ClickHouse/programs/server/Server.cpp

#include "Server.h"

#include <memory>
#include <sys/resource.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <pwd.h>
#include <unistd.h>
#include <Poco/Util/HelpFormatter.h>
#include <Poco/Environment.h>
#include <Common/scope_guard_safe.h>
#include <Common/logger_useful.h>
#include <base/phdr_cache.h>
#include <Common/ErrorHandlers.h>
#include <base/getMemoryAmount.h>
#include <base/getAvailableMemoryAmount.h>
#include <base/errnoToString.h>
#include <base/coverage.h>
#include <base/getFQDNOrHostName.h>
#include <base/safeExit.h>
#include <Common/PoolId.h>
#include <Common/MemoryTracker.h>
#include <Common/ClickHouseRevision.h>
#include <Common/DNSResolver.h>
#include <Common/CgroupsMemoryUsageObserver.h>
#include <Common/CurrentMetrics.h>
#include <Common/ConcurrencyControl.h>
#include <Common/Macros.h>
#include <Common/ShellCommand.h>
#include <Common/ZooKeeper/ZooKeeper.h>
#include <Common/ZooKeeper/ZooKeeperNodeCache.h>
#include <Common/formatReadable.h>
#include <Common/getMultipleKeysFromConfig.h>
#include <Common/getNumberOfPhysicalCPUCores.h>
#include <Common/getExecutablePath.h>
#include <Common/ProfileEvents.h>
#include <Common/Scheduler/IResourceManager.h>
#include <Common/ThreadProfileEvents.h>
#include <Common/ThreadStatus.h>
#include <Common/getMappedArea.h>
#include <Common/remapExecutable.h>
#include <Common/TLDListsHolder.h>
#include <Common/Config/AbstractConfigurationComparison.h>
#include <Common/assertProcessUserMatchesDataOwner.h>
#include <Common/FailPoint.h>
#include <Common/CPUID.h>
#include <Common/HTTPConnectionPool.h>
#include <Interpreters/Cache/FileCacheFactory.h>
#include <Core/ServerUUID.h>
#include <IO/ReadHelpers.h>
#include <IO/ReadBufferFromFile.h>
#include <IO/SharedThreadPools.h>
#include <IO/UseSSL.h>
#include <Interpreters/ServerAsynchronousMetrics.h>
#include <Interpreters/DDLWorker.h>
#include <Interpreters/DNSCacheUpdater.h>
#include <Interpreters/DatabaseCatalog.h>
#include <Interpreters/ExternalDictionariesLoader.h>
#include <Interpreters/ProcessList.h>
#include <Interpreters/loadMetadata.h>
#include <Interpreters/registerInterpreters.h>
#include <Interpreters/JIT/CompiledExpressionCache.h>
#include <Access/AccessControl.h>
#include <Storages/StorageReplicatedMergeTree.h>
#include <Storages/System/attachSystemTables.h>
#include <Storages/System/attachInformationSchemaTables.h>
#include <Storages/Cache/ExternalDataSourceCache.h>
#include <Storages/Cache/registerRemoteFileMetadatas.h>
#include <Common/NamedCollections/NamedCollectionUtils.h>
#include <AggregateFunctions/registerAggregateFunctions.h>
#include <Functions/UserDefined/IUserDefinedSQLObjectsStorage.h>
#include <Functions/registerFunctions.h>
#include <TableFunctions/registerTableFunctions.h>
#include <Formats/registerFormats.h>
#include <Storages/registerStorages.h>
#include <Databases/registerDatabases.h>
#include <Dictionaries/registerDictionaries.h>
#include <Disks/registerDisks.h>
#include <Common/Scheduler/Nodes/registerSchedulerNodes.h>
#include <Common/Scheduler/Nodes/registerResourceManagers.h>
#include <Common/Config/ConfigReloader.h>
#include "MetricsTransmitter.h"
#include <Common/StatusFile.h>
#include <Common/SensitiveDataMasker.h>
#include <Common/ThreadFuzzer.h>
#include <Common/getHashOfLoadedBinary.h>
#include <Common/filesystemHelpers.h>
#include <Compression/CompressionCodecEncrypted.h>
#include <Server/CloudPlacementInfo.h>
#include <Server/ServersManager/InterServersManager.h>
#include <Server/ServersManager/ProtocolServersManager.h>
#include <Interpreters/AsynchronousInsertQueue.h>
#include <Core/ServerSettings.h>
#include <filesystem>

#include "config.h"
#include <Common/config_version.h>

#if defined(OS_LINUX)
#    include <cstdlib>
#    include <sys/un.h>
#    include <sys/mman.h>
#    include <sys/ptrace.h>
#    include <Common/hasLinuxCapability.h>
#endif

#if USE_SSL
#    include <Server/CertificateReloader.h>
#endif

#if USE_JEMALLOC
#    include <jemalloc/jemalloc.h>
#endif

#if USE_AZURE_BLOB_STORAGE
#   include <azure/storage/common/internal/xml_wrapper.hpp>
#   include <azure/core/diagnostics/logger.hpp>
#endif

#include <incbin.h>
/// A minimal file used when the server is run without installation
INCBIN(resource_embedded_xml, SOURCE_DIR "/programs/server/embedded.xml");

namespace CurrentMetrics
{
    extern const Metric Revision;
    extern const Metric VersionInteger;
    extern const Metric MemoryTracking;
    extern const Metric MergesMutationsMemoryTracking;
    extern const Metric MaxDDLEntryID;
    extern const Metric MaxPushedDDLEntryID;
}

namespace ProfileEvents
{
    extern const Event MainConfigLoads;
    extern const Event ServerStartupMilliseconds;
}

namespace fs = std::filesystem;

#if USE_JEMALLOC
static bool jemallocOptionEnabled(const char *name)
{
    bool value;
    size_t size = sizeof(value);

    if (mallctl(name, reinterpret_cast<void *>(&value), &size, /* newp= */ nullptr, /* newlen= */ 0))
        throw Poco::SystemException("mallctl() failed");

    return value;
}
#else
static bool jemallocOptionEnabled(const char *) { return false; }
#endif

int mainEntryClickHouseServer(int argc, char ** argv)
{
    DB::Server app;

    if (jemallocOptionEnabled("opt.background_thread"))
    {
        LOG_ERROR(&app.logger(),
            "jemalloc.background_thread was requested, "
            "however ClickHouse uses percpu_arena and background_thread most likely will not give any benefits, "
            "and also background_thread is not compatible with ClickHouse watchdog "
            "(that can be disabled with CLICKHOUSE_WATCHDOG_ENABLE=0)");
    }

    /// Do not fork separate process from watchdog if we attached to terminal.
    /// Otherwise it breaks gdb usage.
    /// Can be overridden by environment variable (cannot use server config at this moment).
    if (argc > 0)
    {
        const char * env_watchdog = getenv("CLICKHOUSE_WATCHDOG_ENABLE"); // NOLINT(concurrency-mt-unsafe)
        if (env_watchdog)
        {
            if (0 == strcmp(env_watchdog, "1"))
                app.shouldSetupWatchdog(argv[0]);

            /// Other values disable watchdog explicitly.
        }
        else if (!isatty(STDIN_FILENO) && !isatty(STDOUT_FILENO) && !isatty(STDERR_FILENO))
            app.shouldSetupWatchdog(argv[0]);
    }

    try
    {
        return app.run(argc, argv);
    }
    catch (...)
    {
        std::cerr << DB::getCurrentExceptionMessage(true) << "\n";
        auto code = DB::getCurrentExceptionCode();
        return code ? code : 1;
    }
}

namespace DB
{

namespace ErrorCodes
{
    extern const int NO_ELEMENTS_IN_CONFIG;
    extern const int ARGUMENT_OUT_OF_BOUND;
    extern const int EXCESSIVE_ELEMENT_IN_CONFIG;
    extern const int INVALID_CONFIG_PARAMETER;
    extern const int CORRUPTED_DATA;
}


static std::string getCanonicalPath(std::string && path)
{
    Poco::trimInPlace(path);
    if (path.empty())
        throw Exception(ErrorCodes::INVALID_CONFIG_PARAMETER, "path configuration parameter is empty");
    if (path.back() != '/')
        path += '/';
    return std::move(path);
}


#if defined(OS_LINUX)
namespace
{

void setOOMScore(int value, LoggerRawPtr log)
{
    try
    {
        std::string value_string = std::to_string(value);
        DB::WriteBufferFromFile buf("/proc/self/oom_score_adj");
        buf.write(value_string.c_str(), value_string.size());
        buf.next();
        buf.close();
    }
    catch (const Poco::Exception & e)
    {
        LOG_WARNING(log, "Failed to adjust OOM score: '{}'.", e.displayText());
        return;
    }
    LOG_INFO(log, "Set OOM score adjustment to {}", value);
}

}
#endif


void Server::uninitialize()
{
    logger().information("shutting down");
    BaseDaemon::uninitialize();
}

int Server::run()
{
    if (config().hasOption("help"))
    {
        Poco::Util::HelpFormatter help_formatter(Server::options());
        auto header_str = fmt::format("{} [OPTION] [-- [ARG]...]\n"
                                      "positional arguments can be used to rewrite config.xml properties, for example, --http_port=8010",
                                      commandName());
        help_formatter.setHeader(header_str);
        help_formatter.format(std::cout);
        return 0;
    }
    if (config().hasOption("version"))
    {
        std::cout << VERSION_NAME << " server version " << VERSION_STRING << VERSION_OFFICIAL << "." << std::endl;
        return 0;
    }
    return Application::run(); // NOLINT
}

void Server::initialize(Poco::Util::Application & self)
{
    ConfigProcessor::registerEmbeddedConfig("config.xml", std::string_view(reinterpret_cast<const char *>(gresource_embedded_xmlData), gresource_embedded_xmlSize));
    BaseDaemon::initialize(self);
    logger().information("starting up");

    LOG_INFO(&logger(), "OS name: {}, version: {}, architecture: {}",
        Poco::Environment::osName(),
        Poco::Environment::osVersion(),
        Poco::Environment::osArchitecture());
}

std::string Server::getDefaultCorePath() const
{
    return getCanonicalPath(config().getString("path", DBMS_DEFAULT_PATH)) + "cores";
}

void Server::defineOptions(Poco::Util::OptionSet & options)
{
    options.addOption(
        Poco::Util::Option("help", "h", "show help and exit")
            .required(false)
            .repeatable(false)
            .binding("help"));
    options.addOption(
        Poco::Util::Option("version", "V", "show version and exit")
            .required(false)
            .repeatable(false)
            .binding("version"));
    BaseDaemon::defineOptions(options);
}


void checkForUsersNotInMainConfig(
    const Poco::Util::AbstractConfiguration & config,
    const std::string & config_path,
    const std::string & users_config_path,
    LoggerPtr log)
{
    if (config.getBool("skip_check_for_incorrect_settings", false))
        return;

    if (config.has("users") || config.has("profiles") || config.has("quotas"))
    {
        /// We cannot throw exception here, because we have support for obsolete 'conf.d' directory
        /// (that does not correspond to config.d or users.d) but substitute configuration to both of them.

