ClickHouse/src/Interpreters/DDLWorker.cpp

#include <filesystem>

#include <Interpreters/DDLWorker.h>
#include <Interpreters/DDLTask.h>
#include <Parsers/ASTAlterQuery.h>
#include <Parsers/ASTDropQuery.h>
#include <Parsers/ASTOptimizeQuery.h>
#include <Parsers/ASTQueryWithOnCluster.h>
#include <Parsers/ASTQueryWithTableAndOutput.h>
#include <Parsers/ParserQuery.h>
#include <Parsers/parseQuery.h>
#include <Parsers/queryToString.h>
#include <IO/WriteHelpers.h>
#include <IO/ReadHelpers.h>
#include <IO/ReadBufferFromString.h>
#include <Storages/IStorage.h>
#include <Interpreters/executeQuery.h>
#include <Interpreters/Cluster.h>
#include <Interpreters/Context.h>
#include <Common/setThreadName.h>
#include <Common/randomSeed.h>
#include <Common/ZooKeeper/ZooKeeper.h>
#include <Common/ZooKeeper/KeeperException.h>
#include <Common/isLocalAddress.h>
#include <Storages/StorageReplicatedMergeTree.h>
#include <Poco/Timestamp.h>
#include <common/sleep.h>
#include <common/getFQDNOrHostName.h>
#include <common/logger_useful.h>
#include <random>
#include <pcg_random.hpp>
#include <common/scope_guard_safe.h>

#include <Interpreters/ZooKeeperLog.h>

namespace fs = std::filesystem;


namespace DB
{

namespace ErrorCodes
{
    extern const int LOGICAL_ERROR;
    extern const int TIMEOUT_EXCEEDED;
    extern const int UNFINISHED;
    extern const int NOT_A_LEADER;
    extern const int KEEPER_EXCEPTION;
    extern const int CANNOT_ASSIGN_ALTER;
    extern const int CANNOT_ALLOCATE_MEMORY;
    extern const int MEMORY_LIMIT_EXCEEDED;
}

constexpr const char * TASK_PROCESSED_OUT_REASON = "Task has been already processed";

namespace
{

/** Caveats: usage of locks in ZooKeeper is incorrect in 99% of cases,
  *  and highlights your poor understanding of distributed systems.
  *
  * It's only correct if all the operations that are performed under lock
  *  are atomically checking that the lock still holds
  *  or if we ensure that these operations will be undone if lock is lost
  *  (due to ZooKeeper session loss) that's very difficult to achieve.
  *
  * It's Ok if every operation that we perform under lock is actually operation in ZooKeeper.
  *
  * In 1% of cases when you can correctly use Lock, the logic is complex enough, so you don't need this class.
  *
  * TLDR: Don't use this code.
  * We only have a few cases of it's usage and it will be removed.
  */
class ZooKeeperLock
{
public:
    /// lock_prefix - path where the ephemeral lock node will be created
    /// lock_name - the name of the ephemeral lock node
    ZooKeeperLock(
        const zkutil::ZooKeeperPtr & zookeeper_,
        const std::string & lock_prefix_,
        const std::string & lock_name_,
        const std::string & lock_message_ = "")
    :
        zookeeper(zookeeper_),
        lock_path(fs::path(lock_prefix_) / lock_name_),
        lock_message(lock_message_),
        log(&Poco::Logger::get("zkutil::Lock"))
    {
        zookeeper->createIfNotExists(lock_prefix_, "");
    }

    ~ZooKeeperLock()
    {
        try
        {
            unlock();
        }
        catch (...)
        {
            DB::tryLogCurrentException(__PRETTY_FUNCTION__);
        }
    }

    void unlock()
    {
        Coordination::Stat stat;
        std::string dummy;
        bool result = zookeeper->tryGet(lock_path, dummy, &stat);

        if (result && stat.ephemeralOwner == zookeeper->getClientID())
            zookeeper->remove(lock_path, -1);
        else
            LOG_WARNING(log, "Lock is lost. It is normal if session was expired. Path: {}/{}", lock_path, lock_message);
    }

    bool tryLock()
    {
        std::string dummy;
        Coordination::Error code = zookeeper->tryCreate(lock_path, lock_message, zkutil::CreateMode::Ephemeral, dummy);

        if (code == Coordination::Error::ZNODEEXISTS)
        {
            return false;
        }
        else if (code == Coordination::Error::ZOK)
        {
            return true;
        }
        else
        {
            throw Coordination::Exception(code);
        }
    }

private:
    zkutil::ZooKeeperPtr zookeeper;

    std::string lock_path;
    std::string lock_message;
    Poco::Logger * log;

};

std::unique_ptr<ZooKeeperLock> createSimpleZooKeeperLock(
    const zkutil::ZooKeeperPtr & zookeeper, const String & lock_prefix, const String & lock_name, const String & lock_message)
{
    return std::make_unique<ZooKeeperLock>(zookeeper, lock_prefix, lock_name, lock_message);
}

}


DDLWorker::DDLWorker(
    int pool_size_,
    const std::string & zk_root_dir,
    ContextPtr context_,
    const Poco::Util::AbstractConfiguration * config,
    const String & prefix,
    const String & logger_name,
    const CurrentMetrics::Metric * max_entry_metric_,
    const CurrentMetrics::Metric * max_pushed_entry_metric_)
    : context(Context::createCopy(context_))
    , log(&Poco::Logger::get(logger_name))
    , pool_size(pool_size_)
    , max_entry_metric(max_entry_metric_)
    , max_pushed_entry_metric(max_pushed_entry_metric_)
{
    if (max_entry_metric)
        CurrentMetrics::set(*max_entry_metric, 0);

    if (max_pushed_entry_metric)
        CurrentMetrics::set(*max_pushed_entry_metric, 0);

    if (1 < pool_size)
    {
        LOG_WARNING(log, "DDLWorker is configured to use multiple threads. "
                         "It's not recommended because queries can be reordered. Also it may cause some unknown issues to appear.");
        worker_pool = std::make_unique<ThreadPool>(pool_size);
    }

