Merge pull request #41840 from ClickHouse/io-scheduler

IO scheduling subsystem

programs/server/Server.cpp

@@ -70,6 +70,8 @@
 #include <QueryPipeline/ConnectionCollector.h>
 #include <Dictionaries/registerDictionaries.h>
 #include <Disks/registerDisks.h>
+#include <IO/Resource/registerSchedulerNodes.h>
+#include <IO/Resource/registerResourceManagers.h>
 #include <Common/Config/ConfigReloader.h>
 #include <Server/HTTPHandlerFactory.h>
 #include "MetricsTransmitter.h"
@@ -715,6 +717,8 @@ try
     registerDisks(/* global_skip_access_check= */ false);
     registerFormats();
     registerRemoteFileMetadatas();
+    registerSchedulerNodes();
+    registerResourceManagers();
 
     CurrentMetrics::set(CurrentMetrics::Revision, ClickHouseRevision::getVersionRevision());
     CurrentMetrics::set(CurrentMetrics::VersionInteger, ClickHouseRevision::getVersionInteger());
@@ -1289,6 +1293,11 @@ try
                 global_context->getDistributedSchedulePool().increaseThreadsCount(new_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

src/CMakeLists.txt

@@ -88,6 +88,7 @@ add_headers_and_sources(clickhouse_common_io Common)
 add_headers_and_sources(clickhouse_common_io Common/HashTable)
 add_headers_and_sources(clickhouse_common_io IO)
 add_headers_and_sources(clickhouse_common_io IO/Archives)
+add_headers_and_sources(clickhouse_common_io IO/Resource)
 add_headers_and_sources(clickhouse_common_io IO/S3)
 list (REMOVE_ITEM clickhouse_common_io_sources Common/malloc.cpp Common/new_delete.cpp)
 

src/Common/ErrorCodes.cpp

@@ -640,6 +640,9 @@
     M(669, NAMED_COLLECTION_DOESNT_EXIST) \
     M(670, NAMED_COLLECTION_ALREADY_EXISTS) \
     M(671, NAMED_COLLECTION_IS_IMMUTABLE) \
+    M(672, INVALID_SCHEDULER_NODE) \
+    M(673, RESOURCE_ACCESS_DENIED) \
+    M(674, RESOURCE_NOT_FOUND) \
     \
     M(999, KEEPER_EXCEPTION) \
     M(1000, POCO_EXCEPTION) \

src/Core/Settings.h

@@ -651,6 +651,8 @@ static constexpr UInt64 operator""_GiB(unsigned long long value)
     M(Map, additional_table_filters, "", "Additional filter expression which would be applied after reading from specified table. Syntax: {'table1': 'expression', 'database.table2': 'expression'}", 0) \
     M(String, additional_result_filter, "", "Additional filter expression which would be applied to query result", 0) \
     \
+    M(String, workload, "default", "Name of workload to be used to access resources", 0) \
+    \
     /** Experimental functions */ \
     M(Bool, allow_experimental_funnel_functions, false, "Enable experimental functions for funnel analysis.", 0) \
     M(Bool, allow_experimental_nlp_functions, false, "Enable experimental functions for natural language processing.", 0) \

src/IO/IResourceManager.h

@@ -0,0 +1,53 @@
+#pragma once
+
+#include <IO/ResourceRequest.h>
+
+#include <Poco/Util/AbstractConfiguration.h>
+
+#include <boost/noncopyable.hpp>
+
+#include <memory>
+#include <unordered_map>
+
+namespace DB
+{
+
+/*
+ * Instance of derived class holds everything required for resource consumption,
+ * including resources currently registered at `SchedulerRoot`. This is required to avoid
+ * problems during configuration update. Do not hold instances longer than required.
+ * Should be created on query start and destructed when query is done.
+ */
+class IClassifier : private boost::noncopyable
+{
+public:
+    virtual ~IClassifier() {}
+
+    /// Returns ResouceLink that should be used to access resource.
+    /// Returned link is valid until classifier destruction.
+    virtual ResourceLink get(const String & resource_name) = 0;
+};
+
+using ClassifierPtr = std::shared_ptr<IClassifier>;
+
+/*
+ * Represents control plane of resource scheduling. Derived class is responsible for reading
+ * configuration, creating all required `ISchedulerNode` objects and
+ * managing their lifespan.
+ */
+class IResourceManager : private boost::noncopyable
+{
+public:
+    virtual ~IResourceManager() {}
+
+    /// Initialize or reconfigure manager.
+    virtual void updateConfiguration(const Poco::Util::AbstractConfiguration & config) = 0;
+
+    /// Obtain a classifier instance required to get access to resources.
+    /// Note that it holds resource configuration, so should be destructed when query is done.
+    virtual ClassifierPtr acquire(const String & classifier_name) = 0;
+};
+
+using ResourceManagerPtr = std::shared_ptr<IResourceManager>;
+
+}

src/IO/ISchedulerConstraint.h

@@ -0,0 +1,55 @@
+#pragma once
+
+#include <IO/ISchedulerNode.h>
+
+namespace DB
+{
+
+/*
+ * Constraint defined on the set of requests in consumption state.
+ * It allows to track two events:
+ *  - dequeueRequest(): resource consumption begins
+ *  - finishRequest(): resource consumption finishes
+ * This allows to keep track of in-flight requests and implement different constraints (e.g. in-flight limit).
+ * When constraint is violated, node must be deactivated by dequeueRequest() returning `false`.
+ * When constraint is again satisfied, scheduleActivation() is called from finishRequest().
+ *
+ * Derived class behaviour requirements:
+ *  - dequeueRequest() must fill `request->constraint` iff it is nullptr;
+ *  - finishRequest() must be recursive: call to `parent_constraint->finishRequest()`.
+ */
+class ISchedulerConstraint : public ISchedulerNode
+{
+public:
+    ISchedulerConstraint(EventQueue * event_queue_, const Poco::Util::AbstractConfiguration & config = emptyConfig(), const String & config_prefix = {})
+        : ISchedulerNode(event_queue_, config, config_prefix)
+    {}
+
+    /// Resource consumption by `request` is finished.
+    /// Should be called outside of scheduling subsystem, implementation must be thread-safe.
+    virtual void finishRequest(ResourceRequest * request) = 0;
+
+    void setParent(ISchedulerNode * parent_) override
+    {
+        ISchedulerNode::setParent(parent_);
+
+        // Assign `parent_constraint` to the nearest parent derived from ISchedulerConstraint
+        for (ISchedulerNode * node = parent_; node != nullptr; node = node->parent)
+        {
+            if (auto * constraint = dynamic_cast<ISchedulerConstraint *>(node))
+            {
+                parent_constraint = constraint;
+                break;
+            }
+        }
+    }
+
+protected:
+    // Reference to nearest parent that is also derived from ISchedulerConstraint.
+    // Request can traverse through multiple constraints while being dequeue from hierarchy,
+    // while finishing request should traverse the same chain in reverse order.
+    // NOTE: it must be immutable after initialization, because it is accessed in not thread-safe way from finishRequest()
+    ISchedulerConstraint * parent_constraint = nullptr;
+};
+
+}

