BSONObj Helpers::modifiedRangeBound( const BSONObj& bound ,
const BSONObj& keyPattern ,
int minOrMax ){
BSONObjBuilder newBound;
BSONObjIterator src( bound );
BSONObjIterator pat( keyPattern );
while( src.more() ){
massert( 16341 ,
str::stream() << "keyPattern " << keyPattern
<< " shorter than bound " << bound ,
pat.more() );
BSONElement srcElt = src.next();
BSONElement patElt = pat.next();
massert( 16333 ,
str::stream() << "field names of bound " << bound
<< " do not match those of keyPattern " << keyPattern ,
str::equals( srcElt.fieldName() , patElt.fieldName() ) );
newBound.appendAs( srcElt , "" );
}
while( pat.more() ){
BSONElement patElt = pat.next();
verify( patElt.isNumber() );
if( minOrMax * patElt.numberInt() == 1){
newBound.appendMaxKey("");
}
else {
newBound.appendMinKey("");
}
}
return newBound.obj();
}
BSONObj Helpers::modifiedRangeBound( const BSONObj& bound ,
const BSONObj& keyPattern ,
int minOrMax ){
BSONObjBuilder newBound;
BSONObjIterator src( bound );
BSONObjIterator pat( keyPattern );
while( src.more() ){
BSONElement srcElt = src.next();
BSONElement patElt = pat.next();
massert( 16333 , "bound " + bound.toString() +
"not extendible to pattern" + keyPattern.toString(),
strcmp( srcElt.fieldName() , patElt.fieldName() ) == 0 );
newBound.appendAs( srcElt , "" );
}
while( pat.more() ){
BSONElement patElt = pat.next();
verify( patElt.isNumber() );
if( minOrMax * patElt.numberInt() == 1){
newBound.appendMaxKey("");
}
else {
newBound.appendMinKey("");
}
}
return newBound.obj();
}
void run(){
Scope * s = globalScriptEngine->createScope();
BSONObjBuilder b;
b.appendTimestamp( "a" , 123456789 );
b.appendMinKey( "b" );
b.appendMaxKey( "c" );
b.appendTimestamp( "d" , 1234000 , 9876 );
{
BSONObj t = b.done();
ASSERT_EQUALS( 1234000U , t["d"].timestampTime() );
ASSERT_EQUALS( 9876U , t["d"].timestampInc() );
}
s->setObject( "z" , b.obj() );
ASSERT( s->invoke( "y = { a : z.a , b : z.b , c : z.c , d: z.d }" , BSONObj() ) == 0 );
BSONObj out = s->getObject( "y" );
ASSERT_EQUALS( Timestamp , out["a"].type() );
ASSERT_EQUALS( MinKey , out["b"].type() );
ASSERT_EQUALS( MaxKey , out["c"].type() );
ASSERT_EQUALS( Timestamp , out["d"].type() );
ASSERT_EQUALS( 9876U , out["d"].timestampInc() );
ASSERT_EQUALS( 1234000U , out["d"].timestampTime() );
ASSERT_EQUALS( 123456789U , out["a"].date() );
delete s;
}
// static
void IndexBoundsBuilder::allValuesForField(const BSONElement& elt, OrderedIntervalList* out) {
// ARGH, BSONValue would make this shorter.
BSONObjBuilder bob;
bob.appendMinKey("");
bob.appendMaxKey("");
out->name = elt.fieldName();
out->intervals.push_back(makeRangeInterval(bob.obj(), true, true));
}
// static
void IndexBoundsBuilder::allValuesForField(const BSONElement& elt, OrderedIntervalList* out) {
// ARGH, BSONValue would make this shorter.
BSONObjBuilder bob;
if (-1 == elt.number()) {
// Index should go from MaxKey to MinKey as it's descending.
bob.appendMaxKey("");
bob.appendMinKey("");
}
else {
// Index goes from MinKey to MaxKey as it's ascending.
bob.appendMinKey("");
bob.appendMaxKey("");
}
out->name = elt.fieldName();
out->intervals.push_back(makeRangeInterval(bob.obj(), true, true));
}
/**
* "Finishes" the min object for the $min query option by filling in an empty object with
* MinKey/MaxKey and stripping field names.
*
* In the case that 'minObj' is empty, we "finish" it by filling in either MinKey or MaxKey
* instead. Choosing whether to use MinKey or MaxKey is done by comparing against 'maxObj'.