        LOG_ERROR(log, "The <users>, <profiles> and <quotas> elements should be located in users config file: {} not in main config {}."
            " Also note that you should place configuration changes to the appropriate *.d directory like 'users.d'.",
            users_config_path, config_path);
    }
}

/// Unused in other builds
#if defined(OS_LINUX)
static String readLine(const String & path)
{
    ReadBufferFromFile in(path);
    String contents;
    readStringUntilNewlineInto(contents, in);
    return contents;
}

static int readNumber(const String & path)
{
    ReadBufferFromFile in(path);
    int result;
    readText(result, in);
    return result;
}

#endif

static void sanityChecks(Server & server)
{
    std::string data_path = getCanonicalPath(server.config().getString("path", DBMS_DEFAULT_PATH));
    std::string logs_path = server.config().getString("logger.log", "");

    if (server.logger().is(Poco::Message::PRIO_TEST))
        server.context()->addWarningMessage("Server logging level is set to 'test' and performance is degraded. This cannot be used in production.");

#if defined(OS_LINUX)
    try
    {
        const std::unordered_set<std::string> fastClockSources = {
            // ARM clock
            "arch_sys_counter",
            // KVM guest clock
            "kvm-clock",
            // X86 clock
            "tsc",
        };
        const char * filename = "/sys/devices/system/clocksource/clocksource0/current_clocksource";
        if (!fastClockSources.contains(readLine(filename)))
            server.context()->addWarningMessage("Linux is not using a fast clock source. Performance can be degraded. Check " + String(filename));
    }
    catch (...) // NOLINT(bugprone-empty-catch)
    {
    }

    try
    {
        const char * filename = "/proc/sys/vm/overcommit_memory";
        if (readNumber(filename) == 2)
            server.context()->addWarningMessage("Linux memory overcommit is disabled. Check " + String(filename));
    }
    catch (...) // NOLINT(bugprone-empty-catch)
    {
    }

    try
    {
        const char * filename = "/sys/kernel/mm/transparent_hugepage/enabled";
        if (readLine(filename).find("[always]") != std::string::npos)
            server.context()->addWarningMessage("Linux transparent hugepages are set to \"always\". Check " + String(filename));
    }
    catch (...) // NOLINT(bugprone-empty-catch)
    {
    }

    try
    {
        const char * filename = "/proc/sys/kernel/pid_max";
        if (readNumber(filename) < 30000)
            server.context()->addWarningMessage("Linux max PID is too low. Check " + String(filename));
    }
    catch (...) // NOLINT(bugprone-empty-catch)
    {
    }

    try
    {
        const char * filename = "/proc/sys/kernel/threads-max";
        if (readNumber(filename) < 30000)
            server.context()->addWarningMessage("Linux threads max count is too low. Check " + String(filename));
    }
    catch (...) // NOLINT(bugprone-empty-catch)
    {
    }

    try
    {
        const char * filename = "/proc/sys/kernel/task_delayacct";
        if (readNumber(filename) == 0)
            server.context()->addWarningMessage("Delay accounting is not enabled, OSIOWaitMicroseconds will not be gathered. You can enable it using `echo 1 > " + String(filename) + "` or by using sysctl.");
    }
    catch (...) // NOLINT(bugprone-empty-catch)
    {
    }

    std::string dev_id = getBlockDeviceId(data_path);
    if (getBlockDeviceType(dev_id) == BlockDeviceType::ROT && getBlockDeviceReadAheadBytes(dev_id) == 0)
        server.context()->addWarningMessage("Rotational disk with disabled readahead is in use. Performance can be degraded. Used for data: " + String(data_path));
#endif

    try
    {
        if (getAvailableMemoryAmount() < (2l << 30))
            server.context()->addWarningMessage("Available memory at server startup is too low (2GiB).");
    }
    catch (...) // NOLINT(bugprone-empty-catch)
    {
    }

    try
    {
        if (!enoughSpaceInDirectory(data_path, 1ull << 30))
            server.context()->addWarningMessage("Available disk space for data at server startup is too low (1GiB): " + String(data_path));
    }
    catch (...) // NOLINT(bugprone-empty-catch)
    {
    }

    try
    {
        if (!logs_path.empty())
        {
            auto logs_parent = fs::path(logs_path).parent_path();
            if (!enoughSpaceInDirectory(logs_parent, 1ull << 30))
                server.context()->addWarningMessage("Available disk space for logs at server startup is too low (1GiB): " + String(logs_parent));
        }
    }
    catch (...) // NOLINT(bugprone-empty-catch)
    {
    }

    if (server.context()->getMergeTreeSettings().allow_remote_fs_zero_copy_replication)
    {
        server.context()->addWarningMessage("The setting 'allow_remote_fs_zero_copy_replication' is enabled for MergeTree tables."
            " But the feature of 'zero-copy replication' is under development and is not ready for production."
            " The usage of this feature can lead to data corruption and loss. The setting should be disabled in production.");
    }
}

static void initializeAzureSDKLogger(
    [[ maybe_unused ]] const ServerSettings & server_settings,
    [[ maybe_unused ]] int server_logs_level)
{
#if USE_AZURE_BLOB_STORAGE
    if (!server_settings.enable_azure_sdk_logging)
        return;

    using AzureLogsLevel = Azure::Core::Diagnostics::Logger::Level;

    static const std::unordered_map<AzureLogsLevel, std::pair<Poco::Message::Priority, DB::LogsLevel>> azure_to_server_mapping =
    {
        {AzureLogsLevel::Error, {Poco::Message::PRIO_DEBUG, LogsLevel::debug}},
        {AzureLogsLevel::Warning, {Poco::Message::PRIO_DEBUG, LogsLevel::debug}},
        {AzureLogsLevel::Informational, {Poco::Message::PRIO_TRACE, LogsLevel::trace}},
        {AzureLogsLevel::Verbose, {Poco::Message::PRIO_TEST, LogsLevel::test}},
    };

    static const std::map<Poco::Message::Priority, AzureLogsLevel> server_to_azure_mapping =
    {
        {Poco::Message::PRIO_DEBUG, AzureLogsLevel::Warning},
        {Poco::Message::PRIO_TRACE, AzureLogsLevel::Informational},
        {Poco::Message::PRIO_TEST, AzureLogsLevel::Verbose},
    };

    static const LoggerPtr azure_sdk_logger = getLogger("AzureSDK");

    auto it = server_to_azure_mapping.lower_bound(static_cast<Poco::Message::Priority>(server_logs_level));
    chassert(it != server_to_azure_mapping.end());
    Azure::Core::Diagnostics::Logger::SetLevel(it->second);

    Azure::Core::Diagnostics::Logger::SetListener([](AzureLogsLevel level, const std::string & message)
    {
        auto [poco_level, db_level] = azure_to_server_mapping.at(level);
        LOG_IMPL(azure_sdk_logger, db_level, poco_level, fmt::runtime(message));
    });
#endif
}

int Server::main(const std::vector<std::string> & /*args*/)
try
{
    Stopwatch startup_watch;

    Poco::Logger * log = &logger();

    UseSSL use_ssl;

    MainThreadStatus::getInstance();

    ServerSettings server_settings;
    server_settings.loadSettingsFromConfig(config());
    Poco::ThreadPool server_pool(3, server_settings.max_connections);

    ASTAlterCommand::setFormatAlterCommandsWithParentheses(server_settings.format_alter_operations_with_parentheses);

    StackTrace::setShowAddresses(server_settings.show_addresses_in_stack_traces);

#if USE_HDFS
    /// This will point libhdfs3 to the right location for its config.
    /// Note: this has to be done once at server initialization, because 'setenv' is not thread-safe.