    queue_dir = zk_root_dir;
    if (queue_dir.back() == '/')
        queue_dir.resize(queue_dir.size() - 1);

    if (config)
    {
        task_max_lifetime = config->getUInt64(prefix + ".task_max_lifetime", static_cast<UInt64>(task_max_lifetime));
        cleanup_delay_period = config->getUInt64(prefix + ".cleanup_delay_period", static_cast<UInt64>(cleanup_delay_period));
        max_tasks_in_queue = std::max<UInt64>(1, config->getUInt64(prefix + ".max_tasks_in_queue", max_tasks_in_queue));

        if (config->has(prefix + ".profile"))
            context->setSetting("profile", config->getString(prefix + ".profile"));
    }

    if (context->getSettingsRef().readonly)
    {
        LOG_WARNING(log, "Distributed DDL worker is run with readonly settings, it will not be able to execute DDL queries Set appropriate system_profile or distributed_ddl.profile to fix this.");
    }

    host_fqdn = getFQDNOrHostName();
    host_fqdn_id = Cluster::Address::toString(host_fqdn, context->getTCPPort());
}

void DDLWorker::startup()
{
    main_thread = ThreadFromGlobalPool(&DDLWorker::runMainThread, this);
    cleanup_thread = ThreadFromGlobalPool(&DDLWorker::runCleanupThread, this);
}

void DDLWorker::shutdown()
{
    bool prev_stop_flag = stop_flag.exchange(true);
    if (!prev_stop_flag)
    {
        queue_updated_event->set();
        cleanup_event->set();
        main_thread.join();
        cleanup_thread.join();
        worker_pool.reset();
    }
}

DDLWorker::~DDLWorker()
{
    DDLWorker::shutdown();
}


ZooKeeperPtr DDLWorker::tryGetZooKeeper() const
{
    std::lock_guard lock(zookeeper_mutex);
    return current_zookeeper;
}

ZooKeeperPtr DDLWorker::getAndSetZooKeeper()
{
    std::lock_guard lock(zookeeper_mutex);

    if (!current_zookeeper || current_zookeeper->expired())
        current_zookeeper = context->getZooKeeper();

    return current_zookeeper;
}


DDLTaskPtr DDLWorker::initAndCheckTask(const String & entry_name, String & out_reason, const ZooKeeperPtr & zookeeper)
{
    String node_data;
    String entry_path = fs::path(queue_dir) / entry_name;

    auto task = std::make_unique<DDLTask>(entry_name, entry_path);

    if (!zookeeper->tryGet(entry_path, node_data))
    {
        /// It is Ok that node could be deleted just now. It means that there are no current host in node's host list.
        out_reason = "The task was deleted";
        return {};
    }

    auto write_error_status = [&](const String & host_id, const String & error_message, const String & reason)
    {
        LOG_ERROR(log, "Cannot parse DDL task {}: {}. Will try to send error status: {}", entry_name, reason, error_message);
        createStatusDirs(entry_path, zookeeper);
        zookeeper->tryCreate(fs::path(entry_path) / "finished" / host_id, error_message, zkutil::CreateMode::Persistent);
    };

    try
    {
        /// Stage 1: parse entry
        task->entry.parse(node_data);
    }
    catch (...)
    {
        /// What should we do if we even cannot parse host name and therefore cannot properly submit execution status?
        /// We can try to create fail node using FQDN if it equal to host name in cluster config attempt will be successful.
        /// Otherwise, that node will be ignored by DDLQueryStatusInputStream.
        out_reason = "Incorrect task format";
        write_error_status(host_fqdn_id, ExecutionStatus::fromCurrentException().serializeText(), out_reason);
        return {};
    }

    /// Stage 2: resolve host_id and check if we should execute query or not
    /// Multiple clusters can use single DDL queue path in ZooKeeper,
    /// So we should skip task if we cannot find current host in cluster hosts list.
    if (!task->findCurrentHostID(context, log))
    {
        out_reason = "There is no a local address in host list";
        return {};
    }

    try
    {
        /// Stage 3.1: parse query
        task->parseQueryFromEntry(context);
        /// Stage 3.2: check cluster and find the host in cluster
        task->setClusterInfo(context, log);
    }
    catch (...)
    {
        out_reason = "Cannot parse query or obtain cluster info";
        write_error_status(task->host_id_str, ExecutionStatus::fromCurrentException().serializeText(), out_reason);
        return {};
    }

    if (zookeeper->exists(task->getFinishedNodePath()))
    {
        out_reason = TASK_PROCESSED_OUT_REASON;
        return {};
    }

    /// Now task is ready for execution
    return task;
}


static void filterAndSortQueueNodes(Strings & all_nodes)
{
    all_nodes.erase(std::remove_if(all_nodes.begin(), all_nodes.end(), [] (const String & s) { return !startsWith(s, "query-"); }), all_nodes.end());
    std::sort(all_nodes.begin(), all_nodes.end());
}

void DDLWorker::scheduleTasks(bool reinitialized)
{
    LOG_DEBUG(log, "Scheduling tasks");
    auto zookeeper = tryGetZooKeeper();

