Status SubplanRunner::planSubqueries() {
MatchExpression* theOr = _query->root();
for (size_t i = 0; i < _plannerParams.indices.size(); ++i) {
const IndexEntry& ie = _plannerParams.indices[i];
_indexMap[ie.keyPattern] = i;
QLOG() << "Subplanner: index " << i << " is " << ie.toString() << endl;
}
const WhereCallbackReal whereCallback(_collection->ns().db());
for (size_t i = 0; i < theOr->numChildren(); ++i) {
// Turn the i-th child into its own query.
MatchExpression* orChild = theOr->getChild(i);
CanonicalQuery* orChildCQ;
Status childCQStatus = CanonicalQuery::canonicalize(*_query,
orChild,
&orChildCQ,
whereCallback);
if (!childCQStatus.isOK()) {
mongoutils::str::stream ss;
ss << "Subplanner: Can't canonicalize subchild " << orChild->toString()
<< " " << childCQStatus.reason();
return Status(ErrorCodes::BadValue, ss);
}
// Make sure it gets cleaned up.
auto_ptr<CanonicalQuery> safeOrChildCQ(orChildCQ);
// Plan the i-th child.
vector<QuerySolution*> solutions;
// We don't set NO_TABLE_SCAN because peeking at the cache data will keep us from
// considering any plan that's a collscan.
QLOG() << "Subplanner: planning child " << i << " of " << theOr->numChildren();
Status status = QueryPlanner::plan(*safeOrChildCQ, _plannerParams, &solutions);
if (!status.isOK()) {
mongoutils::str::stream ss;
ss << "Subplanner: Can't plan for subchild " << orChildCQ->toString()
<< " " << status.reason();
return Status(ErrorCodes::BadValue, ss);
}
QLOG() << "Subplanner: got " << solutions.size() << " solutions";
if (0 == solutions.size()) {
// If one child doesn't have an indexed solution, bail out.
mongoutils::str::stream ss;
ss << "Subplanner: No solutions for subchild " << orChildCQ->toString();
return Status(ErrorCodes::BadValue, ss);
}
// Hang onto the canonicalized subqueries and the corresponding query solutions
// so that they can be used in subplan running later on.
_cqs.push(safeOrChildCQ.release());
_solutions.push(solutions);
}
return Status::OK();
}
std::string newRunQuery(OperationContext* txn,
Message& m,
QueryMessage& q,
CurOp& curop,
Message &result,
bool fromDBDirectClient) {
// Validate the namespace.
const char *ns = q.ns;
uassert(16332, "can't have an empty ns", ns[0]);
const NamespaceString nsString(ns);
uassert(16256, str::stream() << "Invalid ns [" << ns << "]", nsString.isValid());
// Set curop information.
curop.debug().ns = ns;
curop.debug().ntoreturn = q.ntoreturn;
curop.debug().query = q.query;
curop.setQuery(q.query);
// If the query is really a command, run it.
if (nsString.isCommand()) {
int nToReturn = q.ntoreturn;
uassert(16979, str::stream() << "bad numberToReturn (" << nToReturn
<< ") for $cmd type ns - can only be 1 or -1",
nToReturn == 1 || nToReturn == -1);
curop.markCommand();
BufBuilder bb;
bb.skip(sizeof(QueryResult::Value));
BSONObjBuilder cmdResBuf;
if (!runCommands(txn, ns, q.query, curop, bb, cmdResBuf, false, q.queryOptions)) {
uasserted(13530, "bad or malformed command request?");
}
curop.debug().iscommand = true;
// TODO: Does this get overwritten/do we really need to set this twice?
curop.debug().query = q.query;
QueryResult::View qr = bb.buf();
bb.decouple();
qr.setResultFlagsToOk();
qr.msgdata().setLen(bb.len());
curop.debug().responseLength = bb.len();
qr.msgdata().setOperation(opReply);
qr.setCursorId(0);
qr.setStartingFrom(0);
qr.setNReturned(1);
result.setData(qr.view2ptr(), true);
return "";
}
const NamespaceString nss(q.ns);
// Parse the qm into a CanonicalQuery.
CanonicalQuery* cq;
Status canonStatus = CanonicalQuery::canonicalize(
q, &cq, WhereCallbackReal(txn, StringData(nss.db())));
if (!canonStatus.isOK()) {
uasserted(17287, str::stream() << "Can't canonicalize query: " << canonStatus.toString());
}
QLOG() << "Running query:\n" << cq->toString();
LOG(2) << "Running query: " << cq->toStringShort();
// Parse, canonicalize, plan, transcribe, and get a plan executor.
PlanExecutor* rawExec = NULL;
// We use this a lot below.
const LiteParsedQuery& pq = cq->getParsed();
AutoGetCollectionForRead ctx(txn, nss);
const int dbProfilingLevel = (ctx.getDb() != NULL) ? ctx.getDb()->getProfilingLevel() :
serverGlobalParams.defaultProfile;
Collection* collection = ctx.getCollection();
// We'll now try to get the query executor that will execute this query for us. There
// are a few cases in which we know upfront which executor we should get and, therefore,
// we shortcut the selection process here.
//
// (a) If the query is over a collection that doesn't exist, we use an EOFStage.