* For instance, suppose 'minObj' is empty, 'maxObj' is { a: 3 }, and the key pattern is
* { a: -1 }. According to the key pattern ordering, { a: 3 } < MinKey. This means that the
* proper resulting bounds are
*
* start: { '': MaxKey }, end: { '': 3 }
*
* as opposed to
*
* start: { '': MinKey }, end: { '': 3 }
*
* Suppose instead that the key pattern is { a: 1 }, with the same 'minObj' and 'maxObj'
* (that is, an empty object and { a: 3 } respectively). In this case, { a: 3 } > MinKey,
* which means that we use range [{'': MinKey}, {'': 3}]. The proper 'minObj' in this case is
* MinKey, whereas in the previous example it was MaxKey.
*
* If 'minObj' is non-empty, then all we do is strip its field names (because index keys always
* have empty field names).
*/
static BSONObj finishMinObj(const BSONObj& kp, const BSONObj& minObj, const BSONObj& maxObj) {
BSONObjBuilder bob;
bob.appendMinKey("");
BSONObj minKey = bob.obj();
if (minObj.isEmpty()) {
if (0 > minKey.woCompare(maxObj, kp, false)) {
BSONObjBuilder minKeyBuilder;
minKeyBuilder.appendMinKey("");
return minKeyBuilder.obj();
} else {
BSONObjBuilder maxKeyBuilder;
maxKeyBuilder.appendMaxKey("");
return maxKeyBuilder.obj();
}
} else {
return stripFieldNames(minObj);
}
}
// static
void OrderedIntervalList::complement() {
BSONObjBuilder minBob;
minBob.appendMinKey("");
BSONObj minObj = minBob.obj();
// We complement by scanning the entire range of BSON values
// from MinKey to MaxKey. The value from which we must begin
// the next complemented interval is kept in 'curBoundary'.
BSONElement curBoundary = minObj.firstElement();
// If 'curInclusive' is true, then 'curBoundary' is
// included in one of the original intervals, and hence
// should not be included in the complement (and vice-versa
// if 'curInclusive' is false).
bool curInclusive = false;
// We will build up a list of intervals that represents
// the inversion of those in the OIL.
vector<Interval> newIntervals;
for (size_t j = 0; j < intervals.size(); ++j) {
Interval curInt = intervals[j];
if (0 != curInt.start.woCompare(curBoundary) ||
(!curInclusive && !curInt.startInclusive)) {
// Make a new interval from 'curBoundary' to
// the start of 'curInterval'.
BSONObjBuilder intBob;
intBob.append(curBoundary);
intBob.append(curInt.start);
Interval newInt(intBob.obj(), !curInclusive, !curInt.startInclusive);
newIntervals.push_back(newInt);
}
// Reset the boundary for the next iteration.
curBoundary = curInt.end;
curInclusive = curInt.endInclusive;
}
// We may have to add a final interval which ends in MaxKey.
BSONObjBuilder maxBob;
maxBob.appendMaxKey("");
BSONObj maxObj = maxBob.obj();
BSONElement maxKey = maxObj.firstElement();
if (0 != maxKey.woCompare(curBoundary) || !curInclusive) {
BSONObjBuilder intBob;
intBob.append(curBoundary);
intBob.append(maxKey);
Interval newInt(intBob.obj(), !curInclusive, true);
newIntervals.push_back(newInt);
}
// Replace the old list of intervals with the new one.
intervals.clear();
intervals.insert(intervals.end(), newIntervals.begin(), newIntervals.end());
}
/**
* "Finishes" the min object for the $min query option by filling in an empty object with
* MinKey/MaxKey and stripping field names. Also translates keys according to the collation, if
* necessary.
*
* In the case that 'minObj' is empty, we "finish" it by filling in either MinKey or MaxKey
* instead. Choosing whether to use MinKey or MaxKey is done by comparing against 'maxObj'.
* For instance, suppose 'minObj' is empty, 'maxObj' is { a: 3 }, and the key pattern is
* { a: -1 }. According to the key pattern ordering, { a: 3 } < MinKey. This means that the
* proper resulting bounds are
*
* start: { '': MaxKey }, end: { '': 3 }
*
* as opposed to
*
* start: { '': MinKey }, end: { '': 3 }
*
* Suppose instead that the key pattern is { a: 1 }, with the same 'minObj' and 'maxObj'
* (that is, an empty object and { a: 3 } respectively). In this case, { a: 3 } > MinKey,
* which means that we use range [{'': MinKey}, {'': 3}]. The proper 'minObj' in this case is
* MinKey, whereas in the previous example it was MaxKey.