    String libhdfs3_conf = config().getString("hdfs.libhdfs3_conf", "");
    if (!libhdfs3_conf.empty())
    {
        if (std::filesystem::path{libhdfs3_conf}.is_relative() && !std::filesystem::exists(libhdfs3_conf))
        {
            const String config_path = config().getString("config-file", "config.xml");
            const auto config_dir = std::filesystem::path{config_path}.remove_filename();
            if (std::filesystem::exists(config_dir / libhdfs3_conf))
                libhdfs3_conf = std::filesystem::absolute(config_dir / libhdfs3_conf);
        }
        setenv("LIBHDFS3_CONF", libhdfs3_conf.c_str(), true /* overwrite */); // NOLINT
    }
#endif

    /// When building openssl into clickhouse, clickhouse owns the configuration
    /// Therefore, the clickhouse openssl configuration should be kept separate from
    /// the OS. Default to the one in the standard config directory, unless overridden
    /// by a key in the config.
    /// Note: this has to be done once at server initialization, because 'setenv' is not thread-safe.
    if (config().has("opensslconf"))
    {
        std::string opensslconf_path = config().getString("opensslconf");
        setenv("OPENSSL_CONF", opensslconf_path.c_str(), true); /// NOLINT
    }
    else
    {
        const String config_path = config().getString("config-file", "config.xml");
        const auto config_dir = std::filesystem::path{config_path}.replace_filename("openssl.conf");
        setenv("OPENSSL_CONF", config_dir.c_str(), true); /// NOLINT
    }

    registerInterpreters();
    registerFunctions();
    registerAggregateFunctions();
    registerTableFunctions();
    registerDatabases();
    registerStorages();
    registerDictionaries();
    registerDisks(/* global_skip_access_check= */ false);
    registerFormats();
    registerRemoteFileMetadatas();
    registerSchedulerNodes();
    registerResourceManagers();

    CurrentMetrics::set(CurrentMetrics::Revision, ClickHouseRevision::getVersionRevision());
    CurrentMetrics::set(CurrentMetrics::VersionInteger, ClickHouseRevision::getVersionInteger());

    /** Context contains all that query execution is dependent:
      *  settings, available functions, data types, aggregate functions, databases, ...
      */
    auto shared_context = Context::createShared();
    global_context = Context::createGlobal(shared_context.get());

    global_context->makeGlobalContext();
    global_context->setApplicationType(Context::ApplicationType::SERVER);

#if !defined(NDEBUG) || !defined(__OPTIMIZE__)
    global_context->addWarningMessage("Server was built in debug mode. It will work slowly.");
#endif

    if (ThreadFuzzer::instance().isEffective())
        global_context->addWarningMessage("ThreadFuzzer is enabled. Application will run slowly and unstable.");

#if defined(SANITIZER)
    global_context->addWarningMessage("Server was built with sanitizer. It will work slowly.");
#endif

#if defined(SANITIZE_COVERAGE) || WITH_COVERAGE
    global_context->addWarningMessage("Server was built with code coverage. It will work slowly.");
#endif

    const size_t physical_server_memory = getMemoryAmount();

    LOG_INFO(log, "Available RAM: {}; physical cores: {}; logical cores: {}.",
        formatReadableSizeWithBinarySuffix(physical_server_memory),
        getNumberOfPhysicalCPUCores(),  // on ARM processors it can show only enabled at current moment cores
        std::thread::hardware_concurrency());

#if defined(__x86_64__)
    String cpu_info;
#define COLLECT_FLAG(X) \
    if (CPU::have##X()) \
    {                   \
        if (!cpu_info.empty()) \
            cpu_info += ", ";  \
        cpu_info += #X; \
    }

    CPU_ID_ENUMERATE(COLLECT_FLAG)
#undef COLLECT_FLAG

    LOG_INFO(log, "Available CPU instruction sets: {}", cpu_info);
#endif

    bool will_have_trace_collector = hasPHDRCache() && config().has("trace_log");

    std::mutex servers_lock;
    ProtocolServersManager servers(context(), &logger());
    InterServersManager servers_to_start_before_tables(context(), &logger());

    // Initialize global thread pool. Do it before we fetch configs from zookeeper
    // nodes (`from_zk`), because ZooKeeper interface uses the pool. We will
    // ignore `max_thread_pool_size` in configs we fetch from ZK, but oh well.
    GlobalThreadPool::initialize(
        server_settings.max_thread_pool_size,
        server_settings.max_thread_pool_free_size,
        server_settings.thread_pool_queue_size,
        will_have_trace_collector ? server_settings.global_profiler_real_time_period_ns : 0,
        will_have_trace_collector ? server_settings.global_profiler_cpu_time_period_ns : 0);
    /// Wait for all threads to avoid possible use-after-free (for example logging objects can be already destroyed).
    SCOPE_EXIT({
        Stopwatch watch;
        LOG_INFO(log, "Waiting for background threads");
        GlobalThreadPool::instance().shutdown();
        LOG_INFO(log, "Background threads finished in {} ms", watch.elapsedMilliseconds());
    });

    /// NOTE: global context should be destroyed *before* GlobalThreadPool::shutdown()
    /// Otherwise GlobalThreadPool::shutdown() will hang, since Context holds some threads.
    SCOPE_EXIT({
        /** Ask to cancel background jobs all table engines,
          *  and also query_log.
          * It is important to do early, not in destructor of Context, because
          *  table engines could use Context on destroy.
          */
        LOG_INFO(log, "Shutting down storages.");

        global_context->shutdown();

        LOG_DEBUG(log, "Shut down storages.");

        servers_to_start_before_tables.stopServers(server_settings, servers_lock);

        global_context->shutdownKeeperDispatcher();

        /// Wait server pool to avoid use-after-free of destroyed context in the handlers
        server_pool.joinAll();

        /** Explicitly destroy Context. It is more convenient than in destructor of Server, because logger is still available.
          * At this moment, no one could own shared part of Context.
          */
        global_context.reset();
        shared_context.reset();
        LOG_DEBUG(log, "Destroyed global context.");
    });


#if USE_AZURE_BLOB_STORAGE
    /// It makes sense to deinitialize libxml after joining of all threads
    /// in global pool because libxml uses thread-local memory allocations via
    /// 'pthread_key_create' and 'pthread_setspecific' which should be deallocated
    /// at 'pthread_exit'. Deinitialization of libxml leads to call of 'pthread_key_delete'
    /// and if it is done before joining of threads, allocated memory will not be freed
    /// and there may be memory leaks in threads that used libxml.
    GlobalThreadPool::instance().addOnDestroyCallback([]
    {
        Azure::Storage::_internal::XmlGlobalDeinitialize();
    });
#endif

    getIOThreadPool().initialize(
        server_settings.max_io_thread_pool_size,
        server_settings.max_io_thread_pool_free_size,
        server_settings.io_thread_pool_queue_size);

    getBackupsIOThreadPool().initialize(
        server_settings.max_backups_io_thread_pool_size,
        server_settings.max_backups_io_thread_pool_free_size,
        server_settings.backups_io_thread_pool_queue_size);

    getActivePartsLoadingThreadPool().initialize(
        server_settings.max_active_parts_loading_thread_pool_size,
        0, // We don't need any threads once all the parts will be loaded
        server_settings.max_active_parts_loading_thread_pool_size);

    getOutdatedPartsLoadingThreadPool().initialize(
        server_settings.max_outdated_parts_loading_thread_pool_size,
        0, // We don't need any threads once all the parts will be loaded
        server_settings.max_outdated_parts_loading_thread_pool_size);

    /// It could grow if we need to synchronously wait until all the data parts will be loaded.
    getOutdatedPartsLoadingThreadPool().setMaxTurboThreads(
        server_settings.max_active_parts_loading_thread_pool_size
    );

    getUnexpectedPartsLoadingThreadPool().initialize(
        server_settings.max_unexpected_parts_loading_thread_pool_size,
        0, // We don't need any threads once all the parts will be loaded
        server_settings.max_unexpected_parts_loading_thread_pool_size);

    /// It could grow if we need to synchronously wait until all the data parts will be loaded.
    getUnexpectedPartsLoadingThreadPool().setMaxTurboThreads(
        server_settings.max_active_parts_loading_thread_pool_size
    );

    getPartsCleaningThreadPool().initialize(
        server_settings.max_parts_cleaning_thread_pool_size,
        0, // We don't need any threads one all the parts will be deleted
        server_settings.max_parts_cleaning_thread_pool_size);

    auto max_database_replicated_create_table_thread_pool_size = server_settings.max_database_replicated_create_table_thread_pool_size ?
        server_settings.max_database_replicated_create_table_thread_pool_size : getNumberOfPhysicalCPUCores();
    getDatabaseReplicatedCreateTablesThreadPool().initialize(
        max_database_replicated_create_table_thread_pool_size,
        0, // We don't need any threads once all the tables will be created
        max_database_replicated_create_table_thread_pool_size);

    /// Initialize global local cache for remote filesystem.
    if (config().has("local_cache_for_remote_fs"))
    {
        bool enable = config().getBool("local_cache_for_remote_fs.enable", false);
        if (enable)
        {
            String root_dir = config().getString("local_cache_for_remote_fs.root_dir");
            UInt64 limit_size = config().getUInt64("local_cache_for_remote_fs.limit_size");
            UInt64 bytes_read_before_flush
                = config().getUInt64("local_cache_for_remote_fs.bytes_read_before_flush", DBMS_DEFAULT_BUFFER_SIZE);
            ExternalDataSourceCache::instance().initOnce(global_context, root_dir, limit_size, bytes_read_before_flush);
        }
    }

    /// This object will periodically calculate some metrics.
    ServerAsynchronousMetrics async_metrics(
        global_context,
        server_settings.asynchronous_metrics_update_period_s,
        server_settings.asynchronous_heavy_metrics_update_period_s,
        [&]() -> std::vector<ProtocolServerMetrics>
        {
            std::lock_guard lock(servers_lock);
            std::vector<ProtocolServerMetrics> metrics1 = servers_to_start_before_tables.getMetrics();
            std::vector<ProtocolServerMetrics> metrics2 = servers.getMetrics();
            metrics1.reserve(metrics1.size() + metrics2.size());
            metrics1.insert(metrics1.end(), std::make_move_iterator(metrics2.begin()), std::make_move_iterator(metrics2.end()));
            return metrics1;
        });

    zkutil::validateZooKeeperConfig(config());
    bool has_zookeeper = zkutil::hasZooKeeperConfig(config());

    zkutil::ZooKeeperNodeCache main_config_zk_node_cache([&] { return global_context->getZooKeeper(); });
    zkutil::EventPtr main_config_zk_changed_event = std::make_shared<Poco::Event>();
    if (loaded_config.has_zk_includes)
    {
        auto old_configuration = loaded_config.configuration;
        ConfigProcessor config_processor(config_path);
        loaded_config = config_processor.loadConfigWithZooKeeperIncludes(
            main_config_zk_node_cache, main_config_zk_changed_event, /* fallback_to_preprocessed = */ true);
        config_processor.savePreprocessedConfig(loaded_config, config().getString("path", DBMS_DEFAULT_PATH));
        config().removeConfiguration(old_configuration.get());
        config().add(loaded_config.configuration.duplicate(), PRIO_DEFAULT, false);
        global_context->setConfig(loaded_config.configuration);
    }

    Settings::checkNoSettingNamesAtTopLevel(config(), config_path);

    /// We need to reload server settings because config could be updated via zookeeper.
    server_settings.loadSettingsFromConfig(config());

    /// NOTE: Do sanity checks after we loaded all possible substitutions (for the configuration) from ZK
    sanityChecks(*this);

#if defined(OS_LINUX)
    std::string executable_path = getExecutablePath();

    if (!executable_path.empty())
    {
        /// Integrity check based on checksum of the executable code.
        /// Note: it is not intended to protect from malicious party,
        /// because the reference checksum can be easily modified as well.
        /// And we don't involve asymmetric encryption with PKI yet.
        /// It's only intended to protect from faulty hardware.
        /// Note: it is only based on machine code.
        /// But there are other sections of the binary (e.g. exception handling tables)
        /// that are interpreted (not executed) but can alter the behaviour of the program as well.