//
// (b) if the query is a replication's initial sync one, we use a specifically designed
// stage that skips extents faster (see details in exec/oplogstart.h).
//
// Otherwise we go through the selection of which executor is most suited to the
// query + run-time context at hand.
Status status = Status::OK();
if (NULL != collection && pq.getOptions().oplogReplay) {
// Takes ownership of 'cq'.
status = getOplogStartHack(txn, collection, cq, &rawExec);
}
else {
size_t options = QueryPlannerParams::DEFAULT;
if (shardingState.needCollectionMetadata(pq.ns())) {
options |= QueryPlannerParams::INCLUDE_SHARD_FILTER;
}
// Takes ownership of 'cq'.
status = getExecutor(txn, collection, cq, PlanExecutor::YIELD_AUTO, &rawExec, options);
}
if (!status.isOK()) {
// NOTE: Do not access cq as getExecutor has deleted it.
uasserted(17007, "Unable to execute query: " + status.reason());
}
verify(NULL != rawExec);
auto_ptr<PlanExecutor> exec(rawExec);
// If it's actually an explain, do the explain and return rather than falling through
// to the normal query execution loop.
if (pq.isExplain()) {
BufBuilder bb;
bb.skip(sizeof(QueryResult::Value));
BSONObjBuilder explainBob;
Explain::explainStages(exec.get(), ExplainCommon::EXEC_ALL_PLANS, &explainBob);
// Add the resulting object to the return buffer.
BSONObj explainObj = explainBob.obj();
bb.appendBuf((void*)explainObj.objdata(), explainObj.objsize());
curop.debug().iscommand = true;
// TODO: Does this get overwritten/do we really need to set this twice?
curop.debug().query = q.query;
// Set query result fields.
QueryResult::View qr = bb.buf();
bb.decouple();
qr.setResultFlagsToOk();
qr.msgdata().setLen(bb.len());
curop.debug().responseLength = bb.len();
qr.msgdata().setOperation(opReply);
qr.setCursorId(0);
qr.setStartingFrom(0);
qr.setNReturned(1);
result.setData(qr.view2ptr(), true);
return "";
}
// We freak out later if this changes before we're done with the query.
const ChunkVersion shardingVersionAtStart = shardingState.getVersion(cq->ns());
// Handle query option $maxTimeMS (not used with commands).
curop.setMaxTimeMicros(static_cast<unsigned long long>(pq.getMaxTimeMS()) * 1000);
txn->checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
// uassert if we are not on a primary, and not a secondary with SlaveOk query parameter set.
bool slaveOK = pq.getOptions().slaveOk || pq.hasReadPref();
status = repl::getGlobalReplicationCoordinator()->checkCanServeReadsFor(
txn,
NamespaceString(cq->ns()),
slaveOK);
uassertStatusOK(status);
// If this exists, the collection is sharded.
// If it doesn't exist, we can assume we're not sharded.
// If we're sharded, we might encounter data that is not consistent with our sharding state.
// We must ignore this data.
CollectionMetadataPtr collMetadata;
if (!shardingState.needCollectionMetadata(pq.ns())) {
collMetadata = CollectionMetadataPtr();
}
else {
collMetadata = shardingState.getCollectionMetadata(pq.ns());
}
// Run the query.
// bb is used to hold query results
// this buffer should contain either requested documents per query or
// explain information, but not both
BufBuilder bb(32768);
bb.skip(sizeof(QueryResult::Value));
// How many results have we obtained from the executor?
int numResults = 0;
// If we're replaying the oplog, we save the last time that we read.
OpTime slaveReadTill;
// Do we save the PlanExecutor in a ClientCursor for getMore calls later?
bool saveClientCursor = false;
BSONObj obj;
PlanExecutor::ExecState state;
// uint64_t numMisplacedDocs = 0;
// Get summary info about which plan the executor is using.
curop.debug().planSummary = Explain::getPlanSummary(exec.get());
while (PlanExecutor::ADVANCED == (state = exec->getNext(&obj, NULL))) {
// Add result to output buffer.
bb.appendBuf((void*)obj.objdata(), obj.objsize());
// Count the result.
++numResults;
// Possibly note slave's position in the oplog.
if (pq.getOptions().oplogReplay) {
BSONElement e = obj["ts"];
if (Date == e.type() || Timestamp == e.type()) {
slaveReadTill = e._opTime();
}
}
// TODO: only one type of 2d search doesn't support this. We need a way to pull it out
// of CanonicalQuery. :(
const bool supportsGetMore = true;
if (!supportsGetMore && (enough(pq, numResults)
|| bb.len() >= MaxBytesToReturnToClientAtOnce)) {
break;
}
else if (enoughForFirstBatch(pq, numResults, bb.len())) {
QLOG() << "Enough for first batch, wantMore=" << pq.wantMore()
<< " numToReturn=" << pq.getNumToReturn()
<< " numResults=" << numResults
<< endl;
// If only one result requested assume it's a findOne() and don't save the cursor.
if (pq.wantMore() && 1 != pq.getNumToReturn()) {
QLOG() << " executor EOF=" << exec->isEOF() << endl;
saveClientCursor = !exec->isEOF();
}
break;
}
}
// If we cache the executor later, we want to deregister it as it receives notifications
// anyway by virtue of being cached.
