ClickHouse/programs/copier/ClusterCopier.cpp

1938 lines
80 KiB
C++
Raw Normal View History

#include "ClusterCopier.h"
2020-02-19 15:01:08 +00:00
#include "Internals.h"
#include <Common/ZooKeeper/ZooKeeper.h>
2018-04-03 17:37:30 +00:00
#include <Common/ZooKeeper/KeeperException.h>
namespace DB
2020-02-25 18:58:28 +00:00
{
2020-02-25 18:10:48 +00:00
namespace ErrorCodes
{
extern const int NOT_IMPLEMENTED;
extern const int LOGICAL_ERROR;
extern const int UNFINISHED;
extern const int BAD_ARGUMENTS;
}
2020-02-25 18:58:28 +00:00
2020-02-19 15:01:08 +00:00
void ClusterCopier::init()
{
2020-02-19 15:01:08 +00:00
auto zookeeper = context.getZooKeeper();
2020-02-19 15:01:08 +00:00
task_description_watch_callback = [this] (const Coordination::WatchResponse & response)
{
if (response.error != Coordination::ZOK)
return;
2020-02-20 17:26:20 +00:00
UInt64 version = ++task_description_version;
2020-02-19 15:01:08 +00:00
LOG_DEBUG(log, "Task description should be updated, local version " << version);
};
2020-02-19 15:01:08 +00:00
task_description_path = task_zookeeper_path + "/description";
task_cluster = std::make_unique<TaskCluster>(task_zookeeper_path, working_database_name);
2020-02-19 15:01:08 +00:00
reloadTaskDescription();
task_cluster_initial_config = task_cluster_current_config;
2020-02-19 15:01:08 +00:00
task_cluster->loadTasks(*task_cluster_initial_config);
context.setClustersConfig(task_cluster_initial_config, task_cluster->clusters_prefix);
/// Set up shards and their priority
task_cluster->random_engine.seed(task_cluster->random_device());
for (auto & task_table : task_cluster->table_tasks)
{
2020-02-19 15:01:08 +00:00
task_table.cluster_pull = context.getCluster(task_table.cluster_pull_name);
task_table.cluster_push = context.getCluster(task_table.cluster_push_name);
task_table.initShards(task_cluster->random_engine);
}
2020-02-19 15:01:08 +00:00
LOG_DEBUG(log, "Will process " << task_cluster->table_tasks.size() << " table tasks");
/// Do not initialize tables, will make deferred initialization in process()
zookeeper->createAncestors(getWorkersPathVersion() + "/");
zookeeper->createAncestors(getWorkersPath() + "/");
}
2020-02-19 15:01:08 +00:00
template <typename T>
decltype(auto) ClusterCopier::retry(T && func, UInt64 max_tries)
{
2020-02-19 15:01:08 +00:00
std::exception_ptr exception;
for (UInt64 try_number = 1; try_number <= max_tries; ++try_number)
{
try
{
return func();
}
catch (...)
{
exception = std::current_exception();
if (try_number < max_tries)
{
tryLogCurrentException(log, "Will retry");
std::this_thread::sleep_for(default_sleep_time);
}
}
}
std::rethrow_exception(exception);
}
2020-02-19 15:01:08 +00:00
void ClusterCopier::discoverShardPartitions(const ConnectionTimeouts & timeouts, const TaskShardPtr & task_shard)
{
2020-02-19 15:01:08 +00:00
TaskTable & task_table = task_shard->task_table;
2020-02-19 15:01:08 +00:00
LOG_INFO(log, "Discover partitions of shard " << task_shard->getDescription());
2020-02-19 15:01:08 +00:00
auto get_partitions = [&] () { return getShardPartitions(timeouts, *task_shard); };
auto existing_partitions_names = retry(get_partitions, 60);
Strings filtered_partitions_names;
Strings missing_partitions;
2020-02-19 15:01:08 +00:00
/// Check that user specified correct partition names
auto check_partition_format = [] (const DataTypePtr & type, const String & partition_text_quoted)
{
2020-02-19 15:01:08 +00:00
MutableColumnPtr column_dummy = type->createColumn();
ReadBufferFromString rb(partition_text_quoted);
2020-02-19 15:01:08 +00:00
try
{
type->deserializeAsTextQuoted(*column_dummy, rb, FormatSettings());
}
catch (Exception & e)
{
throw Exception("Partition " + partition_text_quoted + " has incorrect format. " + e.displayText(), ErrorCodes::BAD_ARGUMENTS);
}
};
2020-02-19 15:01:08 +00:00
if (task_table.has_enabled_partitions)
{
2020-02-19 15:01:08 +00:00
/// Process partition in order specified by <enabled_partitions/>
for (const String & partition_name : task_table.enabled_partitions)
{
/// Check that user specified correct partition names
check_partition_format(task_shard->partition_key_column.type, partition_name);
2020-02-19 15:01:08 +00:00
auto it = existing_partitions_names.find(partition_name);
2020-02-19 15:01:08 +00:00
/// Do not process partition if it is not in enabled_partitions list
if (it == existing_partitions_names.end())
{
missing_partitions.emplace_back(partition_name);
continue;
}
filtered_partitions_names.emplace_back(*it);
}
2020-02-19 15:01:08 +00:00
for (const String & partition_name : existing_partitions_names)
{
if (!task_table.enabled_partitions_set.count(partition_name))
{
LOG_DEBUG(log, "Partition " << partition_name << " will not be processed, since it is not in "
<< "enabled_partitions of " << task_table.table_id);
}
}
}
else
{
for (const String & partition_name : existing_partitions_names)
filtered_partitions_names.emplace_back(partition_name);
}
2020-02-19 15:01:08 +00:00
for (const String & partition_name : filtered_partitions_names)
{
2020-03-12 19:46:48 +00:00
const size_t number_of_splits = task_table.number_of_splits;
task_shard->partition_tasks.emplace(partition_name, ShardPartition(*task_shard, partition_name, number_of_splits));
2020-02-19 15:01:08 +00:00
task_shard->checked_partitions.emplace(partition_name, true);
2020-02-18 13:39:22 +00:00
auto shard_partition_it = task_shard->partition_tasks.find(partition_name);
PartitionPieces & shard_partition_pieces = shard_partition_it->second.pieces;
for (size_t piece_number = 0; piece_number < number_of_splits; ++piece_number)
{
bool res = checkPresentPartitionPiecesOnCurrentShard(timeouts, *task_shard, partition_name, piece_number);
shard_partition_pieces.emplace_back(shard_partition_it->second, piece_number, res);
}
2020-02-19 15:01:08 +00:00
}
2020-02-19 15:01:08 +00:00
if (!missing_partitions.empty())
{
std::stringstream ss;
for (const String & missing_partition : missing_partitions)
ss << " " << missing_partition;
2020-02-19 15:01:08 +00:00
LOG_WARNING(log, "There are no " << missing_partitions.size() << " partitions from enabled_partitions in shard "
<< task_shard->getDescription() << " :" << ss.str());
}
2020-02-19 15:01:08 +00:00
LOG_DEBUG(log, "Will copy " << task_shard->partition_tasks.size() << " partitions from shard " << task_shard->getDescription());
}
2018-02-08 11:07:58 +00:00
2020-02-19 15:01:08 +00:00
void ClusterCopier::discoverTablePartitions(const ConnectionTimeouts & timeouts, TaskTable & task_table, UInt64 num_threads)
{
2020-02-19 15:01:08 +00:00
/// Fetch partitions list from a shard
{
ThreadPool thread_pool(num_threads ? num_threads : 2 * getNumberOfPhysicalCPUCores());
2020-02-19 15:01:08 +00:00
for (const TaskShardPtr & task_shard : task_table.all_shards)
thread_pool.scheduleOrThrowOnError([this, timeouts, task_shard]() { discoverShardPartitions(timeouts, task_shard); });
2020-02-19 15:01:08 +00:00
LOG_DEBUG(log, "Waiting for " << thread_pool.active() << " setup jobs");
thread_pool.wait();
}
}
2020-02-19 15:01:08 +00:00
void ClusterCopier::uploadTaskDescription(const std::string & task_path, const std::string & task_file, const bool force)
{
2020-02-19 15:01:08 +00:00
auto local_task_description_path = task_path + "/description";
2020-02-19 15:01:08 +00:00
String task_config_str;
{
2020-02-19 15:01:08 +00:00
ReadBufferFromFile in(task_file);
readStringUntilEOF(task_config_str, in);
}
2020-02-19 15:01:08 +00:00
if (task_config_str.empty())
return;
2020-02-19 15:01:08 +00:00
auto zookeeper = context.getZooKeeper();
2020-02-19 15:01:08 +00:00
zookeeper->createAncestors(local_task_description_path);
auto code = zookeeper->tryCreate(local_task_description_path, task_config_str, zkutil::CreateMode::Persistent);
if (code && force)
zookeeper->createOrUpdate(local_task_description_path, task_config_str, zkutil::CreateMode::Persistent);
2020-02-19 15:01:08 +00:00
LOG_DEBUG(log, "Task description " << ((code && !force) ? "not " : "") << "uploaded to " << local_task_description_path << " with result " << code << " ("<< zookeeper->error2string(code) << ")");
}
2020-02-19 15:01:08 +00:00
void ClusterCopier::reloadTaskDescription()
{
auto zookeeper = context.getZooKeeper();
task_description_watch_zookeeper = zookeeper;
2020-02-19 15:01:08 +00:00
String task_config_str;
2020-02-18 13:39:22 +00:00
Coordination::Stat stat{};
2020-02-19 15:01:08 +00:00
int code;
2020-02-19 15:01:08 +00:00
zookeeper->tryGetWatch(task_description_path, task_config_str, &stat, task_description_watch_callback, &code);
if (code)
throw Exception("Can't get description node " + task_description_path, ErrorCodes::BAD_ARGUMENTS);
2020-02-18 13:39:22 +00:00
LOG_DEBUG(log, "Loading description, zxid=" << task_description_current_stat.czxid);
2020-02-19 15:01:08 +00:00
auto config = getConfigurationFromXMLString(task_config_str);
2020-02-19 15:01:08 +00:00
/// Setup settings
task_cluster->reloadSettings(*config);
context.setSettings(task_cluster->settings_common);
2020-02-19 15:01:08 +00:00
task_cluster_current_config = config;
2020-02-18 13:39:22 +00:00
task_description_current_stat = stat;
2020-02-19 15:01:08 +00:00
}
2020-02-19 15:01:08 +00:00
void ClusterCopier::updateConfigIfNeeded()
{
2020-02-20 17:26:20 +00:00
UInt64 version_to_update = task_description_version;
bool is_outdated_version = task_description_current_version != version_to_update;
bool is_expired_session = !task_description_watch_zookeeper || task_description_watch_zookeeper->expired();
2020-02-19 15:01:08 +00:00
if (!is_outdated_version && !is_expired_session)
return;
2020-02-19 15:01:08 +00:00
LOG_DEBUG(log, "Updating task description");
reloadTaskDescription();
2020-02-20 17:26:20 +00:00
task_description_current_version = version_to_update;
2020-02-19 15:01:08 +00:00
}
2020-02-19 15:01:08 +00:00
void ClusterCopier::process(const ConnectionTimeouts & timeouts)
{
for (TaskTable & task_table : task_cluster->table_tasks)
{
LOG_INFO(log, "Process table task " << task_table.table_id << " with "
2020-02-18 13:39:22 +00:00
<< task_table.all_shards.size() << " shards, "
<< task_table.local_shards.size() << " of them are local ones");
2020-02-19 15:01:08 +00:00
if (task_table.all_shards.empty())
continue;
2020-02-19 15:01:08 +00:00
/// Discover partitions of each shard and total set of partitions
if (!task_table.has_enabled_partitions)
{
/// If there are no specified enabled_partitions, we must discover them manually
discoverTablePartitions(timeouts, task_table);
2020-02-19 15:01:08 +00:00
/// After partitions of each shard are initialized, initialize cluster partitions
for (const TaskShardPtr & task_shard : task_table.all_shards)
{
for (const auto & partition_elem : task_shard->partition_tasks)
{
const String & partition_name = partition_elem.first;
task_table.cluster_partitions.emplace(partition_name, ClusterPartition{});
}
}
2020-02-19 15:01:08 +00:00
for (auto & partition_elem : task_table.cluster_partitions)
{
const String & partition_name = partition_elem.first;
2020-02-19 15:01:08 +00:00
for (const TaskShardPtr & task_shard : task_table.all_shards)
task_shard->checked_partitions.emplace(partition_name);
2020-02-19 15:01:08 +00:00
task_table.ordered_partition_names.emplace_back(partition_name);
}
}
else
{
/// If enabled_partitions are specified, assume that each shard has all partitions
/// We will refine partition set of each shard in future
2020-02-19 15:01:08 +00:00
for (const String & partition_name : task_table.enabled_partitions)
{
task_table.cluster_partitions.emplace(partition_name, ClusterPartition{});
task_table.ordered_partition_names.emplace_back(partition_name);
}
}
2020-02-19 15:01:08 +00:00
task_table.watch.restart();
2018-02-08 11:07:58 +00:00
2020-02-19 15:01:08 +00:00
/// Retry table processing
bool table_is_done = false;
for (UInt64 num_table_tries = 0; num_table_tries < max_table_tries; ++num_table_tries)
{
if (tryProcessTable(timeouts, task_table))
{
table_is_done = true;
break;
}
}
2020-03-13 16:25:07 +00:00
/// Delete helping tables in both cases (whole table is done or not)
dropHelpingTables(task_table);
2020-02-19 15:01:08 +00:00
if (!table_is_done)
{
throw Exception("Too many tries to process table " + task_table.table_id + ". Abort remaining execution",
ErrorCodes::UNFINISHED);
}
}
2020-02-19 15:01:08 +00:00
}
2020-02-19 15:01:08 +00:00
/// Protected section
2020-02-20 17:26:20 +00:00
/*
* Creates task worker node and checks maximum number of workers not to exceed the limit.
