0; i < N; ++i) {
insert(BSON("a" << 1));
insert(BSON("a" << 1 << "b" << 1 << "c" << i));
}
// Indices on 'a' and 'b'.
addIndex(BSON("a" << 1));
addIndex(BSON("b" << 1));
// Solutions using either 'a' or 'b' will take a long time to start producing
// results. However, an index scan on 'b' will start producing results sooner
// than an index scan on 'a'.
CanonicalQuery* cq;
ASSERT(CanonicalQuery::canonicalize(ns,
fromjson("{a: 1, b: 1, c: {$gte: 5000}}"),
&cq).isOK());
ASSERT(NULL != cq);
// Use index on 'b'.
QuerySolution* soln = pickBestPlan(cq);
std::cerr << "PlanRankingWorkPlansLongEnough: soln=" << soln->toString() << std::endl;
ASSERT(QueryPlannerTestLib::solutionMatches(
"{fetch: {node: {ixscan: {pattern: {b: 1}}}}}",
soln->root.get()));
}
Status MultiPlanStage::pickBestPlan(PlanYieldPolicy* yieldPolicy) {
// Adds the amount of time taken by pickBestPlan() to executionTimeMillis. There's lots of
// execution work that happens here, so this is needed for the time accounting to
// make sense.
ScopedTimer timer(&_commonStats.executionTimeMillis);
// Run each plan some number of times. This number is at least as great as
// 'internalQueryPlanEvaluationWorks', but may be larger for big collections.
size_t numWorks = internalQueryPlanEvaluationWorks;
if (NULL != _collection) {
// For large collections, the number of works is set to be this
// fraction of the collection size.
double fraction = internalQueryPlanEvaluationCollFraction;
numWorks = std::max(size_t(internalQueryPlanEvaluationWorks),
size_t(fraction * _collection->numRecords(_txn)));
}
// We treat ntoreturn as though it is a limit during plan ranking.
// This means that ranking might not be great for sort + batchSize.
// But it also means that we don't buffer too much data for sort + limit.
// See SERVER-14174 for details.
size_t numToReturn = _query->getParsed().getNumToReturn();
// Determine the number of results which we will produce during the plan
// ranking phase before stopping.
size_t numResults = (size_t)internalQueryPlanEvaluationMaxResults;
if (numToReturn > 0) {
numResults = std::min(numToReturn, numResults);
}
// Work the plans, stopping when a plan hits EOF or returns some
// fixed number of results.
for (size_t ix = 0; ix < numWorks; ++ix) {
bool moreToDo = workAllPlans(numResults, yieldPolicy);
if (!moreToDo) { break; }
}
if (_failure) {
invariant(WorkingSet::INVALID_ID != _statusMemberId);
WorkingSetMember* member = _candidates[0].ws->get(_statusMemberId);
return WorkingSetCommon::getMemberStatus(*member);
}
// After picking best plan, ranking will own plan stats from
// candidate solutions (winner and losers).
std::auto_ptr<PlanRankingDecision> ranking(new PlanRankingDecision);
_bestPlanIdx = PlanRanker::pickBestPlan(_candidates, ranking.get());
verify(_bestPlanIdx >= 0 && _bestPlanIdx < static_cast<int>(_candidates.size()));
// Copy candidate order. We will need this to sort candidate stats for explain
// after transferring ownership of 'ranking' to plan cache.
std::vector<size_t> candidateOrder = ranking->candidateOrder;
CandidatePlan& bestCandidate = _candidates[_bestPlanIdx];
std::list<WorkingSetID>& alreadyProduced = bestCandidate.results;
QuerySolution* bestSolution = bestCandidate.solution;
LOG(5) << "Winning solution:\n" << bestSolution->toString() << endl;
LOG(2) << "Winning plan: " << Explain::getPlanSummary(bestCandidate.root);
_backupPlanIdx = kNoSuchPlan;
if (bestSolution->hasBlockingStage && (0 == alreadyProduced.size())) {
LOG(5) << "Winner has blocking stage, looking for backup plan...\n";
for (size_t ix = 0; ix < _candidates.size(); ++ix) {
if (!_candidates[ix].solution->hasBlockingStage) {
LOG(5) << "Candidate " << ix << " is backup child\n";
_backupPlanIdx = ix;
break;
}
}
}
// Logging for tied plans.
if (ranking->tieForBest && NULL != _collection) {
// These arrays having two or more entries is implied by 'tieForBest'.
invariant(ranking->scores.size() > 1);
invariant(ranking->candidateOrder.size() > 1);
size_t winnerIdx = ranking->candidateOrder[0];
size_t runnerUpIdx = ranking->candidateOrder[1];
LOG(1) << "Winning plan tied with runner-up."