src/IO/ISchedulerNode.h

@@ -0,0 +1,221 @@
+#pragma once
+
+#include <Common/ErrorCodes.h>
+#include <Common/Exception.h>
+
+#include <IO/ResourceRequest.h>
+#include <Poco/Util/AbstractConfiguration.h>
+#include <Poco/Util/XMLConfiguration.h>
+
+#include <boost/noncopyable.hpp>
+
+#include <deque>
+#include <functional>
+#include <memory>
+#include <mutex>
+
+
+namespace DB
+{
+
+namespace ErrorCodes
+{
+    extern const int INVALID_SCHEDULER_NODE;
+}
+
+class ISchedulerNode;
+
+inline const Poco::Util::AbstractConfiguration & emptyConfig()
+{
+    static Poco::AutoPtr<Poco::Util::XMLConfiguration> config = new Poco::Util::XMLConfiguration();
+    return *config;
+}
+
+/*
+ * Info read and write for scheduling purposes by parent
+ */
+struct SchedulerNodeInfo
+{
+    double weight = 1.0; /// Weight of this node among it's siblings
+    Int64 priority = 0; /// Priority of this node among it's siblings (higher value means higher priority)
+
+    /// Arbitrary data accessed/stored by parent
+    union {
+        size_t idx;
+        void * ptr;
+    } parent;
+
+    SchedulerNodeInfo() = default;
+
+    explicit SchedulerNodeInfo(const Poco::Util::AbstractConfiguration & config = emptyConfig(), const String & config_prefix = {})
+    {
+        setWeight(config.getDouble(config_prefix + ".weight", weight));
+        setPriority(config.getInt64(config_prefix + ".priority", priority));
+    }
+
+    void setWeight(double value)
+    {
+        if (value <= 0 || !isfinite(value))
+            throw Exception(
+                ErrorCodes::INVALID_SCHEDULER_NODE,
+                "Negative and non-finite node weights are not allowed: {}",
+                value);
+        weight = value;
+    }
+
+    void setPriority(Int64 value)
+    {
+        priority = value;
+    }
+};
+
+/*
+ * Simple waitable thread-safe FIFO task queue.
+ * Intended to hold postponed events for later handling (usually by scheduler thread).
+ */
+class EventQueue
+{
+public:
+    using Event = std::function<void()>;
+
+    void enqueue(Event&& event)
+    {
+        std::unique_lock lock{mutex};
+        bool was_empty = queue.empty();
+        queue.emplace_back(event);
+        if (was_empty)
+            pending.notify_one();
+    }
+
+    /// Process single event if it exists
+    /// Returns `true` iff event has been processed
+    bool tryProcess()
+    {
+        std::unique_lock lock{mutex};
+        if (queue.empty())
+            return false;
+        Event event = std::move(queue.front());
+        queue.pop_front();
+        lock.unlock(); // do not hold queue mutext while processing events
+        event();
+        return true;
+    }
+
+    /// Wait for single event (if not available) and process it
+    void process()
+    {
+        std::unique_lock lock{mutex};
+        pending.wait(lock, [&] { return !queue.empty(); });
+        Event event = std::move(queue.front());
+        queue.pop_front();
+        lock.unlock(); // do not hold queue mutext while processing events
+        event();
+    }
+
+private:
+    std::mutex mutex;
+    std::condition_variable pending;
+    std::deque<Event> queue;
+};
+
+/*
+ * Node of hierarchy for scheduling requests for resource. Base class for all
+ * kinds of scheduling elements (queues, policies, constraints and schedulers).
+ *
+ * Root node is a scheduler, which has it's thread to dequeue requests,
+ * execute requests (see ResourceRequest) and process events in a thread-safe manner.
+ * Immediate children of the scheduler represent independent resources.
+ * Each resource has it's own hierarchy to achieve required scheduling policies.
+ * Non-leaf nodes do not hold requests, but keep scheduling state
+ * (e.g. consumption history, amount of in-flight requests, etc).
+ * Leafs of hierarchy are queues capable of holding pending requests.
+ *
+ *        scheduler         (SchedulerRoot)
+ *         /     \
+ *  constraint  constraint  (SemaphoreConstraint)
+ *      |           |
+ *   policy      policy     (PriorityPolicy)
+ *   /    \      /    \
+ *  q1    q2    q3    q4    (FifoQueue)
+ *
+ * Dequeueing request from an inner node will dequeue request from one of active leaf-queues in its subtree.
+ * Node is considered to be active iff:
+ *  - it has at least one pending request in one of leaves of it's subtree;
+ *  - and enforced constraints, if any, are satisfied
+ *    (e.g. amount of concurrent requests is not greater than some number).
+ *
+ * All methods must be called only from scheduler thread for thread-safety.
+ */
+class ISchedulerNode : private boost::noncopyable
+{
+public:
+    ISchedulerNode(EventQueue * event_queue_, const Poco::Util::AbstractConfiguration & config = emptyConfig(), const String & config_prefix = {})
+        : event_queue(event_queue_)
+        , info(config, config_prefix)
+    {}
+
+    virtual ~ISchedulerNode() {}
+
+    // Checks if two nodes configuration is equal
+    virtual bool equals(ISchedulerNode * other) = 0;
+
+    /// Attach new child
+    virtual void attachChild(const std::shared_ptr<ISchedulerNode> & child) = 0;
+
+    /// Detach and destroy child
+    virtual void removeChild(ISchedulerNode * child) = 0;
+
+    /// Get attached child by name
+    virtual ISchedulerNode * getChild(const String & child_name) = 0;
+
+    /// Activation of child due to the first pending request
+    /// Should be called on leaf node (i.e. queue) to propagate activation signal through chain to the root
+    virtual void activateChild(ISchedulerNode * child) = 0;
+
+    /// Returns true iff node is active
+    virtual bool isActive() = 0;
+
+    /// Returns the first request to be executed as the first component of resuting pair.
+    /// The second pair component is `true` iff node is still active after dequeueing.
+    virtual std::pair<ResourceRequest *, bool> dequeueRequest() = 0;
+
+    /// Returns full path string using names of every parent
+    String getPath()
+    {
+        String result;
+        ISchedulerNode * ptr = this;
+        while (ptr->parent)
+        {
+            result = "/" + ptr->basename + result;
+            ptr = ptr->parent;
+        }
+        return result.empty() ? "/" : result;
+    }
+
+    /// Attach to a parent (used by attachChild)
+    virtual void setParent(ISchedulerNode * parent_)
+    {
+        parent = parent_;
+    }
+
+protected:
+    /// Notify parents about the first pending request or constraint becoming satisfied.
+    /// Postponed to be handled in scheduler thread, so it is intended to be called from outside.
+    void scheduleActivation()
+    {
+        if (likely(parent))
+        {
+            event_queue->enqueue([this] { parent->activateChild(this); });
+        }
+    }
+
+public:
+    EventQueue * const event_queue;
+    String basename;
+    SchedulerNodeInfo info;
+    ISchedulerNode * parent = nullptr;
+};
+
+using SchedulerNodePtr = std::shared_ptr<ISchedulerNode>;
+
+}

src/IO/ISchedulerQueue.h

@@ -0,0 +1,26 @@
+#pragma once
+
+#include <IO/ISchedulerNode.h>
+
+#include <memory>
+
+
+namespace DB
+{
+
+/*
+ * Queue for pending requests for specific resource, leaf of hierarchy.
+ */
+class ISchedulerQueue : public ISchedulerNode
+{
+public:
+    ISchedulerQueue(EventQueue * event_queue_, const Poco::Util::AbstractConfiguration & config = emptyConfig(), const String & config_prefix = {})
+        : ISchedulerNode(event_queue_, config, config_prefix)
+    {}
+
+    /// Enqueue new request to be executed using underlying resource.
+    /// Should be called outside of scheduling subsystem, implementation must be thread-safe.
+    virtual void enqueueRequest(ResourceRequest * request) = 0;
+};
+
+}

src/IO/Resource/ClassifiersConfig.cpp

@@ -0,0 +1,40 @@
+#include <IO/Resource/ClassifiersConfig.h>
+
+#include <Common/Exception.h>
+
+namespace DB
+{
+
+namespace ErrorCodes
+{
+    extern const int RESOURCE_NOT_FOUND;
+}
+
+ClassifierDescription::ClassifierDescription(const Poco::Util::AbstractConfiguration & config, const String & config_prefix)
+{
+    Poco::Util::AbstractConfiguration::Keys keys;
+    config.keys(config_prefix, keys);
+    for (const auto & key : keys)
+        emplace(key, config.getString(config_prefix + "." + key));
+}
+
+ClassifiersConfig::ClassifiersConfig(const Poco::Util::AbstractConfiguration & config)
+{
+    Poco::Util::AbstractConfiguration::Keys keys;
+    const String config_prefix = "classifiers";
+    config.keys(config_prefix, keys);
+    for (const auto & key : keys)
+        classifiers.emplace(std::piecewise_construct,
+            std::forward_as_tuple(key),
+            std::forward_as_tuple(config, config_prefix + "." + key));
+}
+
+const ClassifierDescription & ClassifiersConfig::get(const String & classifier_name)
+{
+    if (auto it = classifiers.find(classifier_name); it != classifiers.end())
+        return it->second;
+    else
+        throw Exception(ErrorCodes::RESOURCE_NOT_FOUND, "Unknown classifier '{}' to access resources", classifier_name);
+}
+
+}

