void ReplicatedMergeTreeBlockOutputStream::commitPart(zkutil::ZooKeeperPtr & zookeeper, MergeTreeData::MutableDataPartPtr & part, const String & block_id)
{
storage.check(part->columns);
assertSessionIsNotExpired(zookeeper);
/// Obtain incremental block number and lock it. The lock holds our intention to add the block to the filesystem.
/// We remove the lock just after renaming the part. In case of exception, block number will be marked as abandoned.
/// Also, make deduplication check. If a duplicate is detected, no nodes are created.
/// Allocate new block number and check for duplicates
bool deduplicate_block = !block_id.empty();
String block_id_path = deduplicate_block ? storage.zookeeper_path + "/blocks/" + block_id : "";
auto block_number_lock = storage.allocateBlockNumber(part->info.partition_id, zookeeper, block_id_path);
if (!block_number_lock)
{
LOG_INFO(log, "Block with ID " << block_id << " already exists; ignoring it.");
part->is_duplicate = true;
last_block_is_duplicate = true;
ProfileEvents::increment(ProfileEvents::DuplicatedInsertedBlocks);
return;
}
Int64 block_number = block_number_lock->getNumber();
/// Set part attributes according to part_number. Prepare an entry for log.
part->info.min_block = block_number;
part->info.max_block = block_number;
part->info.level = 0;
String part_name = part->getNewName(part->info);
part->name = part_name;
StorageReplicatedMergeTree::LogEntry log_entry;
log_entry.type = StorageReplicatedMergeTree::LogEntry::GET_PART;
log_entry.create_time = time(nullptr);
log_entry.source_replica = storage.replica_name;
log_entry.new_part_name = part_name;
log_entry.quorum = quorum;
log_entry.block_id = block_id;
/// Simultaneously add information about the part to all the necessary places in ZooKeeper and remove block_number_lock.
/// Information about the part.
Coordination::Requests ops;
storage.getCommitPartOps(ops, part, block_id_path);
/// Replication log.
ops.emplace_back(zkutil::makeCreateRequest(
storage.zookeeper_path + "/log/log-",
log_entry.toString(),
zkutil::CreateMode::PersistentSequential));
/// Deletes the information that the block number is used for writing.
block_number_lock->getUnlockOps(ops);
/** If you need a quorum - create a node in which the quorum is monitored.
* (If such a node already exists, then someone has managed to make another quorum record at the same time, but for it the quorum has not yet been reached.
* You can not do the next quorum record at this time.)
*/
if (quorum)
{
ReplicatedMergeTreeQuorumEntry quorum_entry;
quorum_entry.part_name = part_name;
quorum_entry.required_number_of_replicas = quorum;
quorum_entry.replicas.insert(storage.replica_name);
/** At this point, this node will contain information that the current replica received a part.
* When other replicas will receive this part (in the usual way, processing the replication log),
* they will add themselves to the contents of this node.
* When it contains information about `quorum` number of replicas, this node is deleted,
* which indicates that the quorum has been reached.
*/
ops.emplace_back(
zkutil::makeCreateRequest(
quorum_info.status_path,
quorum_entry.toString(),
zkutil::CreateMode::Persistent));
/// Make sure that during the insertion time, the replica was not reinitialized or disabled (when the server is finished).
ops.emplace_back(
zkutil::makeCheckRequest(
storage.replica_path + "/is_active",
quorum_info.is_active_node_version));
/// Unfortunately, just checking the above is not enough, because `is_active` node can be deleted and reappear with the same version.
/// But then the `host` value will change. We will check this.