        /// Please keep the below log messages in-sync with the ones in daemon/BaseDaemon.cpp
        if (stored_binary_hash.empty())
        {
            LOG_WARNING(log, "Integrity check of the executable skipped because the reference checksum could not be read.");
        }
        else
        {
            String calculated_binary_hash = getHashOfLoadedBinaryHex();
            if (calculated_binary_hash == stored_binary_hash)
            {
                LOG_INFO(log, "Integrity check of the executable successfully passed (checksum: {})", calculated_binary_hash);
            }
            else
            {
                /// If program is run under debugger, ptrace will fail.
                if (ptrace(PTRACE_TRACEME, 0, nullptr, nullptr) == -1)
                {
                    /// Program is run under debugger. Modification of it's binary image is ok for breakpoints.
                    global_context->addWarningMessage(fmt::format(
                        "Server is run under debugger and its binary image is modified (most likely with breakpoints).",
                        calculated_binary_hash));
                }
                else
                {
                    throw Exception(
                        ErrorCodes::CORRUPTED_DATA,
                        "Calculated checksum of the executable ({0}) does not correspond"
                        " to the reference checksum stored in the executable ({1})."
                        " This may indicate one of the following:"
                        " - the executable {2} was changed just after startup;"
                        " - the executable {2} was corrupted on disk due to faulty hardware;"
                        " - the loaded executable was corrupted in memory due to faulty hardware;"
                        " - the file {2} was intentionally modified;"
                        " - a logical error in the code.",
                        calculated_binary_hash,
                        stored_binary_hash,
                        executable_path);
                }
            }
        }
    }
    else
        executable_path = "/usr/bin/clickhouse";    /// It is used for information messages.

    /// After full config loaded
    {
        if (config().getBool("remap_executable", false))
        {
            LOG_DEBUG(log, "Will remap executable in memory.");
            size_t size = remapExecutable();
            LOG_DEBUG(log, "The code ({}) in memory has been successfully remapped.", ReadableSize(size));
        }

        if (config().getBool("mlock_executable", false))
        {
            if (hasLinuxCapability(CAP_IPC_LOCK))
            {
                try
                {
                    /// Get the memory area with (current) code segment.
                    /// It's better to lock only the code segment instead of calling "mlockall",
                    /// because otherwise debug info will be also locked in memory, and it can be huge.
                    auto [addr, len] = getMappedArea(reinterpret_cast<void *>(mainEntryClickHouseServer));

                    LOG_TRACE(log, "Will do mlock to prevent executable memory from being paged out. It may take a few seconds.");
                    if (0 != mlock(addr, len))
                        LOG_WARNING(log, "Failed mlock: {}", errnoToString());
                    else
                        LOG_TRACE(log, "The memory map of clickhouse executable has been mlock'ed, total {}", ReadableSize(len));
                }
                catch (...)
                {
                    LOG_WARNING(log, "Cannot mlock: {}", getCurrentExceptionMessage(false));
                }
            }
            else
            {
                LOG_INFO(log, "It looks like the process has no CAP_IPC_LOCK capability, binary mlock will be disabled."
                    " It could happen due to incorrect ClickHouse package installation."
                    " You could resolve the problem manually with 'sudo setcap cap_ipc_lock=+ep {}'."
                    " Note that it will not work on 'nosuid' mounted filesystems.", executable_path);
            }
        }
    }

    FailPointInjection::enableFromGlobalConfig(config());

    int default_oom_score = 0;

#if !defined(NDEBUG)
    /// In debug version on Linux, increase oom score so that clickhouse is killed
    /// first, instead of some service. Use a carefully chosen random score of 555:
    /// the maximum is 1000, and chromium uses 300 for its tab processes. Ignore
    /// whatever errors that occur, because it's just a debugging aid and we don't
    /// care if it breaks.
    default_oom_score = 555;
#endif

    int oom_score = config().getInt("oom_score", default_oom_score);
    if (oom_score)
        setOOMScore(oom_score, log);
#endif

    global_context->setRemoteHostFilter(config());
    global_context->setHTTPHeaderFilter(config());

    std::string path_str = getCanonicalPath(config().getString("path", DBMS_DEFAULT_PATH));
    fs::path path = path_str;
    std::string default_database = server_settings.default_database.toString();

    /// Check that the process user id matches the owner of the data.
    assertProcessUserMatchesDataOwner(path_str, [&](const std::string & message){ global_context->addWarningMessage(message); });

    global_context->setPath(path_str);

    StatusFile status{path / "status", StatusFile::write_full_info};

    ServerUUID::load(path / "uuid", log);

    /// Try to increase limit on number of open files.
    {
        rlimit rlim;
        if (getrlimit(RLIMIT_NOFILE, &rlim))
            throw Poco::Exception("Cannot getrlimit");

        if (rlim.rlim_cur == rlim.rlim_max)
        {
            LOG_DEBUG(log, "rlimit on number of file descriptors is {}", rlim.rlim_cur);
        }
        else
        {
            rlim_t old = rlim.rlim_cur;
            rlim.rlim_cur = config().getUInt("max_open_files", static_cast<unsigned>(rlim.rlim_max));
            int rc = setrlimit(RLIMIT_NOFILE, &rlim);
            if (rc != 0)
                LOG_WARNING(log, "Cannot set max number of file descriptors to {}. Try to specify max_open_files according to your system limits. error: {}", rlim.rlim_cur, errnoToString());
            else
                LOG_DEBUG(log, "Set max number of file descriptors to {} (was {}).", rlim.rlim_cur, old);
        }
    }

    /// Try to increase limit on number of threads.
    {
        rlimit rlim;
        if (getrlimit(RLIMIT_NPROC, &rlim))
            throw Poco::Exception("Cannot getrlimit");

        if (rlim.rlim_cur == rlim.rlim_max)
        {
            LOG_DEBUG(log, "rlimit on number of threads is {}", rlim.rlim_cur);
        }
        else
        {
            rlim_t old = rlim.rlim_cur;
            rlim.rlim_cur = rlim.rlim_max;
            int rc = setrlimit(RLIMIT_NPROC, &rlim);
            if (rc != 0)
            {
                LOG_WARNING(log, "Cannot set max number of threads to {}. error: {}", rlim.rlim_cur, errnoToString());
                rlim.rlim_cur = old;
            }
            else
            {
                LOG_DEBUG(log, "Set max number of threads to {} (was {}).", rlim.rlim_cur, old);
            }
        }

        if (rlim.rlim_cur < 30000)
        {
            global_context->addWarningMessage("Maximum number of threads is lower than 30000. There could be problems with handling a lot of simultaneous queries.");
        }
    }

    static ServerErrorHandler error_handler;
    Poco::ErrorHandler::set(&error_handler);

    /// Initialize DateLUT early, to not interfere with running time of first query.
    LOG_DEBUG(log, "Initializing DateLUT.");
    DateLUT::serverTimezoneInstance();
    LOG_TRACE(log, "Initialized DateLUT with time zone '{}'.", DateLUT::serverTimezoneInstance().getTimeZone());

    /// Storage with temporary data for processing of heavy queries.
    if (!server_settings.tmp_policy.value.empty())
    {
        global_context->setTemporaryStoragePolicy(server_settings.tmp_policy, server_settings.max_temporary_data_on_disk_size);
    }
    else if (!server_settings.temporary_data_in_cache.value.empty())
    {
        global_context->setTemporaryStorageInCache(server_settings.temporary_data_in_cache, server_settings.max_temporary_data_on_disk_size);
    }
    else
    {
        std::string temporary_path = config().getString("tmp_path", path / "tmp/");
        global_context->setTemporaryStoragePath(temporary_path, server_settings.max_temporary_data_on_disk_size);
    }

    /** Directory with 'flags': files indicating temporary settings for the server set by system administrator.
      * Flags may be cleared automatically after being applied by the server.
      * Examples: do repair of local data; clone all replicated tables from replica.
      */
    {
        auto flags_path = path / "flags/";
        fs::create_directories(flags_path);
        global_context->setFlagsPath(flags_path);
    }

    /** Directory with user provided files that are usable by 'file' table function.
      */
    {

        std::string user_files_path = config().getString("user_files_path", path / "user_files/");
        global_context->setUserFilesPath(user_files_path);
        fs::create_directories(user_files_path);
    }

    {
        std::string dictionaries_lib_path = config().getString("dictionaries_lib_path", path / "dictionaries_lib/");
        global_context->setDictionariesLibPath(dictionaries_lib_path);
        fs::create_directories(dictionaries_lib_path);
    }