*
* If 'minObj' is non-empty, then all we do is strip its field names (because index keys always
* have empty field names).
*/
static BSONObj finishMinObj(const IndexEntry& indexEntry,
const BSONObj& minObj,
const BSONObj& maxObj) {
BSONObjBuilder bob;
bob.appendMinKey("");
BSONObj minKey = bob.obj();
if (minObj.isEmpty()) {
if (0 > minKey.woCompare(maxObj, indexEntry.keyPattern, false)) {
BSONObjBuilder minKeyBuilder;
minKeyBuilder.appendMinKey("");
return minKeyBuilder.obj();
} else {
BSONObjBuilder maxKeyBuilder;
maxKeyBuilder.appendMaxKey("");
return maxKeyBuilder.obj();
}
} else {
return stripFieldNamesAndApplyCollation(minObj, indexEntry.collator);
}
}
bool run(const string& dbname, BSONObj& jsobj, string& errmsg, BSONObjBuilder& result, bool fromRepl ){
const char* ns = jsobj.getStringField( "splitVector" );
BSONObj keyPattern = jsobj.getObjectField( "keyPattern" );
BSONObj min = jsobj.getObjectField( "min" );
BSONObj max = jsobj.getObjectField( "max" );
if ( min.isEmpty() && max.isEmpty() ){
BSONObjBuilder minBuilder;
BSONObjBuilder maxBuilder;
BSONForEach(key, keyPattern){
minBuilder.appendMinKey( key.fieldName() );
maxBuilder.appendMaxKey( key.fieldName() );
}
ShardKeyPattern::ShardKeyPattern( BSONObj p ) : pattern( p.getOwned() ) {
pattern.getFieldNames(patternfields);
BSONObjBuilder min;
BSONObjBuilder max;
BSONObjIterator it(p);
while (it.more()) {
BSONElement e (it.next());
min.appendMinKey(e.fieldName());
max.appendMaxKey(e.fieldName());
}
gMin = min.obj();
gMax = max.obj();
}
BSONObj IndexBounds::toBSON() const {
BSONObjBuilder builder;
if (isSimpleRange) {
// TODO
}
else {
for (vector<OrderedIntervalList>::const_iterator itField = fields.begin();
itField != fields.end();
++itField) {
BSONArrayBuilder fieldBuilder(builder.subarrayStart(itField->name));
for (vector<Interval>::const_iterator itInterval = itField->intervals.begin();
itInterval != itField->intervals.end();
++itInterval) {
BSONArrayBuilder intervalBuilder;
// Careful to output $minElement/$maxElement if we don't have bounds.
if (itInterval->start.eoo()) {
BSONObjBuilder minBuilder;
minBuilder.appendMinKey("");
BSONObj minKeyObj = minBuilder.obj();
intervalBuilder.append(minKeyObj.firstElement());
}
else {
intervalBuilder.append(itInterval->start);
}
if (itInterval->end.eoo()) {
BSONObjBuilder maxBuilder;
maxBuilder.appendMaxKey("");
BSONObj maxKeyObj = maxBuilder.obj();
intervalBuilder.append(maxKeyObj.firstElement());
}
else {
intervalBuilder.append(itInterval->end);
}
fieldBuilder.append(
static_cast<BSONArray>(intervalBuilder.arr().clientReadable()));
}
}
}
return builder.obj();
}
bool run(const string& dbname, BSONObj& jsobj, string& errmsg, BSONObjBuilder& result, bool fromRepl ){
//
// 1.a We'll parse the parameters in two steps. First, make sure the we can use the split index to get
// a good approximation of the size of the chunk -- without needing to access the actual data.
//
const char* ns = jsobj.getStringField( "splitVector" );
BSONObj keyPattern = jsobj.getObjectField( "keyPattern" );
// If min and max are not provided use the "minKey" and "maxKey" for the sharding key pattern.