/// It's great that these two nodes change in the same transaction (see MergeTreeRestartingThread).
ops.emplace_back(
zkutil::makeCheckRequest(
storage.replica_path + "/host",
quorum_info.host_node_version));
}
MergeTreeData::Transaction transaction(storage.data); /// If you can not add a part to ZK, we'll remove it back from the working set.
storage.data.renameTempPartAndAdd(part, nullptr, &transaction);
Coordination::Responses responses;
int32_t multi_code = zookeeper->tryMultiNoThrow(ops, responses); /// 1 RTT
if (multi_code == Coordination::ZOK)
{
transaction.commit();
storage.merge_selecting_task->schedule();
/// Lock nodes have been already deleted, do not delete them in destructor
block_number_lock->assumeUnlocked();
}
else if (multi_code == Coordination::ZCONNECTIONLOSS
|| multi_code == Coordination::ZOPERATIONTIMEOUT)
{
/** If the connection is lost, and we do not know if the changes were applied, we can not delete the local part
* if the changes were applied, the inserted block appeared in `/blocks/`, and it can not be inserted again.
*/
transaction.commit();
storage.enqueuePartForCheck(part->name, MAX_AGE_OF_LOCAL_PART_THAT_WASNT_ADDED_TO_ZOOKEEPER);
/// We do not know whether or not data has been inserted.
throw Exception("Unknown status, client must retry. Reason: " + String(Coordination::errorMessage(multi_code)),
ErrorCodes::UNKNOWN_STATUS_OF_INSERT);
}
else if (Coordination::isUserError(multi_code))
{
String failed_op_path = zkutil::KeeperMultiException(multi_code, ops, responses).getPathForFirstFailedOp();
if (multi_code == Coordination::ZNODEEXISTS && deduplicate_block && failed_op_path == block_id_path)
{
/// Block with the same id have just appeared in table (or other replica), rollback thee insertion.
LOG_INFO(log, "Block with ID " << block_id << " already exists; ignoring it (removing part " << part->name << ")");
part->is_duplicate = true;
transaction.rollback();
last_block_is_duplicate = true;
ProfileEvents::increment(ProfileEvents::DuplicatedInsertedBlocks);
}
else if (multi_code == Coordination::ZNODEEXISTS && failed_op_path == quorum_info.status_path)
{
transaction.rollback();
throw Exception("Another quorum insert has been already started", ErrorCodes::UNSATISFIED_QUORUM_FOR_PREVIOUS_WRITE);
}
else
{
/// NOTE: We could be here if the node with the quorum existed, but was quickly removed.
transaction.rollback();
throw Exception("Unexpected logical error while adding block " + toString(block_number) + " with ID '" + block_id + "': "
+ zkutil::ZooKeeper::error2string(multi_code) + ", path " + failed_op_path,
ErrorCodes::UNEXPECTED_ZOOKEEPER_ERROR);
}
}
else if (Coordination::isHardwareError(multi_code))
{
transaction.rollback();
throw Exception("Unrecoverable network error while adding block " + toString(block_number) + " with ID '" + block_id + "': "
+ zkutil::ZooKeeper::error2string(multi_code), ErrorCodes::UNEXPECTED_ZOOKEEPER_ERROR);
}
else
{
transaction.rollback();
throw Exception("Unexpected ZooKeeper error while adding block " + toString(block_number) + " with ID '" + block_id + "': "
+ zkutil::ZooKeeper::error2string(multi_code), ErrorCodes::UNEXPECTED_ZOOKEEPER_ERROR);
}
if (quorum)
{
/// We are waiting for quorum to be satisfied.
LOG_TRACE(log, "Waiting for quorum");
String quorum_status_path = storage.zookeeper_path + "/quorum/status";
try
{
while (true)
{
zkutil::EventPtr event = std::make_shared<Poco::Event>();
std::string value;
/// `get` instead of `exists` so that `watch` does not leak if the node is no longer there.
if (!zookeeper->tryGet(quorum_status_path, value, nullptr, event))
break;
ReplicatedMergeTreeQuorumEntry quorum_entry(value);
/// If the node has time to disappear, and then appear again for the next insert.
if (quorum_entry.part_name != part_name)
break;
if (!event->tryWait(quorum_timeout_ms))
throw Exception("Timeout while waiting for quorum", ErrorCodes::TIMEOUT_EXCEEDED);
}
/// And what if it is possible that the current replica at this time has ceased to be active and the quorum is marked as failed and deleted?