    {
        std::string user_scripts_path = config().getString("user_scripts_path", path / "user_scripts/");
        global_context->setUserScriptsPath(user_scripts_path);
        fs::create_directories(user_scripts_path);
    }

    /// top_level_domains_lists
    {
        const std::string & top_level_domains_path = config().getString("top_level_domains_path", path / "top_level_domains/");
        TLDListsHolder::getInstance().parseConfig(fs::path(top_level_domains_path) / "", config());
    }

    {
        fs::create_directories(path / "data");
        fs::create_directories(path / "metadata");

        /// Directory with metadata of tables, which was marked as dropped by Atomic database
        fs::create_directories(path / "metadata_dropped");
    }

    if (config().has("interserver_http_port") && config().has("interserver_https_port"))
        throw Exception(ErrorCodes::EXCESSIVE_ELEMENT_IN_CONFIG, "Both http and https interserver ports are specified");

    static const auto interserver_tags =
    {
        std::make_tuple("interserver_http_host", "interserver_http_port", "http"),
        std::make_tuple("interserver_https_host", "interserver_https_port", "https")
    };

    for (auto [host_tag, port_tag, scheme] : interserver_tags)
    {
        if (config().has(port_tag))
        {
            String this_host = config().getString(host_tag, "");

            if (this_host.empty())
            {
                this_host = getFQDNOrHostName();
                LOG_DEBUG(log, "Configuration parameter '{}' doesn't exist or exists and empty. Will use '{}' as replica host.",
                    host_tag, this_host);
            }

            String port_str = config().getString(port_tag);
            int port = parse<int>(port_str);

            if (port < 0 || port > 0xFFFF)
                throw Exception(ErrorCodes::ARGUMENT_OUT_OF_BOUND, "Out of range '{}': {}", String(port_tag), port);

            global_context->setInterserverIOAddress(this_host, port);
            global_context->setInterserverScheme(scheme);
        }
    }

    LOG_DEBUG(log, "Initializing interserver credentials.");
    global_context->updateInterserverCredentials(config());

    if (config().has("macros"))
        global_context->setMacros(std::make_unique<Macros>(config(), "macros", log));

    /// Set up caches.

    const size_t max_cache_size = static_cast<size_t>(physical_server_memory * server_settings.cache_size_to_ram_max_ratio);

    String uncompressed_cache_policy = server_settings.uncompressed_cache_policy;
    size_t uncompressed_cache_size = server_settings.uncompressed_cache_size;
    double uncompressed_cache_size_ratio = server_settings.uncompressed_cache_size_ratio;
    if (uncompressed_cache_size > max_cache_size)
    {
        uncompressed_cache_size = max_cache_size;
        LOG_INFO(log, "Lowered uncompressed cache size to {} because the system has limited RAM", formatReadableSizeWithBinarySuffix(uncompressed_cache_size));
    }
    global_context->setUncompressedCache(uncompressed_cache_policy, uncompressed_cache_size, uncompressed_cache_size_ratio);

    String mark_cache_policy = server_settings.mark_cache_policy;
    size_t mark_cache_size = server_settings.mark_cache_size;
    double mark_cache_size_ratio = server_settings.mark_cache_size_ratio;
    if (mark_cache_size > max_cache_size)
    {
        mark_cache_size = max_cache_size;
        LOG_INFO(log, "Lowered mark cache size to {} because the system has limited RAM", formatReadableSizeWithBinarySuffix(mark_cache_size));
    }
    global_context->setMarkCache(mark_cache_policy, mark_cache_size, mark_cache_size_ratio);

    size_t page_cache_size = server_settings.page_cache_size;
    if (page_cache_size != 0)
        global_context->setPageCache(
            server_settings.page_cache_chunk_size, server_settings.page_cache_mmap_size,
            page_cache_size, server_settings.page_cache_use_madv_free,
            server_settings.page_cache_use_transparent_huge_pages);

    String index_uncompressed_cache_policy = server_settings.index_uncompressed_cache_policy;
    size_t index_uncompressed_cache_size = server_settings.index_uncompressed_cache_size;
    double index_uncompressed_cache_size_ratio = server_settings.index_uncompressed_cache_size_ratio;
    if (index_uncompressed_cache_size > max_cache_size)
    {
        index_uncompressed_cache_size = max_cache_size;
        LOG_INFO(log, "Lowered index uncompressed cache size to {} because the system has limited RAM", formatReadableSizeWithBinarySuffix(uncompressed_cache_size));
    }
    global_context->setIndexUncompressedCache(index_uncompressed_cache_policy, index_uncompressed_cache_size, index_uncompressed_cache_size_ratio);

    String index_mark_cache_policy = server_settings.index_mark_cache_policy;
    size_t index_mark_cache_size = server_settings.index_mark_cache_size;
    double index_mark_cache_size_ratio = server_settings.index_mark_cache_size_ratio;
    if (index_mark_cache_size > max_cache_size)
    {
        index_mark_cache_size = max_cache_size;
        LOG_INFO(log, "Lowered index mark cache size to {} because the system has limited RAM", formatReadableSizeWithBinarySuffix(uncompressed_cache_size));
    }
    global_context->setIndexMarkCache(index_mark_cache_policy, index_mark_cache_size, index_mark_cache_size_ratio);

    size_t mmap_cache_size = server_settings.mmap_cache_size;
    if (mmap_cache_size > max_cache_size)
    {
        mmap_cache_size = max_cache_size;
        LOG_INFO(log, "Lowered mmap file cache size to {} because the system has limited RAM", formatReadableSizeWithBinarySuffix(uncompressed_cache_size));
    }
    global_context->setMMappedFileCache(mmap_cache_size);

    size_t query_cache_max_size_in_bytes = config().getUInt64("query_cache.max_size_in_bytes", DEFAULT_QUERY_CACHE_MAX_SIZE);
    size_t query_cache_max_entries = config().getUInt64("query_cache.max_entries", DEFAULT_QUERY_CACHE_MAX_ENTRIES);
    size_t query_cache_query_cache_max_entry_size_in_bytes = config().getUInt64("query_cache.max_entry_size_in_bytes", DEFAULT_QUERY_CACHE_MAX_ENTRY_SIZE_IN_BYTES);
    size_t query_cache_max_entry_size_in_rows = config().getUInt64("query_cache.max_entry_rows_in_rows", DEFAULT_QUERY_CACHE_MAX_ENTRY_SIZE_IN_ROWS);
    if (query_cache_max_size_in_bytes > max_cache_size)
    {
        query_cache_max_size_in_bytes = max_cache_size;
        LOG_INFO(log, "Lowered query cache size to {} because the system has limited RAM", formatReadableSizeWithBinarySuffix(uncompressed_cache_size));
    }
    global_context->setQueryCache(query_cache_max_size_in_bytes, query_cache_max_entries, query_cache_query_cache_max_entry_size_in_bytes, query_cache_max_entry_size_in_rows);

#if USE_EMBEDDED_COMPILER
    size_t compiled_expression_cache_max_size_in_bytes = config().getUInt64("compiled_expression_cache_size", DEFAULT_COMPILED_EXPRESSION_CACHE_MAX_SIZE);
    size_t compiled_expression_cache_max_elements = config().getUInt64("compiled_expression_cache_elements_size", DEFAULT_COMPILED_EXPRESSION_CACHE_MAX_ENTRIES);
    CompiledExpressionCacheFactory::instance().init(compiled_expression_cache_max_size_in_bytes, compiled_expression_cache_max_elements);
#endif

    NamedCollectionUtils::loadIfNot();

    /// Initialize main config reloader.
    std::string include_from_path = config().getString("include_from", "/etc/metrika.xml");

    if (config().has("query_masking_rules"))
    {
        SensitiveDataMasker::setInstance(std::make_unique<SensitiveDataMasker>(config(), "query_masking_rules"));
    }

    std::optional<CgroupsMemoryUsageObserver> cgroups_memory_usage_observer;
    try
    {
        auto wait_time = server_settings.cgroups_memory_usage_observer_wait_time;
        if (wait_time != 0)
            cgroups_memory_usage_observer.emplace(std::chrono::seconds(wait_time));
    }
    catch (Exception &)
    {
        tryLogCurrentException(log, "Disabling cgroup memory observer because of an error during initialization");
    }

    const std::string cert_path = config().getString("openSSL.server.certificateFile", "");
    const std::string key_path = config().getString("openSSL.server.privateKeyFile", "");

    std::vector<std::string> extra_paths = {include_from_path};
    if (!cert_path.empty())
        extra_paths.emplace_back(cert_path);
    if (!key_path.empty())
        extra_paths.emplace_back(key_path);

    auto main_config_reloader = std::make_unique<ConfigReloader>(
        config_path,
        extra_paths,
        config().getString("path", DBMS_DEFAULT_PATH),
        std::move(main_config_zk_node_cache),
        main_config_zk_changed_event,
        [&](ConfigurationPtr config, bool initial_loading)
        {
            Settings::checkNoSettingNamesAtTopLevel(*config, config_path);

