void Balancer::_doBalanceRound(DBClientBase& conn, vector<CandidateChunkPtr>* candidateChunks) {
verify(candidateChunks);
//
// 1. Check whether there is any sharded collection to be balanced by querying
// the ShardsNS::collections collection
//
auto_ptr<DBClientCursor> cursor = conn.query(CollectionType::ConfigNS, BSONObj());
if (NULL == cursor.get()) {
warning() << "could not query " << CollectionType::ConfigNS << " while trying to balance"
<< endl;
return;
}
vector<string> collections;
while (cursor->more()) {
BSONObj col = cursor->nextSafe();
// sharded collections will have a shard "key".
if (!col[CollectionType::keyPattern()].eoo() &&
!col[CollectionType::noBalance()].trueValue()) {
collections.push_back(col[CollectionType::ns()].String());
} else if (col[CollectionType::noBalance()].trueValue()) {
LOG(1) << "not balancing collection " << col[CollectionType::ns()].String()
<< ", explicitly disabled" << endl;
}
}
cursor.reset();
if (collections.empty()) {
LOG(1) << "no collections to balance" << endl;
return;
}
//
// 2. Get a list of all the shards that are participating in this balance round
// along with any maximum allowed quotas and current utilization. We get the
// latter by issuing db.serverStatus() (mem.mapped) to all shards.
//
// TODO: skip unresponsive shards and mark information as stale.
//
ShardInfoMap shardInfo;
Status loadStatus = DistributionStatus::populateShardInfoMap(&shardInfo);
if (!loadStatus.isOK()) {
warning() << "failed to load shard metadata" << causedBy(loadStatus);
return;
}
if (shardInfo.size() < 2) {
LOG(1) << "can't balance without more active shards";
return;
}
OCCASIONALLY warnOnMultiVersion(shardInfo);
//
// 3. For each collection, check if the balancing policy recommends moving anything around.
//
for (vector<string>::const_iterator it = collections.begin(); it != collections.end(); ++it) {
const string& ns = *it;
OwnedPointerMap<string, OwnedPointerVector<ChunkType>> shardToChunksMap;
cursor = conn.query(ChunkType::ConfigNS, QUERY(ChunkType::ns(ns)).sort(ChunkType::min()));
set<BSONObj> allChunkMinimums;
while (cursor->more()) {
BSONObj chunkDoc = cursor->nextSafe().getOwned();
auto_ptr<ChunkType> chunk(new ChunkType());
string errmsg;
if (!chunk->parseBSON(chunkDoc, &errmsg)) {
error() << "bad chunk format for " << chunkDoc << ": " << errmsg << endl;
return;
}
allChunkMinimums.insert(chunk->getMin().getOwned());
OwnedPointerVector<ChunkType>*& chunkList =
shardToChunksMap.mutableMap()[chunk->getShard()];
if (chunkList == NULL) {
chunkList = new OwnedPointerVector<ChunkType>();
}
chunkList->mutableVector().push_back(chunk.release());
}
cursor.reset();
if (shardToChunksMap.map().empty()) {
LOG(1) << "skipping empty collection (" << ns << ")";
continue;
}
for (ShardInfoMap::const_iterator i = shardInfo.begin(); i != shardInfo.end(); ++i) {
// this just makes sure there is an entry in shardToChunksMap for every shard
OwnedPointerVector<ChunkType>*& chunkList = shardToChunksMap.mutableMap()[i->first];
if (chunkList == NULL) {
chunkList = new OwnedPointerVector<ChunkType>();
}
}
DistributionStatus status(shardInfo, shardToChunksMap.map());
// load tags
cursor = conn.query(TagsType::ConfigNS, QUERY(TagsType::ns(ns)).sort(TagsType::min()));
vector<TagRange> ranges;
while (cursor->more()) {
BSONObj tag = cursor->nextSafe();
TagRange tr(tag[TagsType::min()].Obj().getOwned(),
tag[TagsType::max()].Obj().getOwned(),
tag[TagsType::tag()].String());
ranges.push_back(tr);
uassert(
16356, str::stream() << "tag ranges not valid for: " << ns, status.addTagRange(tr));
}
cursor.reset();
DBConfigPtr cfg = grid.getDBConfig(ns);
if (!cfg) {
warning() << "could not load db config to balance " << ns << " collection" << endl;
continue;
}
// This line reloads the chunk manager once if this process doesn't know the collection
// is sharded yet.