src/IO/Resource/ClassifiersConfig.h

@@ -0,0 +1,39 @@
+#pragma once
+
+#include <base/types.h>
+#include <Poco/Util/AbstractConfiguration.h>
+#include <unordered_map>
+
+namespace DB
+{
+
+/// Mapping of resource name into path string (e.g. "disk1" -> "/path/to/class")
+struct ClassifierDescription : std::unordered_map<String, String>
+{
+    ClassifierDescription(const Poco::Util::AbstractConfiguration & config, const String & config_prefix);
+};
+
+/*
+ * Loads a config with the following format:
+ * <classifiers>
+ *   <classifier1>
+ *     <resource1>/path/to/queue</resource1>
+ *     <resource2>/path/to/another/queue</resource2>
+ *   </classifier1>
+ *   ...
+ *   <classifierN>...</classifierN>
+ * </classifiers>
+ */
+class ClassifiersConfig
+{
+public:
+    ClassifiersConfig() = default;
+    explicit ClassifiersConfig(const Poco::Util::AbstractConfiguration & config);
+
+    const ClassifierDescription & get(const String & classifier_name);
+
+private:
+    std::unordered_map<String, ClassifierDescription> classifiers; // by classifier_name
+};
+
+}

src/IO/Resource/FifoQueue.cpp

@@ -0,0 +1,13 @@
+#include <IO/Resource/FifoQueue.h>
+
+#include <IO/SchedulerNodeFactory.h>
+
+namespace DB
+{
+
+void registerFifoQueue(SchedulerNodeFactory & factory)
+{
+    factory.registerMethod<FifoQueue>("fifo");
+}
+
+}

src/IO/Resource/FifoQueue.h

@@ -0,0 +1,91 @@
+#pragma once
+
+#include <Common/Stopwatch.h>
+
+#include <IO/ISchedulerQueue.h>
+
+#include <Poco/Util/AbstractConfiguration.h>
+
+#include <deque>
+#include <mutex>
+
+
+namespace DB
+{
+
+namespace ErrorCodes
+{
+    extern const int INVALID_SCHEDULER_NODE;
+}
+
+/*
+ * FIFO queue to hold pending resource requests
+ */
+class FifoQueue : public ISchedulerQueue
+{
+public:
+    FifoQueue(EventQueue * event_queue_, const Poco::Util::AbstractConfiguration & config, const String & config_prefix)
+        : ISchedulerQueue(event_queue_, config, config_prefix)
+    {}
+
+    bool equals(ISchedulerNode * other) override
+    {
+        if (auto * o = dynamic_cast<FifoQueue *>(other))
+            return true;
+        return false;
+    }
+
+    void enqueueRequest(ResourceRequest * request) override
+    {
+        std::unique_lock lock(mutex);
+        request->enqueue_ns = clock_gettime_ns();
+        bool was_empty = requests.empty();
+        requests.push_back(request);
+        if (was_empty)
+            scheduleActivation();
+    }
+
+    std::pair<ResourceRequest *, bool> dequeueRequest() override
+    {
+        std::unique_lock lock(mutex);
+        if (requests.empty())
+            return {nullptr, false};
+        ResourceRequest * result = requests.front();
+        requests.pop_front();
+        return {result, !requests.empty()};
+    }
+
+    bool isActive() override
+    {
+        std::unique_lock lock(mutex);
+        return !requests.empty();
+    }
+
+    void activateChild(ISchedulerNode *) override
+    {
+        assert(false); // queue cannot have children
+    }
+
+    void attachChild(const SchedulerNodePtr &) override
+    {
+        throw Exception(
+            ErrorCodes::INVALID_SCHEDULER_NODE,
+            "Cannot add child to leaf scheduler queue: {}",
+            getPath());
+    }
+
+    void removeChild(ISchedulerNode *) override
+    {
+    }
+
+    ISchedulerNode * getChild(const String &) override
+    {
+        return nullptr;
+    }
+
+private:
+    std::mutex mutex;
+    std::deque<ResourceRequest *> requests;
+};
+
+}

src/IO/Resource/PriorityPolicy.cpp

@@ -0,0 +1,13 @@
+#include <IO/Resource/PriorityPolicy.h>
+
+#include <IO/SchedulerNodeFactory.h>
+
+namespace DB
+{
+
+void registerPriorityPolicy(SchedulerNodeFactory & factory)
+{
+    factory.registerMethod<PriorityPolicy>("priority");
+}
+
+}

src/IO/Resource/PriorityPolicy.h

@@ -0,0 +1,143 @@
+#pragma once
+
+#include <IO/ISchedulerQueue.h>
+#include <IO/SchedulerRoot.h>
+
+#include <algorithm>
+#include <unordered_map>
+#include <vector>
+
+
+namespace DB
+{
+
+namespace ErrorCodes
+{
+    extern const int INVALID_SCHEDULER_NODE;
+}
+
+/*
+ * Scheduler node that implements priority scheduling policy.
+ * Requests are scheduled in order of priorities.
+ */
+class PriorityPolicy : public ISchedulerNode
+{
+    /// Scheduling state of a child
+    struct Item
+    {
+        ISchedulerNode * child = nullptr;
+        Int64 priority = 0; // higher value means higher priority
+
+        /// For max-heap by priority
+        bool operator<(const Item& rhs) const noexcept
+        {
+            return priority < rhs.priority;
+        }
+    };
+
+public:
+    PriorityPolicy(EventQueue * event_queue_, const Poco::Util::AbstractConfiguration & config = emptyConfig(), const String & config_prefix = {})
+        : ISchedulerNode(event_queue_, config, config_prefix)
+    {}
+
+    bool equals(ISchedulerNode * other) override
+    {
+        if (auto * o = dynamic_cast<PriorityPolicy *>(other))
+            return true;
+        return false;
+    }
+
+    void attachChild(const SchedulerNodePtr & child) override
+    {
+        // Take ownership
+        chassert(child->parent == nullptr);
+        if (auto [it, inserted] = children.emplace(child->basename, child); !inserted)
+            throw Exception(
+                ErrorCodes::INVALID_SCHEDULER_NODE,
+                "Can't add another child with the same path: {}",
+                it->second->getPath());
+
+        // Attach
+        child->setParent(this);
+
+        // Activate child if it is not empty
+        if (child->isActive())
+            activateChild(child.get());
+    }
+
+    void removeChild(ISchedulerNode * child) override
+    {
+        if (auto iter = children.find(child->basename); iter != children.end())
+        {
+            SchedulerNodePtr removed = iter->second;
+
+            // Deactivate: detach is not very common operation, so we can afford O(N) here
+            for (auto i = items.begin(), e = items.end(); i != e; ++i)
+            {
+                if (i->child == removed.get())
+                {
+                    items.erase(i);
+                    // Element was removed from inside of heap -- heap must be rebuilt
+                    std::make_heap(items.begin(), items.end());
+                    break;
+                }
+            }
+
+            // Detach
+            removed->setParent(nullptr);
+
+            // Get rid of ownership
+            children.erase(iter);
+        }
+    }
+
+    ISchedulerNode * getChild(const String & child_name) override
+    {
+        if (auto iter = children.find(child_name); iter != children.end())
+            return iter->second.get();
+        else
+            return nullptr;
+    }
+
+    std::pair<ResourceRequest *, bool> dequeueRequest() override
+    {
+        if (items.empty())
+            return {nullptr, false};
+
+        // Recursively pull request from child
+        auto [request, child_active] = items.front().child->dequeueRequest();
+        assert(request != nullptr);
+
+        // Deactivate child if it is empty
+        if (!child_active)
+        {
+            std::pop_heap(items.begin(), items.end());
+            items.pop_back();
+        }
+
+        return {request, !items.empty()};
+    }
+
+    bool isActive() override
+    {
+        return !items.empty();
+    }
+
+    void activateChild(ISchedulerNode * child) override
+    {
+        bool activate_parent = items.empty();
+        items.emplace_back(Item{child, child->info.priority});
+        std::push_heap(items.begin(), items.end());
+        if (activate_parent && parent)
+            parent->activateChild(this);
+    }
+
+private:
+    /// Heap of active children
+    std::vector<Item> items;
+
+    /// All children with ownership
+    std::unordered_map<String, SchedulerNodePtr> children; // basename -> child
+};
+
+}