            ServerSettings new_server_settings;
            new_server_settings.loadSettingsFromConfig(*config);

            size_t max_server_memory_usage = new_server_settings.max_server_memory_usage;
            double max_server_memory_usage_to_ram_ratio = new_server_settings.max_server_memory_usage_to_ram_ratio;

            size_t current_physical_server_memory = getMemoryAmount(); /// With cgroups, the amount of memory available to the server can be changed dynamically.
            size_t default_max_server_memory_usage = static_cast<size_t>(current_physical_server_memory * max_server_memory_usage_to_ram_ratio);

            if (max_server_memory_usage == 0)
            {
                max_server_memory_usage = default_max_server_memory_usage;
                LOG_INFO(log, "Setting max_server_memory_usage was set to {}"
                    " ({} available * {:.2f} max_server_memory_usage_to_ram_ratio)",
                    formatReadableSizeWithBinarySuffix(max_server_memory_usage),
                    formatReadableSizeWithBinarySuffix(current_physical_server_memory),
                    max_server_memory_usage_to_ram_ratio);
            }
            else if (max_server_memory_usage > default_max_server_memory_usage)
            {
                max_server_memory_usage = default_max_server_memory_usage;
                LOG_INFO(log, "Setting max_server_memory_usage was lowered to {}"
                    " because the system has low amount of memory. The amount was"
                    " calculated as {} available"
                    " * {:.2f} max_server_memory_usage_to_ram_ratio",
                    formatReadableSizeWithBinarySuffix(max_server_memory_usage),
                    formatReadableSizeWithBinarySuffix(current_physical_server_memory),
                    max_server_memory_usage_to_ram_ratio);
            }

            total_memory_tracker.setHardLimit(max_server_memory_usage);
            total_memory_tracker.setDescription("(total)");
            total_memory_tracker.setMetric(CurrentMetrics::MemoryTracking);

            if (cgroups_memory_usage_observer)
            {
                double hard_limit_ratio = new_server_settings.cgroup_memory_watcher_hard_limit_ratio;
                double soft_limit_ratio = new_server_settings.cgroup_memory_watcher_soft_limit_ratio;
                cgroups_memory_usage_observer->setMemoryUsageLimits(
                    static_cast<uint64_t>(max_server_memory_usage * hard_limit_ratio),
                    static_cast<uint64_t>(max_server_memory_usage * soft_limit_ratio));
            }

            size_t merges_mutations_memory_usage_soft_limit = new_server_settings.merges_mutations_memory_usage_soft_limit;

            size_t default_merges_mutations_server_memory_usage = static_cast<size_t>(current_physical_server_memory * new_server_settings.merges_mutations_memory_usage_to_ram_ratio);
            if (merges_mutations_memory_usage_soft_limit == 0)
            {
                merges_mutations_memory_usage_soft_limit = default_merges_mutations_server_memory_usage;
                LOG_INFO(log, "Setting merges_mutations_memory_usage_soft_limit was set to {}"
                    " ({} available * {:.2f} merges_mutations_memory_usage_to_ram_ratio)",
                    formatReadableSizeWithBinarySuffix(merges_mutations_memory_usage_soft_limit),
                    formatReadableSizeWithBinarySuffix(current_physical_server_memory),
                    new_server_settings.merges_mutations_memory_usage_to_ram_ratio);
            }
            else if (merges_mutations_memory_usage_soft_limit > default_merges_mutations_server_memory_usage)
            {
                merges_mutations_memory_usage_soft_limit = default_merges_mutations_server_memory_usage;
                LOG_WARNING(log, "Setting merges_mutations_memory_usage_soft_limit was set to {}"
                    " ({} available * {:.2f} merges_mutations_memory_usage_to_ram_ratio)",
                    formatReadableSizeWithBinarySuffix(merges_mutations_memory_usage_soft_limit),
                    formatReadableSizeWithBinarySuffix(current_physical_server_memory),
                    new_server_settings.merges_mutations_memory_usage_to_ram_ratio);
            }

            LOG_INFO(log, "Merges and mutations memory limit is set to {}",
                formatReadableSizeWithBinarySuffix(merges_mutations_memory_usage_soft_limit));
            background_memory_tracker.setSoftLimit(merges_mutations_memory_usage_soft_limit);
            background_memory_tracker.setDescription("(background)");
            background_memory_tracker.setMetric(CurrentMetrics::MergesMutationsMemoryTracking);

            total_memory_tracker.setAllowUseJemallocMemory(new_server_settings.allow_use_jemalloc_memory);

            auto * global_overcommit_tracker = global_context->getGlobalOvercommitTracker();
            total_memory_tracker.setOvercommitTracker(global_overcommit_tracker);

            // FIXME logging-related things need synchronization -- see the 'Logger * log' saved
            // in a lot of places. For now, disable updating log configuration without server restart.
            //setTextLog(global_context->getTextLog());
            updateLevels(*config, logger());
            global_context->setClustersConfig(config, has_zookeeper);
            global_context->setMacros(std::make_unique<Macros>(*config, "macros", log));
            global_context->setExternalAuthenticatorsConfig(*config);

            if (global_context->isServerCompletelyStarted())
            {
                /// It does not make sense to reload anything before server has started.
                /// Moreover, it may break initialization order.
                global_context->loadOrReloadDictionaries(*config);
                global_context->loadOrReloadUserDefinedExecutableFunctions(*config);
            }

            global_context->setRemoteHostFilter(*config);
            global_context->setHTTPHeaderFilter(*config);

            global_context->setMaxTableSizeToDrop(new_server_settings.max_table_size_to_drop);
            global_context->setMaxPartitionSizeToDrop(new_server_settings.max_partition_size_to_drop);
            global_context->setMaxTableNumToWarn(new_server_settings.max_table_num_to_warn);
            global_context->setMaxViewNumToWarn(new_server_settings.max_view_num_to_warn);
            global_context->setMaxDictionaryNumToWarn(new_server_settings.max_dictionary_num_to_warn);
            global_context->setMaxDatabaseNumToWarn(new_server_settings.max_database_num_to_warn);
            global_context->setMaxPartNumToWarn(new_server_settings.max_part_num_to_warn);

            SlotCount concurrent_threads_soft_limit = UnlimitedSlots;
            if (new_server_settings.concurrent_threads_soft_limit_num > 0 && new_server_settings.concurrent_threads_soft_limit_num < concurrent_threads_soft_limit)
                concurrent_threads_soft_limit = new_server_settings.concurrent_threads_soft_limit_num;
            if (new_server_settings.concurrent_threads_soft_limit_ratio_to_cores > 0)
            {
                auto value = new_server_settings.concurrent_threads_soft_limit_ratio_to_cores * std::thread::hardware_concurrency();
                if (value > 0 && value < concurrent_threads_soft_limit)
                    concurrent_threads_soft_limit = value;
            }
            ConcurrencyControl::instance().setMaxConcurrency(concurrent_threads_soft_limit);

            global_context->getProcessList().setMaxSize(new_server_settings.max_concurrent_queries);
            global_context->getProcessList().setMaxInsertQueriesAmount(new_server_settings.max_concurrent_insert_queries);
            global_context->getProcessList().setMaxSelectQueriesAmount(new_server_settings.max_concurrent_select_queries);
            global_context->getProcessList().setMaxWaitingQueriesAmount(new_server_settings.max_waiting_queries);

            if (config->has("keeper_server"))
                global_context->updateKeeperConfiguration(*config);

            /// Reload the number of threads for global pools.
            /// Note: If you specified it in the top level config (not it config of default profile)
            /// then ClickHouse will use it exactly.
            /// This is done for backward compatibility.
            if (global_context->areBackgroundExecutorsInitialized())
            {
                auto new_pool_size = new_server_settings.background_pool_size;
                auto new_ratio = new_server_settings.background_merges_mutations_concurrency_ratio;
                global_context->getMergeMutateExecutor()->increaseThreadsAndMaxTasksCount(new_pool_size, static_cast<size_t>(new_pool_size * new_ratio));
                global_context->getMergeMutateExecutor()->updateSchedulingPolicy(new_server_settings.background_merges_mutations_scheduling_policy.toString());
            }

            if (global_context->areBackgroundExecutorsInitialized())
            {
                auto new_pool_size = new_server_settings.background_move_pool_size;
                global_context->getMovesExecutor()->increaseThreadsAndMaxTasksCount(new_pool_size, new_pool_size);
            }

            if (global_context->areBackgroundExecutorsInitialized())
            {
                auto new_pool_size = new_server_settings.background_fetches_pool_size;
                global_context->getFetchesExecutor()->increaseThreadsAndMaxTasksCount(new_pool_size, new_pool_size);
            }

            if (global_context->areBackgroundExecutorsInitialized())
            {
                auto new_pool_size = new_server_settings.background_common_pool_size;
                global_context->getCommonExecutor()->increaseThreadsAndMaxTasksCount(new_pool_size, new_pool_size);
            }

            global_context->getBufferFlushSchedulePool().increaseThreadsCount(new_server_settings.background_buffer_flush_schedule_pool_size);
            global_context->getSchedulePool().increaseThreadsCount(new_server_settings.background_schedule_pool_size);
            global_context->getMessageBrokerSchedulePool().increaseThreadsCount(new_server_settings.background_message_broker_schedule_pool_size);
            global_context->getDistributedSchedulePool().increaseThreadsCount(new_server_settings.background_distributed_schedule_pool_size);

            global_context->getAsyncLoader().setMaxThreads(TablesLoaderForegroundPoolId, new_server_settings.tables_loader_foreground_pool_size);
            global_context->getAsyncLoader().setMaxThreads(TablesLoaderBackgroundLoadPoolId, new_server_settings.tables_loader_background_pool_size);
            global_context->getAsyncLoader().setMaxThreads(TablesLoaderBackgroundStartupPoolId, new_server_settings.tables_loader_background_pool_size);

            getIOThreadPool().reloadConfiguration(
                new_server_settings.max_io_thread_pool_size,
                new_server_settings.max_io_thread_pool_free_size,
                new_server_settings.io_thread_pool_queue_size);

            getBackupsIOThreadPool().reloadConfiguration(
                new_server_settings.max_backups_io_thread_pool_size,
                new_server_settings.max_backups_io_thread_pool_free_size,
                new_server_settings.backups_io_thread_pool_queue_size);

            getActivePartsLoadingThreadPool().reloadConfiguration(
                new_server_settings.max_active_parts_loading_thread_pool_size,
                0, // We don't need any threads once all the parts will be loaded
                new_server_settings.max_active_parts_loading_thread_pool_size);

            getOutdatedPartsLoadingThreadPool().reloadConfiguration(
                new_server_settings.max_outdated_parts_loading_thread_pool_size,
                0, // We don't need any threads once all the parts will be loaded
                new_server_settings.max_outdated_parts_loading_thread_pool_size);