* To achive this we have to check version of workers_version_path node and create current_worker_path
* node atomically.
* */
2020-02-19 15:01:08 +00:00
zkutil::EphemeralNodeHolder::Ptr ClusterCopier::createTaskWorkerNodeAndWaitIfNeed(
const zkutil::ZooKeeperPtr & zookeeper,
const String & description,
bool unprioritized)
{
2020-02-19 15:01:08 +00:00
std::chrono::milliseconds current_sleep_time = default_sleep_time;
static constexpr std::chrono::milliseconds max_sleep_time(30000); // 30 sec
2020-02-19 15:01:08 +00:00
if (unprioritized)
std::this_thread::sleep_for(current_sleep_time);
2020-02-19 15:01:08 +00:00
String workers_version_path = getWorkersPathVersion();
2020-02-20 17:26:20 +00:00
String workers_path = getWorkersPath();
String current_worker_path = getCurrentWorkerNodePath();
2020-02-19 15:01:08 +00:00
UInt64 num_bad_version_errors = 0;
2020-02-19 15:01:08 +00:00
while (true)
{
updateConfigIfNeeded();
2020-02-19 15:01:08 +00:00
Coordination::Stat stat;
zookeeper->get(workers_version_path, &stat);
auto version = stat.version;
zookeeper->get(workers_path, &stat);
2020-02-19 15:01:08 +00:00
if (static_cast<UInt64>(stat.numChildren) >= task_cluster->max_workers)
{
LOG_DEBUG(log, "Too many workers (" << stat.numChildren << ", maximum " << task_cluster->max_workers << ")"
<< ". Postpone processing " << description);
2020-02-19 15:01:08 +00:00
if (unprioritized)
current_sleep_time = std::min(max_sleep_time, current_sleep_time + default_sleep_time);
2020-02-19 15:01:08 +00:00
std::this_thread::sleep_for(current_sleep_time);
num_bad_version_errors = 0;
}
else
{
Coordination::Requests ops;
ops.emplace_back(zkutil::makeSetRequest(workers_version_path, description, version));
ops.emplace_back(zkutil::makeCreateRequest(current_worker_path, description, zkutil::CreateMode::Ephemeral));
Coordination::Responses responses;
auto code = zookeeper->tryMulti(ops, responses);
2020-02-19 15:01:08 +00:00
if (code == Coordination::ZOK || code == Coordination::ZNODEEXISTS)
return std::make_shared<zkutil::EphemeralNodeHolder>(current_worker_path, *zookeeper, false, false, description);
2020-02-19 15:01:08 +00:00
if (code == Coordination::ZBADVERSION)
{
++num_bad_version_errors;
2020-02-19 15:01:08 +00:00
/// Try to make fast retries
if (num_bad_version_errors > 3)
{
LOG_DEBUG(log, "A concurrent worker has just been added, will check free worker slots again");
std::chrono::milliseconds random_sleep_time(std::uniform_int_distribution<int>(1, 1000)(task_cluster->random_engine));
std::this_thread::sleep_for(random_sleep_time);
num_bad_version_errors = 0;
}
}
else
throw Coordination::Exception(code);
}
}
}
2020-02-18 13:39:22 +00:00
2020-02-21 16:00:50 +00:00
bool ClusterCopier::checkPartitionPieceIsClean(
const zkutil::ZooKeeperPtr & zookeeper,
const CleanStateClock & clean_state_clock,
2020-03-27 22:44:13 +00:00
const String & task_status_path)
{
2020-02-21 16:00:50 +00:00
LogicalClock task_start_clock;
2020-02-18 13:39:22 +00:00
2020-02-21 16:00:50 +00:00
Coordination::Stat stat{};
if (zookeeper->exists(task_status_path, &stat))
task_start_clock = LogicalClock(stat.mzxid);
2020-02-18 13:39:22 +00:00
2020-03-17 16:50:22 +00:00
return clean_state_clock.is_clean() && (!task_start_clock.hasHappened() || clean_state_clock.discovery_zxid <= task_start_clock);
2020-02-18 13:39:22 +00:00
}
2020-02-21 16:00:50 +00:00
bool ClusterCopier::checkAllPiecesInPartitionAreDone(const TaskTable & task_table, const String & partition_name, const TasksShard & shards_with_partition)
2020-02-18 13:39:22 +00:00
{
bool answer = true;
2020-03-13 14:19:20 +00:00
for (size_t piece_number = 0; piece_number < task_table.number_of_splits; ++piece_number)
{
bool piece_is_done = checkPartitionPieceIsDone(task_table, partition_name, piece_number, shards_with_partition);
if (!piece_is_done)
LOG_DEBUG(log, "Partition " << partition_name << " piece " + toString(piece_number) + " is not already done.");
answer &= piece_is_done;
}
2020-02-18 13:39:22 +00:00
return answer;
}
/* The same as function above
* Assume that we don't know on which shards do we have partition certain piece.
* We'll check them all (I mean shards that contain the whole partition)
* And shards that don't have certain piece MUST mark that piece is_done true.
* */
bool ClusterCopier::checkPartitionPieceIsDone(const TaskTable & task_table, const String & partition_name,
size_t piece_number, const TasksShard & shards_with_partition)
{
LOG_DEBUG(log, "Check that all shards processed partition " << partition_name
2020-02-21 16:00:50 +00:00
<< " piece " + toString(piece_number) + " successfully");
2020-02-19 15:01:08 +00:00
auto zookeeper = context.getZooKeeper();
2020-02-18 13:39:22 +00:00
/// Collect all shards that contain partition piece number piece_number.
Strings piece_status_paths;
2020-02-19 15:01:08 +00:00
for (auto & shard : shards_with_partition)
{
ShardPartition & task_shard_partition = shard->partition_tasks.find(partition_name)->second;
2020-02-18 13:39:22 +00:00
ShardPartitionPiece & shard_partition_piece = task_shard_partition.pieces[piece_number];
piece_status_paths.emplace_back(shard_partition_piece.getShardStatusPath());
2020-02-19 15:01:08 +00:00
}
2020-02-19 15:01:08 +00:00
std::vector<int64_t> zxid1, zxid2;
2020-02-19 15:01:08 +00:00
try
{
std::vector<zkutil::ZooKeeper::FutureGet> get_futures;
2020-02-18 13:39:22 +00:00
for (const String & path : piece_status_paths)
2020-02-19 15:01:08 +00:00
get_futures.emplace_back(zookeeper->asyncGet(path));
2020-02-19 15:01:08 +00:00
// Check that state is Finished and remember zxid
for (auto & future : get_futures)
{
auto res = future.get();
2020-02-19 15:01:08 +00:00
TaskStateWithOwner status = TaskStateWithOwner::fromString(res.data);
if (status.state != TaskState::Finished)
{
2020-02-21 16:00:50 +00:00
LOG_INFO(log, "The task " << res.data << " is being rewritten by "
<< status.owner << ". Partition piece will be rechecked");
2020-02-19 15:01:08 +00:00
return false;
}
2020-02-19 15:01:08 +00:00
zxid1.push_back(res.stat.pzxid);
}
2020-02-21 16:00:50 +00:00
const String piece_is_dirty_flag_path = task_table.getCertainPartitionPieceIsDirtyPath(partition_name, piece_number);
const String piece_is_dirty_cleaned_path = task_table.getCertainPartitionPieceIsCleanedPath(partition_name, piece_number);
const String piece_task_status_path = task_table.getCertainPartitionPieceTaskStatusPath(partition_name, piece_number);
CleanStateClock clean_state_clock (zookeeper, piece_is_dirty_flag_path, piece_is_dirty_cleaned_path);
const bool is_clean = checkPartitionPieceIsClean(zookeeper, clean_state_clock, piece_task_status_path);
2020-02-18 13:39:22 +00:00
2020-02-21 16:00:50 +00:00
if (!is_clean)
2020-02-19 15:01:08 +00:00
{
2020-02-21 16:00:50 +00:00
LOG_INFO(log, "Partition " << partition_name << " become dirty");
return false;
2020-02-19 15:01:08 +00:00
}
2020-02-19 15:01:08 +00:00
get_futures.clear();
2020-02-18 13:39:22 +00:00
for (const String & path : piece_status_paths)
2020-02-19 15:01:08 +00:00
get_futures.emplace_back(zookeeper->asyncGet(path));
2020-02-19 15:01:08 +00:00
// Remember zxid of states again
for (auto & future : get_futures)
{
auto res = future.get();
zxid2.push_back(res.stat.pzxid);
}
}
catch (const Coordination::Exception & e)
{
2020-02-18 13:39:22 +00:00
LOG_INFO(log, "A ZooKeeper error occurred while checking partition " << partition_name << " piece number "
2020-02-21 16:00:50 +00:00
<< toString(piece_number) << ". Will recheck the partition. Error: " << e.displayText());
2020-02-19 15:01:08 +00:00
return false;
}
2020-02-19 15:01:08 +00:00
// If all task is finished and zxid is not changed then partition could not become dirty again
2020-02-18 13:39:22 +00:00
for (UInt64 shard_num = 0; shard_num < piece_status_paths.size(); ++shard_num)
2020-02-19 15:01:08 +00:00
{
if (zxid1[shard_num] != zxid2[shard_num])
{
2020-02-18 13:39:22 +00:00
LOG_INFO(log, "The task " << piece_status_paths[shard_num] << " is being modified now. Partition piece will be rechecked");
2020-02-19 15:01:08 +00:00
return false;
}
}
2020-03-13 14:19:20 +00:00
LOG_INFO(log, "Partition " << partition_name << " piece number " << toString(piece_number) << " is copied successfully");
2020-02-19 15:01:08 +00:00
return true;
}
2020-03-13 14:19:20 +00:00
2020-03-18 13:25:49 +00:00
TaskStatus ClusterCopier::tryMoveAllPiecesToDestinationTable(const TaskTable & task_table, const String & partition_name)
2020-03-13 14:19:20 +00:00
{
2020-03-16 21:05:38 +00:00
bool inject_fault = false;
if (move_fault_probability > 0)
{
double value = std::uniform_real_distribution<>(0, 1)(task_table.task_cluster.random_engine);
inject_fault = value < move_fault_probability;
}
2020-03-13 14:19:20 +00:00
LOG_DEBUG(log, "Try to move " << partition_name << " to destionation table");
auto zookeeper = context.getZooKeeper();
const auto current_partition_attach_is_active = task_table.getPartitionAttachIsActivePath(partition_name);
const auto current_partition_attach_is_done = task_table.getPartitionAttachIsDonePath(partition_name);
/// Create ephemeral node to mark that we are active and process the partition
zookeeper->createAncestors(current_partition_attach_is_active);
zkutil::EphemeralNodeHolderPtr partition_attach_node_holder;
try
{
partition_attach_node_holder = zkutil::EphemeralNodeHolder::create(current_partition_attach_is_active, *zookeeper, host_id);
}
catch (const Coordination::Exception & e)
{
if (e.code == Coordination::ZNODEEXISTS)
{
LOG_DEBUG(log, "Someone is already moving pieces " << current_partition_attach_is_active);
2020-03-18 13:25:49 +00:00
return TaskStatus::Active;
2020-03-13 14:19:20 +00:00
}
throw;
}
/// Exit if task has been already processed;
/// create blocking node to signal cleaning up if it is abandoned
{
String status_data;
if (zookeeper->tryGet(current_partition_attach_is_done, status_data))
{
TaskStateWithOwner status = TaskStateWithOwner::fromString(status_data);
if (status.state == TaskState::Finished)
{
LOG_DEBUG(log, "All pieces for partition from this task " << current_partition_attach_is_active
<< " has been successfully moved to destination table by " << status.owner);
2020-03-18 13:25:49 +00:00
return TaskStatus::Finished;
2020-03-13 14:19:20 +00:00
}
/// Task is abandoned, because previously we created ephemeral node, possibly in other copier's process.