    /// Main thread of DDLWorker was restarted, probably due to lost connection with ZooKeeper.
    /// We have some unfinished tasks.
    /// To avoid duplication of some queries we should try to write execution status again.
    /// To avoid skipping of some entries which were not executed we should be careful when choosing begin_node to start from.
    /// NOTE: It does not protect from all cases of query duplication, see also comments in processTask(...)
    if (reinitialized)
    {
        if (current_tasks.empty())
            LOG_TRACE(log, "Don't have unfinished tasks after restarting");
        else
            LOG_INFO(log, "Have {} unfinished tasks, will check them", current_tasks.size());
        assert(current_tasks.size() <= pool_size + (worker_pool != nullptr));
        auto task_it = current_tasks.begin();
        while (task_it != current_tasks.end())
        {
            auto & task = *task_it;
            if (task->completely_processed)
            {
                assert(task->was_executed);
                /// Status must be written (but finished/ node may not exist if entry was deleted).
                /// If someone is deleting entry concurrently, then /active status dir must not exist.
                assert(zookeeper->exists(task->getFinishedNodePath()) || !zookeeper->exists(fs::path(task->entry_path) / "active"));
                ++task_it;
            }
            else if (task->was_executed)
            {
                /// Connection was lost on attempt to write status. Will retry.
                bool status_written = zookeeper->exists(task->getFinishedNodePath());
                /// You might think that the following condition is redundant, because status_written implies completely_processed.
                /// But it's wrong. It's possible that (!task->completely_processed && status_written)
                /// if ZooKeeper successfully received and processed our request
                /// but we lost connection while waiting for the response.
                /// Yeah, distributed systems is a zoo.
                if (status_written)
                    task->completely_processed = true;
                else
                    processTask(*task, zookeeper);
                ++task_it;
            }
            else
            {
                /// We didn't even executed a query, so let's just remove it.
                /// We will try to read the task again and execute it from the beginning.
                /// NOTE: We can safely compare entry names as Strings, because they are padded.
                /// Entry name always starts with "query-" and contain exactly 10 decimal digits
                /// of log entry number (with leading zeros).
                if (!first_failed_task_name || task->entry_name < *first_failed_task_name)
                    first_failed_task_name = task->entry_name;
                task_it = current_tasks.erase(task_it);
            }
        }
    }

    Strings queue_nodes = zookeeper->getChildren(queue_dir, &queue_node_stat, queue_updated_event);
    size_t size_before_filtering = queue_nodes.size();
    filterAndSortQueueNodes(queue_nodes);
    /// The following message is too verbose, but it can be useful too debug mysterious test failures in CI
    LOG_TRACE(log, "scheduleTasks: initialized={}, size_before_filtering={}, queue_size={}, "
                   "entries={}..{}, "
                   "first_failed_task_name={}, current_tasks_size={}, "
                   "last_current_task={}, "
                   "last_skipped_entry_name={}",
                   initialized, size_before_filtering, queue_nodes.size(),
                   queue_nodes.empty() ? "none" : queue_nodes.front(), queue_nodes.empty() ? "none" : queue_nodes.back(),
                   first_failed_task_name ? *first_failed_task_name : "none", current_tasks.size(),
                   current_tasks.empty() ? "none" : current_tasks.back()->entry_name,
                   last_skipped_entry_name ? *last_skipped_entry_name : "none");

    if (max_tasks_in_queue < queue_nodes.size())
        cleanup_event->set();

    /// Detect queue start, using:
    /// - skipped tasks
    /// - in memory tasks (that are currently active or were finished recently)
    /// - failed tasks (that should be processed again)
    auto begin_node = queue_nodes.begin();
    if (first_failed_task_name)
    {
        /// If we had failed tasks, then we should start from the first failed task.
        assert(reinitialized);
        begin_node = std::lower_bound(queue_nodes.begin(), queue_nodes.end(), first_failed_task_name);
    }
    else
    {
        /// We had no failed tasks. Let's just choose the maximum entry we have previously seen.
        String last_task_name;
        if (!current_tasks.empty())
            last_task_name = current_tasks.back()->entry_name;
        if (last_skipped_entry_name && last_task_name < *last_skipped_entry_name)
            last_task_name = *last_skipped_entry_name;
        begin_node = std::upper_bound(queue_nodes.begin(), queue_nodes.end(), last_task_name);
    }

    if (begin_node == queue_nodes.end())
        LOG_DEBUG(log, "No tasks to schedule");
    else
        LOG_DEBUG(log, "Will schedule {} tasks starting from {}", std::distance(begin_node, queue_nodes.end()), *begin_node);

    /// Let's ensure that it's exactly the first task we should process.
    /// Maybe such asserts are too paranoid and excessive,
    /// but it's easy enough to break DDLWorker in a very unobvious way by making some minor change in code.
    [[maybe_unused]] bool have_no_tasks_info = !first_failed_task_name && current_tasks.empty() && !last_skipped_entry_name;
    assert(have_no_tasks_info || queue_nodes.end() == std::find_if(queue_nodes.begin(), queue_nodes.end(), [&](const String & entry_name)
    {
        /// We should return true if some invariants are violated.
        String reason;
        auto task = initAndCheckTask(entry_name, reason, zookeeper);
        bool maybe_currently_processing = current_tasks.end() != std::find_if(current_tasks.begin(), current_tasks.end(), [&](const auto & t)
        {
            return t->entry_name == entry_name;
        });
        /// begin_node is something like a log pointer
        if (begin_node == queue_nodes.end() || entry_name < *begin_node)
        {
            /// Return true if entry should be scheduled.
            /// There is a minor race condition: initAndCheckTask(...) may return not null
            /// if someone is deleting outdated entry right now (including finished/ nodes), so we also check active/ status dir.
            bool maybe_concurrently_deleting = task && !zookeeper->exists(fs::path(task->entry_path) / "active");
            return task && !maybe_concurrently_deleting && !maybe_currently_processing;
        }
        else
        {
            /// Return true if entry should not be scheduled.
            bool processed = !task && reason == TASK_PROCESSED_OUT_REASON;
            return processed || maybe_currently_processing;
        }
    }));

    for (auto it = begin_node; it != queue_nodes.end() && !stop_flag; ++it)
    {
        String entry_name = *it;
        LOG_TRACE(log, "Checking task {}", entry_name);

        String reason;
        auto task = initAndCheckTask(entry_name, reason, zookeeper);
        if (!task)
        {
            LOG_DEBUG(log, "Will not execute task {}: {}", entry_name, reason);
            updateMaxDDLEntryID(entry_name);
            last_skipped_entry_name.emplace(entry_name);
            continue;
        }

        auto & saved_task = saveTask(std::move(task));

        if (worker_pool)
        {
            worker_pool->scheduleOrThrowOnError([this, &saved_task, zookeeper]()
            {
                setThreadName("DDLWorkerExec");
                processTask(saved_task, zookeeper);
            });
        }
        else
        {
            processTask(saved_task, zookeeper);
        }
    }
}