            /// It could grow if we need to synchronously wait until all the data parts will be loaded.
            getOutdatedPartsLoadingThreadPool().setMaxTurboThreads(
                new_server_settings.max_active_parts_loading_thread_pool_size
            );

            getPartsCleaningThreadPool().reloadConfiguration(
                new_server_settings.max_parts_cleaning_thread_pool_size,
                0, // We don't need any threads one all the parts will be deleted
                new_server_settings.max_parts_cleaning_thread_pool_size);

            if (config->has("resources"))
            {
                global_context->getResourceManager()->updateConfiguration(*config);
            }

            if (!initial_loading)
            {
                /// We do not load ZooKeeper configuration on the first config loading
                /// because TestKeeper server is not started yet.
                if (zkutil::hasZooKeeperConfig(*config))
                    global_context->reloadZooKeeperIfChanged(config);

                global_context->reloadAuxiliaryZooKeepersConfigIfChanged(config);

                global_context->reloadQueryMaskingRulesIfChanged(config);

                if (global_context->isServerCompletelyStarted())
                {
                    std::lock_guard lock(servers_lock);
                    servers.updateServers(*config, *this, servers_lock, server_pool, async_metrics, latest_config);
                    servers_to_start_before_tables.updateServers(*config, *this, servers_lock, server_pool, async_metrics, latest_config);
                }
            }

            global_context->updateStorageConfiguration(*config);
            global_context->updateInterserverCredentials(*config);

            global_context->updateUncompressedCacheConfiguration(*config);
            global_context->updateMarkCacheConfiguration(*config);
            global_context->updateIndexUncompressedCacheConfiguration(*config);
            global_context->updateIndexMarkCacheConfiguration(*config);
            global_context->updateMMappedFileCacheConfiguration(*config);
            global_context->updateQueryCacheConfiguration(*config);

            CompressionCodecEncrypted::Configuration::instance().tryLoad(*config, "encryption_codecs");
#if USE_SSL
            CertificateReloader::instance().tryLoad(*config);
#endif
            NamedCollectionUtils::reloadFromConfig(*config);

            FileCacheFactory::instance().updateSettingsFromConfig(*config);

            HTTPConnectionPools::instance().setLimits(
                HTTPConnectionPools::Limits{
                    new_server_settings.disk_connections_soft_limit,
                    new_server_settings.disk_connections_warn_limit,
                    new_server_settings.disk_connections_store_limit,
                },
                HTTPConnectionPools::Limits{
                    new_server_settings.storage_connections_soft_limit,
                    new_server_settings.storage_connections_warn_limit,
                    new_server_settings.storage_connections_store_limit,
                },
                HTTPConnectionPools::Limits{
                    new_server_settings.http_connections_soft_limit,
                    new_server_settings.http_connections_warn_limit,
                    new_server_settings.http_connections_store_limit,
                });

            if (global_context->isServerCompletelyStarted())
                CannotAllocateThreadFaultInjector::setFaultProbability(new_server_settings.cannot_allocate_thread_fault_injection_probability);

            ProfileEvents::increment(ProfileEvents::MainConfigLoads);

            /// Must be the last.
            latest_config = config;
        },
        /* already_loaded = */ false); /// Reload it right now (initial loading)

    servers_to_start_before_tables.createServers(
        config(),
        *this,
        servers_lock,
        server_pool,
        async_metrics,
        /* start_servers= */ false,
        ServerType(ServerType::Type::QUERIES_ALL)
    );

    /// Initialize access storages.
    auto & access_control = global_context->getAccessControl();
    try
    {
        access_control.setUpFromMainConfig(config(), config_path, [&] { return global_context->getZooKeeper(); });
    }
    catch (...)
    {
        tryLogCurrentException(log, "Caught exception while setting up access control.");
        throw;
    }

    if (cgroups_memory_usage_observer)
    {
        cgroups_memory_usage_observer->setOnMemoryAmountAvailableChangedFn([&]() { main_config_reloader->reload(); });
        cgroups_memory_usage_observer->startThread();
    }

    /// Reload config in SYSTEM RELOAD CONFIG query.
    global_context->setConfigReloadCallback([&]()
    {
        main_config_reloader->reload();
        access_control.reload(AccessControl::ReloadMode::USERS_CONFIG_ONLY);
    });

    global_context->setStopServersCallback([&](const ServerType & server_type)
    {
        std::lock_guard lock(servers_lock);
        servers.stopServers(server_type);
    });

    global_context->setStartServersCallback([&](const ServerType & server_type)
    {
        std::lock_guard lock(servers_lock);
        servers.createServers(
            config(),
            *this,
            servers_lock,
            server_pool,
            async_metrics,
            /* start_servers= */ true,
            server_type);
    });

    /// Limit on total number of concurrently executed queries.
    global_context->getProcessList().setMaxSize(server_settings.max_concurrent_queries);

    /// Load global settings from default_profile and system_profile.
    global_context->setDefaultProfiles(config());

    /// Initialize background executors after we load default_profile config.
    /// This is needed to load proper values of background_pool_size etc.
    global_context->initializeBackgroundExecutorsIfNeeded();

    if (server_settings.async_insert_threads)
    {
        global_context->setAsynchronousInsertQueue(std::make_shared<AsynchronousInsertQueue>(
            global_context,
            server_settings.async_insert_threads,
            server_settings.async_insert_queue_flush_on_shutdown));
    }

    /// Set path for format schema files
    fs::path format_schema_path(config().getString("format_schema_path", path / "format_schemas/"));
    global_context->setFormatSchemaPath(format_schema_path);
    fs::create_directories(format_schema_path);

    /// Set the path for google proto files
    if (config().has("google_protos_path"))
        global_context->setGoogleProtosPath(fs::weakly_canonical(config().getString("google_protos_path")));

    /// Set path for filesystem caches
    fs::path filesystem_caches_path(config().getString("filesystem_caches_path", ""));
    if (!filesystem_caches_path.empty())
        global_context->setFilesystemCachesPath(filesystem_caches_path);

    /// Check sanity of MergeTreeSettings on server startup
    {
        size_t background_pool_tasks = global_context->getMergeMutateExecutor()->getMaxTasksCount();
        global_context->getMergeTreeSettings().sanityCheck(background_pool_tasks);
        global_context->getReplicatedMergeTreeSettings().sanityCheck(background_pool_tasks);
    }
    /// try set up encryption. There are some errors in config, error will be printed and server wouldn't start.
    CompressionCodecEncrypted::Configuration::instance().load(config(), "encryption_codecs");

    /// DNSCacheUpdater uses BackgroundSchedulePool which lives in shared context
    /// and thus this object must be created after the SCOPE_EXIT object where shared
    /// context is destroyed.
    /// In addition this object has to be created before the loading of the tables.
    std::unique_ptr<DNSCacheUpdater> dns_cache_updater;
    if (server_settings.disable_internal_dns_cache)
    {
        /// Disable DNS caching at all
        DNSResolver::instance().setDisableCacheFlag();
        LOG_DEBUG(log, "DNS caching disabled");
    }
    else
    {
        DNSResolver::instance().setCacheMaxEntries(server_settings.dns_cache_max_entries);

        /// Initialize a watcher periodically updating DNS cache
        dns_cache_updater = std::make_unique<DNSCacheUpdater>(
            global_context, server_settings.dns_cache_update_period, server_settings.dns_max_consecutive_failures);
    }

    if (dns_cache_updater)
        dns_cache_updater->start();

    /// Set current database name before loading tables and databases because
    /// system logs may copy global context.
    global_context->setCurrentDatabaseNameInGlobalContext(default_database);

    LOG_INFO(log, "Loading metadata from {}", path_str);

    LoadTaskPtrs load_metadata_tasks;

    // Make sure that if exception is thrown during startup async, new async loading jobs are not going to be called.
    // This is important for the case when exception is thrown from loading of metadata with `async_load_databases = false`
    // to avoid simultaneously running table startups and destructing databases.
    SCOPE_EXIT_SAFE(
        LOG_INFO(log, "Stopping AsyncLoader.");

        // Waits for all currently running jobs to finish and do not run any other pending jobs.
        global_context->getAsyncLoader().stop();
    );

    try
    {
        auto & database_catalog = DatabaseCatalog::instance();
        /// We load temporary database first, because projections need it.
        database_catalog.initializeAndLoadTemporaryDatabase();
        auto system_startup_tasks = loadMetadataSystem(global_context);
        maybeConvertSystemDatabase(global_context, system_startup_tasks);
        /// This has to be done before the initialization of system logs,
        /// otherwise there is a race condition between the system database initialization
        /// and creation of new tables in the database.
        waitLoad(TablesLoaderForegroundPoolId, system_startup_tasks);
        /// After attaching system databases we can initialize system log.
        global_context->initializeSystemLogs();
        global_context->setSystemZooKeeperLogAfterInitializationIfNeeded();
        /// Build loggers before tables startup to make log messages from tables
        /// attach available in system.text_log
        buildLoggers(config(), logger());
        initializeAzureSDKLogger(server_settings, logger().getLevel());
        /// After the system database is created, attach virtual system tables (in addition to query_log and part_log)
        attachSystemTablesServer(global_context, *database_catalog.getSystemDatabase(), has_zookeeper);
        attachInformationSchema(global_context, *database_catalog.getDatabase(DatabaseCatalog::INFORMATION_SCHEMA));
        attachInformationSchema(global_context, *database_catalog.getDatabase(DatabaseCatalog::INFORMATION_SCHEMA_UPPERCASE));
        /// Firstly remove partially dropped databases, to avoid race with MaterializedMySQLSyncThread,
        /// that may execute DROP before loadMarkedAsDroppedTables() in background,
        /// and so loadMarkedAsDroppedTables() will find it and try to add, and UUID will overlap.
        database_catalog.loadMarkedAsDroppedTables();
        database_catalog.createBackgroundTasks();
        /// Then, load remaining databases (some of them maybe be loaded asynchronously)
        load_metadata_tasks = loadMetadata(global_context, default_database, server_settings.async_load_databases);
        /// If we need to convert database engines, disable async tables loading
        convertDatabasesEnginesIfNeed(load_metadata_tasks, global_context);
        database_catalog.startupBackgroundTasks();
        /// After loading validate that default database exists
        database_catalog.assertDatabaseExists(default_database);
        /// Load user-defined SQL functions.
        global_context->getUserDefinedSQLObjectsStorage().loadObjects();
    }
    catch (...)
    {
        tryLogCurrentException(log, "Caught exception while loading metadata");
        throw;
    }