/// Initialize DROP PARTITION
LOG_DEBUG(log, "Moving piece for partition " << current_partition_attach_is_active
<< " has not been successfully finished by " << status.owner
<< ". Will try to move by myself.");
2020-03-17 16:23:47 +00:00
/// Remove is_done marker.
zookeeper->remove(current_partition_attach_is_done);
2020-03-13 14:19:20 +00:00
}
}
/// Try start processing, create node about it
{
String start_state = TaskStateWithOwner::getData(TaskState::Started, host_id);
zookeeper->create(current_partition_attach_is_done, start_state, zkutil::CreateMode::Persistent);
}
2020-03-16 21:05:38 +00:00
/// Move partition to original destination table.
2020-03-13 14:19:20 +00:00
for (size_t current_piece_number = 0; current_piece_number < task_table.number_of_splits; ++current_piece_number)
{
2020-03-16 21:05:38 +00:00
LOG_DEBUG(log, "Trying to move partition " << partition_name
<< " piece " << toString(current_piece_number)
<< " to original table");
2020-03-13 14:19:20 +00:00
2020-03-16 21:05:38 +00:00
ASTPtr query_alter_ast;
String query_alter_ast_string;
2020-03-13 14:19:20 +00:00
2020-03-16 21:05:38 +00:00
DatabaseAndTableName original_table = task_table.table_push;
DatabaseAndTableName helping_table = DatabaseAndTableName(original_table.first,
original_table.second + "_piece_" +
toString(current_piece_number));
2020-03-13 14:19:20 +00:00
2020-03-20 12:18:26 +00:00
Settings settings_push = task_cluster->settings_push;
/// It is important, ALTER ATTACH PARTITION must be done synchronously
/// And we will execute this ALTER query on each replica of a shard.
/// It is correct, because this query is idempotent.
settings_push.replication_alter_partitions_sync = 2;
2020-03-16 21:05:38 +00:00
query_alter_ast_string += " ALTER TABLE " + getQuotedTable(original_table) +
" ATTACH PARTITION " + partition_name +
2020-03-20 12:18:26 +00:00
" FROM " + getQuotedTable(helping_table);
2020-03-13 14:19:20 +00:00
2020-03-16 21:05:38 +00:00
LOG_DEBUG(log, "Executing ALTER query: " << query_alter_ast_string);
2020-03-13 14:19:20 +00:00
2020-03-16 21:05:38 +00:00
try
{
2020-03-18 13:25:49 +00:00
size_t num_nodes = executeQueryOnCluster(
2020-03-16 21:05:38 +00:00
task_table.cluster_push,
query_alter_ast_string,
nullptr,
2020-03-20 12:18:26 +00:00
&settings_push,
2020-03-16 21:05:38 +00:00
PoolMode::GET_MANY,
ClusterExecutionMode::ON_EACH_NODE);
2020-03-13 14:19:20 +00:00
2020-03-16 21:05:38 +00:00
LOG_INFO(log, "Number of nodes that executed ALTER query successfully : " << toString(num_nodes));
}
catch (...)
{
LOG_DEBUG(log, "Error while moving partition " << partition_name
<< " piece " << toString(current_piece_number)
<< "to original table");
throw;
}
2020-03-13 14:19:20 +00:00
2020-03-16 21:05:38 +00:00
if (inject_fault)
throw Exception("Copy fault injection is activated", ErrorCodes::UNFINISHED);
2020-03-13 14:19:20 +00:00
2020-03-16 21:05:38 +00:00
try
{
String query_deduplicate_ast_string;
2020-03-17 18:07:54 +00:00
if (!task_table.isReplicatedTable())
{
2020-03-16 21:05:38 +00:00
query_deduplicate_ast_string += " OPTIMIZE TABLE " + getQuotedTable(original_table) +
" PARTITION " + partition_name + " DEDUPLICATE;";
LOG_DEBUG(log, "Executing OPTIMIZE DEDUPLICATE query: " << query_alter_ast_string);
UInt64 num_nodes = executeQueryOnCluster(
task_table.cluster_push,
query_deduplicate_ast_string,
nullptr,
&task_cluster->settings_push,
PoolMode::GET_MANY);
LOG_INFO(log, "Number of shard that executed OPTIMIZE DEDUPLICATE query successfully : "
<< toString(num_nodes));
2020-03-13 14:19:20 +00:00
}
}
2020-03-16 21:05:38 +00:00
catch (...)
{
LOG_DEBUG(log, "Error while executing OPTIMIZE DEDUPLICATE partition " << partition_name
<< "in the original table");
throw;
}
2020-03-13 14:19:20 +00:00
}
/// Create node to signal that we finished moving
{
String state_finished = TaskStateWithOwner::getData(TaskState::Finished, host_id);
zookeeper->set(current_partition_attach_is_done, state_finished, 0);
}
2020-03-18 13:25:49 +00:00
return TaskStatus::Finished;
2020-03-13 14:19:20 +00:00
}
2020-02-18 13:39:22 +00:00
/// Removes MATERIALIZED and ALIAS columns from create table query
2020-02-19 15:01:08 +00:00
ASTPtr ClusterCopier::removeAliasColumnsFromCreateQuery(const ASTPtr & query_ast)
{
2020-02-19 15:01:08 +00:00
const ASTs & column_asts = query_ast->as<ASTCreateQuery &>().columns_list->columns->children;
auto new_columns = std::make_shared<ASTExpressionList>();
2020-02-19 15:01:08 +00:00
for (const ASTPtr & column_ast : column_asts)
{
const auto & column = column_ast->as<ASTColumnDeclaration &>();
2020-02-19 15:01:08 +00:00
if (!column.default_specifier.empty())
{
ColumnDefaultKind kind = columnDefaultKindFromString(column.default_specifier);
if (kind == ColumnDefaultKind::Materialized || kind == ColumnDefaultKind::Alias)
continue;
}
2020-02-19 15:01:08 +00:00
new_columns->children.emplace_back(column_ast->clone());
}
2020-02-19 15:01:08 +00:00
ASTPtr new_query_ast = query_ast->clone();
auto & new_query = new_query_ast->as<ASTCreateQuery &>();
2020-02-19 15:01:08 +00:00
auto new_columns_list = std::make_shared<ASTColumns>();
new_columns_list->set(new_columns_list->columns, new_columns);
if (auto indices = query_ast->as<ASTCreateQuery>()->columns_list->indices)
new_columns_list->set(new_columns_list->indices, indices->clone());
2020-02-19 15:01:08 +00:00
new_query.replace(new_query.columns_list, new_columns_list);
return new_query_ast;
}
2020-03-18 00:57:00 +00:00
/// Replaces ENGINE and table name in a create query
std::shared_ptr<ASTCreateQuery> rewriteCreateQueryStorage(const ASTPtr & create_query_ast,
const DatabaseAndTableName & new_table,
const ASTPtr & new_storage_ast)
{
2020-02-19 15:01:08 +00:00
const auto & create = create_query_ast->as<ASTCreateQuery &>();
auto res = std::make_shared<ASTCreateQuery>(create);
2020-02-19 15:01:08 +00:00
if (create.storage == nullptr || new_storage_ast == nullptr)
throw Exception("Storage is not specified", ErrorCodes::LOGICAL_ERROR);
2020-02-19 15:01:08 +00:00
res->database = new_table.first;
res->table = new_table.second;
2020-02-19 15:01:08 +00:00
res->children.clear();
res->set(res->columns_list, create.columns_list->clone());
res->set(res->storage, new_storage_ast->clone());
2020-02-19 15:01:08 +00:00
return res;
}
2020-02-20 18:58:00 +00:00
bool ClusterCopier::tryDropPartitionPiece(
ShardPartition & task_partition,
const size_t current_piece_number,
const zkutil::ZooKeeperPtr & zookeeper,
const CleanStateClock & clean_state_clock)
{
2020-02-19 15:01:08 +00:00
if (is_safe_mode)
throw Exception("DROP PARTITION is prohibited in safe mode", ErrorCodes::NOT_IMPLEMENTED);
2020-02-19 15:01:08 +00:00
TaskTable & task_table = task_partition.task_shard.task_table;
2020-02-20 18:58:00 +00:00
ShardPartitionPiece & partition_piece = task_partition.pieces[current_piece_number];
2020-02-20 18:58:00 +00:00
const String current_shards_path = partition_piece.getPartitionPieceShardsPath();
const String current_partition_active_workers_dir = partition_piece.getPartitionPieceActiveWorkersPath();
const String is_dirty_flag_path = partition_piece.getPartitionPieceIsDirtyPath();
const String dirty_cleaner_path = partition_piece.getPartitionPieceCleanerPath();
const String is_dirty_cleaned_path = partition_piece.getPartitionPieceIsCleanedPath();
2020-02-19 15:01:08 +00:00
zkutil::EphemeralNodeHolder::Ptr cleaner_holder;
try
{
2020-02-18 13:39:22 +00:00
cleaner_holder = zkutil::EphemeralNodeHolder::create(dirty_cleaner_path, *zookeeper, host_id);
}
2020-02-19 15:01:08 +00:00
catch (const Coordination::Exception & e)
{
2020-02-19 15:01:08 +00:00
if (e.code == Coordination::ZNODEEXISTS)
{
2020-02-20 18:58:00 +00:00
LOG_DEBUG(log, "Partition " << task_partition.name << " piece "
<< toString(current_piece_number) << " is cleaning now by somebody, sleep");
2020-02-19 15:01:08 +00:00
std::this_thread::sleep_for(default_sleep_time);
return false;
}
2020-02-19 15:01:08 +00:00
throw;
}
2020-02-18 13:39:22 +00:00
Coordination::Stat stat{};
2020-02-19 15:01:08 +00:00
if (zookeeper->exists(current_partition_active_workers_dir, &stat))
{
2020-02-19 15:01:08 +00:00
if (stat.numChildren != 0)
{
2020-02-20 18:58:00 +00:00
LOG_DEBUG(log, "Partition " << task_partition.name << " contains " << stat.numChildren
<< " active workers while trying to drop it. Going to sleep.");
2020-02-19 15:01:08 +00:00
std::this_thread::sleep_for(default_sleep_time);
return false;
}
else
{
2020-02-19 15:01:08 +00:00
zookeeper->remove(current_partition_active_workers_dir);
}
}
2020-02-19 15:01:08 +00:00
{
zkutil::EphemeralNodeHolder::Ptr active_workers_lock;
try
{
active_workers_lock = zkutil::EphemeralNodeHolder::create(current_partition_active_workers_dir, *zookeeper, host_id);
}
catch (const Coordination::Exception & e)
{
if (e.code == Coordination::ZNODEEXISTS)
{
LOG_DEBUG(log, "Partition " << task_partition.name << " is being filled now by somebody, sleep");
return false;
}
2020-02-19 15:01:08 +00:00
throw;
}
2020-02-19 15:01:08 +00:00
// Lock the dirty flag
zookeeper->set(is_dirty_flag_path, host_id, clean_state_clock.discovery_version.value());
2020-02-20 18:58:00 +00:00
zookeeper->tryRemove(partition_piece.getPartitionPieceCleanStartPath());
2020-02-18 13:39:22 +00:00
CleanStateClock my_clock(zookeeper, is_dirty_flag_path, is_dirty_cleaned_path);
2020-02-19 15:01:08 +00:00
/// Remove all status nodes
{
2020-02-19 15:01:08 +00:00
Strings children;
if (zookeeper->tryGetChildren(current_shards_path, children) == Coordination::ZOK)
