const Shard& DBConfig::getShard( const string& ns ){
if ( isSharded( ns ) )
return Shard::EMPTY;
uassert( 10178 , "no primary!" , _primary.ok() );
return _primary;
}
string DBConfig::getShard( const string& ns ){
if ( isSharded( ns ) )
return "";
uassert( 10178 , "no primary!" , _primary.size() );
return _primary;
}
ChunkManager* DBConfig::getChunkManager( const string& ns , bool reload ){
ChunkManager* m = _shards[ns];
if ( m && ! reload )
return m;
uassert( 10181 , (string)"not sharded:" + ns , isSharded( ns ) );
if ( m && reload )
log() << "reloading shard info for: " << ns << endl;
m = new ChunkManager( this , ns , _sharded[ ns ].key , _sharded[ns].unique );
_shards[ns] = m;
return m;
}
ChunkManager* DBConfig::shardCollection( const string& ns , ShardKeyPattern fieldsAndOrder , bool unique ){
if ( ! _shardingEnabled )
throw UserException( 8042 , "db doesn't have sharding enabled" );
ChunkManager * info = _shards[ns];
if ( info )
return info;
if ( isSharded( ns ) )
throw UserException( 8043 , "already sharded" );
log() << "enable sharding on: " << ns << " with shard key: " << fieldsAndOrder << endl;
_sharded[ns] = CollectionInfo( fieldsAndOrder , unique );
info = new ChunkManager( this , ns , fieldsAndOrder , unique );
_shards[ns] = info;
return info;
}
Status ClusterAggregate::runAggregate(OperationContext* txn,
const Namespaces& namespaces,
BSONObj cmdObj,
int options,
BSONObjBuilder* result) {
auto scopedShardDbStatus = ScopedShardDatabase::getExisting(txn, namespaces.executionNss.db());
if (!scopedShardDbStatus.isOK()) {
appendEmptyResultSet(
*result, scopedShardDbStatus.getStatus(), namespaces.requestedNss.ns());
return Status::OK();
}
auto request = AggregationRequest::parseFromBSON(namespaces.executionNss, cmdObj);
if (!request.isOK()) {
return request.getStatus();
}
const auto conf = scopedShardDbStatus.getValue().db();
// Determine the appropriate collation and 'resolve' involved namespaces to make the
// ExpressionContext.
// We won't try to execute anything on a mongos, but we still have to populate this map so that
// any $lookups, etc. will be able to have a resolved view definition. It's okay that this is
// incorrect, we will repopulate the real resolved namespace map on the mongod. Note that we
// need to check if any involved collections are sharded before forwarding an aggregation
// command on an unsharded collection.
StringMap<ExpressionContext::ResolvedNamespace> resolvedNamespaces;
LiteParsedPipeline liteParsedPipeline(request.getValue());
for (auto&& ns : liteParsedPipeline.getInvolvedNamespaces()) {
uassert(28769, str::stream() << ns.ns() << " cannot be sharded", !conf->isSharded(ns.ns()));
resolvedNamespaces[ns.coll()] = {ns, std::vector<BSONObj>{}};
}
if (!conf->isSharded(namespaces.executionNss.ns())) {
return aggPassthrough(txn, namespaces, conf, cmdObj, result, options);
}
auto chunkMgr = conf->getChunkManager(txn, namespaces.executionNss.ns());
std::unique_ptr<CollatorInterface> collation;
if (!request.getValue().getCollation().isEmpty()) {
collation = uassertStatusOK(CollatorFactoryInterface::get(txn->getServiceContext())
->makeFromBSON(request.getValue().getCollation()));
} else if (chunkMgr->getDefaultCollator()) {
collation = chunkMgr->getDefaultCollator()->clone();
}
boost::intrusive_ptr<ExpressionContext> mergeCtx = new ExpressionContext(
txn, request.getValue(), std::move(collation), std::move(resolvedNamespaces));
mergeCtx->inRouter = true;
// explicitly *not* setting mergeCtx->tempDir
// Parse and optimize the pipeline specification.
auto pipeline = Pipeline::parse(request.getValue().getPipeline(), mergeCtx);
if (!pipeline.isOK()) {
return pipeline.getStatus();
}
pipeline.getValue()->optimizePipeline();
// If the first $match stage is an exact match on the shard key (with a simple collation or
// no string matching), we only have to send it to one shard, so send the command to that
// shard.