<< " ns: " << _collection->ns()
<< " " << _query->toStringShort()
<< " winner score: " << ranking->scores[0]
<< " winner summary: "
<< Explain::getPlanSummary(_candidates[winnerIdx].root)
<< " runner-up score: " << ranking->scores[1]
<< " runner-up summary: "
<< Explain::getPlanSummary(_candidates[runnerUpIdx].root);
// There could be more than a 2-way tie, so log the stats for the remaining plans
// involved in the tie.
static const double epsilon = 1e-10;
for (size_t i = 2; i < ranking->scores.size(); i++) {
if (fabs(ranking->scores[i] - ranking->scores[0]) >= epsilon) {
break;
}
size_t planIdx = ranking->candidateOrder[i];
LOG(1) << "Plan " << i << " involved in multi-way tie."
<< " ns: " << _collection->ns()
<< " " << _query->toStringShort()
<< " score: " << ranking->scores[i]
<< " summary: "
<< Explain::getPlanSummary(_candidates[planIdx].root);
}
}
// If the winning plan produced no results during the ranking period (and, therefore, no
// plan produced results during the ranking period), then we will not create a plan cache
// entry.
if (alreadyProduced.empty() && NULL != _collection) {
size_t winnerIdx = ranking->candidateOrder[0];
LOG(1) << "Winning plan had zero results. Not caching."
<< " ns: " << _collection->ns()
<< " " << _query->toStringShort()
<< " winner score: " << ranking->scores[0]
<< " winner summary: "
<< Explain::getPlanSummary(_candidates[winnerIdx].root);
}
// Store the choice we just made in the cache. In order to do so,
// 1) the query must be of a type that is safe to cache,
// 2) two or more plans cannot have tied for the win. Caching in the case of ties can
// cause successive queries of the same shape to use a bad index.
// 3) Furthermore, the winning plan must have returned at least one result. Plans which
// return zero results cannot be reliably ranked. Such query shapes are generally
// existence type queries, and a winning plan should get cached once the query finds a
// result.
if (PlanCache::shouldCacheQuery(*_query)
&& !ranking->tieForBest
&& !alreadyProduced.empty()) {
// Create list of candidate solutions for the cache with
// the best solution at the front.
std::vector<QuerySolution*> solutions;
// Generate solutions and ranking decisions sorted by score.
for (size_t orderingIndex = 0;
orderingIndex < candidateOrder.size(); ++orderingIndex) {
// index into candidates/ranking
size_t ix = candidateOrder[orderingIndex];
solutions.push_back(_candidates[ix].solution);
}
// Check solution cache data. Do not add to cache if
// we have any invalid SolutionCacheData data.
// XXX: One known example is 2D queries
bool validSolutions = true;
for (size_t ix = 0; ix < solutions.size(); ++ix) {
if (NULL == solutions[ix]->cacheData.get()) {
LOG(5) << "Not caching query because this solution has no cache data: "
<< solutions[ix]->toString();
validSolutions = false;
break;
}
}
if (validSolutions) {
_collection->infoCache()->getPlanCache()->add(*_query, solutions, ranking.release());
}
}
return Status::OK();
}
// 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
Status QueryPlanner::planFromCache(const CanonicalQuery& query,
const QueryPlannerParams& params,
const CachedSolution& cachedSoln,
QuerySolution** out) {
invariant(!cachedSoln.plannerData.empty());
invariant(out);
// A query not suitable for caching should not have made its way into the cache.
invariant(PlanCache::shouldCacheQuery(query));
// Look up winning solution in cached solution's array.
const SolutionCacheData& winnerCacheData = *cachedSoln.plannerData[0];
if (SolutionCacheData::WHOLE_IXSCAN_SOLN == winnerCacheData.solnType) {
// The solution can be constructed by a scan over the entire index.