for (const auto & child : children)
{
zookeeper->removeRecursive(current_shards_path + "/" + child);
}
}
2020-03-12 19:46:48 +00:00
DatabaseAndTableName original_table = task_table.table_push;
DatabaseAndTableName helping_table = DatabaseAndTableName(original_table.first, original_table.second + "_piece_" + toString(current_piece_number));
String query = "ALTER TABLE " + getQuotedTable(helping_table);
2020-02-19 15:01:08 +00:00
query += " DROP PARTITION " + task_partition.name + "";
2020-02-19 15:01:08 +00:00
/// TODO: use this statement after servers will be updated up to 1.1.54310
// query += " DROP PARTITION ID '" + task_partition.name + "'";
2020-02-19 15:01:08 +00:00
ClusterPtr & cluster_push = task_table.cluster_push;
Settings settings_push = task_cluster->settings_push;
2020-02-19 15:01:08 +00:00
/// It is important, DROP PARTITION must be done synchronously
settings_push.replication_alter_partitions_sync = 2;
2020-02-19 15:01:08 +00:00
LOG_DEBUG(log, "Execute distributed DROP PARTITION: " << query);
2020-03-12 16:48:28 +00:00
/// We have to drop partition_piece on each replica
2020-03-17 18:07:54 +00:00
size_t num_shards = executeQueryOnCluster(
2020-03-10 20:04:08 +00:00
cluster_push, query,
nullptr,
&settings_push,
2020-03-12 16:48:28 +00:00
PoolMode::GET_MANY,
ClusterExecutionMode::ON_EACH_NODE);
2020-03-17 18:07:54 +00:00
LOG_INFO(log, "DROP PARTITION was successfully executed on " << num_shards << " nodes of a cluster.");
2020-02-19 15:01:08 +00:00
/// Update the locking node
if (!my_clock.is_stale())
{
zookeeper->set(is_dirty_flag_path, host_id, my_clock.discovery_version.value());
if (my_clock.clean_state_version)
2020-02-18 13:39:22 +00:00
zookeeper->set(is_dirty_cleaned_path, host_id, my_clock.clean_state_version.value());
2020-02-19 15:01:08 +00:00
else
2020-02-18 13:39:22 +00:00
zookeeper->create(is_dirty_cleaned_path, host_id, zkutil::CreateMode::Persistent);
2020-02-19 15:01:08 +00:00
}
else
{
2020-02-19 15:01:08 +00:00
LOG_DEBUG(log, "Clean state is altered when dropping the partition, cowardly bailing");
/// clean state is stale
return false;
}
2020-02-20 18:58:00 +00:00
LOG_INFO(log, "Partition " << task_partition.name << " piece " << toString(current_piece_number)
<< " was dropped on cluster " << task_table.cluster_push_name);
2020-02-19 15:01:08 +00:00
if (zookeeper->tryCreate(current_shards_path, host_id, zkutil::CreateMode::Persistent) == Coordination::ZNODEEXISTS)
zookeeper->set(current_shards_path, host_id);
}
2020-03-12 16:48:28 +00:00
LOG_INFO(log, "Partition " << task_partition.name << " piece " << toString(current_piece_number) << " is safe for work now.");
2020-02-19 15:01:08 +00:00
return true;
}
2020-02-19 15:01:08 +00:00
bool ClusterCopier::tryProcessTable(const ConnectionTimeouts & timeouts, TaskTable & task_table)
{
2020-02-19 15:01:08 +00:00
/// An heuristic: if previous shard is already done, then check next one without sleeps due to max_workers constraint
bool previous_shard_is_instantly_finished = false;
2020-02-19 15:01:08 +00:00
/// Process each partition that is present in cluster
for (const String & partition_name : task_table.ordered_partition_names)
{
if (!task_table.cluster_partitions.count(partition_name))
throw Exception("There are no expected partition " + partition_name + ". It is a bug", ErrorCodes::LOGICAL_ERROR);
2020-02-19 15:01:08 +00:00
ClusterPartition & cluster_partition = task_table.cluster_partitions[partition_name];
2020-02-19 15:01:08 +00:00
Stopwatch watch;
2020-02-18 13:39:22 +00:00
/// We will check all the shards of the table and check if they contain current partition.
2020-02-19 15:01:08 +00:00
TasksShard expected_shards;
UInt64 num_failed_shards = 0;
2020-02-19 15:01:08 +00:00
++cluster_partition.total_tries;
2020-02-19 15:01:08 +00:00
LOG_DEBUG(log, "Processing partition " << partition_name << " for the whole cluster");
2020-02-19 15:01:08 +00:00
/// Process each source shard having current partition and copy current partition
/// NOTE: shards are sorted by "distance" to current host
bool has_shard_to_process = false;
for (const TaskShardPtr & shard : task_table.all_shards)
{
2020-02-19 15:01:08 +00:00
/// Does shard have a node with current partition?
if (shard->partition_tasks.count(partition_name) == 0)
{
2020-02-19 15:01:08 +00:00
/// If not, did we check existence of that partition previously?
if (shard->checked_partitions.count(partition_name) == 0)
{
2020-02-19 15:01:08 +00:00
auto check_shard_has_partition = [&] () { return checkShardHasPartition(timeouts, *shard, partition_name); };
bool has_partition = retry(check_shard_has_partition);
2020-02-19 15:01:08 +00:00
shard->checked_partitions.emplace(partition_name);
2020-02-19 15:01:08 +00:00
if (has_partition)
{
2020-03-12 19:46:48 +00:00
const size_t number_of_splits = task_table.number_of_splits;
shard->partition_tasks.emplace(partition_name, ShardPartition(*shard, partition_name, number_of_splits));
2020-02-19 15:01:08 +00:00
LOG_DEBUG(log, "Discovered partition " << partition_name << " in shard " << shard->getDescription());
2020-02-18 13:39:22 +00:00
/// To save references in the future.
auto shard_partition_it = shard->partition_tasks.find(partition_name);
PartitionPieces & shard_partition_pieces = shard_partition_it->second.pieces;
for (size_t piece_number = 0; piece_number < number_of_splits; ++piece_number)
{
auto res = checkPresentPartitionPiecesOnCurrentShard(timeouts, *shard, partition_name, piece_number);
shard_partition_pieces.emplace_back(shard_partition_it->second, piece_number, res);
}
2020-02-19 15:01:08 +00:00
}
else
{
LOG_DEBUG(log, "Found that shard " << shard->getDescription() << " does not contain current partition " << partition_name);
continue;
}
}
else
{
2020-02-19 15:01:08 +00:00
/// We have already checked that partition, but did not discover it
previous_shard_is_instantly_finished = true;
continue;
}
2020-02-19 15:01:08 +00:00
}
2020-02-19 15:01:08 +00:00
auto it_shard_partition = shard->partition_tasks.find(partition_name);
2020-02-18 13:39:22 +00:00
/// Previously when we discovered that shard does not contain current partition, we skipped it.
/// At this moment partition have to be present.
2020-02-19 15:01:08 +00:00
if (it_shard_partition == shard->partition_tasks.end())
2020-02-18 13:39:22 +00:00
throw Exception("There are no such partition in a shard. This is a bug.", ErrorCodes::LOGICAL_ERROR);
2020-02-19 15:01:08 +00:00
auto & partition = it_shard_partition->second;
2020-02-19 15:01:08 +00:00
expected_shards.emplace_back(shard);
2020-02-19 15:01:08 +00:00
/// Do not sleep if there is a sequence of already processed shards to increase startup
bool is_unprioritized_task = !previous_shard_is_instantly_finished && shard->priority.is_remote;
2020-03-18 13:25:49 +00:00
TaskStatus task_status = TaskStatus::Error;
2020-02-19 15:01:08 +00:00
bool was_error = false;
has_shard_to_process = true;
for (UInt64 try_num = 0; try_num < max_shard_partition_tries; ++try_num)
{
2020-02-19 15:01:08 +00:00
task_status = tryProcessPartitionTask(timeouts, partition, is_unprioritized_task);
2020-02-19 15:01:08 +00:00
/// Exit if success
2020-03-18 13:25:49 +00:00
if (task_status == TaskStatus::Finished)
2020-02-19 15:01:08 +00:00
break;
2020-02-19 15:01:08 +00:00
was_error = true;
2020-02-19 15:01:08 +00:00
/// Skip if the task is being processed by someone
2020-03-18 13:25:49 +00:00
if (task_status == TaskStatus::Active)
2020-02-19 15:01:08 +00:00
break;
2020-02-19 15:01:08 +00:00
/// Repeat on errors
std::this_thread::sleep_for(default_sleep_time);
}
2020-02-19 15:01:08 +00:00
2020-03-18 13:25:49 +00:00
if (task_status == TaskStatus::Error)
2020-02-19 15:01:08 +00:00
++num_failed_shards;
previous_shard_is_instantly_finished = !was_error;
}
2020-02-19 15:01:08 +00:00
cluster_partition.elapsed_time_seconds += watch.elapsedSeconds();
/// Check that whole cluster partition is done
/// Firstly check the number of failed partition tasks, then look into ZooKeeper and ensure that each partition is done
2020-03-13 14:19:20 +00:00
bool partition_copying_is_done = num_failed_shards == 0;
2020-02-19 15:01:08 +00:00
try
{
2020-03-13 14:19:20 +00:00
partition_copying_is_done =
2020-02-18 13:39:22 +00:00
!has_shard_to_process
2020-03-13 14:19:20 +00:00
|| (partition_copying_is_done && checkAllPiecesInPartitionAreDone(task_table, partition_name, expected_shards));
}
2020-02-19 15:01:08 +00:00
catch (...)
{
2020-02-19 15:01:08 +00:00
tryLogCurrentException(log);
2020-03-13 14:19:20 +00:00
partition_copying_is_done = false;
}
bool partition_moving_is_done = false;
/// Try to move only if all pieces were copied.
if (partition_copying_is_done)
{
2020-03-16 21:05:38 +00:00
for (UInt64 try_num = 0; try_num < max_shard_partition_piece_tries_for_alter; ++try_num)
2020-03-13 14:19:20 +00:00
{
2020-03-16 21:05:38 +00:00
try
{
auto res = tryMoveAllPiecesToDestinationTable(task_table, partition_name);
2020-03-18 13:25:49 +00:00
/// Exit and mark current task is done.
if (res == TaskStatus::Finished)
2020-03-16 21:05:38 +00:00
{
partition_moving_is_done = true;
break;
}
2020-03-18 13:25:49 +00:00
/// Exit if this task is active.
if (res == TaskStatus::Active)
2020-03-17 16:23:47 +00:00
break;
2020-03-16 21:05:38 +00:00
2020-03-18 13:25:49 +00:00
/// Repeat on errors.
2020-03-17 16:23:47 +00:00
std::this_thread::sleep_for(default_sleep_time);
2020-03-16 21:05:38 +00:00
}
2020-03-17 18:07:54 +00:00
catch (...)