    LOG_DEBUG(log, "Loaded metadata.");

    /// Init trace collector only after trace_log system table was created
    /// Disable it if we collect test coverage information, because it will work extremely slow.
#if !WITH_COVERAGE
    /// Profilers cannot work reliably with any other libunwind or without PHDR cache.
    if (hasPHDRCache())
    {
        global_context->initializeTraceCollector();

        /// Set up server-wide memory profiler (for total memory tracker).
        if (server_settings.total_memory_profiler_step)
        {
            total_memory_tracker.setProfilerStep(server_settings.total_memory_profiler_step);
        }

        if (server_settings.total_memory_tracker_sample_probability > 0.0)
        {
            total_memory_tracker.setSampleProbability(server_settings.total_memory_tracker_sample_probability);
        }

        if (server_settings.total_memory_profiler_sample_min_allocation_size)
        {
            total_memory_tracker.setSampleMinAllocationSize(server_settings.total_memory_profiler_sample_min_allocation_size);
        }

        if (server_settings.total_memory_profiler_sample_max_allocation_size)
        {
            total_memory_tracker.setSampleMaxAllocationSize(server_settings.total_memory_profiler_sample_max_allocation_size);
        }
    }
#endif

    /// Describe multiple reasons when query profiler cannot work.

#if WITH_COVERAGE
    LOG_INFO(log, "Query Profiler and TraceCollector are disabled because they work extremely slow with test coverage.");
#endif

#if defined(SANITIZER)
    LOG_INFO(log, "Query Profiler disabled because they cannot work under sanitizers"
        " when two different stack unwinding methods will interfere with each other.");
#endif

    if (!hasPHDRCache())
        LOG_INFO(log, "Query Profiler and TraceCollector are disabled because they require PHDR cache to be created"
            " (otherwise the function 'dl_iterate_phdr' is not lock free and not async-signal safe).");

#if defined(OS_LINUX)
    auto tasks_stats_provider = TasksStatsCounters::findBestAvailableProvider();
    if (tasks_stats_provider == TasksStatsCounters::MetricsProvider::None)
    {
        LOG_INFO(log, "It looks like this system does not have procfs mounted at /proc location,"
            " neither clickhouse-server process has CAP_NET_ADMIN capability."
            " 'taskstats' performance statistics will be disabled."
            " It could happen due to incorrect ClickHouse package installation."
            " You can try to resolve the problem manually with 'sudo setcap cap_net_admin=+ep {}'."
            " Note that it will not work on 'nosuid' mounted filesystems."
            " It also doesn't work if you run clickhouse-server inside network namespace as it happens in some containers.",
            executable_path);
    }
    else
    {
        LOG_INFO(log, "Tasks stats provider: {}", TasksStatsCounters::metricsProviderString(tasks_stats_provider));
    }

    if (!hasLinuxCapability(CAP_SYS_NICE))
    {
        LOG_INFO(log, "It looks like the process has no CAP_SYS_NICE capability, the setting 'os_thread_priority' will have no effect."
            " It could happen due to incorrect ClickHouse package installation."
            " You could resolve the problem manually with 'sudo setcap cap_sys_nice=+ep {}'."
            " Note that it will not work on 'nosuid' mounted filesystems.",
            executable_path);
    }
#else
    LOG_INFO(log, "TaskStats is not implemented for this OS. IO accounting will be disabled.");
#endif

    {
        attachSystemTablesAsync(global_context, *DatabaseCatalog::instance().getSystemDatabase(), async_metrics);

        {
            std::lock_guard lock(servers_lock);
            servers.createServers(
                config(),
                *this,
                servers_lock,
                server_pool,
                async_metrics,
                false,
                ServerType(ServerType::Type::QUERIES_ALL));
            if (servers.empty())
                throw Exception(
                    ErrorCodes::NO_ELEMENTS_IN_CONFIG,
                    "No servers started (add valid listen_host and 'tcp_port' "
                    "or 'http_port' to configuration file.)");
        }

#if USE_SSL
        CertificateReloader::instance().tryLoad(config());
#endif

        /// Must be done after initialization of `servers`, because async_metrics will access `servers` variable from its thread.
        async_metrics.start();
        global_context->setAsynchronousMetrics(&async_metrics);

        main_config_reloader->start();
        access_control.startPeriodicReloading();

        /// try to load dictionaries immediately, throw on error and die
        try
        {
            global_context->loadOrReloadDictionaries(config());

            if (!config().getBool("dictionaries_lazy_load", true) && config().getBool("wait_dictionaries_load_at_startup", true))
                global_context->waitForDictionariesLoad();
        }
        catch (...)
        {
            tryLogCurrentException(log, "Caught exception while loading dictionaries.");
            throw;
        }

        /// try to load embedded dictionaries immediately, throw on error and die
        try
        {
            global_context->tryCreateEmbeddedDictionaries(config());
        }
        catch (...)
        {
            tryLogCurrentException(log, "Caught exception while loading embedded dictionaries.");
            throw;
        }

        /// try to load user defined executable functions, throw on error and die
        try
        {
            global_context->loadOrReloadUserDefinedExecutableFunctions(config());
        }
        catch (...)
        {
            tryLogCurrentException(log, "Caught exception while loading user defined executable functions.");
            throw;
        }

        if (has_zookeeper && config().has("distributed_ddl"))
        {
            /// DDL worker should be started after all tables were loaded
            String ddl_zookeeper_path = config().getString("distributed_ddl.path", "/clickhouse/task_queue/ddl/");
            int pool_size = config().getInt("distributed_ddl.pool_size", 1);
            if (pool_size < 1)
                throw Exception(ErrorCodes::ARGUMENT_OUT_OF_BOUND, "distributed_ddl.pool_size should be greater then 0");
            global_context->setDDLWorker(std::make_unique<DDLWorker>(pool_size, ddl_zookeeper_path, global_context, &config(),
                                                                     "distributed_ddl", "DDLWorker",
                                                                     &CurrentMetrics::MaxDDLEntryID, &CurrentMetrics::MaxPushedDDLEntryID),
                                         load_metadata_tasks);
        }

        if (config().has(DB::PlacementInfo::PLACEMENT_CONFIG_PREFIX))
        {
            PlacementInfo::PlacementInfo::instance().initialize(config());
        }

        /// Do not keep tasks in server, they should be kept inside databases. Used here to make dependent tasks only.
        load_metadata_tasks.clear();
        load_metadata_tasks.shrink_to_fit();

        {
            std::lock_guard lock(servers_lock);
            servers.startServers();
            global_context->setServerCompletelyStarted();
            LOG_INFO(log, "Ready for connections.");
        }

        startup_watch.stop();
        ProfileEvents::increment(ProfileEvents::ServerStartupMilliseconds, startup_watch.elapsedMilliseconds());

        CannotAllocateThreadFaultInjector::setFaultProbability(server_settings.cannot_allocate_thread_fault_injection_probability);

        try
        {
            global_context->startClusterDiscovery();
        }
        catch (...)
        {
            tryLogCurrentException(log, "Caught exception while starting cluster discovery");
        }

#if defined(OS_LINUX)
        /// Tell the service manager that service startup is finished.
        /// NOTE: the parent clickhouse-watchdog process must do systemdNotify("MAINPID={}\n", child_pid); before
        /// the child process notifies 'READY=1'.
        systemdNotify("READY=1\n");
#endif

        SCOPE_EXIT_SAFE({
            LOG_DEBUG(log, "Received termination signal.");

            /// Stop reloading of the main config. This must be done before everything else because it
            /// can try to access/modify already deleted objects.
            /// E.g. it can recreate new servers or it may pass a changed config to some destroyed parts of ContextSharedPart.
            main_config_reloader.reset();
            access_control.stopPeriodicReloading();

            is_cancelled = true;
            const auto remaining_connections = servers.stopServers(server_settings, servers_lock);
            dns_cache_updater.reset();

            if (remaining_connections)
            {
                /// There is no better way to force connections to close in Poco.
                /// Otherwise connection handlers will continue to live
                /// (they are effectively dangling objects, but they use global thread pool
                ///  and global thread pool destructor will wait for threads, preventing server shutdown).

                /// Dump coverage here, because std::atexit callback would not be called.
                dumpCoverageReportIfPossible();
                LOG_WARNING(log, "Will shutdown forcefully.");
                safeExit(0);
            }
        });

        std::vector<std::unique_ptr<MetricsTransmitter>> metrics_transmitters;
        for (const auto & graphite_key : DB::getMultipleKeysFromConfig(config(), "", "graphite"))
        {
            metrics_transmitters.emplace_back(std::make_unique<MetricsTransmitter>(
                global_context->getConfigRef(), graphite_key, async_metrics));
        }

        waitForTerminationRequest();
    }

    return Application::EXIT_OK;
}
catch (...)
{
    /// Poco does not provide stacktrace.
    tryLogCurrentException("Application");
    auto code = getCurrentExceptionCode();
    return code ? code : -1;
}

}