String value;
if (!zookeeper->tryGet(storage.replica_path + "/is_active", value, nullptr)
|| value != quorum_info.is_active_node_value)
throw Exception("Replica become inactive while waiting for quorum", ErrorCodes::NO_ACTIVE_REPLICAS);
}
catch (...)
{
/// We do not know whether or not data has been inserted
/// - whether other replicas have time to download the part and mark the quorum as done.
throw Exception("Unknown status, client must retry. Reason: " + getCurrentExceptionMessage(false),
ErrorCodes::UNKNOWN_STATUS_OF_INSERT);
}
LOG_TRACE(log, "Quorum satisfied");
}
}
void ReplicatedMergeTreeBlockOutputStream::write(const Block & block)
{
/// TODO Can I not lock the table structure here?
storage.data.delayInsertIfNeeded(&storage.restarting_thread->getWakeupEvent());
auto zookeeper = storage.getZooKeeper();
assertSessionIsNotExpired(zookeeper);
/** If write is with quorum, then we check that the required number of replicas is now live,
* and also that for all previous pieces for which quorum is required, this quorum is reached.
* And also check that during the insertion, the replica was not reinitialized or disabled (by the value of `is_active` node).
* TODO Too complex logic, you can do better.
*/
String quorum_status_path = storage.zookeeper_path + "/quorum/status";
String is_active_node_value;
int is_active_node_version = -1;
int host_node_version = -1;
if (quorum)
{
zkutil::ZooKeeper::TryGetFuture quorum_status_future = zookeeper->asyncTryGet(quorum_status_path);
zkutil::ZooKeeper::TryGetFuture is_active_future = zookeeper->asyncTryGet(storage.replica_path + "/is_active");
zkutil::ZooKeeper::TryGetFuture host_future = zookeeper->asyncTryGet(storage.replica_path + "/host");
/// List of live replicas. All of them register an ephemeral node for leader_election.
zkutil::Stat leader_election_stat;
zookeeper->get(storage.zookeeper_path + "/leader_election", &leader_election_stat);
if (leader_election_stat.numChildren < static_cast<int32_t>(quorum))
throw Exception("Number of alive replicas ("
+ toString(leader_election_stat.numChildren) + ") is less than requested quorum (" + toString(quorum) + ").",
ErrorCodes::TOO_LESS_LIVE_REPLICAS);
/** Is there a quorum for the last piece for which a quorum is needed?
* Write of all the pieces with the included quorum is linearly ordered.
* This means that at any time there can be only one piece,
* for which you need, but not yet reach the quorum.
* Information about this piece will be located in `/quorum/status` node.
* If the quorum is reached, then the node is deleted.
*/
auto quorum_status = quorum_status_future.get();
if (quorum_status.exists)
throw Exception("Quorum for previous write has not been satisfied yet. Status: " + quorum_status.value, ErrorCodes::UNSATISFIED_QUORUM_FOR_PREVIOUS_WRITE);
/// Both checks are implicitly made also later (otherwise there would be a race condition).
auto is_active = is_active_future.get();
auto host = host_future.get();
if (!is_active.exists || !host.exists)
throw Exception("Replica is not active right now", ErrorCodes::READONLY);
is_active_node_value = is_active.value;
is_active_node_version = is_active.stat.version;
host_node_version = host.stat.version;
}
auto part_blocks = storage.writer.splitBlockIntoParts(block);
for (auto & current_block : part_blocks)
{
assertSessionIsNotExpired(zookeeper);
++block_index;
String block_id = insert_id.empty() ? "" : insert_id + "__" + toString(block_index);
String month_name = toString(DateLUT::instance().toNumYYYYMMDD(DayNum_t(current_block.min_date)) / 100);
AbandonableLockInZooKeeper block_number_lock = storage.allocateBlockNumber(month_name); /// 2 RTT
Int64 part_number = block_number_lock.getNumber();
MergeTreeData::MutableDataPartPtr part = storage.writer.writeTempPart(current_block, part_number);
String part_name = ActiveDataPartSet::getPartName(part->left_date, part->right_date, part->left, part->right, part->level);
/// Hash from the data.