DDLTaskBase & DDLWorker::saveTask(DDLTaskPtr && task)
{
    current_tasks.remove_if([](const DDLTaskPtr & t) { return t->completely_processed.load(); });
    /// Tasks are scheduled and executed in main thread <==> Parallel execution is disabled
    assert((worker_pool != nullptr) == (1 < pool_size));
    /// Parallel execution is disabled ==> All previous tasks are failed to start or finished,
    /// so current tasks list must be empty when we are ready to process new one.
    assert(worker_pool || current_tasks.empty());
    /// Parallel execution is enabled ==> Not more than pool_size tasks are currently executing.
    /// Note: If current_tasks.size() == pool_size, then all worker threads are busy,
    /// so we will wait on worker_pool->scheduleOrThrowOnError(...)
    assert(!worker_pool || current_tasks.size() <= pool_size);
    current_tasks.emplace_back(std::move(task));
    if (first_failed_task_name && *first_failed_task_name == current_tasks.back()->entry_name)
        first_failed_task_name.reset();
    return *current_tasks.back();
}

bool DDLWorker::tryExecuteQuery(const String & query, DDLTaskBase & task, const ZooKeeperPtr & zookeeper)
{
    /// Add special comment at the start of query to easily identify DDL-produced queries in query_log
    String query_prefix = "/* ddl_entry=" + task.entry_name + " */ ";
    String query_to_execute = query_prefix + query;

    ReadBufferFromString istr(query_to_execute);
    String dummy_string;
    WriteBufferFromString ostr(dummy_string);
    std::optional<CurrentThread::QueryScope> query_scope;

    try
    {
        auto query_context = task.makeQueryContext(context, zookeeper);
        if (!task.is_initial_query)
            query_scope.emplace(query_context);
        executeQuery(istr, ostr, !task.is_initial_query, query_context, {});

        if (auto txn = query_context->getZooKeeperMetadataTransaction())
        {
            /// Most queries commit changes to ZooKeeper right before applying local changes,
            /// but some queries does not support it, so we have to do it here.
            if (!txn->isExecuted())
                txn->commit();
        }
    }
    catch (const DB::Exception & e)
    {
        if (task.is_initial_query)
            throw;

        task.execution_status = ExecutionStatus::fromCurrentException();
        tryLogCurrentException(log, "Query " + query + " wasn't finished successfully");

        /// We use return value of tryExecuteQuery(...) in tryExecuteQueryOnLeaderReplica(...) to determine
        /// if replica has stopped being leader and we should retry query.
        /// However, for the majority of exceptions there is no sense to retry, because most likely we will just
        /// get the same exception again. So we return false only for several special exception codes,
        /// and consider query as executed with status "failed" and return true in other cases.
        bool no_sense_to_retry = e.code() != ErrorCodes::KEEPER_EXCEPTION &&
                                 e.code() != ErrorCodes::NOT_A_LEADER &&
                                 e.code() != ErrorCodes::CANNOT_ASSIGN_ALTER &&
                                 e.code() != ErrorCodes::CANNOT_ALLOCATE_MEMORY &&
                                 e.code() != ErrorCodes::MEMORY_LIMIT_EXCEEDED;
        return no_sense_to_retry;
    }
    catch (...)
    {
        if (task.is_initial_query)
            throw;

        task.execution_status = ExecutionStatus::fromCurrentException();
        tryLogCurrentException(log, "Query " + query + " wasn't finished successfully");

        /// We don't know what exactly happened, but maybe it's Poco::NetException or std::bad_alloc,
        /// so we consider unknown exception as retryable error.
        return false;
    }

    task.execution_status = ExecutionStatus(0);
    LOG_DEBUG(log, "Executed query: {}", query);

    return true;
}

void DDLWorker::updateMaxDDLEntryID(const String & entry_name)
{
    UInt64 id = DDLTaskBase::getLogEntryNumber(entry_name);
    auto prev_id = max_id.load(std::memory_order_relaxed);
    while (prev_id < id)
    {
        if (max_id.compare_exchange_weak(prev_id, id))
        {
            if (max_entry_metric)
                CurrentMetrics::set(*max_entry_metric, id);
            break;
        }
    }
}

void DDLWorker::processTask(DDLTaskBase & task, const ZooKeeperPtr & zookeeper)
{
    LOG_DEBUG(log, "Processing task {} ({})", task.entry_name, task.entry.query);
    assert(!task.completely_processed);

    String active_node_path = task.getActiveNodePath();
    String finished_node_path = task.getFinishedNodePath();

    /// Step 1: Create ephemeral node in active/ status dir.
    /// It allows other hosts to understand that task is currently executing (useful for system.distributed_ddl_queue)
    /// and protects from concurrent deletion or the task.

    /// It will tryRemove(...) on exception
    auto active_node = zkutil::EphemeralNodeHolder::existing(active_node_path, *zookeeper);

    /// Try fast path
    auto create_active_res = zookeeper->tryCreate(active_node_path, {}, zkutil::CreateMode::Ephemeral);
    if (create_active_res != Coordination::Error::ZOK)
    {
        if (create_active_res != Coordination::Error::ZNONODE && create_active_res != Coordination::Error::ZNODEEXISTS)
        {
            assert(Coordination::isHardwareError(create_active_res));
            throw Coordination::Exception(create_active_res, active_node_path);
        }

