StatusWith<VersionType> CatalogManagerReplicaSet::_getConfigVersion(OperationContext* txn) {
const auto configShard = grid.shardRegistry()->getShard(txn, "config");
const auto readHostStatus = configShard->getTargeter()->findHost(kConfigReadSelector);
if (!readHostStatus.isOK()) {
return readHostStatus.getStatus();
}
auto readHost = readHostStatus.getValue();
auto findStatus = _exhaustiveFindOnConfig(readHost,
NamespaceString(VersionType::ConfigNS),
BSONObj(),
BSONObj(),
boost::none /* no limit */);
if (!findStatus.isOK()) {
return findStatus.getStatus();
}
auto queryResults = findStatus.getValue().value;
if (queryResults.size() > 1) {
return {ErrorCodes::RemoteValidationError,
str::stream() << "should only have 1 document in " << VersionType::ConfigNS};
}
if (queryResults.empty()) {
auto countStatus =
_runCountCommandOnConfig(readHost, NamespaceString(ShardType::ConfigNS), BSONObj());
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
const auto& shardCount = countStatus.getValue();
if (shardCount > 0) {
// Version document doesn't exist, but config.shards is not empty. Assuming that
// the current config metadata is pre v2.4.
VersionType versionInfo;
versionInfo.setMinCompatibleVersion(UpgradeHistory_UnreportedVersion);
versionInfo.setCurrentVersion(UpgradeHistory_UnreportedVersion);
return versionInfo;
}
VersionType versionInfo;
versionInfo.setMinCompatibleVersion(UpgradeHistory_EmptyVersion);
versionInfo.setCurrentVersion(UpgradeHistory_EmptyVersion);
return versionInfo;
}
BSONObj versionDoc = queryResults.front();
auto versionTypeResult = VersionType::fromBSON(versionDoc);
if (!versionTypeResult.isOK()) {
return Status(ErrorCodes::UnsupportedFormat,
str::stream() << "invalid config version document: " << versionDoc
<< versionTypeResult.getStatus().toString());
}
return versionTypeResult.getValue();
}
StatusWith<ShardDrainingStatus> ShardingCatalogManager::removeShard(OperationContext* opCtx,
const ShardId& shardId) {
// Check preconditions for removing the shard
std::string name = shardId.toString();
auto countStatus = _runCountCommandOnConfig(
opCtx,
ShardType::ConfigNS,
BSON(ShardType::name() << NE << name << ShardType::draining(true)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
if (countStatus.getValue() > 0) {
return Status(ErrorCodes::ConflictingOperationInProgress,
"Can't have more than one draining shard at a time");
}
countStatus =
_runCountCommandOnConfig(opCtx, ShardType::ConfigNS, BSON(ShardType::name() << NE << name));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
if (countStatus.getValue() == 0) {
return Status(ErrorCodes::IllegalOperation, "Can't remove last shard");
}
// Figure out if shard is already draining
countStatus = _runCountCommandOnConfig(
opCtx, ShardType::ConfigNS, BSON(ShardType::name() << name << ShardType::draining(true)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
auto* const shardRegistry = Grid::get(opCtx)->shardRegistry();
if (countStatus.getValue() == 0) {
log() << "going to start draining shard: " << name;
// Record start in changelog
const Status logStatus = Grid::get(opCtx)->catalogClient()->logChangeChecked(
opCtx,
"removeShard.start",
"",
BSON("shard" << name),
ShardingCatalogClient::kLocalWriteConcern);
if (!logStatus.isOK()) {
return logStatus;
}
auto updateStatus = Grid::get(opCtx)->catalogClient()->updateConfigDocument(
opCtx,
ShardType::ConfigNS,
BSON(ShardType::name() << name),
BSON("$set" << BSON(ShardType::draining(true))),
false,
ShardingCatalogClient::kLocalWriteConcern);
if (!updateStatus.isOK()) {
log() << "error starting removeShard: " << name
<< causedBy(redact(updateStatus.getStatus()));
return updateStatus.getStatus();
}
shardRegistry->reload(opCtx);
return ShardDrainingStatus::STARTED;
}
// Draining has already started, now figure out how many chunks and databases are still on the
// shard.