QuerySolution* soln = buildWholeIXSoln(
*winnerCacheData.tree->entry, query, params, winnerCacheData.wholeIXSolnDir);
if (soln == NULL) {
return Status(ErrorCodes::BadValue,
"plan cache error: soln that uses index to provide sort");
} else {
*out = soln;
return Status::OK();
}
} else if (SolutionCacheData::COLLSCAN_SOLN == winnerCacheData.solnType) {
// The cached solution is a collection scan. We don't cache collscans
// with tailable==true, hence the false below.
QuerySolution* soln = buildCollscanSoln(query, false, params);
if (soln == NULL) {
return Status(ErrorCodes::BadValue, "plan cache error: collection scan soln");
} else {
*out = soln;
return Status::OK();
}
}
// SolutionCacheData::USE_TAGS_SOLN == cacheData->solnType
// If we're here then this is neither the whole index scan or collection scan
// cases, and we proceed by using the PlanCacheIndexTree to tag the query tree.
// Create a copy of the expression tree. We use cachedSoln to annotate this with indices.
unique_ptr<MatchExpression> clone = std::move(query.root()->shallowClone());
LOG(5) << "Tagging the match expression according to cache data: " << endl
<< "Filter:" << endl
<< clone->toString() << "Cache data:" << endl
<< winnerCacheData.toString();
// Map from index name to index number.
// TODO: can we assume that the index numbering has the same lifetime
// as the cache state?
map<BSONObj, size_t> indexMap;
for (size_t i = 0; i < params.indices.size(); ++i) {
const IndexEntry& ie = params.indices[i];
indexMap[ie.keyPattern] = i;
LOG(5) << "Index " << i << ": " << ie.keyPattern.toString() << endl;
}
Status s = tagAccordingToCache(clone.get(), winnerCacheData.tree.get(), indexMap);
if (!s.isOK()) {
return s;
}
// The planner requires a defined sort order.
sortUsingTags(clone.get());
LOG(5) << "Tagged tree:" << endl
<< clone->toString();
// Use the cached index assignments to build solnRoot.
QuerySolutionNode* solnRoot = QueryPlannerAccess::buildIndexedDataAccess(
query, clone.release(), false, params.indices, params);
if (!solnRoot) {
return Status(ErrorCodes::BadValue,
str::stream() << "Failed to create data access plan from cache. Query: "
<< query.toStringShort());
}
// Takes ownership of 'solnRoot'.
QuerySolution* soln = QueryPlannerAnalysis::analyzeDataAccess(query, params, solnRoot);
if (!soln) {
return Status(ErrorCodes::BadValue,
str::stream()
<< "Failed to analyze plan from cache. Query: " << query.toStringShort());
}
LOG(5) << "Planner: solution constructed from the cache:\n" << soln->toString();
*out = soln;
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;
}
}
}
Status MultiPlanStage::pickBestPlan(PlanYieldPolicy* yieldPolicy) {
// Adds the amount of time taken by pickBestPlan() to executionTimeMillis. There's lots of
// execution work that happens here, so this is needed for the time accounting to
// make sense.
ScopedTimer timer(&_commonStats.executionTimeMillis);
size_t numWorks = getTrialPeriodWorks(_txn, _collection);
size_t numResults = getTrialPeriodNumToReturn(*_query);
// Work the plans, stopping when a plan hits EOF or returns some
// fixed number of results.
for (size_t ix = 0; ix < numWorks; ++ix) {
bool moreToDo = workAllPlans(numResults, yieldPolicy);
if (!moreToDo) { break; }
}
if (_failure) {
invariant(WorkingSet::INVALID_ID != _statusMemberId);
WorkingSetMember* member = _candidates[0].ws->get(_statusMemberId);
return WorkingSetCommon::getMemberStatus(*member);
}
// After picking best plan, ranking will own plan stats from
// candidate solutions (winner and losers).