BSONObj min = jsobj.getObjectField( "min" );
BSONObj max = jsobj.getObjectField( "max" );
if ( min.isEmpty() && max.isEmpty() ){
BSONObjBuilder minBuilder;
BSONObjBuilder maxBuilder;
BSONForEach(key, keyPattern){
minBuilder.appendMinKey( key.fieldName() );
maxBuilder.appendMaxKey( key.fieldName() );
}
bool appendSpecialDBObject( Convertor * c , BSONObjBuilder& b , const string& name , JSObject * o ){
if ( JS_InstanceOf( c->_context , o , &object_id_class , 0 ) ){
OID oid;
oid.init( c->getString( o , "str" ) );
b.append( name.c_str() , oid );
return true;
}
if ( JS_InstanceOf( c->_context , o , &minkey_class , 0 ) ){
b.appendMinKey( name.c_str() );
return true;
}
if ( JS_InstanceOf( c->_context , o , &maxkey_class , 0 ) ){
b.appendMaxKey( name.c_str() );
return true;
}
if ( JS_InstanceOf( c->_context , o , ×tamp_class , 0 ) ){
b.appendTimestamp( name.c_str() , (unsigned long long)c->getNumber( o , "t" ) , (unsigned int )c->getNumber( o , "i" ) );
return true;
}
{
jsdouble d = js_DateGetMsecSinceEpoch( c->_context , o );
if ( d ){
b.appendDate( name.c_str() , (unsigned long long)d );
return true;
}
}
return false;
}
Interval IndexBoundsBuilder::allValues() {
BSONObjBuilder bob;
bob.appendMinKey("");
bob.appendMaxKey("");
return makeRangeInterval(bob.obj(), true, true);
}
// static
void IndexBoundsBuilder::translate(const MatchExpression* expr, int direction,
OrderedIntervalList* oilOut, bool* exactOut) {
Interval interval;
bool exact = false;
if (expr->isLeaf()) {
if (MatchExpression::EQ == expr->matchType()) {
const EqualityMatchExpression* node = static_cast<const EqualityMatchExpression*>(expr);
// We have to copy the data out of the parse tree and stuff it into the index bounds.
// BSONValue will be useful here.
BSONObj dataObj = objFromElement(node->getData());
if (dataObj.couldBeArray()) {
// XXX: build better bounds
warning() << "building lazy bounds for " << expr->toString() << endl;
interval = allValues();
exact = false;
}
else {
verify(dataObj.isOwned());
interval = makePointInterval(dataObj);
exact = true;
}
}
else if (MatchExpression::LTE == expr->matchType()) {
const LTEMatchExpression* node = static_cast<const LTEMatchExpression*>(expr);
BSONObjBuilder bob;
bob.appendMinKey("");
bob.append(node->getData());
BSONObj dataObj = bob.obj();
verify(dataObj.isOwned());
interval = makeRangeInterval(dataObj, true, true);
exact = true;
}
else if (MatchExpression::LT == expr->matchType()) {
const LTMatchExpression* node = static_cast<const LTMatchExpression*>(expr);
BSONObjBuilder bob;
bob.appendMinKey("");
bob.append(node->getData());
BSONObj dataObj = bob.obj();
verify(dataObj.isOwned());
interval = makeRangeInterval(dataObj, true, false);
exact = true;
}
else if (MatchExpression::GT == expr->matchType()) {
const GTMatchExpression* node = static_cast<const GTMatchExpression*>(expr);
BSONObjBuilder bob;
bob.append(node->getData());
bob.appendMaxKey("");
BSONObj dataObj = bob.obj();
verify(dataObj.isOwned());
interval = makeRangeInterval(dataObj, false, true);
exact = true;
}
else if (MatchExpression::GTE == expr->matchType()) {
const GTEMatchExpression* node = static_cast<const GTEMatchExpression*>(expr);
BSONObjBuilder bob;
bob.append(node->getData());
bob.appendMaxKey("");
BSONObj dataObj = bob.obj();
verify(dataObj.isOwned());
interval = makeRangeInterval(dataObj, true, true);
exact = true;
}
else {
// XXX: build better bounds
warning() << "building lazy bounds for " << expr->toString() << endl;
interval = allValues();
exact = false;
}
}
else {
// XXX: build better bounds
verify(expr->isArray());
warning() << "building lazy bounds for " << expr->toString() << endl;
interval = allValues();
exact = false;
}
if (-1 == direction) {
reverseInterval(&interval);
}
oilOut->intervals.push_back(interval);
*exactOut = exact;
}
// Returns the min and max keys which bound a particular location.
// The only time these may be equal is when we actually equal the location
// itself, otherwise our expanding algorithm will fail.
// static
bool BtreeLocation::initial(IndexDescriptor* descriptor, const TwoDIndexingParams& params,
BtreeLocation& min, BtreeLocation& max, GeoHash start) {
verify(descriptor);
min._eof = false;
max._eof = false;
// Add the range for the 2d indexed field to the keys used.
// Two scans: one for min one for max.