countStatus =
_runCountCommandOnConfig(opCtx, ChunkType::ConfigNS, BSON(ChunkType::shard(name)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
const long long chunkCount = countStatus.getValue();
countStatus =
_runCountCommandOnConfig(opCtx, DatabaseType::ConfigNS, BSON(DatabaseType::primary(name)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
const long long databaseCount = countStatus.getValue();
if (chunkCount > 0 || databaseCount > 0) {
// Still more draining to do
LOG(0) << "chunkCount: " << chunkCount;
LOG(0) << "databaseCount: " << databaseCount;
return ShardDrainingStatus::ONGOING;
}
// Draining is done, now finish removing the shard.
log() << "going to remove shard: " << name;
audit::logRemoveShard(opCtx->getClient(), name);
Status status = Grid::get(opCtx)->catalogClient()->removeConfigDocuments(
opCtx,
ShardType::ConfigNS,
BSON(ShardType::name() << name),
ShardingCatalogClient::kLocalWriteConcern);
if (!status.isOK()) {
log() << "Error concluding removeShard operation on: " << name
<< "; err: " << status.reason();
return status;
}
shardConnectionPool.removeHost(name);
ReplicaSetMonitor::remove(name);
shardRegistry->reload(opCtx);
// Record finish in changelog
Grid::get(opCtx)->catalogClient()->logChange(
opCtx, "removeShard", "", BSON("shard" << name), ShardingCatalogClient::kLocalWriteConcern);
return ShardDrainingStatus::COMPLETED;
}
StatusWith<ShardDrainingStatus> CatalogManagerReplicaSet::removeShard(OperationContext* txn,
const std::string& name) {
// Check preconditions for removing the shard
auto countStatus = _runCountCommandOnConfig(
txn,
NamespaceString(ShardType::ConfigNS),
BSON(ShardType::name() << NE << name << ShardType::draining(true)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
if (countStatus.getValue() > 0) {
return Status(ErrorCodes::ConflictingOperationInProgress,
"Can't have more than one draining shard at a time");
}
countStatus = _runCountCommandOnConfig(
txn, NamespaceString(ShardType::ConfigNS), BSON(ShardType::name() << NE << name));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
if (countStatus.getValue() == 0) {
return Status(ErrorCodes::IllegalOperation, "Can't remove last shard");
}
// Figure out if shard is already draining
countStatus =
_runCountCommandOnConfig(txn,
NamespaceString(ShardType::ConfigNS),
BSON(ShardType::name() << name << ShardType::draining(true)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
if (countStatus.getValue() == 0) {
log() << "going to start draining shard: " << name;
Status status = update(txn,
ShardType::ConfigNS,
BSON(ShardType::name() << name),
BSON("$set" << BSON(ShardType::draining(true))),
false, // upsert
false, // multi
NULL);
if (!status.isOK()) {
log() << "error starting removeShard: " << name << "; err: " << status.reason();
return status;
}
grid.shardRegistry()->reload(txn);
// Record start in changelog
logChange(txn,
txn->getClient()->clientAddress(true),
"removeShard.start",
"",
BSON("shard" << name));
return ShardDrainingStatus::STARTED;
}
// Draining has already started, now figure out how many chunks and databases are still on the
// shard.
countStatus = _runCountCommandOnConfig(
txn, NamespaceString(ChunkType::ConfigNS), BSON(ChunkType::shard(name)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
const long long chunkCount = countStatus.getValue();
countStatus = _runCountCommandOnConfig(
txn, NamespaceString(DatabaseType::ConfigNS), BSON(DatabaseType::primary(name)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
const long long databaseCount = countStatus.getValue();
if (chunkCount > 0 || databaseCount > 0) {
// Still more draining to do
return ShardDrainingStatus::ONGOING;
}
// Draining is done, now finish removing the shard.