src/IO/Resource/SemaphoreConstraint.cpp

@@ -0,0 +1,13 @@
+#include <IO/Resource/SemaphoreConstraint.h>
+
+#include <IO/SchedulerNodeFactory.h>
+
+namespace DB
+{
+
+void registerSemaphoreConstraint(SchedulerNodeFactory & factory)
+{
+    factory.registerMethod<SemaphoreConstraint>("inflight_limit");
+}
+
+}

src/IO/Resource/SemaphoreConstraint.h

@@ -0,0 +1,138 @@
+#pragma once
+
+#include <IO/ISchedulerConstraint.h>
+#include <IO/SchedulerRoot.h>
+
+#include <mutex>
+#include <limits>
+#include <utility>
+
+namespace DB
+{
+
+/*
+ * Limited concurrency constraint.
+ * Blocks if either number of concurrent in-flight requests exceeds `max_requests`, or their total cost exceeds `max_cost`
+ */
+class SemaphoreConstraint : public ISchedulerConstraint
+{
+    static constexpr Int64 default_max_requests = std::numeric_limits<Int64>::max();
+    static constexpr Int64 default_max_cost = std::numeric_limits<Int64>::max();
+public:
+    SemaphoreConstraint(EventQueue * event_queue_, const Poco::Util::AbstractConfiguration & config = emptyConfig(), const String & config_prefix = {})
+        : ISchedulerConstraint(event_queue_, config, config_prefix)
+        , max_requests(config.getInt64(config_prefix + ".max_requests", default_max_requests))
+        , max_cost(config.getInt64(config_prefix + ".max_cost", config.getInt64(config_prefix + ".max_bytes", default_max_cost)))
+    {}
+
+    bool equals(ISchedulerNode * other) override
+    {
+        if (auto * o = dynamic_cast<SemaphoreConstraint *>(other))
+            return max_requests == o->max_requests && max_cost == o->max_cost;
+        return false;
+    }
+
+    void attachChild(const std::shared_ptr<ISchedulerNode> & child_) override
+    {
+        // Take ownership
+        child = child_;
+        child->setParent(this);
+
+        // Activate if required
+        if (child->isActive())
+            activateChild(child.get());
+    }
+
+    void removeChild(ISchedulerNode * child_) override
+    {
+        if (child.get() == child_)
+        {
+            child_active = false; // deactivate
+            child->setParent(nullptr); // detach
+            child.reset();
+        }
+    }
+
+    ISchedulerNode * getChild(const String & child_name) override
+    {
+        if (child->basename == child_name)
+            return child.get();
+        else
+            return nullptr;
+    }
+
+    std::pair<ResourceRequest *, bool> dequeueRequest() override
+    {
+        // Dequeue request from the child
+        auto [request, child_now_active] = child->dequeueRequest();
+        if (!request)
+            return {nullptr, false};
+
+        // Request has reference to the first (closest to leaf) `constraint`, which can have `parent_constraint`.
+        // The former is initialized here dynamically and the latter is initialized once during hierarchy construction.
+        if (!request->constraint)
+            request->constraint = this;
+
+        // Update state on request arrival
+        std::unique_lock lock(mutex);
+        requests++;
+        cost += request->cost;
+        child_active = child_now_active;
+
+        return {request, active()};
+    }
+
+    void finishRequest(ResourceRequest * request) override
+    {
+        // Recursive traverse of parent flow controls in reverse order
+        if (parent_constraint)
+            parent_constraint->finishRequest(request);
+
+        // Update state on request departure
+        std::unique_lock lock(mutex);
+        bool was_active = active();
+        requests--;
+        cost -= request->cost;
+
+        // Schedule activation on transition from inactive state
+        if (!was_active && active())
+            scheduleActivation();
+    }
+
+    void activateChild(ISchedulerNode * child_) override
+    {
+        std::unique_lock lock(mutex);
+        if (child_ == child.get())
+            if (!std::exchange(child_active, true) && satisfied() && parent)
+                parent->activateChild(this);
+    }
+
+    bool isActive() override
+    {
+        std::unique_lock lock(mutex);
+        return active();
+    }
+
+private:
+    bool satisfied() const
+    {
+        return requests < max_requests && cost < max_cost;
+    }
+
+    bool active() const
+    {
+        return satisfied() && child_active;
+    }
+
+private:
+    std::mutex mutex;
+    Int64 requests = 0;
+    Int64 cost = 0;
+    bool child_active = false;
+
+    SchedulerNodePtr child;
+    Int64 max_requests = default_max_requests;
+    Int64 max_cost = default_max_cost;
+};
+
+}

src/IO/Resource/StaticResourceManager.cpp

@@ -0,0 +1,138 @@
+#include <IO/Resource/StaticResourceManager.h>
+
+#include <IO/SchedulerNodeFactory.h>
+#include <IO/ResourceManagerFactory.h>
+#include <IO/ISchedulerQueue.h>
+
+#include <Common/Exception.h>
+#include <Common/StringUtils/StringUtils.h>
+
+#include <map>
+#include <tuple>
+#include <algorithm>
+
+namespace DB
+{
+
+namespace ErrorCodes
+{
+    extern const int RESOURCE_ACCESS_DENIED;
+    extern const int RESOURCE_NOT_FOUND;
+    extern const int INVALID_SCHEDULER_NODE;
+}
+
+StaticResourceManager::Resource::Resource(
+    const String & name,
+    EventQueue * event_queue,
+    const Poco::Util::AbstractConfiguration & config,
+    const std::string & config_prefix)
+{
+    // Initialize scheduler nodes
+    Poco::Util::AbstractConfiguration::Keys keys;
+    std::sort(keys.begin(), keys.end()); // for parents to appear before children
+    config.keys(config_prefix, keys);
+    for (const auto & key : keys)
+    {
+        if (!startsWith(key, "node"))
+            continue;
+
+        // Validate path
+        String path = config.getString(config_prefix + "." + key + "[@path]", "");
+        if (path.empty())
+            throw Exception(ErrorCodes::INVALID_SCHEDULER_NODE, "Attribute 'path' must be specified in all nodes for resource '{}'", name);
+        if (path[0] != '/')
+            throw Exception(ErrorCodes::INVALID_SCHEDULER_NODE, "path must start with '/' for resource '{}'", name);
+
+        // Create node
+        String type = config.getString(config_prefix + "." + key + ".type", "fifo");
+        SchedulerNodePtr node = SchedulerNodeFactory::instance().get(type, event_queue, config, config_prefix + "." + key);
+        node->basename = path.substr(1);
+
+        // Take ownership
+        if (auto [_, inserted] = nodes.emplace(path, node); !inserted)
+            throw Exception(ErrorCodes::INVALID_SCHEDULER_NODE, "Duplicate path '{}' for resource '{}'", path, name);
+
+        // Attach created node to parent (if not root)
+        if (path != "/")
+        {
+            String parent_path = path.substr(0, path.rfind('/'));
+            if (parent_path.empty())
+                parent_path = "/";
+            if (auto parent = nodes.find(parent_path); parent != nodes.end())
+                parent->second->attachChild(node);
+            else
+                throw Exception(ErrorCodes::INVALID_SCHEDULER_NODE, "Parent doesn't exist for path '{}' for resource '{}'", path, name);
+        }
+    }
+
+    if (nodes.find("/") == nodes.end())
+        throw Exception(ErrorCodes::INVALID_SCHEDULER_NODE, "undefined root node path '/' for resource '{}'", name);
+}
+
+StaticResourceManager::Classifier::Classifier(const StaticResourceManager & manager, const ClassifierDescription & cfg)
+{
+    for (auto [resource_name, path] : cfg)
+    {
+        if (auto resource_iter = manager.resources.find(resource_name); resource_iter != manager.resources.end())
+        {
+            const Resource & resource = resource_iter->second;
+            if (auto node_iter = resource.nodes.find(path); node_iter != resource.nodes.end())
+            {
+                if (auto * queue = dynamic_cast<ISchedulerQueue *>(node_iter->second.get()))
+                    resources.emplace(resource_name, ResourceLink{.queue = queue});
+                else
+                    throw Exception(ErrorCodes::RESOURCE_NOT_FOUND, "Unable to access non-queue node at path '{}' for resource '{}'", path, resource_name);
+            }
+            else
+                throw Exception(ErrorCodes::RESOURCE_NOT_FOUND, "Path '{}' for resource '{}' does not exist", path, resource_name);
+        }
+        else
+            resources.emplace(resource_name, ResourceLink{}); // resource not configured - unlimited
+    }
+}
+
+ResourceLink StaticResourceManager::Classifier::get(const String & resource_name)
+{
+    if (auto iter = resources.find(resource_name); iter != resources.end())
+        return iter->second;
+    else
+        throw Exception(ErrorCodes::RESOURCE_ACCESS_DENIED, "Access denied to resource '{}'", resource_name);
+}
+
+void StaticResourceManager::updateConfiguration(const Poco::Util::AbstractConfiguration & config)
+{
+    if (!resources.empty())
+        return; // already initialized, configuration update is not supported
+
+    Poco::Util::AbstractConfiguration::Keys keys;
+    const String config_prefix = "resources";
+    config.keys(config_prefix, keys);
+
+    // Create resource for every element under <resources> tag
+    for (const auto & key : keys)
+    {
+        auto [iter, _] = resources.emplace(std::piecewise_construct,
+            std::forward_as_tuple(key),
+            std::forward_as_tuple(key, scheduler.event_queue, config, config_prefix + "." + key));
+        // Attach root of resource to scheduler
+        scheduler.attachChild(iter->second.nodes.find("/")->second);
+    }
+
+    // Initialize classifiers
+    classifiers = std::make_unique<ClassifiersConfig>(config);
+
+    // Run scheduler thread
+    scheduler.start();
+}
+
+ClassifierPtr StaticResourceManager::acquire(const String & classifier_name)
+{
+    return std::make_shared<Classifier>(*this, classifiers->get(classifier_name));
+}
+
+void registerStaticResourceManager(ResourceManagerFactory & factory)
+{
+    factory.registerMethod<StaticResourceManager>("static");
+}
+
+}