        /// Status dirs were not created in enqueueQuery(...) or someone is removing entry
        if (create_active_res == Coordination::Error::ZNONODE)
        {
            assert(dynamic_cast<DatabaseReplicatedTask *>(&task) == nullptr);
            if (task.was_executed)
            {
                /// Special case:
                /// Task was executed (and we are trying to write status after connection loss) ==> Status dirs were previously created.
                /// (Status dirs were previously created AND active/ does not exist) ==> Task was removed.
                /// We cannot write status, but it's not required anymore, because no one will try to execute it again.
                /// So we consider task as completely processed.
                LOG_WARNING(log, "Task {} is executed, but looks like entry {} was deleted, cannot write status", task.entry_name, task.entry_path);
                task.completely_processed = true;
                return;
            }
            createStatusDirs(task.entry_path, zookeeper);
        }

        if (create_active_res == Coordination::Error::ZNODEEXISTS)
        {
            /// Connection has been lost and now we are retrying,
            /// but our previous ephemeral node still exists.
            zkutil::EventPtr eph_node_disappeared = std::make_shared<Poco::Event>();
            String dummy;
            if (zookeeper->tryGet(active_node_path, dummy, nullptr, eph_node_disappeared))
            {
                constexpr int timeout_ms = 30 * 1000;
                if (!eph_node_disappeared->tryWait(timeout_ms))
                    throw Exception(ErrorCodes::LOGICAL_ERROR, "Ephemeral node {} still exists, "
                                    "probably it's owned by someone else", active_node_path);
            }
        }

        zookeeper->create(active_node_path, {}, zkutil::CreateMode::Ephemeral);
    }

    /// Step 2: Execute query from the task.
    if (!task.was_executed)
    {
        /// If table and database engine supports it, they will execute task.ops by their own in a single transaction
        /// with other zk operations (such as appending something to ReplicatedMergeTree log, or
        /// updating metadata in Replicated database), so we make create request for finished_node_path with status "0",
        /// which means that query executed successfully.
        task.ops.emplace_back(zkutil::makeRemoveRequest(active_node_path, -1));
        task.ops.emplace_back(zkutil::makeCreateRequest(finished_node_path, ExecutionStatus(0).serializeText(), zkutil::CreateMode::Persistent));

        try
        {
            String rewritten_query = queryToString(task.query);
            LOG_DEBUG(log, "Executing query: {}", rewritten_query);

            StoragePtr storage;
            if (auto * query_with_table = dynamic_cast<ASTQueryWithTableAndOutput *>(task.query.get()); query_with_table)
            {
                if (!query_with_table->table.empty())
                {
                    /// It's not CREATE DATABASE
                    auto table_id = context->tryResolveStorageID(*query_with_table, Context::ResolveOrdinary);
                    storage = DatabaseCatalog::instance().tryGetTable(table_id, context);
                }

                task.execute_on_leader = storage && taskShouldBeExecutedOnLeader(task.query, storage) && !task.is_circular_replicated;
            }

            if (task.execute_on_leader)
            {
                tryExecuteQueryOnLeaderReplica(task, storage, rewritten_query, task.entry_path, zookeeper);
            }
            else
            {
                storage.reset();
                tryExecuteQuery(rewritten_query, task, zookeeper);
            }
        }
        catch (const Coordination::Exception &)
        {
            throw;
        }
        catch (...)
        {
            if (task.is_initial_query)
                throw;
            tryLogCurrentException(log, "An error occurred before execution of DDL task: ");
            task.execution_status = ExecutionStatus::fromCurrentException("An error occurred before execution");
        }

        if (task.execution_status.code != 0)
        {
            bool status_written_by_table_or_db = task.ops.empty();
            if (status_written_by_table_or_db)
            {
                throw Exception(ErrorCodes::UNFINISHED, "Unexpected error: {}", task.execution_status.serializeText());
            }
            else
            {
                /// task.ops where not executed by table or database engine, so DDLWorker is responsible for
                /// writing query execution status into ZooKeeper.
                task.ops.emplace_back(zkutil::makeSetRequest(finished_node_path, task.execution_status.serializeText(), -1));
            }
        }

        /// We need to distinguish ZK errors occurred before and after query executing
        task.was_executed = true;
    }

    updateMaxDDLEntryID(task.entry_name);

    /// Step 3: Create node in finished/ status dir and write execution status.
    /// FIXME: if server fails right here, the task will be executed twice. We need WAL here.
    /// NOTE: If ZooKeeper connection is lost here, we will try again to write query status.
    /// NOTE: If both table and database are replicated, task is executed in single ZK transaction.

    bool status_written = task.ops.empty();
    if (!status_written)
    {
        zookeeper->multi(task.ops);
        task.ops.clear();
    }

    /// Active node was removed in multi ops
    active_node->setAlreadyRemoved();

    task.completely_processed = true;
}


bool DDLWorker::taskShouldBeExecutedOnLeader(const ASTPtr & ast_ddl, const StoragePtr storage)
{
    /// Pure DROP queries have to be executed on each node separately
    if (auto * query = ast_ddl->as<ASTDropQuery>(); query && query->kind != ASTDropQuery::Kind::Truncate)
        return false;

    if (!ast_ddl->as<ASTAlterQuery>() && !ast_ddl->as<ASTOptimizeQuery>() && !ast_ddl->as<ASTDropQuery>())
        return false;

    if (auto * alter = ast_ddl->as<ASTAlterQuery>())
    {
        // Setting alters should be executed on all replicas
        if (alter->isSettingsAlter())
            return false;

        if (alter->isFreezeAlter())
            return false;
    }

    return storage->supportsReplication();
}

bool DDLWorker::tryExecuteQueryOnLeaderReplica(
    DDLTaskBase & task,
    StoragePtr storage,
    const String & rewritten_query,
    const String & /*node_path*/,
    const ZooKeeperPtr & zookeeper)
{
    StorageReplicatedMergeTree * replicated_storage = dynamic_cast<StorageReplicatedMergeTree *>(storage.get());

    /// If we will develop new replicated storage
    if (!replicated_storage)
        throw Exception(ErrorCodes::LOGICAL_ERROR, "Storage type '{}' is not supported by distributed DDL", storage->getName());

    String shard_path = task.getShardNodePath();
    String is_executed_path = fs::path(shard_path) / "executed";
    String tries_to_execute_path = fs::path(shard_path) / "tries_to_execute";
    zookeeper->createAncestors(fs::path(shard_path) / ""); /* appends "/" at the end of shard_path */

    /// Leader replica creates is_executed_path node on successful query execution.
    /// We will remove create_shard_flag from zk operations list, if current replica is just waiting for leader to execute the query.
    auto create_shard_flag = zkutil::makeCreateRequest(is_executed_path, task.host_id_str, zkutil::CreateMode::Persistent);