//
// If we don't cache the executor later, we are deleting it, so it must be deregistered.
//
// So, no matter what, deregister the executor.
exec->deregisterExec();
// Caller expects exceptions thrown in certain cases.
if (PlanExecutor::EXEC_ERROR == state) {
scoped_ptr<PlanStageStats> stats(exec->getStats());
error() << "Plan executor error, stats: "
<< Explain::statsToBSON(*stats);
uasserted(17144, "Executor error: " + WorkingSetCommon::toStatusString(obj));
}
// Why save a dead executor?
if (PlanExecutor::DEAD == state) {
saveClientCursor = false;
}
else if (pq.getOptions().tailable) {
// If we're tailing a capped collection, we don't bother saving the cursor if the
// collection is empty. Otherwise, the semantics of the tailable cursor is that the
// client will keep trying to read from it. So we'll keep it around.
if (collection && collection->numRecords(txn) != 0 && pq.getNumToReturn() != 1) {
saveClientCursor = true;
}
}
// TODO(greg): This will go away soon.
if (!shardingState.getVersion(pq.ns()).isWriteCompatibleWith(shardingVersionAtStart)) {
// if the version changed during the query we might be missing some data and its safe to
// send this as mongos can resend at this point
throw SendStaleConfigException(pq.ns(), "version changed during initial query",
shardingVersionAtStart,
shardingState.getVersion(pq.ns()));
}
const logger::LogComponent queryLogComponent = logger::LogComponent::kQuery;
const logger::LogSeverity logLevelOne = logger::LogSeverity::Debug(1);
PlanSummaryStats summaryStats;
Explain::getSummaryStats(exec.get(), &summaryStats);
curop.debug().ntoskip = pq.getSkip();
curop.debug().nreturned = numResults;
curop.debug().scanAndOrder = summaryStats.hasSortStage;
curop.debug().nscanned = summaryStats.totalKeysExamined;
curop.debug().nscannedObjects = summaryStats.totalDocsExamined;
curop.debug().idhack = summaryStats.isIdhack;
// Set debug information for consumption by the profiler.
if (dbProfilingLevel > 0 ||
curop.elapsedMillis() > serverGlobalParams.slowMS ||
logger::globalLogDomain()->shouldLog(queryLogComponent, logLevelOne)) {
// Get BSON stats.
scoped_ptr<PlanStageStats> execStats(exec->getStats());
BSONObjBuilder statsBob;
Explain::statsToBSON(*execStats, &statsBob);
curop.debug().execStats.set(statsBob.obj());
// Replace exec stats with plan summary if stats cannot fit into CachedBSONObj.
if (curop.debug().execStats.tooBig() && !curop.debug().planSummary.empty()) {
BSONObjBuilder bob;
bob.append("summary", curop.debug().planSummary.toString());
curop.debug().execStats.set(bob.done());
}
}
long long ccId = 0;
if (saveClientCursor) {
// We won't use the executor until it's getMore'd.
exec->saveState();
// Allocate a new ClientCursor. We don't have to worry about leaking it as it's
// inserted into a global map by its ctor.
ClientCursor* cc = new ClientCursor(collection, exec.get(),
cq->getParsed().getOptions().toInt(),
cq->getParsed().getFilter());
ccId = cc->cursorid();
if (fromDBDirectClient) {
cc->setUnownedRecoveryUnit(txn->recoveryUnit());
}
else if (state == PlanExecutor::IS_EOF && pq.getOptions().tailable) {
// Don't stash the RU for tailable cursors at EOF, let them get a new RU on their
// next getMore.
}
else {
// We stash away the RecoveryUnit in the ClientCursor. It's used for subsequent
// getMore requests. The calling OpCtx gets a fresh RecoveryUnit.
cc->setOwnedRecoveryUnit(txn->releaseRecoveryUnit());
StorageEngine* storageEngine = getGlobalEnvironment()->getGlobalStorageEngine();
txn->setRecoveryUnit(storageEngine->newRecoveryUnit(txn));
}
QLOG() << "caching executor with cursorid " << ccId
<< " after returning " << numResults << " results" << endl;
// ClientCursor takes ownership of executor. Release to make sure it's not deleted.
exec.release();
// TODO document
if (pq.getOptions().oplogReplay && !slaveReadTill.isNull()) {
cc->slaveReadTill(slaveReadTill);
}
// TODO document
if (pq.getOptions().exhaust) {
curop.debug().exhaust = true;
}
// Set attributes for getMore.
cc->setCollMetadata(collMetadata);
cc->setPos(numResults);
// If the query had a time limit, remaining time is "rolled over" to the cursor (for
// use by future getmore ops).
cc->setLeftoverMaxTimeMicros(curop.getRemainingMaxTimeMicros());
}
else {
QLOG() << "Not caching executor but returning " << numResults << " results.\n";
}
// Add the results from the query into the output buffer.
result.appendData(bb.buf(), bb.len());
bb.decouple();
// Fill out the output buffer's header.