IndexScanParams minParams;
minParams.direction = -1;
minParams.forceBtreeAccessMethod = true;
minParams.descriptor = descriptor->clone();
minParams.bounds.fields.resize(descriptor->keyPattern().nFields());
minParams.doNotDedup = true;
// First field of start key goes (MINKEY, start] (in reverse)
BSONObjBuilder firstBob;
firstBob.appendMinKey("");
start.appendToBuilder(&firstBob, "");
minParams.bounds.fields[0].intervals.push_back(Interval(firstBob.obj(), false, true));
IndexScanParams maxParams;
maxParams.forceBtreeAccessMethod = true;
maxParams.direction = 1;
maxParams.descriptor = descriptor->clone();
maxParams.bounds.fields.resize(descriptor->keyPattern().nFields());
// Don't have the ixscan dedup since we want dup DiskLocs because of multi-point docs.
maxParams.doNotDedup = true;
// First field of end key goes (start, MAXKEY)
BSONObjBuilder secondBob;
start.appendToBuilder(&secondBob, "");
secondBob.appendMaxKey("");
maxParams.bounds.fields[0].intervals.push_back(Interval(secondBob.obj(), false, false));
BSONObjIterator it(descriptor->keyPattern());
BSONElement kpElt = it.next();
maxParams.bounds.fields[0].name = kpElt.fieldName();
minParams.bounds.fields[0].name = kpElt.fieldName();
// Fill out the non-2d indexed fields with the "all values" interval, aligned properly.
size_t idx = 1;
while (it.more()) {
kpElt = it.next();
maxParams.bounds.fields[idx].intervals.push_back(IndexBoundsBuilder::allValues());
minParams.bounds.fields[idx].intervals.push_back(IndexBoundsBuilder::allValues());
maxParams.bounds.fields[idx].name = kpElt.fieldName();
minParams.bounds.fields[idx].name = kpElt.fieldName();
if (kpElt.number() == -1) {
IndexBoundsBuilder::reverseInterval(&minParams.bounds.fields[idx].intervals[0]);
IndexBoundsBuilder::reverseInterval(&maxParams.bounds.fields[idx].intervals[0]);
}
++idx;
}
for (size_t i = 0; i < minParams.bounds.fields.size(); ++i) {
IndexBoundsBuilder::reverseInterval(&minParams.bounds.fields[i].intervals[0]);
}
//cout << "keyPattern " << descriptor->keyPattern().toString() << endl;
//cout << "minBounds " << minParams.bounds.toString() << endl;
//cout << "maxBounds " << maxParams.bounds.toString() << endl;
verify(minParams.bounds.isValidFor(descriptor->keyPattern(), -1));
verify(maxParams.bounds.isValidFor(descriptor->keyPattern(), 1));
min._ws.reset(new WorkingSet());
min._scan.reset(new IndexScan(minParams, min._ws.get(), NULL));
max._ws.reset(new WorkingSet());
max._scan.reset(new IndexScan(maxParams, max._ws.get(), NULL));
min.advance();
max.advance();
return !max._eof || !min._eof;
}
// 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();
}
BSONObj objWithMinKey(int start) {
BSONObjBuilder startKeyBob;
startKeyBob.append("", start);
startKeyBob.appendMinKey("");
return startKeyBob.obj();
}
void run() {
int numShards = 10;
int numInitialChunks = 5;
int maxChunks = 100000; // Needed to not overflow the BSONArray's max bytes
int keySize = 2;
BSONArrayBuilder chunksB;
BSONObj lastSplitPt;
ShardChunkVersion version( 1, 0, OID() );
//
// Generate numChunks with a given key size over numShards
// All chunks have double key values, so we can split them a bunch
//
for( int i = -1; i < numInitialChunks; i++ ){
BSONObjBuilder splitPtB;
for( int k = 0; k < keySize; k++ ){
string field = string( "k" ) + string( 1, (char)('0' + k) );
if( i < 0 )
splitPtB.appendMinKey( field );
else if( i < numInitialChunks - 1 )
splitPtB.append( field, (double)i );
else
splitPtB.appendMaxKey( field );
}
BSONObj splitPt = splitPtB.obj();
if( i >= 0 ){
BSONObjBuilder chunkB;
chunkB.append( "min", lastSplitPt );
chunkB.append( "max", splitPt );