{
tryLogCurrentException(log, "Some error occured while moving pieces to destination table for partition " + partition_name);
2020-03-16 21:05:38 +00:00
}
2020-03-13 14:19:20 +00:00
}
}
2020-03-13 14:19:20 +00:00
if (partition_copying_is_done && partition_moving_is_done)
{
2020-02-19 15:01:08 +00:00
task_table.finished_cluster_partitions.emplace(partition_name);
2020-02-19 15:01:08 +00:00
task_table.bytes_copied += cluster_partition.bytes_copied;
task_table.rows_copied += cluster_partition.rows_copied;
double elapsed = cluster_partition.elapsed_time_seconds;
2020-02-19 15:01:08 +00:00
LOG_INFO(log, "It took " << std::fixed << std::setprecision(2) << elapsed << " seconds to copy partition " << partition_name
2020-02-18 13:39:22 +00:00
<< ": " << formatReadableSizeWithDecimalSuffix(cluster_partition.bytes_copied) << " uncompressed bytes"
<< ", " << formatReadableQuantity(cluster_partition.rows_copied) << " rows"
<< " and " << cluster_partition.blocks_copied << " source blocks are copied");
2020-02-19 15:01:08 +00:00
if (cluster_partition.rows_copied)
{
2020-02-19 15:01:08 +00:00
LOG_INFO(log, "Average partition speed: "
2020-02-18 13:39:22 +00:00
<< formatReadableSizeWithDecimalSuffix(cluster_partition.bytes_copied / elapsed) << " per second.");
}
2020-02-19 15:01:08 +00:00
if (task_table.rows_copied)
{
2020-02-19 15:01:08 +00:00
LOG_INFO(log, "Average table " << task_table.table_id << " speed: "
2020-02-18 13:39:22 +00:00
<< formatReadableSizeWithDecimalSuffix(task_table.bytes_copied / elapsed) << " per second.");
2020-02-19 15:01:08 +00:00
}
}
}
2020-02-19 15:01:08 +00:00
UInt64 required_partitions = task_table.cluster_partitions.size();
UInt64 finished_partitions = task_table.finished_cluster_partitions.size();
bool table_is_done = finished_partitions >= required_partitions;
2020-02-19 15:01:08 +00:00
if (!table_is_done)
{
LOG_INFO(log, "Table " + task_table.table_id + " is not processed yet."
2020-02-18 13:39:22 +00:00
<< "Copied " << finished_partitions << " of " << required_partitions << ", will retry");
2020-02-19 15:01:08 +00:00
}
2020-02-19 15:01:08 +00:00
return table_is_done;
}
2020-02-18 13:39:22 +00:00
/// Job for copying partition from particular shard.
2020-03-18 13:25:49 +00:00
TaskStatus ClusterCopier::tryProcessPartitionTask(const ConnectionTimeouts & timeouts, ShardPartition & task_partition, bool is_unprioritized_task)
2020-02-19 15:01:08 +00:00
{
2020-03-18 13:25:49 +00:00
TaskStatus res;
2020-02-19 15:01:08 +00:00
try
{
2020-02-18 13:39:22 +00:00
res = iterateThroughAllPiecesInPartition(timeouts, task_partition, is_unprioritized_task);
2020-02-19 15:01:08 +00:00
}
catch (...)
{
tryLogCurrentException(log, "An error occurred while processing partition " + task_partition.name);
2020-03-18 13:25:49 +00:00
res = TaskStatus::Error;
2020-02-19 15:01:08 +00:00
}
2020-02-19 15:01:08 +00:00
/// At the end of each task check if the config is updated
try
{
updateConfigIfNeeded();
}
catch (...)
{
tryLogCurrentException(log, "An error occurred while updating the config");
}
2020-02-19 15:01:08 +00:00
return res;
}
2020-03-18 13:25:49 +00:00
TaskStatus ClusterCopier::iterateThroughAllPiecesInPartition(const ConnectionTimeouts & timeouts, ShardPartition & task_partition,
2020-02-18 13:39:22 +00:00
bool is_unprioritized_task)
{
const size_t total_number_of_pieces = task_partition.task_shard.task_table.number_of_splits;
2020-03-18 13:25:49 +00:00
TaskStatus res{TaskStatus::Finished};
2020-03-07 00:05:49 +00:00
bool was_failed_pieces = false;
bool was_active_pieces = false;
2020-02-18 13:39:22 +00:00
2020-02-25 12:38:11 +00:00
for (size_t piece_number = 0; piece_number < total_number_of_pieces; piece_number++)
{
2020-03-07 00:05:49 +00:00
for (UInt64 try_num = 0; try_num < max_shard_partition_tries; ++try_num)
{
2020-03-12 16:48:28 +00:00
LOG_INFO(log, "Attempt number " << try_num << " to process partition " << task_partition.name
<< " piece number " << piece_number << " on shard number " << task_partition.task_shard.numberInCluster()
<< " with index " << task_partition.task_shard.indexInCluster());
2020-03-07 00:05:49 +00:00
res = processPartitionPieceTaskImpl(timeouts, task_partition, piece_number, is_unprioritized_task);
/// Exit if success
2020-03-18 13:25:49 +00:00
if (res == TaskStatus::Finished)
2020-03-07 00:05:49 +00:00
break;
/// Skip if the task is being processed by someone
2020-03-18 13:25:49 +00:00
if (res == TaskStatus::Active)
2020-03-07 00:05:49 +00:00
break;
/// Repeat on errors
std::this_thread::sleep_for(default_sleep_time);
}
2020-03-18 13:25:49 +00:00
was_active_pieces = (res == TaskStatus::Active);
was_failed_pieces = (res == TaskStatus::Error);
2020-02-25 12:38:11 +00:00
}
2020-02-18 13:39:22 +00:00
2020-03-07 00:05:49 +00:00
if (was_failed_pieces)
2020-03-18 13:25:49 +00:00
return TaskStatus::Error;
2020-03-07 00:05:49 +00:00
if (was_active_pieces)
2020-03-18 13:25:49 +00:00
return TaskStatus::Active;
2020-03-07 00:05:49 +00:00
2020-03-18 13:25:49 +00:00
return TaskStatus::Finished;
2020-02-18 13:39:22 +00:00
}
2020-02-20 17:26:20 +00:00
2020-03-18 13:25:49 +00:00
TaskStatus ClusterCopier::processPartitionPieceTaskImpl(
2020-02-20 17:26:20 +00:00
const ConnectionTimeouts & timeouts, ShardPartition & task_partition,
const size_t current_piece_number, bool is_unprioritized_task)
2020-02-19 15:01:08 +00:00
{
TaskShard & task_shard = task_partition.task_shard;
TaskTable & task_table = task_shard.task_table;
2020-02-20 17:26:20 +00:00
ClusterPartition & cluster_partition = task_table.getClusterPartition(task_partition.name);
2020-02-18 13:39:22 +00:00
ShardPartitionPiece & partition_piece = task_partition.pieces[current_piece_number];
const size_t number_of_splits = task_table.number_of_splits;
const String primary_key_comma_separated = task_table.primary_key_comma_separated;
2020-02-19 15:01:08 +00:00
/// We need to update table definitions for each partition, it could be changed after ALTER
2020-02-21 16:00:50 +00:00
createShardInternalTables(timeouts, task_shard, true);
auto split_table_for_current_piece = task_shard.list_of_split_tables_on_shard[current_piece_number];
2020-02-19 15:01:08 +00:00
auto zookeeper = context.getZooKeeper();
2020-02-20 17:26:20 +00:00
const String piece_is_dirty_flag_path = partition_piece.getPartitionPieceIsDirtyPath();
const String piece_is_dirty_cleaned_path = partition_piece.getPartitionPieceIsCleanedPath();
2020-02-18 13:39:22 +00:00
const String current_task_piece_is_active_path = partition_piece.getActiveWorkerPath();
2020-02-20 17:26:20 +00:00
const String current_task_piece_status_path = partition_piece.getShardStatusPath();
2020-02-19 15:01:08 +00:00
/// Auxiliary functions:
2020-02-19 15:01:08 +00:00
/// Creates is_dirty node to initialize DROP PARTITION
2020-02-20 17:26:20 +00:00
auto create_is_dirty_node = [&] (const CleanStateClock & clock)
2020-02-19 15:01:08 +00:00
{
if (clock.is_stale())
LOG_DEBUG(log, "Clean state clock is stale while setting dirty flag, cowardly bailing");
else if (!clock.is_clean())
LOG_DEBUG(log, "Thank you, Captain Obvious");
else if (clock.discovery_version)
{
LOG_DEBUG(log, "Updating clean state clock");
2020-02-18 13:39:22 +00:00
zookeeper->set(piece_is_dirty_flag_path, host_id, clock.discovery_version.value());
2020-02-19 15:01:08 +00:00
}
else
{
2020-02-19 15:01:08 +00:00
LOG_DEBUG(log, "Creating clean state clock");
2020-02-18 13:39:22 +00:00
zookeeper->create(piece_is_dirty_flag_path, host_id, zkutil::CreateMode::Persistent);
}
2020-02-19 15:01:08 +00:00
};
2020-02-19 15:01:08 +00:00
/// Returns SELECT query filtering current partition and applying user filter
2020-02-20 17:26:20 +00:00
auto get_select_query = [&] (const DatabaseAndTableName & from_table, const String & fields, bool enable_splitting, String limit = "")
{
2020-02-19 15:01:08 +00:00
String query;
query += "SELECT " + fields + " FROM " + getQuotedTable(from_table);
/// TODO: Bad, it is better to rewrite with ASTLiteral(partition_key_field)
query += " WHERE (" + queryToString(task_table.engine_push_partition_key_ast) + " = (" + task_partition.name + " AS partition_key))";
2020-02-18 13:39:22 +00:00
if (enable_splitting)
2020-03-17 18:07:54 +00:00
query += " AND ( cityHash64(" + primary_key_comma_separated + ") %" + toString(number_of_splits) + " = " + toString(current_piece_number) + " )";
2020-02-18 13:39:22 +00:00
2020-02-19 15:01:08 +00:00
if (!task_table.where_condition_str.empty())
query += " AND (" + task_table.where_condition_str + ")";
if (!limit.empty())
query += " LIMIT " + limit;
2020-02-19 15:01:08 +00:00
ParserQuery p_query(query.data() + query.size());
const auto & settings = context.getSettingsRef();
return parseQuery(p_query, query, settings.max_query_size, settings.max_parser_depth);
2020-02-19 15:01:08 +00:00
};
2020-02-19 15:01:08 +00:00
/// Load balancing
2020-02-18 13:39:22 +00:00
auto worker_node_holder = createTaskWorkerNodeAndWaitIfNeed(zookeeper, current_task_piece_status_path, is_unprioritized_task);
2020-02-18 13:39:22 +00:00
LOG_DEBUG(log, "Processing " << current_task_piece_status_path);
2020-02-21 16:00:50 +00:00
const String piece_status_path = partition_piece.getPartitionPieceShardsPath();
2020-02-21 16:00:50 +00:00
CleanStateClock clean_state_clock(zookeeper, piece_is_dirty_flag_path, piece_is_dirty_cleaned_path);
const bool is_clean = checkPartitionPieceIsClean(zookeeper, clean_state_clock, piece_status_path);
2020-02-20 17:26:20 +00:00
/// Do not start if partition piece is dirty, try to clean it
2020-02-21 16:00:50 +00:00
if (is_clean)
{
2020-02-20 17:26:20 +00:00
LOG_DEBUG(log, "Partition " << task_partition.name
<< " piece " + toString(current_piece_number) + " appears to be clean");
2020-02-18 13:39:22 +00:00
zookeeper->createAncestors(current_task_piece_status_path);
}
2020-02-19 15:01:08 +00:00
else
{
2020-02-20 17:26:20 +00:00
LOG_DEBUG(log, "Partition " << task_partition.name
<< " piece " + toString(current_piece_number) + " is dirty, try to drop it");
2020-02-19 15:01:08 +00:00
try
{
2020-02-20 18:58:00 +00:00
tryDropPartitionPiece(task_partition, current_piece_number, zookeeper, clean_state_clock);
2020-02-19 15:01:08 +00:00
}
catch (...)