QueryResult::View qr = result.header().view2ptr();
qr.setCursorId(ccId);
curop.debug().cursorid = (0 == ccId ? -1 : ccId);
qr.setResultFlagsToOk();
qr.msgdata().setOperation(opReply);
qr.setStartingFrom(0);
qr.setNReturned(numResults);
// curop.debug().exhaust is set above.
return curop.debug().exhaust ? pq.ns() : "";
}
// 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;
}
}
}
std::string newRunQuery(Message& m, QueryMessage& q, CurOp& curop, Message &result) {
// Validate the namespace.
const char *ns = q.ns;
uassert(16332, "can't have an empty ns", ns[0]);
const NamespaceString nsString(ns);
uassert(16256, str::stream() << "Invalid ns [" << ns << "]", nsString.isValid());
// Set curop information.
curop.debug().ns = ns;
curop.debug().ntoreturn = q.ntoreturn;
curop.debug().query = q.query;
curop.setQuery(q.query);
// If the query is really a command, run it.
if (nsString.isCommand()) {
int nToReturn = q.ntoreturn;
uassert(16979, str::stream() << "bad numberToReturn (" << nToReturn
<< ") for $cmd type ns - can only be 1 or -1",
nToReturn == 1 || nToReturn == -1);
curop.markCommand();
BufBuilder bb;
bb.skip(sizeof(QueryResult));
BSONObjBuilder cmdResBuf;
if (!runCommands(ns, q.query, curop, bb, cmdResBuf, false, q.queryOptions)) {
uasserted(13530, "bad or malformed command request?");
}
curop.debug().iscommand = true;
// TODO: Does this get overwritten/do we really need to set this twice?
curop.debug().query = q.query;
QueryResult* qr = reinterpret_cast<QueryResult*>(bb.buf());
bb.decouple();
qr->setResultFlagsToOk();
qr->len = bb.len();
curop.debug().responseLength = bb.len();
qr->setOperation(opReply);
qr->cursorId = 0;
qr->startingFrom = 0;
qr->nReturned = 1;
result.setData(qr, true);
return "";
}
// This is a read lock. We require this because if we're parsing a $where, the
// where-specific parsing code assumes we have a lock and creates execution machinery that
// requires it.
Client::ReadContext ctx(q.ns);
Collection* collection = ctx.ctx().db()->getCollection( ns );
// Parse the qm into a CanonicalQuery.
CanonicalQuery* cq;
Status canonStatus = CanonicalQuery::canonicalize(q, &cq);
if (!canonStatus.isOK()) {
uasserted(17287, str::stream() << "Can't canonicalize query: " << canonStatus.toString());
}
verify(cq);
QLOG() << "Running query:\n" << cq->toString();
LOG(2) << "Running query: " << cq->toStringShort();
// Parse, canonicalize, plan, transcribe, and get a runner.
Runner* rawRunner = NULL;
// We use this a lot below.
const LiteParsedQuery& pq = cq->getParsed();
// We'll now try to get the query runner that will execute this query for us. There
// are a few cases in which we know upfront which runner we should get and, therefore,
// we shortcut the selection process here.
//
// (a) If the query is over a collection that doesn't exist, we get a special runner
// that's is so (a runner) which doesn't return results, the EOFRunner.
//
// (b) if the query is a replication's initial sync one, we get a SingleSolutinRunner
// that uses a specifically designed stage that skips extents faster (see details in
// exec/oplogstart.h)
//
// Otherwise we go through the selection of which runner is most suited to the
// query + run-time context at hand.
Status status = Status::OK();
if (collection == NULL) {
rawRunner = new EOFRunner(cq, cq->ns());
}
else if (pq.hasOption(QueryOption_OplogReplay)) {
status = getOplogStartHack(collection, cq, &rawRunner);
}
else {
// Takes ownership of cq.
size_t options = QueryPlannerParams::DEFAULT;
if (shardingState.needCollectionMetadata(pq.ns())) {
options |= QueryPlannerParams::INCLUDE_SHARD_FILTER;
}
status = getRunner(cq, &rawRunner, options);
}
if (!status.isOK()) {
// NOTE: Do not access cq as getRunner has deleted it.
uasserted(17007, "Unable to execute query: " + status.reason());
}
verify(NULL != rawRunner);
auto_ptr<Runner> runner(rawRunner);
// We freak out later if this changes before we're done with the query.
const ChunkVersion shardingVersionAtStart = shardingState.getVersion(cq->ns());
// Handle query option $maxTimeMS (not used with commands).
curop.setMaxTimeMicros(static_cast<unsigned long long>(pq.getMaxTimeMS()) * 1000);
killCurrentOp.checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
// uassert if we are not on a primary, and not a secondary with SlaveOk query parameter set.
replVerifyReadsOk(&pq);
// If this exists, the collection is sharded.
// If it doesn't exist, we can assume we're not sharded.
// If we're sharded, we might encounter data that is not consistent with our sharding state.
// We must ignore this data.
CollectionMetadataPtr collMetadata;
if (!shardingState.needCollectionMetadata(pq.ns())) {
collMetadata = CollectionMetadataPtr();
}
else {
collMetadata = shardingState.getCollectionMetadata(pq.ns());
}
// Run the query.
// bb is used to hold query results
// this buffer should contain either requested documents per query or
// explain information, but not both
BufBuilder bb(32768);
bb.skip(sizeof(QueryResult));
// How many results have we obtained from the runner?
int numResults = 0;
// If we're replaying the oplog, we save the last time that we read.