src/IO/Resource/StaticResourceManager.h

@@ -0,0 +1,49 @@
+#pragma once
+
+#include <IO/IResourceManager.h>
+#include <IO/SchedulerRoot.h>
+#include <IO/Resource/ClassifiersConfig.h>
+
+#include <mutex>
+
+namespace DB
+{
+
+/*
+ * Reads `<resources>` from config at startup and registers them in single `SchedulerRoot`.
+ * Do not support configuration updates, server restart is required.
+ */
+class StaticResourceManager : public IResourceManager
+{
+public:
+    // Just initialization, any further updates are ignored for the sake of simplicity
+    // NOTE: manager must be initialized before any acquire() calls to avoid races
+    void updateConfiguration(const Poco::Util::AbstractConfiguration & config) override;
+
+    ClassifierPtr acquire(const String & classifier_name) override;
+
+private:
+    struct Resource
+    {
+        std::unordered_map<String, SchedulerNodePtr> nodes; // by paths
+
+        Resource(
+            const String & name,
+            EventQueue * event_queue,
+            const Poco::Util::AbstractConfiguration & config,
+            const std::string & config_prefix);
+    };
+
+    struct Classifier : public IClassifier
+    {
+        Classifier(const StaticResourceManager & manager, const ClassifierDescription & cfg);
+        ResourceLink get(const String & resource_name) override;
+        std::unordered_map<String, ResourceLink> resources; // accessible resources by names
+    };
+
+    SchedulerRoot scheduler;
+    std::unordered_map<String, Resource> resources; // by name
+    std::unique_ptr<ClassifiersConfig> classifiers;
+};
+
+}

src/IO/Resource/registerResourceManagers.cpp

@@ -0,0 +1,15 @@
+#include <IO/Resource/registerResourceManagers.h>
+#include <IO/ResourceManagerFactory.h>
+
+namespace DB
+{
+
+void registerStaticResourceManager(ResourceManagerFactory &);
+
+void registerResourceManagers()
+{
+    auto & factory = ResourceManagerFactory::instance();
+    registerStaticResourceManager(factory);
+}
+
+}

src/IO/Resource/registerResourceManagers.h

@@ -0,0 +1,8 @@
+#pragma once
+
+namespace DB
+{
+
+void registerResourceManagers();
+
+}

src/IO/Resource/registerSchedulerNodes.cpp

@@ -0,0 +1,28 @@
+#include <IO/Resource/registerSchedulerNodes.h>
+
+#include <IO/ISchedulerNode.h>
+#include <IO/ISchedulerConstraint.h>
+#include <IO/SchedulerNodeFactory.h>
+
+namespace DB
+{
+
+void registerPriorityPolicy(SchedulerNodeFactory &);
+void registerSemaphoreConstraint(SchedulerNodeFactory &);
+void registerFifoQueue(SchedulerNodeFactory &);
+
+void registerSchedulerNodes()
+{
+    auto & factory = SchedulerNodeFactory::instance();
+
+    // ISchedulerNode
+    registerPriorityPolicy(factory);
+
+    // ISchedulerConstraint
+    registerSemaphoreConstraint(factory);
+
+    // ISchedulerQueue
+    registerFifoQueue(factory);
+}
+
+}

src/IO/Resource/registerSchedulerNodes.h

@@ -0,0 +1,8 @@
+#pragma once
+
+namespace DB
+{
+
+void registerSchedulerNodes();
+
+}