{
tryLogCurrentException(log, "An error occurred when clean partition");
}
2020-03-18 13:25:49 +00:00
return TaskStatus::Error;
}
2020-02-19 15:01:08 +00:00
/// Create ephemeral node to mark that we are active and process the partition
2020-02-18 13:39:22 +00:00
zookeeper->createAncestors(current_task_piece_is_active_path);
2020-02-19 15:01:08 +00:00
zkutil::EphemeralNodeHolderPtr partition_task_node_holder;
try
{
2020-02-18 13:39:22 +00:00
partition_task_node_holder = zkutil::EphemeralNodeHolder::create(current_task_piece_is_active_path, *zookeeper, host_id);
2020-02-19 15:01:08 +00:00
}
catch (const Coordination::Exception & e)
{
if (e.code == Coordination::ZNODEEXISTS)
{
2020-02-18 13:39:22 +00:00
LOG_DEBUG(log, "Someone is already processing " << current_task_piece_is_active_path);
2020-03-18 13:25:49 +00:00
return TaskStatus::Active;
}
2020-02-19 15:01:08 +00:00
throw;
}
2020-02-19 15:01:08 +00:00
/// Exit if task has been already processed;
/// create blocking node to signal cleaning up if it is abandoned
{
String status_data;
2020-02-18 13:39:22 +00:00
if (zookeeper->tryGet(current_task_piece_status_path, status_data))
{
2020-02-19 15:01:08 +00:00
TaskStateWithOwner status = TaskStateWithOwner::fromString(status_data);
if (status.state == TaskState::Finished)
{
2020-02-20 17:26:20 +00:00
LOG_DEBUG(log, "Task " << current_task_piece_status_path
<< " has been successfully executed by " << status.owner);
2020-03-18 13:25:49 +00:00
return TaskStatus::Finished;
}
2020-02-20 17:26:20 +00:00
/// Task is abandoned, because previously we created ephemeral node, possibly in other copier's process.
/// Initialize DROP PARTITION
LOG_DEBUG(log, "Task " << current_task_piece_status_path
<< " has not been successfully finished by " << status.owner
<< ". Partition will be dropped and refilled.");
2020-02-19 15:01:08 +00:00
create_is_dirty_node(clean_state_clock);
2020-03-18 13:25:49 +00:00
return TaskStatus::Error;
2020-02-19 15:01:08 +00:00
}
}
2020-03-10 20:04:08 +00:00
/// Exit if current piece is absent on this shard. Also mark it as finished, because we will check
/// whether each shard have processed each partitition (and its pieces).
if (partition_piece.is_absent_piece)
{
String state_finished = TaskStateWithOwner::getData(TaskState::Finished, host_id);
auto res = zookeeper->tryCreate(current_task_piece_status_path, state_finished, zkutil::CreateMode::Persistent);
if (res == Coordination::ZNODEEXISTS)
LOG_DEBUG(log, "Partition " << task_partition.name << " piece "
+ toString(current_piece_number) + " is absent on current replica of a shard. But other replicas have already marked it as done.");
if (res == Coordination::ZOK)
LOG_DEBUG(log, "Partition " << task_partition.name << " piece "
+ toString(current_piece_number) + " is absent on current replica of a shard. Will mark it as done. Other replicas will do the same.");
2020-03-18 13:25:49 +00:00
return TaskStatus::Finished;
2020-03-10 20:04:08 +00:00
}
2020-02-19 15:01:08 +00:00
/// Check that destination partition is empty if we are first worker
/// NOTE: this check is incorrect if pull and push tables have different partition key!
String clean_start_status;
2020-02-20 17:26:20 +00:00
if (!zookeeper->tryGet(partition_piece.getPartitionPieceCleanStartPath(), clean_start_status) || clean_start_status != "ok")
{
2020-02-20 17:26:20 +00:00
zookeeper->createIfNotExists(partition_piece.getPartitionPieceCleanStartPath(), "");
auto checker = zkutil::EphemeralNodeHolder::create(partition_piece.getPartitionPieceCleanStartPath() + "/checker",
2020-02-18 13:39:22 +00:00
*zookeeper, host_id);
2020-02-19 15:01:08 +00:00
// Maybe we are the first worker
2020-02-21 16:00:50 +00:00
ASTPtr query_select_ast = get_select_query(split_table_for_current_piece, "count()", /*enable_splitting*/ true);
2020-02-19 15:01:08 +00:00
UInt64 count;
{
Context local_context = context;
// Use pull (i.e. readonly) settings, but fetch data from destination servers
local_context.setSettings(task_cluster->settings_pull);
local_context.setSetting("skip_unavailable_shards", true);
2020-02-19 15:01:08 +00:00
Block block = getBlockWithAllStreamData(InterpreterFactory::get(query_select_ast, local_context)->execute().in);
count = (block) ? block.safeGetByPosition(0).column->getUInt(0) : 0;
}
2020-02-19 15:01:08 +00:00
if (count != 0)
{
2020-03-12 16:48:28 +00:00
LOG_INFO(log, "Partition " << task_partition.name << " piece "
<< current_piece_number << "is not empty. In contains " << count << " rows.");
2020-02-18 13:39:22 +00:00
Coordination::Stat stat_shards{};
2020-02-20 17:26:20 +00:00
zookeeper->get(partition_piece.getPartitionPieceShardsPath(), &stat_shards);
2020-02-19 15:01:08 +00:00
/// NOTE: partition is still fresh if dirt discovery happens before cleaning
if (stat_shards.numChildren == 0)
{
LOG_WARNING(log, "There are no workers for partition " << task_partition.name
2020-02-20 17:26:20 +00:00
<< " piece " << toString(current_piece_number)
<< ", but destination table contains " << count << " rows"
<< ". Partition will be dropped and refilled.");
2020-02-19 15:01:08 +00:00
create_is_dirty_node(clean_state_clock);
2020-03-18 13:25:49 +00:00
return TaskStatus::Error;
2020-02-19 15:01:08 +00:00
}
}
2020-02-20 17:26:20 +00:00
zookeeper->set(partition_piece.getPartitionPieceCleanStartPath(), "ok");
2020-02-19 15:01:08 +00:00
}
/// At this point, we need to sync that the destination table is clean
/// before any actual work
2020-02-19 15:01:08 +00:00
/// Try start processing, create node about it
{
String start_state = TaskStateWithOwner::getData(TaskState::Started, host_id);
2020-02-18 13:39:22 +00:00
CleanStateClock new_clean_state_clock (zookeeper, piece_is_dirty_flag_path, piece_is_dirty_cleaned_path);
2020-02-19 15:01:08 +00:00
if (clean_state_clock != new_clean_state_clock)
{
2020-02-20 17:26:20 +00:00
LOG_INFO(log, "Partition " << task_partition.name << " piece "
<< toString(current_piece_number) << " clean state changed, cowardly bailing");
2020-03-18 13:25:49 +00:00
return TaskStatus::Error;
2020-02-19 15:01:08 +00:00
}
else if (!new_clean_state_clock.is_clean())
{
2020-02-20 17:26:20 +00:00
LOG_INFO(log, "Partition " << task_partition.name << " piece "
<< toString(current_piece_number) << " is dirty and will be dropped and refilled");
2020-02-19 15:01:08 +00:00
create_is_dirty_node(new_clean_state_clock);
2020-03-18 13:25:49 +00:00
return TaskStatus::Error;
2020-02-19 15:01:08 +00:00
}
2020-02-18 13:39:22 +00:00
zookeeper->create(current_task_piece_status_path, start_state, zkutil::CreateMode::Persistent);
}
2020-02-19 15:01:08 +00:00
/// Try create table (if not exists) on each shard
{
2020-02-18 13:39:22 +00:00
/// Define push table for current partition piece
auto database_and_table_for_current_piece= std::pair<String, String>(
task_table.table_push.first,
2020-02-20 17:26:20 +00:00
task_table.table_push.second + "_piece_" + toString(current_piece_number));
2020-02-18 13:39:22 +00:00
2020-03-03 13:15:23 +00:00
auto new_engine_push_ast = task_table.engine_push_ast;
if (task_table.isReplicatedTable())
{
new_engine_push_ast = task_table.rewriteReplicatedCreateQueryToPlain();
2020-03-03 13:15:23 +00:00
}
auto create_query_push_ast = rewriteCreateQueryStorage(
task_shard.current_pull_table_create_query,
database_and_table_for_current_piece, new_engine_push_ast);
2020-02-19 15:01:08 +00:00
create_query_push_ast->as<ASTCreateQuery &>().if_not_exists = true;
String query = queryToString(create_query_push_ast);
LOG_DEBUG(log, "Create destination tables. Query: " << query);
2020-02-18 13:39:22 +00:00
UInt64 shards = executeQueryOnCluster(task_table.cluster_push, query,
create_query_push_ast, &task_cluster->settings_push,
PoolMode::GET_MANY);
2020-02-20 17:26:20 +00:00
LOG_DEBUG(log, "Destination tables " << getQuotedTable(task_table.table_push)
<< " have been created on " << shards << " shards of " << task_table.cluster_push->getShardCount());
2020-02-19 15:01:08 +00:00
}
2020-02-19 15:01:08 +00:00
/// Do the copying
{
bool inject_fault = false;
if (copy_fault_probability > 0)
{
2020-02-19 15:01:08 +00:00
double value = std::uniform_real_distribution<>(0, 1)(task_table.task_cluster.random_engine);
inject_fault = value < copy_fault_probability;
}
2020-02-19 15:01:08 +00:00
// Select all fields
2020-03-11 19:55:27 +00:00
ASTPtr query_select_ast = get_select_query(task_shard.table_read_shard, "*", /*enable_splitting*/ true, inject_fault ? "1" : "");
2020-02-19 15:01:08 +00:00
LOG_DEBUG(log, "Executing SELECT query and pull from " << task_shard.getDescription()
2020-02-18 13:39:22 +00:00
<< " : " << queryToString(query_select_ast));
2020-02-19 15:01:08 +00:00
ASTPtr query_insert_ast;
{
2020-02-19 15:01:08 +00:00
String query;
2020-02-21 16:00:50 +00:00
query += "INSERT INTO " + getQuotedTable(split_table_for_current_piece) + " VALUES ";
2020-02-19 15:01:08 +00:00
ParserQuery p_query(query.data() + query.size());
const auto & settings = context.getSettingsRef();
query_insert_ast = parseQuery(p_query, query, settings.max_query_size, settings.max_parser_depth);
2020-02-19 15:01:08 +00:00
LOG_DEBUG(log, "Executing INSERT query: " << query);
}
try
{
/// Custom INSERT SELECT implementation
Context context_select = context;
context_select.setSettings(task_cluster->settings_pull);
2020-02-19 15:01:08 +00:00
Context context_insert = context;
context_insert.setSettings(task_cluster->settings_push);
2020-02-19 15:01:08 +00:00
BlockInputStreamPtr input;
BlockOutputStreamPtr output;
{
2020-02-19 15:01:08 +00:00
BlockIO io_select = InterpreterFactory::get(query_select_ast, context_select)->execute();
BlockIO io_insert = InterpreterFactory::get(query_insert_ast, context_insert)->execute();
2020-02-19 15:01:08 +00:00
input = io_select.in;
output = io_insert.out;
}
2020-02-19 15:01:08 +00:00
/// Fail-fast optimization to abort copying when the current clean state expires
std::future<Coordination::ExistsResponse> future_is_dirty_checker;
Stopwatch watch(CLOCK_MONOTONIC_COARSE);
constexpr UInt64 check_period_milliseconds = 500;
/// Will asynchronously check that ZooKeeper connection and is_dirty flag appearing while copying data
auto cancel_check = [&] ()
{
2020-02-19 15:01:08 +00:00
if (zookeeper->expired())
throw Exception("ZooKeeper session is expired, cancel INSERT SELECT", ErrorCodes::UNFINISHED);
2020-02-19 15:01:08 +00:00
if (!future_is_dirty_checker.valid())
2020-02-18 13:39:22 +00:00
future_is_dirty_checker = zookeeper->asyncExists(piece_is_dirty_flag_path);
2020-02-19 15:01:08 +00:00
/// check_period_milliseconds should less than average insert time of single block
/// Otherwise, the insertion will slow a little bit
if (watch.elapsedMilliseconds() >= check_period_milliseconds)
{
2020-02-19 15:01:08 +00:00
Coordination::ExistsResponse status = future_is_dirty_checker.get();
2020-02-19 15:01:08 +00:00
if (status.error != Coordination::ZNONODE)
{
2020-02-19 15:01:08 +00:00
LogicalClock dirt_discovery_epoch (status.stat.mzxid);
if (dirt_discovery_epoch == clean_state_clock.discovery_zxid)
return false;
throw Exception("Partition is dirty, cancel INSERT SELECT", ErrorCodes::UNFINISHED);
}
}
2020-02-19 15:01:08 +00:00
return false;
};
2020-02-19 15:01:08 +00:00
/// Update statistics
/// It is quite rough: bytes_copied don't take into account DROP PARTITION.
auto update_stats = [&cluster_partition] (const Block & block)
{
cluster_partition.bytes_copied += block.bytes();
cluster_partition.rows_copied += block.rows();
cluster_partition.blocks_copied += 1;
};
2020-02-19 15:01:08 +00:00
/// Main work is here
copyData(*input, *output, cancel_check, update_stats);
2020-02-19 15:01:08 +00:00
// Just in case
if (future_is_dirty_checker.valid())
future_is_dirty_checker.get();
2020-02-19 15:01:08 +00:00
if (inject_fault)
throw Exception("Copy fault injection is activated", ErrorCodes::UNFINISHED);
}
catch (...)