ChunkManagerPtr cm = cfg->getChunkManagerIfExists(ns, true);
if (!cm) {
warning() << "could not load chunks to balance " << ns << " collection" << endl;
continue;
}
// loop through tags to make sure no chunk spans tags; splits on tag min. for all chunks
bool didAnySplits = false;
for (unsigned i = 0; i < ranges.size(); i++) {
BSONObj min = ranges[i].min;
min = cm->getShardKeyPattern().getKeyPattern().extendRangeBound(min, false);
if (allChunkMinimums.count(min) > 0)
continue;
didAnySplits = true;
log() << "ns: " << ns << " need to split on " << min
<< " because there is a range there" << endl;
ChunkPtr c = cm->findIntersectingChunk(min);
vector<BSONObj> splitPoints;
splitPoints.push_back(min);
Status status = c->multiSplit(splitPoints, NULL);
if (!status.isOK()) {
error() << "split failed: " << status << endl;
} else {
LOG(1) << "split worked" << endl;
}
break;
}
if (didAnySplits) {
// state change, just wait till next round
continue;
}
CandidateChunk* p = _policy->balance(ns, status, _balancedLastTime);
if (p)
candidateChunks->push_back(CandidateChunkPtr(p));
}
}
Status ClusterAggregate::runAggregate(OperationContext* txn,
const Namespaces& namespaces,
BSONObj cmdObj,
int options,
BSONObjBuilder* result) {
auto dbname = namespaces.executionNss.db().toString();
auto status = grid.catalogCache()->getDatabase(txn, dbname);
if (!status.isOK()) {
appendEmptyResultSet(*result, status.getStatus(), namespaces.requestedNss.ns());
return Status::OK();
}
std::shared_ptr<DBConfig> conf = status.getValue();
if (!conf->isShardingEnabled()) {
return aggPassthrough(txn, namespaces, conf, cmdObj, result, options);
}
auto request = AggregationRequest::parseFromBSON(namespaces.executionNss, cmdObj);
if (!request.isOK()) {
return request.getStatus();
}
boost::intrusive_ptr<ExpressionContext> mergeCtx =
new ExpressionContext(txn, request.getValue());
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();
}
for (auto&& ns : pipeline.getValue()->getInvolvedCollections()) {
uassert(28769, str::stream() << ns.ns() << " cannot be sharded", !conf->isSharded(ns.ns()));
// 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.
// TODO SERVER-25038 This should become unnecessary once we can get the involved
// namespaces before parsing.
mergeCtx->resolvedNamespaces[ns.coll()] = {ns, std::vector<BSONObj>{}};
}
if (!conf->isSharded(namespaces.executionNss.ns())) {
return aggPassthrough(txn, namespaces, conf, cmdObj, result, options);
}
ChunkManagerPtr chunkMgr = conf->getChunkManager(txn, namespaces.executionNss.ns());
// If there was no collation specified, but there is a default collation for the collation,
// use that.
if (request.getValue().getCollation().isEmpty() && chunkMgr->getDefaultCollator()) {
mergeCtx->setCollator(chunkMgr->getDefaultCollator()->clone());
}
// Now that we know the collation we'll be using, inject the ExpressionContext and optimize.
// TODO SERVER-25038: this must happen before we parse the pipeline, since we can make
// string comparisons during parse time.
pipeline.getValue()->injectExpressionContext(mergeCtx);
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,
dbname,
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.getExecutorPool();
const BSONObj reply =
uassertStatusOK(storePossibleCursor(shardResults[0].target.getServers()[0],
shardResults[0].result,
namespaces.requestedNss,
executorPool->getArbitraryExecutor(),
grid.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"]));
// Not propagating readConcern to merger since it doesn't do local reads.
// 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
? conf->getPrimaryId()
: shardResults[prng.nextInt32(shardResults.size())].shardTargetId;
const auto mergingShard = uassertStatusOK(grid.shardRegistry()->getShard(txn, mergingShardId));
ShardConnection conn(mergingShard->getConnString(), outputNsOrEmpty);
BSONObj mergedResults =
aggRunCommand(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());
}