    /// Node exists, or we will create or we will get an exception
    zookeeper->tryCreate(tries_to_execute_path, "0", zkutil::CreateMode::Persistent);

    static constexpr int MAX_TRIES_TO_EXECUTE = 3;
    static constexpr int MAX_EXECUTION_TIMEOUT_SEC = 3600;

    String executed_by;

    zkutil::EventPtr event = std::make_shared<Poco::Event>();
    /// We must use exists request instead of get, because zookeeper will not setup event
    /// for non existing node after get request
    if (zookeeper->exists(is_executed_path, nullptr, event))
    {
        LOG_DEBUG(log, "Task {} has already been executed by replica ({}) of the same shard.", task.entry_name, zookeeper->get(is_executed_path));
        return true;
    }

    pcg64 rng(randomSeed());

    auto lock = createSimpleZooKeeperLock(zookeeper, shard_path, "lock", task.host_id_str);

    Stopwatch stopwatch;

    bool executed_by_us = false;
    bool executed_by_other_leader = false;

    /// Defensive programming. One hour is more than enough to execute almost all DDL queries.
    /// If it will be very long query like ALTER DELETE for a huge table it's still will be executed,
    /// but DDL worker can continue processing other queries.
    while (stopwatch.elapsedSeconds() <= MAX_EXECUTION_TIMEOUT_SEC)
    {
        StorageReplicatedMergeTree::Status status;
        // Has to get with zk fields to get active replicas field
        replicated_storage->getStatus(status, true);

        // Should return as soon as possible if the table is dropped.
        bool replica_dropped = replicated_storage->is_dropped;
        bool all_replicas_likely_detached = status.active_replicas == 0 && !DatabaseCatalog::instance().isTableExist(replicated_storage->getStorageID(), context);
        if (replica_dropped || all_replicas_likely_detached)
        {
            LOG_WARNING(log, ", task {} will not be executed.", task.entry_name);
            task.execution_status = ExecutionStatus(ErrorCodes::UNFINISHED, "Cannot execute replicated DDL query, table is dropped or detached permanently");
            return false;
        }

        if (task.is_initial_query && !status.is_leader)
            throw Exception(ErrorCodes::NOT_A_LEADER, "Cannot execute initial query on non-leader replica");

        /// Any replica which is leader tries to take lock
        if (status.is_leader && lock->tryLock())
        {
            /// In replicated merge tree we can have multiple leaders. So we can
            /// be "leader" and took lock, but another "leader" replica may have
            /// already executed this task.
            if (zookeeper->tryGet(is_executed_path, executed_by))
            {
                LOG_DEBUG(log, "Task {} has already been executed by replica ({}) of the same shard.", task.entry_name, executed_by);
                executed_by_other_leader = true;
                break;
            }

            /// Checking and incrementing counter exclusively.
            size_t counter = parse<int>(zookeeper->get(tries_to_execute_path));
            if (counter > MAX_TRIES_TO_EXECUTE)
                break;

            zookeeper->set(tries_to_execute_path, toString(counter + 1));

            task.ops.push_back(create_shard_flag);
            SCOPE_EXIT_MEMORY({ if (!executed_by_us && !task.ops.empty()) task.ops.pop_back(); });

            /// If the leader will unexpectedly changed this method will return false
            /// and on the next iteration new leader will take lock
            if (tryExecuteQuery(rewritten_query, task, zookeeper))
            {
                executed_by_us = true;
                break;
            }

            lock->unlock();
        }

        /// Waiting for someone who will execute query and change is_executed_path node
        if (event->tryWait(std::uniform_int_distribution<int>(0, 1000)(rng)))
        {
            LOG_DEBUG(log, "Task {} has already been executed by replica ({}) of the same shard.", task.entry_name, zookeeper->get(is_executed_path));
            executed_by_other_leader = true;
            break;
        }
        else
        {
            String tries_count;
            zookeeper->tryGet(tries_to_execute_path, tries_count);
            if (parse<int>(tries_count) > MAX_TRIES_TO_EXECUTE)
            {
                /// Nobody will try to execute query again
                LOG_WARNING(log, "Maximum retries count for task {} exceeded, cannot execute replicated DDL query", task.entry_name);
                break;
            }
            else
            {
                /// Will try to wait or execute
                LOG_TRACE(log, "Task {} still not executed, will try to wait for it or execute ourselves, tries count {}", task.entry_name, tries_count);
            }
        }
    }

    assert(!(executed_by_us && executed_by_other_leader));

    /// Not executed by leader so was not executed at all
    if (!executed_by_us && !executed_by_other_leader)
    {
        /// If we failed with timeout
        if (stopwatch.elapsedSeconds() >= MAX_EXECUTION_TIMEOUT_SEC)
        {
            LOG_WARNING(log, "Task {} was not executed by anyone, maximum timeout {} seconds exceeded", task.entry_name, MAX_EXECUTION_TIMEOUT_SEC);
            task.execution_status = ExecutionStatus(ErrorCodes::TIMEOUT_EXCEEDED, "Cannot execute replicated DDL query, timeout exceeded");
        }
        else /// If we exceeded amount of tries
        {
            LOG_WARNING(log, "Task {} was not executed by anyone, maximum number of retries exceeded", task.entry_name);
            task.execution_status = ExecutionStatus(ErrorCodes::UNFINISHED, "Cannot execute replicated DDL query, maximum retries exceeded");
        }
        return false;
    }

    if (executed_by_us)
        LOG_DEBUG(log, "Task {} executed by current replica", task.entry_name);
    else // if (executed_by_other_leader)
        LOG_DEBUG(log, "Task {} has already been executed by replica ({}) of the same shard.", task.entry_name, zookeeper->get(is_executed_path));

    return true;
}


void DDLWorker::cleanupQueue(Int64, const ZooKeeperPtr & zookeeper)
{
    LOG_DEBUG(log, "Cleaning queue");

    Strings queue_nodes = zookeeper->getChildren(queue_dir);
    filterAndSortQueueNodes(queue_nodes);

    for (auto it = queue_nodes.cbegin(); it < queue_nodes.cend(); ++it)
    {
        if (stop_flag)
            return;