int shardNum = rand( numShards );
chunkB.append( "shard", "shard" + string( 1, (char)('A' + shardNum) ) );
rand( 2 ) ? version.incMajor() : version.incMinor();
version.addToBSON( chunkB, "lastmod" );
chunksB.append( chunkB.obj() );
}
lastSplitPt = splitPt;
}
BSONArray chunks = chunksB.arr();
// log() << "Chunks generated : " << chunks << endl;
DBClientMockCursor chunksCursor( chunks );
// Setup the empty ranges and versions first
RangeMap ranges;
ShardChunkVersion maxVersion = ShardChunkVersion( 0, 0, OID() );
VersionMap maxShardVersions;
// Create a differ which will track our progress
boost::shared_ptr< DefaultDiffAdapter > differ( _inverse ? new InverseDiffAdapter() : new DefaultDiffAdapter() );
differ->attach( "test", ranges, maxVersion, maxShardVersions );
// Validate initial load
differ->calculateConfigDiff( chunksCursor );
validate( chunks, ranges, maxVersion, maxShardVersions );
// Generate a lot of diffs, and keep validating that updating from the diffs always
// gives us the right ranges and versions
int numDiffs = 135; // Makes about 100000 chunks overall
int numChunks = numInitialChunks;
for( int i = 0; i < numDiffs; i++ ){
// log() << "Generating new diff... " << i << endl;
BSONArrayBuilder diffsB;
BSONArrayBuilder newChunksB;
BSONObjIterator chunksIt( chunks );
while( chunksIt.more() ){
BSONObj chunk = chunksIt.next().Obj();
int randChoice = rand( 10 );
if( randChoice < 2 && numChunks < maxChunks ){
// Simulate a split
// log() << " ...starting a split with chunk " << chunk << endl;
BSONObjBuilder leftB;
BSONObjBuilder rightB;
BSONObjBuilder midB;
for( int k = 0; k < keySize; k++ ){
string field = string( "k" ) + string( 1, (char)('0' + k) );
BSONType maxType = chunk["max"].Obj()[field].type();
double max = maxType == NumberDouble ? chunk["max"].Obj()[field].Number() : 0.0;
BSONType minType = chunk["min"].Obj()[field].type();
double min = minType == NumberDouble ? chunk["min"].Obj()[field].Number() : 0.0;
if( minType == MinKey ){
midB.append( field, max - 1.0 );
}
else if( maxType == MaxKey ){
midB.append( field, min + 1.0 );
}
else {
midB.append( field, ( max + min ) / 2.0 );
}
}
BSONObj midPt = midB.obj();
// Only happens if we can't split the min chunk
if( midPt.isEmpty() ) continue;
leftB.append( chunk["min"] );
leftB.append( "max", midPt );
rightB.append( "min", midPt );
rightB.append( chunk["max"] );
leftB.append( chunk["shard"] );
rightB.append( chunk["shard"] );
version.incMajor();
version._minor = 0;
version.addToBSON( leftB, "lastmod" );
version.incMinor();
version.addToBSON( rightB, "lastmod" );
BSONObj left = leftB.obj();
BSONObj right = rightB.obj();
// log() << " ... split into " << left << " and " << right << endl;
newChunksB.append( left );
newChunksB.append( right );
diffsB.append( right );
diffsB.append( left );
numChunks++;
}
else if( randChoice < 4 && chunksIt.more() ){
// Simulate a migrate
// log() << " ...starting a migrate with chunk " << chunk << endl;
BSONObj prevShardChunk;
while( chunksIt.more() ){
prevShardChunk = chunksIt.next().Obj();
if( prevShardChunk["shard"].String() == chunk["shard"].String() ) break;
// log() << "... appending chunk from diff shard: " << prevShardChunk << endl;
newChunksB.append( prevShardChunk );
prevShardChunk = BSONObj();
}
// We need to move between different shards, hence the weirdness in logic here
if( ! prevShardChunk.isEmpty() ){
BSONObjBuilder newShardB;
BSONObjBuilder prevShardB;
newShardB.append( chunk["min"] );
newShardB.append( chunk["max"] );
prevShardB.append( prevShardChunk["min"] );
prevShardB.append( prevShardChunk["max"] );
int shardNum = rand( numShards );