std::auto_ptr<PlanRankingDecision> ranking(new PlanRankingDecision);
_bestPlanIdx = PlanRanker::pickBestPlan(_candidates, ranking.get());
verify(_bestPlanIdx >= 0 && _bestPlanIdx < static_cast<int>(_candidates.size()));
// Copy candidate order. We will need this to sort candidate stats for explain
// after transferring ownership of 'ranking' to plan cache.
std::vector<size_t> candidateOrder = ranking->candidateOrder;
CandidatePlan& bestCandidate = _candidates[_bestPlanIdx];
std::list<WorkingSetID>& alreadyProduced = bestCandidate.results;
QuerySolution* bestSolution = bestCandidate.solution;
LOG(5) << "Winning solution:\n" << bestSolution->toString() << endl;
LOG(2) << "Winning plan: " << Explain::getPlanSummary(bestCandidate.root);
_backupPlanIdx = kNoSuchPlan;
if (bestSolution->hasBlockingStage && (0 == alreadyProduced.size())) {
LOG(5) << "Winner has blocking stage, looking for backup plan...\n";
for (size_t ix = 0; ix < _candidates.size(); ++ix) {
if (!_candidates[ix].solution->hasBlockingStage) {
LOG(5) << "Candidate " << ix << " is backup child\n";
_backupPlanIdx = ix;
break;
}
}
}
// Store the choice we just made in the cache, if the query is of a type that is safe to
// cache.
if (PlanCache::shouldCacheQuery(*_query) && _shouldCache) {
// Create list of candidate solutions for the cache with
// the best solution at the front.
std::vector<QuerySolution*> solutions;
// Generate solutions and ranking decisions sorted by score.
for (size_t orderingIndex = 0;
orderingIndex < candidateOrder.size(); ++orderingIndex) {
// index into candidates/ranking
size_t ix = candidateOrder[orderingIndex];
solutions.push_back(_candidates[ix].solution);
}
// Check solution cache data. Do not add to cache if
// we have any invalid SolutionCacheData data.
// XXX: One known example is 2D queries
bool validSolutions = true;
for (size_t ix = 0; ix < solutions.size(); ++ix) {
if (NULL == solutions[ix]->cacheData.get()) {
LOG(5) << "Not caching query because this solution has no cache data: "
<< solutions[ix]->toString();
validSolutions = false;
break;
}
}
if (validSolutions) {
_collection->infoCache()->getPlanCache()->add(*_query, solutions, ranking.release());
}
}
return Status::OK();
}
void MultiPlanStage::pickBestPlan() {
// Run each plan some number of times. This number is at least as great as
// 'internalQueryPlanEvaluationWorks', but may be larger for big collections.
size_t numWorks = internalQueryPlanEvaluationWorks;
if (NULL != _collection) {
// For large collections, the number of works is set to be this
// fraction of the collection size.
double fraction = internalQueryPlanEvaluationCollFraction;
numWorks = std::max(size_t(internalQueryPlanEvaluationWorks),
size_t(fraction * _collection->numRecords()));
}
// We treat ntoreturn as though it is a limit during plan ranking.
// This means that ranking might not be great for sort + batchSize.
// But it also means that we don't buffer too much data for sort + limit.
// See SERVER-14174 for details.
size_t numToReturn = _query->getParsed().getNumToReturn();
// Determine the number of results which we will produce during the plan
// ranking phase before stopping.
size_t numResults = (size_t)internalQueryPlanEvaluationMaxResults;
if (numToReturn > 0) {
numResults = std::min(numToReturn, numResults);
}
// Work the plans, stopping when a plan hits EOF or returns some
// fixed number of results.
for (size_t ix = 0; ix < numWorks; ++ix) {
bool moreToDo = workAllPlans(numResults);
if (!moreToDo) { break; }
}
if (_failure) { return; }
// After picking best plan, ranking will own plan stats from
// candidate solutions (winner and losers).
std::auto_ptr<PlanRankingDecision> ranking(new PlanRankingDecision);
_bestPlanIdx = PlanRanker::pickBestPlan(_candidates, ranking.get());
verify(_bestPlanIdx >= 0 && _bestPlanIdx < static_cast<int>(_candidates.size()));
// Copy candidate order. We will need this to sort candidate stats for explain
// after transferring ownership of 'ranking' to plan cache.