        String node_name = *it;
        String node_path = fs::path(queue_dir) / node_name;

        Coordination::Stat stat;
        String dummy;

        try
        {
            /// Already deleted
            if (!zookeeper->exists(node_path, &stat))
                continue;

            if (!canRemoveQueueEntry(node_name, stat))
                continue;

            /// At first we remove entry/active node to prevent staled hosts from executing entry concurrently
            auto rm_active_res = zookeeper->tryRemove(fs::path(node_path) / "active");
            if (rm_active_res != Coordination::Error::ZOK && rm_active_res != Coordination::Error::ZNONODE)
            {
                if (rm_active_res == Coordination::Error::ZNOTEMPTY)
                    LOG_DEBUG(log, "Task {} should be deleted, but there are active workers. Skipping it.", node_name);
                else
                    LOG_WARNING(log, "Unexpected status code {} on attempt to remove {}/active", rm_active_res, node_name);
                continue;
            }

            /// Now we can safely delete entry
            LOG_INFO(log, "Task {} is outdated, deleting it", node_name);

            /// We recursively delete all nodes except node_path/finished to prevent staled hosts from
            /// creating node_path/active node (see createStatusDirs(...))
            zookeeper->tryRemoveChildrenRecursive(node_path, "finished");

            /// And then we remove node_path and node_path/finished in a single transaction
            Coordination::Requests ops;
            Coordination::Responses res;
            ops.emplace_back(zkutil::makeCheckRequest(node_path, -1));  /// See a comment below
            ops.emplace_back(zkutil::makeRemoveRequest(fs::path(node_path) / "finished", -1));
            ops.emplace_back(zkutil::makeRemoveRequest(node_path, -1));
            auto rm_entry_res = zookeeper->tryMulti(ops, res);

            if (rm_entry_res == Coordination::Error::ZNONODE)
            {
                /// Most likely both node_path/finished and node_path were removed concurrently.
                bool entry_removed_concurrently = res[0]->error == Coordination::Error::ZNONODE;
                if (entry_removed_concurrently)
                    continue;

                /// Possible rare case: initiator node has lost connection after enqueueing entry and failed to create status dirs.
                /// No one has started to process the entry, so node_path/active and node_path/finished nodes were never created, node_path has no children.
                /// Entry became outdated, but we cannot remove remove it in a transaction with node_path/finished.
                assert(res[0]->error == Coordination::Error::ZOK && res[1]->error == Coordination::Error::ZNONODE);
                rm_entry_res = zookeeper->tryRemove(node_path);
                assert(rm_entry_res != Coordination::Error::ZNOTEMPTY);
                continue;
            }
            zkutil::KeeperMultiException::check(rm_entry_res, ops, res);
        }
        catch (...)
        {
            LOG_INFO(log, "An error occurred while checking and cleaning task {} from queue: {}", node_name, getCurrentExceptionMessage(false));
        }
    }
}

bool DDLWorker::canRemoveQueueEntry(const String & entry_name, const Coordination::Stat & stat)
{
    /// Delete node if its lifetime is expired (according to task_max_lifetime parameter)
    constexpr UInt64 zookeeper_time_resolution = 1000;
    Int64 zookeeper_time_seconds = stat.ctime / zookeeper_time_resolution;
    bool node_lifetime_is_expired = zookeeper_time_seconds + task_max_lifetime < Poco::Timestamp().epochTime();

    /// If too many nodes in task queue (> max_tasks_in_queue), delete oldest one
    UInt32 entry_number = DDLTaskBase::getLogEntryNumber(entry_name);
    bool node_is_outside_max_window = entry_number + max_tasks_in_queue < max_id.load(std::memory_order_relaxed);

    return node_lifetime_is_expired || node_is_outside_max_window;
}

/// Try to create nonexisting "status" dirs for a node
void DDLWorker::createStatusDirs(const std::string & node_path, const ZooKeeperPtr & zookeeper)
{
    Coordination::Requests ops;
    ops.emplace_back(zkutil::makeCreateRequest(fs::path(node_path) / "active", {}, zkutil::CreateMode::Persistent));
    ops.emplace_back(zkutil::makeCreateRequest(fs::path(node_path) / "finished", {}, zkutil::CreateMode::Persistent));

    Coordination::Responses responses;
    Coordination::Error code = zookeeper->tryMulti(ops, responses);

    bool both_created = code == Coordination::Error::ZOK;

    /// Failed on attempt to create node_path/active because it exists, so node_path/finished must exist too
    bool both_already_exists = responses.size() == 2 && responses[0]->error == Coordination::Error::ZNODEEXISTS
                                                     && responses[1]->error == Coordination::Error::ZRUNTIMEINCONSISTENCY;
    assert(!both_already_exists || (zookeeper->exists(fs::path(node_path) / "active") && zookeeper->exists(fs::path(node_path) / "finished")));

    /// Failed on attempt to create node_path/finished, but node_path/active does not exist
    bool is_currently_deleting = responses.size() == 2 && responses[0]->error == Coordination::Error::ZOK
                                                       && responses[1]->error == Coordination::Error::ZNODEEXISTS;
    if (both_created || both_already_exists)
        return;

    if (is_currently_deleting)
    {
        cleanup_event->set();
        throw Exception(ErrorCodes::UNFINISHED, "Cannot create status dirs for {}, "
                        "most likely because someone is deleting it concurrently", node_path);
    }

    /// Connection lost or entry was removed
    assert(Coordination::isHardwareError(code) || code == Coordination::Error::ZNONODE);
    zkutil::KeeperMultiException::check(code, ops, responses);
}


String DDLWorker::enqueueQuery(DDLLogEntry & entry)
{
    if (entry.hosts.empty())
        throw Exception("Empty host list in a distributed DDL task", ErrorCodes::LOGICAL_ERROR);

    auto zookeeper = getAndSetZooKeeper();

    String query_path_prefix = fs::path(queue_dir) / "query-";
    zookeeper->createAncestors(query_path_prefix);