{
tryLogCurrentException(log, "An error occurred during copying, partition will be marked as dirty");
2020-03-12 16:48:28 +00:00
create_is_dirty_node(clean_state_clock);
2020-03-18 13:25:49 +00:00
return TaskStatus::Error;
2020-02-19 15:01:08 +00:00
}
}
2020-02-21 16:00:50 +00:00
LOG_INFO(log, "Partition " << task_partition.name << " piece "
<< toString(current_piece_number) << " copied. But not moved to original destination table.");
/// Try create original table (if not exists) on each shard
2020-03-07 00:05:49 +00:00
try
2020-02-21 16:00:50 +00:00
{
auto create_query_push_ast = rewriteCreateQueryStorage(task_shard.current_pull_table_create_query,
task_table.table_push, task_table.engine_push_ast);
create_query_push_ast->as<ASTCreateQuery &>().if_not_exists = true;
String query = queryToString(create_query_push_ast);
LOG_DEBUG(log, "Create destination tables. Query: " << query);
UInt64 shards = executeQueryOnCluster(task_table.cluster_push, query,
create_query_push_ast, &task_cluster->settings_push,
PoolMode::GET_MANY);
LOG_DEBUG(log, "Destination tables " << getQuotedTable(task_table.table_push)
<< " have been created on " << shards << " shards of " << task_table.cluster_push->getShardCount());
}
2020-03-07 00:05:49 +00:00
catch (...)
{
tryLogCurrentException(log, "Error while creating original table. Maybe we are not first.");
}
2020-02-21 16:00:50 +00:00
2020-02-19 15:01:08 +00:00
/// Finalize the processing, change state of current partition task (and also check is_dirty flag)
{
String state_finished = TaskStateWithOwner::getData(TaskState::Finished, host_id);
2020-02-18 13:39:22 +00:00
CleanStateClock new_clean_state_clock (zookeeper, piece_is_dirty_flag_path, piece_is_dirty_cleaned_path);
2020-02-19 15:01:08 +00:00
if (clean_state_clock != new_clean_state_clock)
{
2020-02-21 16:00:50 +00:00
LOG_INFO(log, "Partition " << task_partition.name << " piece "
<< toString(current_piece_number) << " clean state changed, cowardly bailing");
2020-03-18 13:25:49 +00:00
return TaskStatus::Error;
2020-02-19 15:01:08 +00:00
}
else if (!new_clean_state_clock.is_clean())
{
2020-02-21 16:00:50 +00:00
LOG_INFO(log, "Partition " << task_partition.name << " piece "
<< toString(current_piece_number) << " became dirty and will be dropped and refilled");
2020-02-19 15:01:08 +00:00
create_is_dirty_node(new_clean_state_clock);
2020-03-18 13:25:49 +00:00
return TaskStatus::Error;
2020-02-19 15:01:08 +00:00
}
2020-02-18 13:39:22 +00:00
zookeeper->set(current_task_piece_status_path, state_finished, 0);
2020-02-19 15:01:08 +00:00
}
2020-03-18 13:25:49 +00:00
return TaskStatus::Finished;
2020-02-19 15:01:08 +00:00
}
2020-02-19 15:01:08 +00:00
void ClusterCopier::dropAndCreateLocalTable(const ASTPtr & create_ast)
{
const auto & create = create_ast->as<ASTCreateQuery &>();
dropLocalTableIfExists({create.database, create.table});
2020-02-19 15:01:08 +00:00
InterpreterCreateQuery interpreter(create_ast, context);
interpreter.execute();
}
2020-02-19 15:01:08 +00:00
void ClusterCopier::dropLocalTableIfExists(const DatabaseAndTableName & table_name) const
{
auto drop_ast = std::make_shared<ASTDropQuery>();
drop_ast->if_exists = true;
drop_ast->database = table_name.first;
drop_ast->table = table_name.second;
2020-02-19 15:01:08 +00:00
InterpreterDropQuery interpreter(drop_ast, context);
interpreter.execute();
}
2020-03-13 16:25:07 +00:00
void ClusterCopier::dropHelpingTables(const TaskTable & task_table)
{
LOG_DEBUG(log, "Removing helping tables");
for (size_t current_piece_number = 0; current_piece_number < task_table.number_of_splits; ++current_piece_number)
{
DatabaseAndTableName original_table = task_table.table_push;
DatabaseAndTableName helping_table = DatabaseAndTableName(original_table.first, original_table.second + "_piece_" + toString(current_piece_number));
String query = "DROP TABLE IF EXISTS " + getQuotedTable(helping_table);
const ClusterPtr & cluster_push = task_table.cluster_push;
Settings settings_push = task_cluster->settings_push;
LOG_DEBUG(log, "Execute distributed DROP TABLE: " << query);
/// We have to drop partition_piece on each replica
UInt64 num_nodes = executeQueryOnCluster(
cluster_push, query,
nullptr,
&settings_push,
PoolMode::GET_MANY,
ClusterExecutionMode::ON_EACH_NODE);
LOG_DEBUG(log, "DROP TABLE query was successfully executed on " << toString(num_nodes) << " nodes.");
}
}
2020-03-18 18:35:58 +00:00
void ClusterCopier::dropParticularPartitionPieceFromAllHelpingTables(const TaskTable & task_table, const String & partition_name)
{
LOG_DEBUG(log, "Try drop partition partition from all helping tables.");
for (size_t current_piece_number = 0; current_piece_number < task_table.number_of_splits; ++current_piece_number)
{
DatabaseAndTableName original_table = task_table.table_push;
DatabaseAndTableName helping_table = DatabaseAndTableName(original_table.first, original_table.second + "_piece_" + toString(current_piece_number));
String query = "ALTER TABLE " + getQuotedTable(helping_table) + " DROP PARTITION " + partition_name;
const ClusterPtr & cluster_push = task_table.cluster_push;
Settings settings_push = task_cluster->settings_push;
LOG_DEBUG(log, "Execute distributed DROP PARTITION: " << query);
/// We have to drop partition_piece on each replica
UInt64 num_nodes = executeQueryOnCluster(
cluster_push, query,
nullptr,
&settings_push,
PoolMode::GET_MANY,
ClusterExecutionMode::ON_EACH_NODE);
LOG_DEBUG(log, "DROP PARTITION query was successfully executed on " << toString(num_nodes) << " nodes.");
}
LOG_DEBUG(log, "All helping tables dropped partition " << partition_name);
}
2020-02-19 15:01:08 +00:00
String ClusterCopier::getRemoteCreateTable(const DatabaseAndTableName & table, Connection & connection, const Settings * settings)
{
String query = "SHOW CREATE TABLE " + getQuotedTable(table);
Block block = getBlockWithAllStreamData(std::make_shared<RemoteBlockInputStream>(
2020-02-18 13:39:22 +00:00
connection, query, InterpreterShowCreateQuery::getSampleBlock(), context, settings));
2020-02-19 15:01:08 +00:00
return typeid_cast<const ColumnString &>(*block.safeGetByPosition(0).column).getDataAt(0).toString();
}
2020-02-19 15:01:08 +00:00
ASTPtr ClusterCopier::getCreateTableForPullShard(const ConnectionTimeouts & timeouts, TaskShard & task_shard)
{
/// Fetch and parse (possibly) new definition
2020-03-18 03:27:32 +00:00
auto connection_entry = task_shard.info.pool->get(timeouts, &task_cluster->settings_pull, true);
2020-02-19 15:01:08 +00:00
String create_query_pull_str = getRemoteCreateTable(
2020-02-18 13:39:22 +00:00
task_shard.task_table.table_pull,
*connection_entry,
&task_cluster->settings_pull);
2020-02-19 15:01:08 +00:00
ParserCreateQuery parser_create_query;
const auto & settings = context.getSettingsRef();
return parseQuery(parser_create_query, create_query_pull_str, settings.max_query_size, settings.max_parser_depth);
2020-02-19 15:01:08 +00:00
}
2020-02-18 13:39:22 +00:00
/// If it is implicitly asked to create split Distributed table for certain piece on current shard, we will do it.