std::vector<size_t> candidateOrder = ranking->candidateOrder;
CandidatePlan& bestCandidate = _candidates[_bestPlanIdx];
std::list<WorkingSetID>& alreadyProduced = bestCandidate.results;
QuerySolution* bestSolution = bestCandidate.solution;
QLOG() << "Winning solution:\n" << bestSolution->toString() << endl;
LOG(2) << "Winning plan: " << getPlanSummary(*bestSolution);
_backupPlanIdx = kNoSuchPlan;
if (bestSolution->hasBlockingStage && (0 == alreadyProduced.size())) {
QLOG() << "Winner has blocking stage, looking for backup plan...\n";
for (size_t ix = 0; ix < _candidates.size(); ++ix) {
if (!_candidates[ix].solution->hasBlockingStage) {
QLOG() << "Candidate " << ix << " is backup child\n";
_backupPlanIdx = ix;
break;
}
}
}
// Store the choice we just made in the cache. In order to do so,
// 1) the query must be of a type that is safe to cache, and
// 2) two or more plans cannot have tied for the win. Caching in the
// case of ties can cause successive queries of the same shape to
// use a bad index.
if (PlanCache::shouldCacheQuery(*_query) && !ranking->tieForBest) {
// Create list of candidate solutions for the cache with
// the best solution at the front.
std::vector<QuerySolution*> solutions;
// Generate solutions and ranking decisions sorted by score.
for (size_t orderingIndex = 0;
orderingIndex < candidateOrder.size(); ++orderingIndex) {
// index into candidates/ranking
size_t ix = candidateOrder[orderingIndex];
solutions.push_back(_candidates[ix].solution);
}
// Check solution cache data. Do not add to cache if
// we have any invalid SolutionCacheData data.
// XXX: One known example is 2D queries
bool validSolutions = true;
for (size_t ix = 0; ix < solutions.size(); ++ix) {
if (NULL == solutions[ix]->cacheData.get()) {
QLOG() << "Not caching query because this solution has no cache data: "
<< solutions[ix]->toString();
validSolutions = false;
break;
}
}
if (validSolutions) {
_collection->infoCache()->getPlanCache()->add(*_query, solutions, ranking.release());
}
}
}
bool SubplanRunner::runSubplans() {
// This is what we annotate with the index selections and then turn into a solution.
auto_ptr<OrMatchExpression> theOr(
static_cast<OrMatchExpression*>(_query->root()->shallowClone()));
// This is the skeleton of index selections that is inserted into the cache.
auto_ptr<PlanCacheIndexTree> cacheData(new PlanCacheIndexTree());
for (size_t i = 0; i < theOr->numChildren(); ++i) {
MatchExpression* orChild = theOr->getChild(i);
auto_ptr<CanonicalQuery> orChildCQ(_cqs.front());
_cqs.pop();
// 'solutions' is owned by the SubplanRunner instance until
// it is popped from the queue.
vector<QuerySolution*> solutions = _solutions.front();
_solutions.pop();
// We already checked for zero solutions in planSubqueries(...).
invariant(!solutions.empty());
if (1 == solutions.size()) {
// There is only one solution. Transfer ownership to an auto_ptr.
auto_ptr<QuerySolution> autoSoln(solutions[0]);
// We want a well-formed *indexed* solution.
if (NULL == autoSoln->cacheData.get()) {
// For example, we don't cache things for 2d indices.
QLOG() << "Subplanner: No cache data for subchild " << orChild->toString();
return false;
}
if (SolutionCacheData::USE_INDEX_TAGS_SOLN != autoSoln->cacheData->solnType) {
QLOG() << "Subplanner: No indexed cache data for subchild "
<< orChild->toString();
return false;
}
// Add the index assignments to our original query.