OpTime slaveReadTill;
// Do we save the Runner in a ClientCursor for getMore calls later?
bool saveClientCursor = false;
// We turn on auto-yielding for the runner here. The runner registers itself with the
// active runners list in ClientCursor.
auto_ptr<ScopedRunnerRegistration> safety(new ScopedRunnerRegistration(runner.get()));
runner->setYieldPolicy(Runner::YIELD_AUTO);
BSONObj obj;
Runner::RunnerState state;
// uint64_t numMisplacedDocs = 0;
// set this outside loop. we will need to use this both within loop and when deciding
// to fill in explain information
const bool isExplain = pq.isExplain();
// Have we retrieved info about which plan the runner will
// use to execute the query yet?
bool gotPlanInfo = false;
PlanInfo* rawInfo;
boost::scoped_ptr<PlanInfo> planInfo;
while (Runner::RUNNER_ADVANCED == (state = runner->getNext(&obj, NULL))) {
// Add result to output buffer. This is unnecessary if explain info is requested
if (!isExplain) {
bb.appendBuf((void*)obj.objdata(), obj.objsize());
}
// Count the result.
++numResults;
// In the case of the multi plan runner, we may not be able to
// successfully retrieve plan info until after the query starts
// to run. This is because the multi plan runner doesn't know what
// plan it will end up using until it runs candidates and selects
// the best.
//
// TODO: Do we ever want to output what the MPR is comparing?
if (!gotPlanInfo) {
Status infoStatus = runner->getInfo(NULL, &rawInfo);
if (infoStatus.isOK()) {
gotPlanInfo = true;
planInfo.reset(rawInfo);
// planSummary is really a ThreadSafeString which copies the data from
// the provided pointer.
curop.debug().planSummary = planInfo->planSummary.c_str();
}
}
// Possibly note slave's position in the oplog.
if (pq.hasOption(QueryOption_OplogReplay)) {
BSONElement e = obj["ts"];
if (Date == e.type() || Timestamp == e.type()) {
slaveReadTill = e._opTime();
}
}
// TODO: only one type of 2d search doesn't support this. We need a way to pull it out
// of CanonicalQuery. :(
const bool supportsGetMore = true;
if (isExplain) {
if (enoughForExplain(pq, numResults)) {
break;
}
}
else if (!supportsGetMore && (enough(pq, numResults)
|| bb.len() >= MaxBytesToReturnToClientAtOnce)) {
break;
}
else if (enoughForFirstBatch(pq, numResults, bb.len())) {
QLOG() << "Enough for first batch, wantMore=" << pq.wantMore()
<< " numToReturn=" << pq.getNumToReturn()
<< " numResults=" << numResults
<< endl;
// If only one result requested assume it's a findOne() and don't save the cursor.
if (pq.wantMore() && 1 != pq.getNumToReturn()) {
QLOG() << " runner EOF=" << runner->isEOF() << endl;
saveClientCursor = !runner->isEOF();
}
break;
}
}
// Try to get information about the plan which the runner
// will use to execute the query, it we don't have it already.
if (!gotPlanInfo) {
Status infoStatus = runner->getInfo(NULL, &rawInfo);
if (infoStatus.isOK()) {
gotPlanInfo = true;
planInfo.reset(rawInfo);
// planSummary is really a ThreadSafeString which copies the data from
// the provided pointer.
curop.debug().planSummary = planInfo->planSummary.c_str();
}
}
// If we cache the runner later, we want to deregister it as it receives notifications
// anyway by virtue of being cached.
//
// If we don't cache the runner later, we are deleting it, so it must be deregistered.
//
// So, no matter what, deregister the runner.
safety.reset();
// Caller expects exceptions thrown in certain cases.
if (Runner::RUNNER_ERROR == state) {
TypeExplain* bareExplain;
Status res = runner->getInfo(&bareExplain, NULL);
if (res.isOK()) {
boost::scoped_ptr<TypeExplain> errorExplain(bareExplain);
error() << "Runner error, stats:\n"
<< errorExplain->stats.jsonString(Strict, true);
}
uasserted(17144, "Runner error: " + WorkingSetCommon::toStatusString(obj));
}
// Why save a dead runner?
if (Runner::RUNNER_DEAD == state) {
saveClientCursor = false;
}
else if (pq.hasOption(QueryOption_CursorTailable)) {
// If we're tailing a capped collection, we don't bother saving the cursor if the
// collection is empty. Otherwise, the semantics of the tailable cursor is that the
// client will keep trying to read from it. So we'll keep it around.
Collection* collection = ctx.ctx().db()->getCollection(cq->ns());
if (collection && collection->numRecords() != 0 && pq.getNumToReturn() != 1) {
saveClientCursor = true;
}
}
// TODO(greg): This will go away soon.
if (!shardingState.getVersion(pq.ns()).isWriteCompatibleWith(shardingVersionAtStart)) {
// if the version changed during the query we might be missing some data and its safe to
// send this as mongos can resend at this point
throw SendStaleConfigException(pq.ns(), "version changed during initial query",
shardingVersionAtStart,
shardingState.getVersion(pq.ns()));
}
// Used to fill in explain and to determine if the query is slow enough to be logged.
int elapsedMillis = curop.elapsedMillis();
// Get explain information if:
// 1) it is needed by an explain query;
// 2) profiling is enabled; or
// 3) profiling is disabled but we still need explain details to log a "slow" query.