SipHash hash;
part->checksums.summaryDataChecksum(hash);
union
{
char bytes[16];
UInt64 words[2];
} hash_value;
hash.get128(hash_value.bytes);
String checksum(hash_value.bytes, 16);
/// If no ID is specified in query, we take the hash from the data as ID. That is, do not insert the same data twice.
/// NOTE: If you do not need this deduplication, you can leave `block_id` empty instead.
/// Setting or syntax in the query (for example, `ID = null`) could be done for this.
if (block_id.empty())
{
block_id = toString(hash_value.words[0]) + "_" + toString(hash_value.words[1]);
if (block_id.empty())
throw Exception("Logical error: block_id is empty.", ErrorCodes::LOGICAL_ERROR);
}
LOG_DEBUG(log, "Wrote block " << part_number << " with ID " << block_id << ", " << current_block.block.rows() << " rows");
StorageReplicatedMergeTree::LogEntry log_entry;
log_entry.type = StorageReplicatedMergeTree::LogEntry::GET_PART;
log_entry.create_time = time(0);
log_entry.source_replica = storage.replica_name;
log_entry.new_part_name = part_name;
log_entry.quorum = quorum;
log_entry.block_id = block_id;
/// Simultaneously add information about the part to all the necessary places in ZooKeeper and remove block_number_lock.
/// Information about the block.
zkutil::Ops ops;
auto acl = zookeeper->getDefaultACL();
ops.emplace_back(
std::make_unique<zkutil::Op::Create>(
storage.zookeeper_path + "/blocks/" + block_id,
"",
acl,
zkutil::CreateMode::Persistent));
ops.emplace_back(
std::make_unique<zkutil::Op::Create>(
storage.zookeeper_path + "/blocks/" + block_id + "/checksum",
checksum,
acl,
zkutil::CreateMode::Persistent));
ops.emplace_back(
std::make_unique<zkutil::Op::Create>(
storage.zookeeper_path + "/blocks/" + block_id + "/number",
toString(part_number),
acl,
zkutil::CreateMode::Persistent));
/// Information about the part, in the replica data.
storage.addNewPartToZooKeeper(part, ops, part_name);
/// Replication log.
ops.emplace_back(std::make_unique<zkutil::Op::Create>(
storage.zookeeper_path + "/log/log-",
log_entry.toString(),
acl,
zkutil::CreateMode::PersistentSequential));
/// Deletes the information that the block number is used for writing.
block_number_lock.getUnlockOps(ops);
/** If you need a quorum - create a node in which the quorum is monitored.
* (If such a node already exists, then someone has managed to make another quorum record at the same time, but for it the quorum has not yet been reached.
* You can not do the next quorum record at this time.)
*/
if (quorum)
{
ReplicatedMergeTreeQuorumEntry quorum_entry;
quorum_entry.part_name = part_name;
quorum_entry.required_number_of_replicas = quorum;
quorum_entry.replicas.insert(storage.replica_name);
/** At this point, this node will contain information that the current replica received a piece.
* When other replicas will receive this piece (in the usual way, processing the replication log),
* they will add themselves to the contents of this node.
* When it contains information about `quorum` number of replicas, this node is deleted,
* which indicates that the quorum has been reached.
*/
ops.emplace_back(
std::make_unique<zkutil::Op::Create>(
quorum_status_path,
quorum_entry.toString(),
acl,
zkutil::CreateMode::Persistent));
/// Make sure that during the insertion time, the replica was not reinitialized or disabled (when the server is finished).
ops.emplace_back(
std::make_unique<zkutil::Op::Check>(
storage.replica_path + "/is_active",
is_active_node_version));
/// Unfortunately, just checking the above is not enough, because `is_active` node can be deleted and reappear with the same version.
/// But then the `host` value will change. We will check this.
/// It's great that these two nodes change in the same transaction (see MergeTreeRestartingThread).
ops.emplace_back(
std::make_unique<zkutil::Op::Check>(
storage.replica_path + "/host",
host_node_version));
}
MergeTreeData::Transaction transaction; /// If you can not add a piece to ZK, we'll remove it again from the working set.
storage.data.renameTempPartAndAdd(part, nullptr, &transaction);
try
{
auto code = zookeeper->tryMulti(ops);
if (code == ZOK)
{
transaction.commit();
storage.merge_selecting_event.set();
}
else if (code == ZNODEEXISTS)
{
/// If the block with such ID already exists in the table, roll back its insertion.