src/IO/Resource/tests/ResourceTest.h

@@ -0,0 +1,309 @@
+#pragma once
+
+#include <IO/IResourceManager.h>
+#include <IO/SchedulerRoot.h>
+#include <IO/ResourceGuard.h>
+#include <IO/SchedulerNodeFactory.h>
+#include <IO/Resource/PriorityPolicy.h>
+#include <IO/Resource/FifoQueue.h>
+#include <IO/Resource/SemaphoreConstraint.h>
+#include <IO/Resource/registerSchedulerNodes.h>
+#include <IO/Resource/registerResourceManagers.h>
+
+#include <Poco/Util/XMLConfiguration.h>
+
+#include <atomic>
+#include <barrier>
+#include <unordered_map>
+#include <mutex>
+#include <set>
+#include <sstream>
+
+namespace DB
+{
+
+struct ResourceTestBase
+{
+    ResourceTestBase()
+    {
+        [[maybe_unused]] static bool typesRegistered = [] { registerSchedulerNodes(); registerResourceManagers(); return true; }();
+    }
+
+    template <class TClass>
+    static TClass * add(EventQueue * event_queue, SchedulerNodePtr & root_node, const String & path, const String & xml = {})
+    {
+        std::stringstream stream; // STYLE_CHECK_ALLOW_STD_STRING_STREAM
+        stream << "<resource><node path=\"" << path << "\">" << xml << "</node></resource>";
+        Poco::AutoPtr config{new Poco::Util::XMLConfiguration(stream)};
+        String config_prefix = "node";
+
+        if (path == "/")
+        {
+            EXPECT_TRUE(root_node.get() == nullptr);
+            root_node.reset(new TClass(event_queue, *config, config_prefix));
+            return static_cast<TClass *>(root_node.get());
+        }
+
+        EXPECT_TRUE(root_node.get() != nullptr); // root should be initialized first
+        ISchedulerNode * parent = root_node.get();
+        size_t pos = 1;
+        String child_name;
+        while (pos < path.length())
+        {
+            size_t slash = path.find('/', pos);
+            if (slash != String::npos)
+            {
+                parent = parent->getChild(path.substr(pos, slash - pos));
+                EXPECT_TRUE(parent != nullptr); // parent does not exist
+                pos = slash + 1;
+            }
+            else
+            {
+                child_name = path.substr(pos);
+                pos = String::npos;
+            }
+        }
+
+        EXPECT_TRUE(!child_name.empty()); // wrong path
+        SchedulerNodePtr node = std::make_shared<TClass>(event_queue, *config, config_prefix);
+        node->basename = child_name;
+        parent->attachChild(node);
+        return static_cast<TClass *>(node.get());
+    }
+};
+
+
+struct ConstraintTest : public SemaphoreConstraint
+{
+    ConstraintTest(EventQueue * event_queue_, const Poco::Util::AbstractConfiguration & config = emptyConfig(), const String & config_prefix = {})
+        : SemaphoreConstraint(event_queue_, config, config_prefix)
+    {}
+
+    std::pair<ResourceRequest *, bool> dequeueRequest() override
+    {
+        auto [request, active] = SemaphoreConstraint::dequeueRequest();
+        if (request)
+        {
+            std::unique_lock lock(mutex);
+            requests.insert(request);
+        }
+        return {request, active};
+    }
+
+    void finishRequest(ResourceRequest * request) override
+    {
+        {
+            std::unique_lock lock(mutex);
+            requests.erase(request);
+        }
+        SemaphoreConstraint::finishRequest(request);
+    }
+
+    std::mutex mutex;
+    std::set<ResourceRequest *> requests;
+};
+
+class ResourceTestClass : public ResourceTestBase
+{
+    struct Request : public ResourceRequest
+    {
+        String name;
+
+        Request(ResourceCost cost_, const String & name_)
+            : ResourceRequest(cost_)
+            , name(name_)
+        {}
+
+        void execute() override
+        {
+        }
+    };
+
+public:
+    template <class TClass>
+    void add(const String & path, const String & xml = {})
+    {
+        ResourceTestBase::add<TClass>(&event_queue, root_node, path, xml);
+    }
+
+    void enqueue(const String & path, const std::vector<ResourceCost> & costs)
+    {
+        ASSERT_TRUE(root_node.get() != nullptr); // root should be initialized first
+        ISchedulerNode * node = root_node.get();
+        size_t pos = 1;
+        while (pos < path.length())
+        {
+            size_t slash = path.find('/', pos);
+            if (slash != String::npos)
+            {
+                node = node->getChild(path.substr(pos, slash - pos));
+                ASSERT_TRUE(node != nullptr); // does not exist
+                pos = slash + 1;
+            }
+            else
+            {
+                node = node->getChild(path.substr(pos));
+                pos = String::npos;
+            }
+        }
+        ISchedulerQueue * queue = dynamic_cast<ISchedulerQueue *>(node);
+        ASSERT_TRUE(queue != nullptr); // not a queue
+
+        for (ResourceCost cost : costs)
+        {
+            queue->enqueueRequest(new Request(cost, queue->basename));
+        }
+        processEvents(); // to activate queues
+    }
+
+    void dequeue(size_t count_limit = size_t(-1), ResourceCost cost_limit = ResourceCostMax)
+    {
+        while (count_limit > 0 && cost_limit > 0)
+        {
+            if (auto [request, _] = root_node->dequeueRequest(); request)
+            {
+                count_limit--;
+                cost_limit -= request->cost;
+                handle(static_cast<Request *>(request));
+            }
+            else
+            {
+                break;
+            }
+        }
+    }
+
+    void handle(Request * request)
+    {
+        consumed_cost[request->name] += request->cost;
+        delete request;
+    }
+
+    void consumed(const String & name, ResourceCost value, ResourceCost error = 0)
+    {
+        EXPECT_TRUE(consumed_cost[name] >= value - error);
+        EXPECT_TRUE(consumed_cost[name] <= value + error);
+        consumed_cost[name] -= value;
+    }
+
+    void processEvents()
+    {
+        while (event_queue.tryProcess()) {}
+    }
+
+private:
+    EventQueue event_queue;
+    SchedulerNodePtr root_node;
+    std::unordered_map<String, ResourceCost> consumed_cost;
+};
+
+template <class TManager>
+struct ResourceTestManager : public ResourceTestBase
+{
+    ResourceManagerPtr manager;
+
+    std::vector<ThreadFromGlobalPool> threads;
+    std::barrier<> busy_period;
+
+    struct Guard : public ResourceGuard
+    {
+        ResourceTestManager & t;
+
+        Guard(ResourceTestManager & t_, ResourceLink link_, ResourceCost cost)
+            : ResourceGuard(link_, cost, PostponeLocking)
+            , t(t_)
+        {
+            t.onEnqueue(link);
+            lock();
+            t.onExecute(link);
+        }
+    };
+
+    struct TItem
+    {
+        std::atomic<Int64> enqueued = 0; // number of enqueued requests
+        std::atomic<Int64> left = 0; // number of requests left to be executed
+    };
+
+    struct ResourceQueueHash
+    {
+        size_t operator()(const ResourceLink & link) const
+        {
+            return std::hash<ISchedulerQueue*>()(link.queue);
+        }
+    };
+
+    std::mutex link_data_mutex;
+    std::unordered_map<ResourceLink, TItem, ResourceQueueHash> link_data;
+
+    explicit ResourceTestManager(size_t thread_count = 1)
+        : manager(new TManager)
+        , busy_period(thread_count)
+    {}
+
+    ~ResourceTestManager()
+    {
+        for (auto & thread : threads)
+            thread.join();
+    }
+
+    void update(const String & xml)
+    {
+        std::istringstream stream(xml); // STYLE_CHECK_ALLOW_STD_STRING_STREAM
+        Poco::AutoPtr config{new Poco::Util::XMLConfiguration(stream)};
+        manager->updateConfiguration(*config);
+    }
+
+    auto & getLinkData(ResourceLink link)
+    {
+        std::unique_lock lock{link_data_mutex};
+        return link_data[link];
+    }
+
+    // Use at least two threads for each queue to avoid queue being deactivated:
+    // while the first request is executing, the second request is in queue - holding it active.
+    // use onEnqueue() and onExecute() functions for this purpose.
+    void onEnqueue(ResourceLink link)
+    {
+        getLinkData(link).enqueued.fetch_add(1, std::memory_order_relaxed);
+    }
+    void onExecute(ResourceLink link)
+    {
+        auto & data = getLinkData(link);
+        Int64 left = data.left.fetch_sub(1, std::memory_order_relaxed) - 1;
+        Int64 enqueued = data.enqueued.fetch_sub(1, std::memory_order_relaxed) - 1;
+        while (left > 0 && enqueued <= 0) // Ensure at least one thread has already enqueued itself (or there is no more requests)
+        {
+            std::this_thread::yield();
+            left = data.left.load();
+            enqueued = data.enqueued.load();
+        }
+    }
+
+    // This is required for proper busy period start, i.e. everyone to be seen by scheduler as appeared at the same time:
+    //  - resource is blocked with queries by leader thread;
+    //  - leader thread notifies followers to enqueue their requests;
+    //  - leader thread unblocks resource;
+    //  - busy period begins.
+    // NOTE: actually leader's request(s) make their own small busy period.
+    void blockResource(ResourceLink link)
+    {
+        ResourceGuard g(link, 1, ResourceGuard::PostponeLocking);
+        g.lock();
+        // NOTE: at this point we assume resource to be blocked by single request (<max_requests>1</max_requests>)
+        busy_period.arrive_and_wait(); // (1) notify all followers that resource is blocked
+        busy_period.arrive_and_wait(); // (2) wait all followers to enqueue their requests
+    }
+    void startBusyPeriod(ResourceLink link, ResourceCost cost, Int64 total_requests)
+    {
+        getLinkData(link).left += total_requests + 1;
+        busy_period.arrive_and_wait(); // (1) wait leader to block resource
+        ResourceGuard g(link, cost, ResourceGuard::PostponeLocking);
+        onEnqueue(link);
+        busy_period.arrive_and_wait(); // (2) notify leader to unblock
+        g.lock();
+        onExecute(link);
+    }
+};
+
+}