BSONObj firstMatchQuery = pipeline.getValue()->getInitialQuery();
BSONObj shardKeyMatches;
shardKeyMatches = uassertStatusOK(
chunkMgr->getShardKeyPattern().extractShardKeyFromQuery(txn, firstMatchQuery));
bool singleShard = false;
if (!shardKeyMatches.isEmpty()) {
auto chunk = chunkMgr->findIntersectingChunk(
txn, shardKeyMatches, request.getValue().getCollation());
if (chunk.isOK()) {
singleShard = true;
}
}
// Don't need to split pipeline if the first $match is an exact match on shard key, unless
// there is a stage that needs to be run on the primary shard.
const bool needPrimaryShardMerger = pipeline.getValue()->needsPrimaryShardMerger();
const bool needSplit = !singleShard || needPrimaryShardMerger;
// Split the pipeline into pieces for mongod(s) and this mongos. If needSplit is true,
// 'pipeline' will become the merger side.
boost::intrusive_ptr<Pipeline> shardPipeline(needSplit ? pipeline.getValue()->splitForSharded()
: pipeline.getValue());
// Create the command for the shards. The 'fromRouter' field means produce output to be
// merged.
MutableDocument commandBuilder(request.getValue().serializeToCommandObj());
commandBuilder[AggregationRequest::kPipelineName] = Value(shardPipeline->serialize());
if (needSplit) {
commandBuilder[AggregationRequest::kFromRouterName] = Value(true);
commandBuilder[AggregationRequest::kCursorName] =
Value(DOC(AggregationRequest::kBatchSizeName << 0));
}
// These fields are not part of the AggregationRequest since they are not handled by the
// aggregation subsystem, so we serialize them separately.
const std::initializer_list<StringData> fieldsToPropagateToShards = {
"$queryOptions", "readConcern", QueryRequest::cmdOptionMaxTimeMS,
};
for (auto&& field : fieldsToPropagateToShards) {
commandBuilder[field] = Value(cmdObj[field]);
}
BSONObj shardedCommand = commandBuilder.freeze().toBson();
BSONObj shardQuery = shardPipeline->getInitialQuery();
// Run the command on the shards
// TODO need to make sure cursors are killed if a retry is needed
std::vector<Strategy::CommandResult> shardResults;
Strategy::commandOp(txn,
namespaces.executionNss.db().toString(),
shardedCommand,
options,
namespaces.executionNss.ns(),
shardQuery,
request.getValue().getCollation(),
&shardResults);
if (mergeCtx->isExplain) {
// This must be checked before we start modifying result.
uassertAllShardsSupportExplain(shardResults);
if (needSplit) {
*result << "needsPrimaryShardMerger" << needPrimaryShardMerger << "splitPipeline"
<< DOC("shardsPart" << shardPipeline->writeExplainOps() << "mergerPart"
<< pipeline.getValue()->writeExplainOps());
} else {
*result << "splitPipeline" << BSONNULL;
}
BSONObjBuilder shardExplains(result->subobjStart("shards"));
for (size_t i = 0; i < shardResults.size(); i++) {
shardExplains.append(shardResults[i].shardTargetId,
BSON("host" << shardResults[i].target.toString() << "stages"
<< shardResults[i].result["stages"]));
}
return Status::OK();
}
if (!needSplit) {
invariant(shardResults.size() == 1);
invariant(shardResults[0].target.getServers().size() == 1);
auto executorPool = Grid::get(txn)->getExecutorPool();
const BSONObj reply =
uassertStatusOK(storePossibleCursor(shardResults[0].target.getServers()[0],
shardResults[0].result,
namespaces.requestedNss,
executorPool->getArbitraryExecutor(),
Grid::get(txn)->getCursorManager()));
result->appendElements(reply);
return getStatusFromCommandResult(reply);
}
pipeline.getValue()->addInitialSource(
DocumentSourceMergeCursors::create(parseCursors(shardResults), mergeCtx));
MutableDocument mergeCmd(request.getValue().serializeToCommandObj());
mergeCmd["pipeline"] = Value(pipeline.getValue()->serialize());
mergeCmd["cursor"] = Value(cmdObj["cursor"]);
if (cmdObj.hasField("$queryOptions")) {
mergeCmd["$queryOptions"] = Value(cmdObj["$queryOptions"]);
}
if (cmdObj.hasField(QueryRequest::cmdOptionMaxTimeMS)) {
mergeCmd[QueryRequest::cmdOptionMaxTimeMS] =
Value(cmdObj[QueryRequest::cmdOptionMaxTimeMS]);
}
mergeCmd.setField("writeConcern", Value(cmdObj["writeConcern"]));
mergeCmd.setField("readConcern", Value(cmdObj["readConcern"]));
// If the user didn't specify a collation already, make sure there's a collation attached to
// the merge command, since the merging shard may not have the collection metadata.