String expected_checksum;
if (!block_id.empty() && zookeeper->tryGet(
storage.zookeeper_path + "/blocks/" + block_id + "/checksum", expected_checksum))
{
LOG_INFO(log, "Block with ID " << block_id << " already exists; ignoring it (removing part " << part->name << ")");
/// If the data is different from the ones that were inserted earlier with the same ID, throw an exception.
if (expected_checksum != checksum)
{
if (!insert_id.empty())
throw Exception("Attempt to insert block with same ID but different checksum", ErrorCodes::CHECKSUM_DOESNT_MATCH);
else
throw Exception("Logical error: got ZNODEEXISTS while inserting data, block ID is derived from checksum but checksum doesn't match", ErrorCodes::LOGICAL_ERROR);
}
transaction.rollback();
}
else if (zookeeper->exists(quorum_status_path))
{
transaction.rollback();
throw Exception("Another quorum insert has been already started", ErrorCodes::UNSATISFIED_QUORUM_FOR_PREVIOUS_WRITE);
}
else
{
/// if the node with the quorum existed, but was quickly removed.
throw Exception("Unexpected ZNODEEXISTS while adding block " + toString(part_number) + " with ID " + block_id + ": "
+ zkutil::ZooKeeper::error2string(code), ErrorCodes::UNEXPECTED_ZOOKEEPER_ERROR);
}
}
else
{
throw Exception("Unexpected error while adding block " + toString(part_number) + " with ID " + block_id + ": "
+ zkutil::ZooKeeper::error2string(code), ErrorCodes::UNEXPECTED_ZOOKEEPER_ERROR);
}
}
catch (const zkutil::KeeperException & e)
{
/** If the connection is lost, and we do not know if the changes were applied, you can not delete the local chunk
* if the changes were applied, the inserted block appeared in `/blocks/`, and it can not be inserted again.
*/
if (e.code == ZOPERATIONTIMEOUT ||
e.code == ZCONNECTIONLOSS)
{
transaction.commit();
storage.enqueuePartForCheck(part->name, MAX_AGE_OF_LOCAL_PART_THAT_WASNT_ADDED_TO_ZOOKEEPER);
/// We do not know whether or not data has been inserted.
throw Exception("Unknown status, client must retry. Reason: " + e.displayText(), ErrorCodes::UNKNOWN_STATUS_OF_INSERT);
}
throw;
}
if (quorum)
{
/// We are waiting for the quorum to be reached.
LOG_TRACE(log, "Waiting for quorum");
try
{
while (true)
{
zkutil::EventPtr event = std::make_shared<Poco::Event>();
std::string value;
/// `get` instead of `exists` so that `watch` does not leak if the node is no longer there.
if (!zookeeper->tryGet(quorum_status_path, value, nullptr, event))
break;
ReplicatedMergeTreeQuorumEntry quorum_entry(value);
/// If the node has time to disappear, and then appear again for the next insert.
if (quorum_entry.part_name != part_name)
break;
if (!event->tryWait(quorum_timeout_ms))
throw Exception("Timeout while waiting for quorum");
}
/// And what if it is possible that the current replica at this time has ceased to be active and the quorum is marked as failed and deleted?
String value;
if (!zookeeper->tryGet(storage.replica_path + "/is_active", value, nullptr)
|| value != is_active_node_value)
throw Exception("Replica become inactive while waiting for quorum");
}
catch (...)
{
/// We do not know whether or not data has been inserted
/// - whether other replicas have time to download the part and mark the quorum as done.
throw Exception("Unknown status, client must retry. Reason: " + getCurrentExceptionMessage(false),
ErrorCodes::UNKNOWN_STATUS_OF_INSERT);
}
LOG_TRACE(log, "Quorum satisfied");
}
}
}