newShardB.append( "shard", "shard" + string( 1, (char)('A' + shardNum) ) );
prevShardB.append( prevShardChunk["shard"] );
version.incMajor();
version._minor = 0;
version.addToBSON( newShardB, "lastmod" );
version.incMinor();
version.addToBSON( prevShardB, "lastmod" );
BSONObj newShard = newShardB.obj();
BSONObj prevShard = prevShardB.obj();
// log() << " ... migrated to " << newShard << " and updated " << prevShard << endl;
newChunksB.append( newShard );
newChunksB.append( prevShard );
diffsB.append( newShard );
diffsB.append( prevShard );
}
else{
// log() << "... appending chunk, no more left: " << chunk << endl;
newChunksB.append( chunk );
}
}
else{
// log() << "Appending chunk : " << chunk << endl;
newChunksB.append( chunk );
}
}
BSONArray diffs = diffsB.arr();
chunks = newChunksB.arr();
// log() << "Diffs generated : " << diffs << endl;
// log() << "All chunks : " << chunks << endl;
// Rarely entirely clear out our data
if( rand( 10 ) < 1 ){
diffs = chunks;
ranges.clear();
maxVersion = ShardChunkVersion( 0, 0, OID() );
maxShardVersions.clear();
}
// log() << "Total number of chunks : " << numChunks << " iteration " << i << endl;
DBClientMockCursor diffCursor( diffs );
differ->calculateConfigDiff( diffCursor );
validate( chunks, ranges, maxVersion, maxShardVersions );
}
}
void v8ToMongoElement( BSONObjBuilder & b , v8::Handle<v8::String> name , const string sname , v8::Handle<v8::Value> value ){
if ( value->IsString() ){
b.append( sname.c_str() , toSTLString( value ).c_str() );
return;
}
if ( value->IsFunction() ){
b.appendCode( sname.c_str() , toSTLString( value ).c_str() );
return;
}
if ( value->IsNumber() ){
if ( value->IsInt32() )
b.append( sname.c_str(), int( value->ToInt32()->Value() ) );
else
b.append( sname.c_str() , value->ToNumber()->Value() );
return;
}
if ( value->IsArray() ){
BSONObj sub = v8ToMongo( value->ToObject() );
b.appendArray( sname.c_str() , sub );
return;
}
if ( value->IsDate() ){
b.appendDate( sname.c_str() , Date_t(v8::Date::Cast( *value )->NumberValue()) );
return;
}
if ( value->IsExternal() )
return;
if ( value->IsObject() ){
// The user could potentially modify the fields of these special objects,
// wreaking havoc when we attempt to reinterpret them. Not doing any validation
// for now...
Local< v8::Object > obj = value->ToObject();
if ( obj->InternalFieldCount() && obj->GetInternalField( 0 )->IsNumber() ) {
switch( obj->GetInternalField( 0 )->ToInt32()->Value() ) { // NOTE Uint32's Value() gave me a linking error, so going with this instead
case Timestamp:
b.appendTimestamp( sname.c_str(),
Date_t( v8::Date::Cast( *obj->Get( v8::String::New( "time" ) ) )->NumberValue() ),
obj->Get( v8::String::New( "i" ) )->ToInt32()->Value() );
return;
case MinKey:
b.appendMinKey( sname.c_str() );
return;
case MaxKey:
b.appendMaxKey( sname.c_str() );
return;
default:
assert( "invalid internal field" == 0 );
}
}
string s = toSTLString( value );
if ( s.size() && s[0] == '/' ){
s = s.substr( 1 );
string r = s.substr( 0 , s.rfind( "/" ) );
string o = s.substr( s.rfind( "/" ) + 1 );
b.appendRegex( sname.c_str() , r.c_str() , o.c_str() );
}
else if ( value->ToObject()->GetPrototype()->IsObject() &&
value->ToObject()->GetPrototype()->ToObject()->HasRealNamedProperty( v8::String::New( "isObjectId" ) ) ){
OID oid;
oid.init( toSTLString( value ) );
b.appendOID( sname.c_str() , &oid );
}
else if ( !value->ToObject()->GetHiddenValue( v8::String::New( "__NumberLong" ) ).IsEmpty() ) {
// TODO might be nice to potentially speed this up with an indexed internal
// field, but I don't yet know how to use an ObjectTemplate with a
// constructor.