// Producing explain information is expensive and should be done only if we are certain
// the information will be used.
boost::scoped_ptr<TypeExplain> explain(NULL);
if (isExplain ||
ctx.ctx().db()->getProfilingLevel() > 0 ||
elapsedMillis > serverGlobalParams.slowMS) {
// Ask the runner to produce explain information.
TypeExplain* bareExplain;
Status res = runner->getInfo(&bareExplain, NULL);
if (res.isOK()) {
explain.reset(bareExplain);
}
else if (isExplain) {
error() << "could not produce explain of query '" << pq.getFilter()
<< "', error: " << res.reason();
// If numResults and the data in bb don't correspond, we'll crash later when rooting
// through the reply msg.
BSONObj emptyObj;
bb.appendBuf((void*)emptyObj.objdata(), emptyObj.objsize());
// The explain output is actually a result.
numResults = 1;
// TODO: we can fill out millis etc. here just fine even if the plan screwed up.
}
}
// Fill in the missing run-time fields in explain, starting with propeties of
// the process running the query.
if (isExplain && NULL != explain.get()) {
std::string server = mongoutils::str::stream()
<< getHostNameCached() << ":" << serverGlobalParams.port;
explain->setServer(server);
// We might have skipped some results due to chunk migration etc. so our count is
// correct.
explain->setN(numResults);
// Clock the whole operation.
explain->setMillis(elapsedMillis);
BSONObj explainObj = explain->toBSON();
bb.appendBuf((void*)explainObj.objdata(), explainObj.objsize());
// The explain output is actually a result.
numResults = 1;
}
long long ccId = 0;
if (saveClientCursor) {
// We won't use the runner until it's getMore'd.
runner->saveState();
// Allocate a new ClientCursor. We don't have to worry about leaking it as it's
// inserted into a global map by its ctor.
ClientCursor* cc = new ClientCursor(collection, runner.get(),
cq->getParsed().getOptions(),
cq->getParsed().getFilter());
ccId = cc->cursorid();
QLOG() << "caching runner with cursorid " << ccId
<< " after returning " << numResults << " results" << endl;
// ClientCursor takes ownership of runner. Release to make sure it's not deleted.
runner.release();
// TODO document
if (pq.hasOption(QueryOption_OplogReplay) && !slaveReadTill.isNull()) {
cc->slaveReadTill(slaveReadTill);
}
// TODO document
if (pq.hasOption(QueryOption_Exhaust)) {
curop.debug().exhaust = true;
}
// Set attributes for getMore.
cc->setCollMetadata(collMetadata);
cc->setPos(numResults);
// If the query had a time limit, remaining time is "rolled over" to the cursor (for
// use by future getmore ops).
cc->setLeftoverMaxTimeMicros(curop.getRemainingMaxTimeMicros());
}
else {
QLOG() << "Not caching runner but returning " << numResults << " results.\n";
}
// Add the results from the query into the output buffer.
result.appendData(bb.buf(), bb.len());
bb.decouple();
// Fill out the output buffer's header.
QueryResult* qr = static_cast<QueryResult*>(result.header());
qr->cursorId = ccId;
curop.debug().cursorid = (0 == ccId ? -1 : ccId);
qr->setResultFlagsToOk();
qr->setOperation(opReply);
qr->startingFrom = 0;
qr->nReturned = numResults;
// Set debug information for consumption by the profiler.
curop.debug().ntoskip = pq.getSkip();
curop.debug().nreturned = numResults;
if (NULL != explain.get()) {
if (explain->isScanAndOrderSet()) {
curop.debug().scanAndOrder = explain->getScanAndOrder();
}
else {
curop.debug().scanAndOrder = false;
}
if (explain->isNScannedSet()) {
curop.debug().nscanned = explain->getNScanned();
}
if (explain->isNScannedObjectsSet()) {
curop.debug().nscannedObjects = explain->getNScannedObjects();
}
if (explain->isIDHackSet()) {
curop.debug().idhack = explain->getIDHack();
}
if (!explain->stats.isEmpty()) {
// execStats is a CachedBSONObj because it lives in the race-prone
// curop.
curop.debug().execStats.set(explain->stats);
// Replace exec stats with plan summary if stats cannot fit into CachedBSONObj.
if (curop.debug().execStats.tooBig() && !curop.debug().planSummary.empty()) {
BSONObjBuilder bob;
bob.append("summary", curop.debug().planSummary.toString());
curop.debug().execStats.set(bob.done());
}
}
}
// curop.debug().exhaust is set above.
return curop.debug().exhaust ? pq.ns() : "";
}
// 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();
}
// static
StatusWith<std::vector<std::unique_ptr<QuerySolution>>> QueryPlanner::plan(
const CanonicalQuery& query, const QueryPlannerParams& params) {
LOG(5) << "Beginning planning..." << endl
<< "=============================" << endl
<< "Options = " << optionString(params.options) << endl
<< "Canonical query:" << endl
<< redact(query.toString()) << "=============================";
std::vector<std::unique_ptr<QuerySolution>> out;
for (size_t i = 0; i < params.indices.size(); ++i) {
LOG(5) << "Index " << i << " is " << params.indices[i].toString();
}
const bool canTableScan = !(params.options & QueryPlannerParams::NO_TABLE_SCAN);
const bool isTailable = query.getQueryRequest().isTailable();
// If the query requests a tailable cursor, the only solution is a collscan + filter with
// tailable set on the collscan.