2020-02-20 10:01:02 +00:00
void ClusterCopier::createShardInternalTables(const ConnectionTimeouts & timeouts,
2020-02-21 16:00:50 +00:00
TaskShard & task_shard, bool create_split)
{
2020-02-19 15:01:08 +00:00
TaskTable & task_table = task_shard.task_table;
2020-02-19 15:01:08 +00:00
/// We need to update table definitions for each part, it could be changed after ALTER
task_shard.current_pull_table_create_query = getCreateTableForPullShard(timeouts, task_shard);
2020-02-19 15:01:08 +00:00
/// Create local Distributed tables:
/// a table fetching data from current shard and a table inserting data to the whole destination cluster
String read_shard_prefix = ".read_shard_" + toString(task_shard.indexInCluster()) + ".";
String split_shard_prefix = ".split.";
task_shard.table_read_shard = DatabaseAndTableName(working_database_name, read_shard_prefix + task_table.table_id);
2020-02-21 16:00:50 +00:00
task_shard.main_table_split_shard = DatabaseAndTableName(working_database_name, split_shard_prefix + task_table.table_id);
for (const auto & piece_number : ext::range(0, task_table.number_of_splits))
{
task_shard.list_of_split_tables_on_shard[piece_number] =
DatabaseAndTableName(working_database_name, split_shard_prefix + task_table.table_id + "_piece_" + toString(piece_number));
}
2020-02-19 15:01:08 +00:00
/// Create special cluster with single shard
String shard_read_cluster_name = read_shard_prefix + task_table.cluster_pull_name;
ClusterPtr cluster_pull_current_shard = task_table.cluster_pull->getClusterWithSingleShard(task_shard.indexInCluster());
context.setCluster(shard_read_cluster_name, cluster_pull_current_shard);
2020-02-19 15:01:08 +00:00
auto storage_shard_ast = createASTStorageDistributed(shard_read_cluster_name, task_table.table_pull.first, task_table.table_pull.second);
2020-02-19 15:01:08 +00:00
auto create_query_ast = removeAliasColumnsFromCreateQuery(task_shard.current_pull_table_create_query);
2020-02-21 16:00:50 +00:00
auto create_table_pull_ast = rewriteCreateQueryStorage(create_query_ast, task_shard.table_read_shard, storage_shard_ast);
2020-02-19 15:01:08 +00:00
dropAndCreateLocalTable(create_table_pull_ast);
2020-02-19 15:01:08 +00:00
if (create_split)
2020-02-21 16:00:50 +00:00
{
auto create_table_split_piece_ast = rewriteCreateQueryStorage(
create_query_ast,
task_shard.main_table_split_shard,
task_table.main_engine_split_ast);
2020-02-18 13:39:22 +00:00
dropAndCreateLocalTable(create_table_split_piece_ast);
2020-02-21 16:00:50 +00:00
/// Create auxilary split tables for each piece
for (const auto & piece_number : ext::range(0, task_table.number_of_splits))
{
const auto & storage_piece_split_ast = task_table.auxiliary_engine_split_asts[piece_number];
create_table_split_piece_ast = rewriteCreateQueryStorage(
create_query_ast,
task_shard.list_of_split_tables_on_shard[piece_number],
storage_piece_split_ast);
dropAndCreateLocalTable(create_table_split_piece_ast);
}
}
}
2020-02-19 15:01:08 +00:00
std::set<String> ClusterCopier::getShardPartitions(const ConnectionTimeouts & timeouts, TaskShard & task_shard)
{
2020-02-19 15:01:08 +00:00
createShardInternalTables(timeouts, task_shard, false);
2020-02-19 15:01:08 +00:00
TaskTable & task_table = task_shard.task_table;
2020-02-19 15:01:08 +00:00
String query;
{
WriteBufferFromOwnString wb;
wb << "SELECT DISTINCT " << queryToString(task_table.engine_push_partition_key_ast) << " AS partition FROM"
<< " " << getQuotedTable(task_shard.table_read_shard) << " ORDER BY partition DESC";
query = wb.str();
}
2020-02-19 15:01:08 +00:00
ParserQuery parser_query(query.data() + query.size());
const auto & settings = context.getSettingsRef();
ASTPtr query_ast = parseQuery(parser_query, query, settings.max_query_size, settings.max_parser_depth);
2020-02-19 15:01:08 +00:00
LOG_DEBUG(log, "Computing destination partition set, executing query: " << query);
Context local_context = context;
local_context.setSettings(task_cluster->settings_pull);
Block block = getBlockWithAllStreamData(InterpreterFactory::get(query_ast, local_context)->execute().in);
std::set<String> res;
if (block)
{
ColumnWithTypeAndName & column = block.getByPosition(0);
task_shard.partition_key_column = column;
for (size_t i = 0; i < column.column->size(); ++i)
{
WriteBufferFromOwnString wb;
column.type->serializeAsTextQuoted(*column.column, i, wb, FormatSettings());
res.emplace(wb.str());
}
}
LOG_DEBUG(log, "There are " << res.size() << " destination partitions in shard " << task_shard.getDescription());
2020-02-19 15:01:08 +00:00
return res;
}
2017-11-15 17:09:16 +00:00
2020-02-20 10:01:02 +00:00
bool ClusterCopier::checkShardHasPartition(const ConnectionTimeouts & timeouts,
TaskShard & task_shard, const String & partition_quoted_name)
{
2020-02-19 15:01:08 +00:00
createShardInternalTables(timeouts, task_shard, false);
TaskTable & task_table = task_shard.task_table;
std::string query = "SELECT 1 FROM " + getQuotedTable(task_shard.table_read_shard)
2020-02-20 10:01:02 +00:00
+ " WHERE (" + queryToString(task_table.engine_push_partition_key_ast) +
" = (" + partition_quoted_name + " AS partition_key))";
2020-02-19 15:01:08 +00:00
if (!task_table.where_condition_str.empty())
query += " AND (" + task_table.where_condition_str + ")";
query += " LIMIT 1";
LOG_DEBUG(log, "Checking shard " << task_shard.getDescription() << " for partition "
2020-02-18 13:39:22 +00:00
<< partition_quoted_name << " existence, executing query: " << query);
2020-02-19 15:01:08 +00:00
ParserQuery parser_query(query.data() + query.size());
const auto & settings = context.getSettingsRef();
ASTPtr query_ast = parseQuery(parser_query, query, settings.max_query_size, settings.max_parser_depth);
2020-02-19 15:01:08 +00:00
Context local_context = context;
local_context.setSettings(task_cluster->settings_pull);
return InterpreterFactory::get(query_ast, local_context)->execute().in->read().rows() != 0;
}
2020-02-18 13:39:22 +00:00
bool ClusterCopier::checkPresentPartitionPiecesOnCurrentShard(const ConnectionTimeouts & timeouts,
2020-02-20 10:01:02 +00:00
TaskShard & task_shard, const String & partition_quoted_name, size_t current_piece_number)
2020-02-18 13:39:22 +00:00
{
createShardInternalTables(timeouts, task_shard, false);
TaskTable & task_table = task_shard.task_table;
std::string query = "SELECT 1 FROM " + getQuotedTable(task_shard.table_read_shard)
2020-02-20 10:01:02 +00:00
+ " WHERE (" + queryToString(task_table.engine_push_partition_key_ast)
+ " = (" + partition_quoted_name + " AS partition_key))";
2020-02-18 13:39:22 +00:00
const size_t number_of_splits = task_table.number_of_splits;
const String & primary_key_comma_separated = task_table.primary_key_comma_separated;
query += " AND (cityHash64(" + primary_key_comma_separated + ") % "
+ std::to_string(number_of_splits) + " = " + std::to_string(current_piece_number) + " )";
if (!task_table.where_condition_str.empty())
query += " AND (" + task_table.where_condition_str + ")";
query += " LIMIT 1";
LOG_DEBUG(log, "Checking shard " << task_shard.getDescription() << " for partition "
2020-03-17 18:07:54 +00:00
<< partition_quoted_name << " piece " << std::to_string(current_piece_number)
<< "existence, executing query: " << query);
2020-02-18 13:39:22 +00:00
ParserQuery parser_query(query.data() + query.size());
const auto & settings = context.getSettingsRef();
ASTPtr query_ast = parseQuery(parser_query, query, settings.max_query_size, settings.max_parser_depth);
2020-02-18 13:39:22 +00:00
Context local_context = context;
local_context.setSettings(task_cluster->settings_pull);
auto result = InterpreterFactory::get(query_ast, local_context)->execute().in->read().rows();
if (result != 0)
LOG_DEBUG(log, "Partition " << partition_quoted_name << " piece number "
2020-03-17 18:07:54 +00:00
<< std::to_string(current_piece_number) << " is PRESENT on shard " << task_shard.getDescription());
2020-02-18 13:39:22 +00:00
else
LOG_DEBUG(log, "Partition " << partition_quoted_name << " piece number "
2020-03-17 18:07:54 +00:00
<< std::to_string(current_piece_number) << " is ABSENT on shard " << task_shard.getDescription());
2020-02-18 13:39:22 +00:00
return result != 0;
}
2020-02-18 13:39:22 +00:00
/** Executes simple query (without output streams, for example DDL queries) on each shard of the cluster
* Returns number of shards for which at least one replica executed query successfully
*/
2020-02-19 15:01:08 +00:00
UInt64 ClusterCopier::executeQueryOnCluster(
2020-02-18 13:39:22 +00:00
const ClusterPtr & cluster,
const String & query,
2020-02-20 10:01:02 +00:00
const ASTPtr & query_ast_,
const Settings * settings,
PoolMode pool_mode,
2020-03-10 20:04:08 +00:00
ClusterExecutionMode execution_mode,
2020-02-20 10:01:02 +00:00
UInt64 max_successful_executions_per_shard) const
{
2020-02-19 15:01:08 +00:00
auto num_shards = cluster->getShardsInfo().size();
std::vector<UInt64> per_shard_num_successful_replicas(num_shards, 0);
2020-02-19 15:01:08 +00:00
ASTPtr query_ast;
if (query_ast_ == nullptr)
{
2020-02-19 15:01:08 +00:00
ParserQuery p_query(query.data() + query.size());
const auto & settings = context.getSettingsRef();
query_ast = parseQuery(p_query, query, settings.max_query_size, settings.max_parser_depth);
}
2020-02-19 15:01:08 +00:00
else
query_ast = query_ast_;
2020-03-10 20:04:08 +00:00
/// We will have to execute query on each replica of a shard.
if (execution_mode == ClusterExecutionMode::ON_EACH_NODE)
max_successful_executions_per_shard = 0;
std::atomic<size_t> origin_replicas_number;
2020-02-19 15:01:08 +00:00
/// We need to execute query on one replica at least
auto do_for_shard = [&] (UInt64 shard_index)
{
2020-02-19 15:01:08 +00:00
const Cluster::ShardInfo & shard = cluster->getShardsInfo().at(shard_index);
UInt64 & num_successful_executions = per_shard_num_successful_replicas.at(shard_index);
num_successful_executions = 0;
2020-02-18 13:39:22 +00:00
auto increment_and_check_exit = [&] () -> bool
2020-02-19 15:01:08 +00:00
{
++num_successful_executions;
return max_successful_executions_per_shard && num_successful_executions >= max_successful_executions_per_shard;
};
2020-02-19 15:01:08 +00:00
UInt64 num_replicas = cluster->getShardsAddresses().at(shard_index).size();
2020-03-10 20:04:08 +00:00
origin_replicas_number += num_replicas;
2020-02-19 15:01:08 +00:00
UInt64 num_local_replicas = shard.getLocalNodeCount();
UInt64 num_remote_replicas = num_replicas - num_local_replicas;
2020-02-19 15:01:08 +00:00
/// In that case we don't have local replicas, but do it just in case
for (UInt64 i = 0; i < num_local_replicas; ++i)
{
auto interpreter = InterpreterFactory::get(query_ast, context);
interpreter->execute();
2020-02-19 15:01:08 +00:00
if (increment_and_check_exit())
return;
}
2020-02-19 15:01:08 +00:00
/// Will try to make as many as possible queries
if (shard.hasRemoteConnections())
{
Settings current_settings = settings ? *settings : task_cluster->settings_common;
current_settings.max_parallel_replicas = num_remote_replicas ? num_remote_replicas : 1;
auto timeouts = ConnectionTimeouts::getTCPTimeoutsWithFailover(current_settings).getSaturated(current_settings.max_execution_time);
auto connections = shard.pool->getMany(timeouts, &current_settings, pool_mode);
for (auto & connection : connections)
{
if (connection.isNull())
continue;
try
{
/// CREATE TABLE and DROP PARTITION queries return empty block
RemoteBlockInputStream stream{*connection, query, Block{}, context, &current_settings};
NullBlockOutputStream output{Block{}};
copyData(stream, output);
if (increment_and_check_exit())
return;
}
catch (const Exception &)
{
LOG_INFO(log, getCurrentExceptionMessage(false, true));
}
}
}
};
{
2020-02-19 15:01:08 +00:00
ThreadPool thread_pool(std::min<UInt64>(num_shards, getNumberOfPhysicalCPUCores()));
for (UInt64 shard_index = 0; shard_index < num_shards; ++shard_index)
thread_pool.scheduleOrThrowOnError([=] { do_for_shard(shard_index); });
2020-02-19 15:01:08 +00:00
thread_pool.wait();
}
2020-02-19 15:01:08 +00:00
2020-03-10 20:04:08 +00:00
UInt64 successful_nodes = 0;
2020-02-19 15:01:08 +00:00
for (UInt64 num_replicas : per_shard_num_successful_replicas)
2020-03-10 20:04:08 +00:00
{
if (execution_mode == ClusterExecutionMode::ON_EACH_NODE)
successful_nodes += num_replicas;
else
/// Count only successful shards
successful_nodes += (num_replicas > 0);
}
2020-02-19 15:01:08 +00:00
2020-03-11 19:55:27 +00:00
if (execution_mode == ClusterExecutionMode::ON_EACH_NODE && successful_nodes != origin_replicas_number)
2020-03-10 20:04:08 +00:00
{
LOG_INFO(log, "There was an error while executing ALTER on each node. Query was executed on "
<< toString(successful_nodes) << " nodes. But had to be executed on " << toString(origin_replicas_number.load()));
}
2020-02-19 15:01:08 +00:00
2020-02-18 13:39:22 +00:00
2020-03-10 20:04:08 +00:00
return successful_nodes;
2020-02-18 13:39:22 +00:00
}
}