src/IO/Resource/tests/gtest_resource_class_priority.cpp

@@ -0,0 +1,122 @@
+#include <gtest/gtest.h>
+
+#include <IO/Resource/tests/ResourceTest.h>
+
+#include <IO/Resource/PriorityPolicy.h>
+
+using namespace DB;
+
+using ResourceTest = ResourceTestClass;
+
+TEST(IOResourcePriorityPolicy, Factory)
+{
+    ResourceTest t;
+
+    Poco::AutoPtr cfg = new Poco::Util::XMLConfiguration();
+    SchedulerNodePtr prio = SchedulerNodeFactory::instance().get("priority", /* event_queue = */ nullptr, *cfg, "");
+    EXPECT_TRUE(dynamic_cast<PriorityPolicy *>(prio.get()) != nullptr);
+}
+
+TEST(IOResourcePriorityPolicy, Priorities)
+{
+    ResourceTest t;
+
+    t.add<PriorityPolicy>("/");
+    t.add<FifoQueue>("/A", "<priority>1</priority>");
+    t.add<FifoQueue>("/B", "<priority>2</priority>");
+    t.add<FifoQueue>("/C", "<priority>3</priority>");
+
+    t.enqueue("/A", {10, 10, 10});
+    t.enqueue("/B", {10, 10, 10});
+    t.enqueue("/C", {10, 10, 10});
+
+    t.dequeue(2);
+    t.consumed("A", 0);
+    t.consumed("B", 0);
+    t.consumed("C", 20);
+
+    t.dequeue(2);
+    t.consumed("A", 0);
+    t.consumed("B", 10);
+    t.consumed("C", 10);
+
+    t.dequeue(2);
+    t.consumed("A", 0);
+    t.consumed("B", 20);
+    t.consumed("C", 0);
+
+    t.dequeue();
+    t.consumed("A", 30);
+    t.consumed("B", 0);
+    t.consumed("C", 0);
+}
+
+TEST(IOResourcePriorityPolicy, Activation)
+{
+    ResourceTest t;
+
+    t.add<PriorityPolicy>("/");
+    t.add<FifoQueue>("/A", "<priority>1</priority>");
+    t.add<FifoQueue>("/B", "<priority>2</priority>");
+    t.add<FifoQueue>("/C", "<priority>3</priority>");
+
+    t.enqueue("/A", {10, 10, 10, 10, 10, 10});
+    t.enqueue("/B", {10});
+    t.enqueue("/C", {10, 10});
+
+    t.dequeue(3);
+    t.consumed("A", 0);
+    t.consumed("B", 10);
+    t.consumed("C", 20);
+
+    t.dequeue(2);
+    t.consumed("A", 20);
+    t.consumed("B", 0);
+    t.consumed("C", 0);
+
+    t.enqueue("/B", {10, 10, 10});
+    t.dequeue(2);
+    t.consumed("A", 0);
+    t.consumed("B", 20);
+    t.consumed("C", 0);
+
+    t.enqueue("/C", {10, 10});
+    t.dequeue(3);
+    t.consumed("A", 0);
+    t.consumed("B", 10);
+    t.consumed("C", 20);
+
+    t.dequeue(2);
+    t.consumed("A", 20);
+    t.consumed("B", 0);
+    t.consumed("C", 0);
+}
+
+TEST(IOResourcePriorityPolicy, SinglePriority)
+{
+    ResourceTest t;
+
+    t.add<PriorityPolicy>("/");
+    t.add<FifoQueue>("/A");
+
+    for (int i = 0; i < 3; i++)
+    {
+        t.enqueue("/A", {10, 10});
+        t.dequeue(1);
+        t.consumed("A", 10);
+
+        for (int j = 0; j < 3; j++)
+        {
+            t.enqueue("/A", {10, 10, 10});
+            t.dequeue(1);
+            t.consumed("A", 10);
+            t.dequeue(1);
+            t.consumed("A", 10);
+            t.dequeue(1);
+            t.consumed("A", 10);
+        }
+
+        t.dequeue(1);
+        t.consumed("A", 10);
+    }
+}

src/IO/Resource/tests/gtest_resource_manager_static.cpp

@@ -0,0 +1,103 @@
+#include <gtest/gtest.h>
+
+#include <IO/Resource/tests/ResourceTest.h>
+
+#include <IO/Resource/StaticResourceManager.h>
+#include <Poco/Util/XMLConfiguration.h>
+
+using namespace DB;
+
+using ResourceTest = ResourceTestManager<StaticResourceManager>;
+using TestGuard = ResourceTest::Guard;
+
+TEST(IOResourceStaticResourceManager, Smoke)
+{
+    ResourceTest t;
+
+    t.update(R"CONFIG(
+        <clickhouse>
+            <resources>
+                <res1>
+                    <node path="/">      <type>inflight_limit</type><max_requests>10</max_requests></node>
+                    <node path="/prio">  <type>priority</type></node>
+                    <node path="/prio/A"></node>
+                    <node path="/prio/B"><priority>1</priority></node>
+                </res1>
+            </resources>
+            <classifiers>
+                <A><res1>/prio/A</res1></A>
+                <B><res1>/prio/B</res1></B>
+            </classifiers>
+        </clickhouse>
+    )CONFIG");
+
+    ClassifierPtr ca = t.manager->acquire("A");
+    ClassifierPtr cb = t.manager->acquire("B");
+
+    for (int i = 0; i < 10; i++)
+    {
+        ResourceGuard ga(ca->get("res1"));
+        ResourceGuard gb(cb->get("res1"));
+    }
+}
+
+TEST(IOResourceStaticResourceManager, Prioritization)
+{
+    constexpr size_t threads_per_queue = 2;
+    int requests_per_thead = 100;
+    ResourceTest t(4 * threads_per_queue + 1);
+
+    t.update(R"CONFIG(
+        <clickhouse>
+            <resources>
+                <res1>
+                    <node path="/">           <type>inflight_limit</type><max_requests>1</max_requests></node>
+                    <node path="/prio">       <type>priority</type></node>
+                    <node path="/prio/A">     <priority>-1</priority></node>
+                    <node path="/prio/B">     <priority>1</priority></node>
+                    <node path="/prio/C">     </node>
+                    <node path="/prio/D">     </node>
+                    <node path="/prio/leader"></node>
+                </res1>
+            </resources>
+            <classifiers>
+                <A><res1>/prio/A</res1></A>
+                <B><res1>/prio/B</res1></B>
+                <C><res1>/prio/C</res1></C>
+                <D><res1>/prio/D</res1></D>
+                <leader><res1>/prio/leader</res1></leader>
+            </classifiers>
+        </clickhouse>
+    )CONFIG");
+
+    std::optional<Int64> last_priority;
+    auto check = [&] (Int64 priority)
+    {
+        // Lock is not required here because this is called during request execution and we have max_requests = 1
+        if (last_priority)
+            EXPECT_TRUE(priority <= *last_priority); // Should be true if every queue arrived at the same time at busy period start
+        last_priority = priority;
+    };
+
+    for (String name : {"A", "B", "C", "D"})
+    {
+        for (int thr = 0; thr < threads_per_queue; thr++)
+        {
+            t.threads.emplace_back([&, name]
+            {
+                ClassifierPtr c = t.manager->acquire(name);
+                ResourceLink link = c->get("res1");
+                t.startBusyPeriod(link, 1, requests_per_thead);
+                for (int req = 0; req < requests_per_thead; req++)
+                {
+                    TestGuard g(t, link, 1);
+                    check(link.queue->info.priority);
+                }
+            });
+        }
+    }
+
+    ClassifierPtr c = t.manager->acquire("leader");
+    ResourceLink link = c->get("res1");
+    t.blockResource(link);
+}

src/IO/Resource/tests/gtest_resource_scheduler.cpp

@@ -0,0 +1,113 @@
+#include <gtest/gtest.h>
+
+#include <IO/SchedulerRoot.h>
+
+#include <IO/Resource/tests/ResourceTest.h>
+
+#include <future>
+
+using namespace DB;
+
+struct ResourceTest : public ResourceTestBase
+{
+    SchedulerRoot scheduler;
+
+    ResourceTest()
+    {
+        scheduler.start();
+    }
+
+    ~ResourceTest()
+    {
+        scheduler.stop(true);
+    }
+};
+
+struct ResourceHolder
+{
+    ResourceTest & t;
+    SchedulerNodePtr root_node;
+
+    explicit ResourceHolder(ResourceTest & t_)
+        : t(t_)
+    {}
+
+    ~ResourceHolder()
+    {
+        unregisterResource();
+    }
+
+    template <class TClass>
+    TClass * add(const String & path, const String & xml = {})
+    {
+        return ResourceTest::add<TClass>(t.scheduler.event_queue, root_node, path, xml);
+    }
+
+    ResourceLink addQueue(const String & path, const String & xml = {})
+    {
+        return {.queue = static_cast<ISchedulerQueue *>(ResourceTest::add<FifoQueue>(t.scheduler.event_queue, root_node, path, xml))};
+    }
+
+    void registerResource()
+    {
+        std::promise<void> p;
+        auto f = p.get_future();
+        t.scheduler.event_queue->enqueue([this, &p]
+        {
+            t.scheduler.attachChild(root_node);
+            p.set_value();
+        });
+        f.get();
+    }
+
+    void unregisterResource()
+    {
+        std::promise<void> p;
+        auto f = p.get_future();
+        t.scheduler.event_queue->enqueue([this, &p]
+        {
+            t.scheduler.removeChild(root_node.get());
+            p.set_value();
+        });
+        f.get();
+    }
+};
+
+TEST(IOSchedulerRoot, Smoke)
+{
+    ResourceTest t;
+
+    ResourceHolder r1(t);
+    auto * fc1 = r1.add<ConstraintTest>("/", "<max_requests>1</max_requests>");
+    r1.add<PriorityPolicy>("/prio");
+    auto a = r1.addQueue("/prio/A", "<priority>1</priority>");
+    auto b = r1.addQueue("/prio/B", "<priority>2</priority>");
+    r1.registerResource();
+
+    ResourceHolder r2(t);
+    auto * fc2 = r2.add<ConstraintTest>("/", "<max_requests>1</max_requests>");
+    r2.add<PriorityPolicy>("/prio");
+    auto c = r2.addQueue("/prio/C", "<priority>-1</priority>");
+    auto d = r2.addQueue("/prio/D", "<priority>-2</priority>");
+    r2.registerResource();
+
+    {
+        ResourceGuard rg(a);
+        EXPECT_TRUE(fc1->requests.contains(&rg.request));
+    }
+
+    {
+        ResourceGuard rg(b);
+        EXPECT_TRUE(fc1->requests.contains(&rg.request));
+    }
+
+    {
+        ResourceGuard rg(c);
+        EXPECT_TRUE(fc2->requests.contains(&rg.request));
+    }
+
+    {
+        ResourceGuard rg(d);
+        EXPECT_TRUE(fc2->requests.contains(&rg.request));
+    }
+}