Status tagStatus = QueryPlanner::tagAccordingToCache(
orChild, autoSoln->cacheData->tree.get(), _indexMap);
if (!tagStatus.isOK()) {
QLOG() << "Subplanner: Failed to extract indices from subchild "
<< orChild->toString();
return false;
}
// Add the child's cache data to the cache data we're creating for the main query.
cacheData->children.push_back(autoSoln->cacheData->tree->clone());
}
else {
// N solutions, rank them. Takes ownership of orChildCQ.
// the working set will be shared by the candidate plans and owned by the runner
WorkingSet* sharedWorkingSet = new WorkingSet();
MultiPlanStage* multiPlanStage = new MultiPlanStage(_collection,
orChildCQ.get());
// Dump all the solutions into the MPR.
for (size_t ix = 0; ix < solutions.size(); ++ix) {
PlanStage* nextPlanRoot;
verify(StageBuilder::build(_txn,
_collection,
*solutions[ix],
sharedWorkingSet,
&nextPlanRoot));
// Owns first two arguments
multiPlanStage->addPlan(solutions[ix], nextPlanRoot, sharedWorkingSet);
}
multiPlanStage->pickBestPlan();
if (! multiPlanStage->bestPlanChosen()) {
QLOG() << "Subplanner: Failed to pick best plan for subchild "
<< orChildCQ->toString();
return false;
}
Runner* mpr = new SingleSolutionRunner(_collection,
orChildCQ.release(),
multiPlanStage->bestSolution(),
multiPlanStage,
sharedWorkingSet);
_underlyingRunner.reset(mpr);
if (_killed) {
QLOG() << "Subplanner: Killed while picking best plan for subchild "
<< orChild->toString();
return false;
}
QuerySolution* bestSoln = multiPlanStage->bestSolution();
if (SolutionCacheData::USE_INDEX_TAGS_SOLN != bestSoln->cacheData->solnType) {
QLOG() << "Subplanner: No indexed cache data for subchild "
<< orChild->toString();
return false;
}
// Add the index assignments to our original query.
Status tagStatus = QueryPlanner::tagAccordingToCache(
orChild, bestSoln->cacheData->tree.get(), _indexMap);
if (!tagStatus.isOK()) {
QLOG() << "Subplanner: Failed to extract indices from subchild "
<< orChild->toString();
return false;
}
cacheData->children.push_back(bestSoln->cacheData->tree->clone());
}
}
// Must do this before using the planner functionality.
sortUsingTags(theOr.get());
// Use the cached index assignments to build solnRoot. Takes ownership of 'theOr'
QuerySolutionNode* solnRoot = QueryPlannerAccess::buildIndexedDataAccess(
*_query, theOr.release(), false, _plannerParams.indices);
if (NULL == solnRoot) {
QLOG() << "Subplanner: Failed to build indexed data path for subplanned query\n";
return false;
}
QLOG() << "Subplanner: fully tagged tree is " << solnRoot->toString();
// Takes ownership of 'solnRoot'
QuerySolution* soln = QueryPlannerAnalysis::analyzeDataAccess(*_query,
_plannerParams,
solnRoot);
if (NULL == soln) {
QLOG() << "Subplanner: Failed to analyze subplanned query";
return false;
}
// We want our franken-solution to be cached.
SolutionCacheData* scd = new SolutionCacheData();
scd->tree.reset(cacheData.release());
soln->cacheData.reset(scd);
QLOG() << "Subplanner: Composite solution is " << soln->toString() << endl;
// We use one of these even if there is one plan. We do this so that the entry is cached
// with stats obtained in the same fashion as a competitive ranking would have obtained
// them.
MultiPlanStage* multiPlanStage = new MultiPlanStage(_collection, _query.get());
WorkingSet* ws = new WorkingSet();
PlanStage* root;
verify(StageBuilder::build(_txn, _collection, *soln, ws, &root));
multiPlanStage->addPlan(soln, root, ws); // Takes ownership first two arguments.
multiPlanStage->pickBestPlan();
if (! multiPlanStage->bestPlanChosen()) {
QLOG() << "Subplanner: Failed to pick best plan for subchild "
<< _query->toString();
return false;
}
Runner* mpr = new SingleSolutionRunner(_collection,
_query.release(),
multiPlanStage->bestSolution(),
multiPlanStage,
ws);
_underlyingRunner.reset(mpr);
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();
}