    String node_path = zookeeper->create(query_path_prefix, entry.toString(), zkutil::CreateMode::PersistentSequential);
    if (max_pushed_entry_metric)
    {
        String str_buf = node_path.substr(query_path_prefix.length());
        DB::ReadBufferFromString in(str_buf);
        CurrentMetrics::Metric id;
        readText(id, in);
        id = std::max(*max_pushed_entry_metric, id);
        CurrentMetrics::set(*max_pushed_entry_metric, id);
    }

    /// We cannot create status dirs in a single transaction with previous request,
    /// because we don't know node_path until previous request is executed.
    /// Se we try to create status dirs here or later when we will execute entry.
    try
    {
        createStatusDirs(node_path, zookeeper);
    }
    catch (...)
    {
        LOG_INFO(log, "An error occurred while creating auxiliary ZooKeeper directories in {} . They will be created later. Error : {}", node_path, getCurrentExceptionMessage(true));
    }

    return node_path;
}


bool DDLWorker::initializeMainThread()
{
    assert(!initialized);
    setThreadName("DDLWorker");
    LOG_DEBUG(log, "Initializing DDLWorker thread");

    while (!stop_flag)
    {
        try
        {
            auto zookeeper = getAndSetZooKeeper();
            zookeeper->createAncestors(fs::path(queue_dir) / "");
            initialized = true;
            return true;
        }
        catch (const Coordination::Exception & e)
        {
            if (!Coordination::isHardwareError(e.code))
            {
                /// A logical error.
                LOG_ERROR(log, "ZooKeeper error: {}. Failed to start DDLWorker.", getCurrentExceptionMessage(true));
                assert(false);  /// Catch such failures in tests with debug build
            }

            tryLogCurrentException(__PRETTY_FUNCTION__);
        }
        catch (...)
        {
            tryLogCurrentException(log, "Cannot initialize DDL queue.");
        }

        /// Avoid busy loop when ZooKeeper is not available.
        sleepForSeconds(5);
    }

    return false;
}

void DDLWorker::runMainThread()
{
    auto reset_state = [&]()
    {
        initialized = false;
        /// It will wait for all threads in pool to finish and will not rethrow exceptions (if any).
        /// We create new thread pool to forget previous exceptions.
        if (1 < pool_size)
            worker_pool = std::make_unique<ThreadPool>(pool_size);
        /// Clear other in-memory state, like server just started.
        current_tasks.clear();
        last_skipped_entry_name.reset();
        max_id = 0;
        LOG_INFO(log, "Cleaned DDLWorker state");
    };

    setThreadName("DDLWorker");
    LOG_DEBUG(log, "Starting DDLWorker thread");

    while (!stop_flag)
    {
        try
        {
            bool reinitialized = !initialized;

            /// Reinitialize DDLWorker state (including ZooKeeper connection) if required
            if (!initialized)
            {
                /// Stopped
                if (!initializeMainThread())
                    break;
                LOG_DEBUG(log, "Initialized DDLWorker thread");
            }

            cleanup_event->set();
            scheduleTasks(reinitialized);

            LOG_DEBUG(log, "Waiting for queue updates (stat: {}, {}, {}, {})",
                      queue_node_stat.version, queue_node_stat.cversion, queue_node_stat.numChildren, queue_node_stat.pzxid);
            /// FIXME It may hang for unknown reason. Timeout is just a hotfix.
            constexpr int queue_wait_timeout_ms = 10000;
            bool updated = queue_updated_event->tryWait(queue_wait_timeout_ms);
            if (!updated)
            {
                Coordination::Stat new_stat;
                tryGetZooKeeper()->get(queue_dir, &new_stat);
                bool queue_changed = memcmp(&queue_node_stat, &new_stat, sizeof(Coordination::Stat)) != 0;
                bool watch_triggered = queue_updated_event->tryWait(0);
                if (queue_changed && !watch_triggered)
                {
                    /// It should never happen.
                    /// Maybe log message, abort() and system.zookeeper_log will help to debug it and remove timeout (#26036).
                    LOG_TRACE(
                        log,
                        "Queue was not updated (stat: {}, {}, {}, {})",
                        new_stat.version,
                        new_stat.cversion,
                        new_stat.numChildren,
                        new_stat.pzxid);
                    context->getZooKeeperLog()->flush();
                }
            }
        }
        catch (const Coordination::Exception & e)
        {
            if (Coordination::isHardwareError(e.code))
            {
                initialized = false;
                /// Wait for pending async tasks
                if (1 < pool_size)
                    worker_pool = std::make_unique<ThreadPool>(pool_size);
                LOG_INFO(log, "Lost ZooKeeper connection, will try to connect again: {}", getCurrentExceptionMessage(true));
            }
            else
            {
                LOG_ERROR(log, "Unexpected ZooKeeper error, will try to restart main thread: {}", getCurrentExceptionMessage(true));
                reset_state();
            }
            sleepForSeconds(1);
        }
        catch (...)
        {
            tryLogCurrentException(log, "Unexpected error, will try to restart main thread:");
            reset_state();
            sleepForSeconds(5);
        }
    }
}


void DDLWorker::runCleanupThread()
{
    setThreadName("DDLWorkerClnr");
    LOG_DEBUG(log, "Started DDLWorker cleanup thread");

    Int64 last_cleanup_time_seconds = 0;
    while (!stop_flag)
    {
        try
        {
            cleanup_event->wait();
            if (stop_flag)
                break;

            Int64 current_time_seconds = Poco::Timestamp().epochTime();
            if (last_cleanup_time_seconds && current_time_seconds < last_cleanup_time_seconds + cleanup_delay_period)
            {
                LOG_TRACE(log, "Too early to clean queue, will do it later.");
                continue;
            }

            /// ZooKeeper connection is recovered by main thread. We will wait for it on cleanup_event.
            auto zookeeper = tryGetZooKeeper();
            if (zookeeper->expired())
                continue;

            cleanupQueue(current_time_seconds, zookeeper);
            last_cleanup_time_seconds = current_time_seconds;
        }
        catch (...)
        {
            tryLogCurrentException(log, __PRETTY_FUNCTION__);
        }
    }
}

}