if (isTailable) {
if (!QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR) && canTableScan) {
auto soln = buildCollscanSoln(query, isTailable, params);
if (soln) {
out.push_back(std::move(soln));
}
}
return {std::move(out)};
}
// The hint or sort can be $natural: 1. If this happens, output a collscan. If both
// a $natural hint and a $natural sort are specified, then the direction of the collscan
// is determined by the sign of the sort (not the sign of the hint).
if (!query.getQueryRequest().getHint().isEmpty() ||
!query.getQueryRequest().getSort().isEmpty()) {
BSONObj hintObj = query.getQueryRequest().getHint();
BSONObj sortObj = query.getQueryRequest().getSort();
BSONElement naturalHint = dps::extractElementAtPath(hintObj, "$natural");
BSONElement naturalSort = dps::extractElementAtPath(sortObj, "$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";
// min/max are incompatible with $natural.
if (canTableScan && query.getQueryRequest().getMin().isEmpty() &&
query.getQueryRequest().getMax().isEmpty()) {
auto soln = buildCollscanSoln(query, isTailable, params);
if (soln) {
out.push_back(std::move(soln));
}
}
return {std::move(out)};
}
}
// 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 << "'";
}
// 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.getQueryRequest().getHint();
}
// If snapshot is set, default to collscanning. If the query param SNAPSHOT_USE_ID is set,
// snapshot is a form of a hint, so try to use _id index to make a real plan. If that fails,
// just scan the _id index.
//
// Don't do this if the query is a geonear or text as as text search queries must be answered
// using full text indices and geoNear queries must be answered using geospatial indices.
if (query.getQueryRequest().isSnapshot()) {
RARELY {
warning() << "The snapshot option is deprecated. See "
"http://dochub.mongodb.org/core/snapshot-deprecation";
}
if (!QueryPlannerCommon::hasNode(query.root(), MatchExpression::GEO_NEAR) &&
!QueryPlannerCommon::hasNode(query.root(), MatchExpression::TEXT)) {
const bool useIXScan = params.options & QueryPlannerParams::SNAPSHOT_USE_ID;
if (!useIXScan) {
auto soln = buildCollscanSoln(query, isTailable, params);
if (soln) {
out.push_back(std::move(soln));
}
return {std::move(out)};
} else {
// 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;
}
}
}
}
}
/**
* This is called by db/ops/query.cpp. This is the entry point for answering a query.
*/
string newRunQuery(Message& m, QueryMessage& q, CurOp& curop, Message &result) {
// This is a read lock.
Client::ReadContext ctx(q.ns, dbpath);
// Parse, canonicalize, plan, transcribe, and get a runner.
Runner* rawRunner;
CanonicalQuery* cq;
Status status = getRunner(q, &rawRunner, &cq);
if (!status.isOK()) {
uasserted(17007, "Couldn't process query " + q.query.toString()
+ " why: " + status.reason());
}
verify(NULL != rawRunner);
auto_ptr<Runner> runner(rawRunner);
log() << "Running query on new system: " << cq->toString();
// We freak out later if this changes before we're done with the query.
const ChunkVersion shardingVersionAtStart = shardingState.getVersion(q.ns);
// We use this a lot below.
const LiteParsedQuery& pq = cq->getParsed();
// TODO: Remove when impl'd
if (pq.hasOption(QueryOption_OplogReplay)) {
warning() << "haven't implemented findingstartcursor yet\n";
}
// Handle query option $maxTimeMS (not used with commands).
curop.setMaxTimeMicros(static_cast<unsigned long long>(pq.getMaxTimeMS()) * 1000);
killCurrentOp.checkForInterrupt(); // May trigger maxTimeAlwaysTimeOut fail point.
// uassert if we are not on a primary, and not a secondary with SlaveOk query parameter set.
replVerifyReadsOk(&pq);
// If this exists, the collection is sharded.
// If it doesn't exist, we can assume we're not sharded.
// If we're sharded, we might encounter data that is not consistent with our sharding state.
// We must ignore this data.
CollectionMetadataPtr collMetadata;
if (!shardingState.needCollectionMetadata(pq.ns())) {
collMetadata = CollectionMetadataPtr();
}
else {
collMetadata = shardingState.getCollectionMetadata(pq.ns());
}
// Run the query.
// bb is used to hold query results
// this buffer should contain either requested documents per query or
// explain information, but not both
BufBuilder bb(32768);
bb.skip(sizeof(QueryResult));
// How many results have we obtained from the runner?
int numResults = 0;
// If we're replaying the oplog, we save the last time that we read.
OpTime slaveReadTill;
// Do we save the Runner in a ClientCursor for getMore calls later?
bool saveClientCursor = false;
// We turn on auto-yielding for the runner here. The runner registers itself with the
// active runners list in ClientCursor.