log() << "going to remove shard: " << name;
audit::logRemoveShard(txn->getClient(), name);
Status status = remove(txn, ShardType::ConfigNS, BSON(ShardType::name() << name), 0, NULL);
if (!status.isOK()) {
log() << "Error concluding removeShard operation on: " << name
<< "; err: " << status.reason();
return status;
}
grid.shardRegistry()->remove(name);
grid.shardRegistry()->reload(txn);
// Record finish in changelog
logChange(txn, txn->getClient()->clientAddress(true), "removeShard", "", BSON("shard" << name));
return ShardDrainingStatus::COMPLETED;
}
Status CatalogManagerReplicaSet::shardCollection(OperationContext* txn,
const string& ns,
const ShardKeyPattern& fieldsAndOrder,
bool unique,
const vector<BSONObj>& initPoints,
const set<ShardId>& initShardIds) {
// Lock the collection globally so that no other mongos can try to shard or drop the collection
// at the same time.
auto scopedDistLock = getDistLockManager()->lock(ns, "shardCollection");
if (!scopedDistLock.isOK()) {
return scopedDistLock.getStatus();
}
auto status = getDatabase(txn, nsToDatabase(ns));
if (!status.isOK()) {
return status.getStatus();
}
ShardId dbPrimaryShardId = status.getValue().value.getPrimary();
const auto primaryShard = grid.shardRegistry()->getShard(txn, dbPrimaryShardId);
{
// In 3.0 and prior we include this extra safety check that the collection is not getting
// sharded concurrently by two different mongos instances. It is not 100%-proof, but it
// reduces the chance that two invocations of shard collection will step on each other's
// toes. Now we take the distributed lock so going forward this check won't be necessary
// but we leave it around for compatibility with other mongoses from 3.0.
// TODO(spencer): Remove this after 3.2 ships.
auto countStatus = _runCountCommandOnConfig(
txn, NamespaceString(ChunkType::ConfigNS), BSON(ChunkType::ns(ns)));
if (!countStatus.isOK()) {
return countStatus.getStatus();
}
if (countStatus.getValue() > 0) {
return Status(ErrorCodes::AlreadyInitialized,
str::stream() << "collection " << ns << " already sharded with "
<< countStatus.getValue() << " chunks.");
}
}
// Record start in changelog
{
BSONObjBuilder collectionDetail;
collectionDetail.append("shardKey", fieldsAndOrder.toBSON());
collectionDetail.append("collection", ns);
collectionDetail.append("primary", primaryShard->toString());
{
BSONArrayBuilder initialShards(collectionDetail.subarrayStart("initShards"));
for (const ShardId& shardId : initShardIds) {
initialShards.append(shardId);
}
}
collectionDetail.append("numChunks", static_cast<int>(initPoints.size() + 1));
logChange(txn,
txn->getClient()->clientAddress(true),
"shardCollection.start",
ns,
collectionDetail.obj());
}
shared_ptr<ChunkManager> manager(new ChunkManager(ns, fieldsAndOrder, unique));
manager->createFirstChunks(txn, dbPrimaryShardId, &initPoints, &initShardIds);
manager->loadExistingRanges(txn, nullptr);
CollectionInfo collInfo;
collInfo.useChunkManager(manager);
collInfo.save(txn, ns);
manager->reload(txn, true);
// Tell the primary mongod to refresh its data
// TODO: Think the real fix here is for mongos to just
// assume that all collections are sharded, when we get there
SetShardVersionRequest ssv = SetShardVersionRequest::makeForVersioningNoPersist(
grid.shardRegistry()->getConfigServerConnectionString(),
dbPrimaryShardId,
primaryShard->getConnString(),
NamespaceString(ns),
manager->getVersion(),
true);
auto ssvStatus = grid.shardRegistry()->runCommandWithNotMasterRetries(
txn, dbPrimaryShardId, "admin", ssv.toBSON());
if (!ssvStatus.isOK()) {
warning() << "could not update initial version of " << ns << " on shard primary "
<< dbPrimaryShardId << ssvStatus.getStatus();
}
logChange(txn,
txn->getClient()->clientAddress(true),
"shardCollection",
ns,
BSON("version" << manager->getVersion().toString()));
return Status::OK();
}