if (mergeCmd.peek()["collation"].missing()) {
mergeCmd.setField("collation",
mergeCtx->getCollator()
? Value(mergeCtx->getCollator()->getSpec().toBSON())
: Value(Document{CollationSpec::kSimpleSpec}));
}
std::string outputNsOrEmpty;
if (DocumentSourceOut* out =
dynamic_cast<DocumentSourceOut*>(pipeline.getValue()->getSources().back().get())) {
outputNsOrEmpty = out->getOutputNs().ns();
}
// Run merging command on random shard, unless a stage needs the primary shard. Need to use
// ShardConnection so that the merging mongod is sent the config servers on connection init.
auto& prng = txn->getClient()->getPrng();
const auto& mergingShardId =
(needPrimaryShardMerger || internalQueryAlwaysMergeOnPrimaryShard.load())
? conf->getPrimaryId()
: shardResults[prng.nextInt32(shardResults.size())].shardTargetId;
const auto mergingShard =
uassertStatusOK(Grid::get(txn)->shardRegistry()->getShard(txn, mergingShardId));
ShardConnection conn(mergingShard->getConnString(), outputNsOrEmpty);
BSONObj mergedResults =
aggRunCommand(txn, conn.get(), namespaces, mergeCmd.freeze().toBson(), options);
conn.done();
if (auto wcErrorElem = mergedResults["writeConcernError"]) {
appendWriteConcernErrorToCmdResponse(mergingShardId, wcErrorElem, *result);
}
// Copy output from merging (primary) shard to the output object from our command.
// Also, propagates errmsg and code if ok == false.
result->appendElementsUnique(mergedResults);
return getStatusFromCommandResult(result->asTempObj());
}
void CollectionShardingState::checkShardVersionOrThrow(OperationContext* opCtx) {
const auto optReceivedShardVersion = getOperationReceivedVersion(opCtx, _nss);
if (!optReceivedShardVersion)
return;
const auto& receivedShardVersion = *optReceivedShardVersion;
if (ChunkVersion::isIgnoredVersion(receivedShardVersion)) {
return;
}
// An operation with read concern 'available' should never have shardVersion set.
invariant(repl::ReadConcernArgs::get(opCtx).getLevel() !=
repl::ReadConcernLevel::kAvailableReadConcern);
const auto metadata = getCurrentMetadata();
const auto wantedShardVersion =
metadata->isSharded() ? metadata->getShardVersion() : ChunkVersion::UNSHARDED();
auto criticalSectionSignal = [&] {
auto csrLock = CSRLock::lock(opCtx, this);
return _critSec.getSignal(opCtx->lockState()->isWriteLocked()
? ShardingMigrationCriticalSection::kWrite
: ShardingMigrationCriticalSection::kRead);
}();
if (criticalSectionSignal) {
// Set migration critical section on operation sharding state: operation will wait for the
// migration to finish before returning failure and retrying.
auto& oss = OperationShardingState::get(opCtx);
oss.setMigrationCriticalSectionSignal(criticalSectionSignal);
uasserted(StaleConfigInfo(_nss, receivedShardVersion, wantedShardVersion),
str::stream() << "migration commit in progress for " << _nss.ns());
}
if (receivedShardVersion.isWriteCompatibleWith(wantedShardVersion)) {
return;
}
//
// Figure out exactly why not compatible, send appropriate error message
// The versions themselves are returned in the error, so not needed in messages here
//
StaleConfigInfo sci(_nss, receivedShardVersion, wantedShardVersion);
uassert(std::move(sci),
str::stream() << "epoch mismatch detected for " << _nss.ns() << ", "
<< "the collection may have been dropped and recreated",
wantedShardVersion.epoch() == receivedShardVersion.epoch());
if (!wantedShardVersion.isSet() && receivedShardVersion.isSet()) {
uasserted(std::move(sci),
str::stream() << "this shard no longer contains chunks for " << _nss.ns() << ", "
<< "the collection may have been dropped");
}
if (wantedShardVersion.isSet() && !receivedShardVersion.isSet()) {
uasserted(std::move(sci),
str::stream() << "this shard contains chunks for " << _nss.ns() << ", "
<< "but the client expects unsharded collection");
}
if (wantedShardVersion.majorVersion() != receivedShardVersion.majorVersion()) {
// Could be > or < - wanted is > if this is the source of a migration, wanted < if this is
// the target of a migration
uasserted(std::move(sci), str::stream() << "version mismatch detected for " << _nss.ns());
}
// Those are all the reasons the versions can mismatch
MONGO_UNREACHABLE;
}