ClientCursor::registerRunner(runner.get());
runner->setYieldPolicy(Runner::YIELD_AUTO);
auto_ptr<DeregisterEvenIfUnderlyingCodeThrows> safety(
new DeregisterEvenIfUnderlyingCodeThrows(runner.get()));
BSONObj obj;
Runner::RunnerState state;
// set this outside loop. we will need to use this both within loop and when deciding
// to fill in explain information
const bool isExplain = pq.isExplain();
while (Runner::RUNNER_ADVANCED == (state = runner->getNext(&obj, NULL))) {
// If we're sharded make sure that we don't return any data that hasn't been migrated
// off of our shared yet.
if (collMetadata) {
// This information can change if we yield and as such we must make sure to re-fetch
// it if we yield.
KeyPattern kp(collMetadata->getKeyPattern());
// This performs excessive BSONObj creation but that's OK for now.
if (!collMetadata->keyBelongsToMe(kp.extractSingleKey(obj))) { continue; }
}
// Add result to output buffer. This is unnecessary if explain info is requested
if (!isExplain) {
bb.appendBuf((void*)obj.objdata(), obj.objsize());
}
// Count the result.
++numResults;
// Possibly note slave's position in the oplog.
if (pq.hasOption(QueryOption_OplogReplay)) {
BSONElement e = obj["ts"];
if (Date == e.type() || Timestamp == e.type()) {
slaveReadTill = e._opTime();
}
}
// TODO: only one type of 2d search doesn't support this. We need a way to pull it out
// of CanonicalQuery. :(
const bool supportsGetMore = true;
if (isExplain) {
if (enoughForExplain(pq, numResults)) {
break;
}
}
else if (!supportsGetMore && (enough(pq, numResults)
|| bb.len() >= MaxBytesToReturnToClientAtOnce)) {
break;
}
else if (enoughForFirstBatch(pq, numResults, bb.len())) {
// If only one result requested assume it's a findOne() and don't save the cursor.
if (pq.wantMore() && 1 != pq.getNumToReturn()) {
saveClientCursor = true;
}
break;
}
}
// If we cache the runner later, we want to deregister it as it receives notifications
// anyway by virtue of being cached.
//
// If we don't cache the runner later, we are deleting it, so it must be deregistered.
//
// So, no matter what, deregister the runner.
safety.reset();
// Caller expects exceptions thrown in certain cases:
// * in-memory sort using too much RAM.
if (Runner::RUNNER_ERROR == state) {
uasserted(17144, "Runner error, memory limit for sort probably exceeded");
}
// Why save a dead runner?
if (Runner::RUNNER_DEAD == state) {
saveClientCursor = false;
}
else if (pq.hasOption(QueryOption_CursorTailable) && (1 != pq.getNumToReturn())) {
// If pq.hasOption(tailable) the only plan the planner will output is a collscan with
// tailable set.
saveClientCursor = true;
}
// TODO(greg): This will go away soon.
if (!shardingState.getVersion(pq.ns()).isWriteCompatibleWith(shardingVersionAtStart)) {
// if the version changed during the query we might be missing some data and its safe to
// send this as mongos can resend at this point
throw SendStaleConfigException(pq.ns(), "version changed during initial query",
shardingVersionAtStart,
shardingState.getVersion(pq.ns()));
}
long long ccId = 0;
if (saveClientCursor) {
// We won't use the runner until it's getMore'd.
runner->saveState();
// Allocate a new ClientCursor. We don't have to worry about leaking it as it's
// inserted into a global map by its ctor.
ClientCursor* cc = new ClientCursor(runner.get(), cq->getParsed().getOptions(),
cq->getParsed().getFilter());
ccId = cc->cursorid();
log() << "caching runner with cursorid " << ccId
<< " after returning " << numResults << " results" << endl;
// ClientCursor takes ownership of runner. Release to make sure it's not deleted.
runner.release();
// TODO document
if (pq.hasOption(QueryOption_OplogReplay) && !slaveReadTill.isNull()) {
cc->slaveReadTill(slaveReadTill);
}
// TODO document
if (pq.hasOption(QueryOption_Exhaust)) {
curop.debug().exhaust = true;
}
// Set attributes for getMore.
cc->setCollMetadata(collMetadata);
cc->setPos(numResults);
// If the query had a time limit, remaining time is "rolled over" to the cursor (for
// use by future getmore ops).
cc->setLeftoverMaxTimeMicros(curop.getRemainingMaxTimeMicros());
}
// append explain information to query results
if (isExplain) {
BSONObjBuilder bob;
bob.append("n", numResults);
BSONObj obj = bob.done();
bb.appendBuf((void*)obj.objdata(), obj.objsize());
// The explain output is actually a result.
numResults = 1;
}
// Add the results from the query into the output buffer.
result.appendData(bb.buf(), bb.len());
bb.decouple();
// Fill out the output buffer's header.
QueryResult* qr = static_cast<QueryResult*>(result.header());
qr->cursorId = ccId;
curop.debug().cursorid = (0 == ccId ? -1 : ccId);
qr->setResultFlagsToOk();
qr->setOperation(opReply);
qr->startingFrom = 0;
qr->nReturned = numResults;
// TODO: nscanned is bogus.
// curop.debug().nscanned = ( cursor ? cursor->nscanned() : 0LL );
curop.debug().ntoskip = pq.getSkip();
curop.debug().nreturned = numResults;
// curop.debug().exhaust is set above.
return curop.debug().exhaust ? pq.ns() : "";
}