unsigned long long val =
( (unsigned long long)( value->ToObject()->Get( v8::String::New( "top" ) )->ToInt32()->Value() ) << 32 ) +
(unsigned)( value->ToObject()->Get( v8::String::New( "bottom" ) )->ToInt32()->Value() );
b.append( sname.c_str(), (long long)val );
}
else if ( !value->ToObject()->GetHiddenValue( v8::String::New( "__DBPointer" ) ).IsEmpty() ) {
OID oid;
oid.init( toSTLString( value->ToObject()->Get( v8::String::New( "id" ) ) ) );
string ns = toSTLString( value->ToObject()->Get( v8::String::New( "ns" ) ) );
b.appendDBRef( sname.c_str(), ns.c_str(), oid );
}
else if ( !value->ToObject()->GetHiddenValue( v8::String::New( "__BinData" ) ).IsEmpty() ) {
int len = obj->Get( v8::String::New( "len" ) )->ToInt32()->Value();
v8::String::Utf8Value data( obj->Get( v8::String::New( "data" ) ) );
const char *dataArray = *data;
assert( data.length() == len );
b.appendBinData( sname.c_str(),
len,
mongo::BinDataType( obj->Get( v8::String::New( "type" ) )->ToInt32()->Value() ),
dataArray );
} else {
BSONObj sub = v8ToMongo( value->ToObject() );
b.append( sname.c_str() , sub );
}
return;
}
if ( value->IsBoolean() ){
b.appendBool( sname.c_str() , value->ToBoolean()->Value() );
return;
}
else if ( value->IsUndefined() ){
b.appendUndefined( sname.c_str() );
return;
}
else if ( value->IsNull() ){
b.appendNull( sname.c_str() );
return;
}
cout << "don't know how to convert to mongo field [" << name << "]\t" << value << endl;
}
// static
Status QueryPlanner::plan(const CanonicalQuery& query,
const QueryPlannerParams& params,
std::vector<QuerySolution*>* out) {
QLOG() << "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) {
QLOG() << "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().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 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())) {
QLOG() << "Forcing a table scan due to hinted $natural\n";
// min/max are incompatible with $natural.
if (canTableScan && query.getParsed().getMin().isEmpty()
&& query.getParsed().getMax().isEmpty()) {
QuerySolution* soln = buildCollscanSoln(query, false, params);
if (NULL != soln) {
out->push_back(soln);
}
}
return Status::OK();
}
}
// Figure out what fields we care about.
unordered_set<string> fields;
QueryPlannerIXSelect::getFields(query.root(), "", &fields);
for (unordered_set<string>::const_iterator it = fields.begin(); it != fields.end(); ++it) {
QLOG() << "Predicate over field '" << *it << "'" << endl;
}
// Filter our indices so we only look at indices that are over our predicates.
vector<IndexEntry> relevantIndices;
// Hints require us to only consider the hinted index.
// If index filters in the query settings were used to override
// the allowed indices for planning, we should not use the hinted index
// requested in the query.
BSONObj hintIndex;
if (!params.indexFiltersApplied) {
hintIndex = query.getParsed().getHint();
}
// Snapshot is a form of a hint. If snapshot is set, try to use _id index to make a real
// plan. If that fails, just scan the _id index.
if (query.getParsed().isSnapshot()) {
// Find the ID index in indexKeyPatterns. It's our hint.
for (size_t i = 0; i < params.indices.size(); ++i) {
if (isIdIndex(params.indices[i].keyPattern)) {
hintIndex = params.indices[i].keyPattern;
break;
}
}
}
size_t hintIndexNumber = numeric_limits<size_t>::max();
if (hintIndex.isEmpty()) {
QueryPlannerIXSelect::findRelevantIndices(fields, params.indices, &relevantIndices);
}
else {
// Sigh. If the hint is specified it might be using the index name.
BSONElement firstHintElt = hintIndex.firstElement();
if (str::equals("$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) {
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 doesn't work with hint";
return Status(ErrorCodes::BadValue,
"hint provided does not work with min query");
}
if (!maxObj.isEmpty() && !indexCompatibleMaxMin(maxObj, hintIndex)) {
QLOG() << "Maxobj doesn't work with 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 index " << i << " is " << relevantIndices[i].toString() << endl;
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);
// query.root() is now annotated with RelevantTag(s).
QLOG() << "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()) {
QLOG() << "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");
}
QLOG() << "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());
QLOG() << "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)) {
QLOG() << "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.
boost::scoped_ptr<MatchExpression> clone(rawTree->shallowClone());
CanonicalQuery::sortTree(clone.get());
PlanCacheIndexTree* cacheData;
Status indexTreeStatus = cacheDataFromTaggedTree(clone.get(), relevantIndices, &cacheData);
if (!indexTreeStatus.isOK()) {
QLOG() << "Query is not cachable: " << indexTreeStatus.reason() << endl;
}
auto_ptr<PlanCacheIndexTree> autoData(cacheData);
// 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:" << 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();
QLOG() << "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);
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.
// 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.
if (index.type != INDEX_BTREE) {
continue;
}
// Only non-sparse indexes can be used to provide a sort.
if (index.sparse) {
continue;
}
// 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.
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) {
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))) {
QLOG() << "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);
QLOG() << "Planner: outputting a collscan:" << endl
<< collscan->toString();
}
}
return Status::OK();
}