src/IO/ResourceGuard.h

@@ -0,0 +1,93 @@
+#pragma once
+
+#include <base/types.h>
+
+#include <IO/ResourceRequest.h>
+#include <IO/ISchedulerQueue.h>
+#include <IO/ISchedulerConstraint.h>
+
+#include <future>
+
+namespace DB
+{
+
+/*
+ * Scoped resource guard.
+ * Waits for resource to be available in constructor and releases resource in destructor
+ */
+class ResourceGuard
+{
+public:
+    enum ResourceGuardCtor
+    {
+        LockStraightAway, /// Lock inside constructor (default)
+        PostponeLocking /// Don't lock in constructor, but during later `lock()` call
+    };
+
+    struct Request : public ResourceRequest
+    {
+        /// Promise to be set on request execution
+        std::promise<void> dequeued;
+
+        explicit Request(ResourceCost cost_ = 1)
+            : ResourceRequest(cost_)
+        {}
+
+        void execute() override
+        {
+            // This function is executed inside scheduler thread and wakes thread issued this `request` (using ResourceGuard)
+            // That thread will continue execution and do real consumption of requested resource synchronously.
+            dequeued.set_value();
+        }
+    };
+
+    /// Creates pending request for resource; blocks while resource is not available (unless `PostponeLocking`)
+    explicit ResourceGuard(ResourceLink link_, ResourceCost cost = 1, ResourceGuardCtor ctor = LockStraightAway)
+        : link(link_)
+        , request(cost)
+    {
+        if (link.queue)
+        {
+            dequeued_future = request.dequeued.get_future();
+            link.queue->enqueueRequest(&request);
+            if (ctor == LockStraightAway)
+                lock();
+        }
+    }
+
+    ~ResourceGuard()
+    {
+        unlock();
+    }
+
+    /// Blocks until resource is available
+    void lock()
+    {
+        if (link.queue)
+            dequeued_future.get();
+    }
+
+    /// Report request execution has finished
+    void unlock()
+    {
+        if (link.queue)
+        {
+            assert(!dequeued_future.valid()); // unlock must be called only after lock()
+            if (request.constraint)
+                request.constraint->finishRequest(&request);
+        }
+    }
+
+    /// Mark request as unsuccessful; by default request is considered to be successful
+    void setFailure()
+    {
+        request.successful = false;
+    }
+
+public:
+    ResourceLink link;
+    Request request;
+    std::future<void> dequeued_future;
+};
+
+}

src/IO/ResourceManagerFactory.h

@@ -0,0 +1,55 @@
+#pragma once
+
+#include <Common/ErrorCodes.h>
+#include <Common/Exception.h>
+
+#include <IO/IResourceManager.h>
+
+#include <boost/noncopyable.hpp>
+
+#include <memory>
+#include <mutex>
+#include <unordered_map>
+
+namespace DB
+{
+
+namespace ErrorCodes
+{
+    extern const int INVALID_SCHEDULER_NODE;
+}
+
+class ResourceManagerFactory : private boost::noncopyable
+{
+public:
+    static ResourceManagerFactory & instance()
+    {
+        static ResourceManagerFactory ret;
+        return ret;
+    }
+
+    ResourceManagerPtr get(const String & name)
+    {
+        std::lock_guard lock{mutex};
+        if (auto iter = methods.find(name); iter != methods.end())
+            return iter->second();
+        throw Exception(ErrorCodes::INVALID_SCHEDULER_NODE, "Unknown scheduler node type: {}", name);
+    }
+
+    template <class TDerived>
+    void registerMethod(const String & name)
+    {
+        std::lock_guard lock{mutex};
+        methods[name] = [] ()
+        {
+            return std::make_shared<TDerived>();
+        };
+    }
+
+private:
+    std::mutex mutex;
+    using Method = std::function<ResourceManagerPtr()>;
+    std::unordered_map<String, Method> methods;
+};
+
+}

src/IO/ResourceRequest.h

@@ -0,0 +1,92 @@
+#pragma once
+
+#include <base/types.h>
+#include <limits>
+
+namespace DB
+{
+
+// Forward declarations
+class ISchedulerQueue;
+class ISchedulerNode;
+class ISchedulerConstraint;
+
+/// Cost in terms of used resource (e.g. bytes for network IO)
+using ResourceCost = Int64;
+constexpr ResourceCost ResourceCostMax = std::numeric_limits<int>::max();
+
+/// Internal identifier of a resource (for arrays; unique per scheduler)
+using ResourceIdx = size_t;
+constexpr ResourceIdx ResourceIdxNotSet = ResourceIdx(-1);
+
+/// Timestamps (nanoseconds since epoch)
+using ResourceNs = UInt64;
+
+/*
+ * Info required for resource consumption.
+ */
+struct ResourceLink
+{
+    ISchedulerQueue * queue = nullptr;
+    bool operator==(const ResourceLink &) const = default;
+};
+
+/*
+ * Request for a resource consumption. The main moving part of the scheduling subsystem.
+ * Resource requests processing workflow:
+ *
+ * ----1=2222222222222=3=4=555555555555555=6-----> time
+ *     ^     ^         ^ ^          ^      ^
+ *     |     |         | |          |      |
+ *  enqueue wait dequeue execute consume finish
+ *
+ *  1) Request is enqueued using ISchedulerQueue::enqueueRequest().
+ *  2) Request competes with others for access to a resource; effectively just waiting in a queue.
+ *  3) Scheduler calls ISchedulerNode::dequeueRequest() that returns the request.
+ *  4) Callback ResourceRequest::execute() is called to provide access to the resource.
+ *  5) The resource consumption is happening outside of the scheduling subsystem.
+ *  6) request->constraint->finishRequest() is called when consumption is finished.
+ *
+ * Steps (5) and (6) can be omitted if constraint is not used by the resource.
+ *
+ * Request can be created on stack or heap.
+ * Request ownership is done outside of the scheduling subsystem.
+ * After (6) request can be destructed safely.
+ *
+ * Request cancelling is not supported yet.
+ */
+class ResourceRequest
+{
+public:
+    /// Cost of request execution; should be filled before request enqueueing.
+    /// NOTE: If cost is not known in advance, credit model can be used:
+    /// NOTE: for the first request use 1 and
+    ResourceCost cost;
+
+    /// Request outcome
+    /// Should be filled during resource consumption
+    bool successful = true;
+
+    /// Scheduler node to be notified on consumption finish
+    /// Auto-filled during request enqueue/dequeue
+    ISchedulerConstraint * constraint = nullptr;
+
+    /// Timestamps for introspection
+    ResourceNs enqueue_ns = 0;
+    ResourceNs execute_ns = 0;
+    ResourceNs finish_ns = 0;
+
+    explicit ResourceRequest(ResourceCost cost_ = 1)
+        : cost(cost_)
+    {}
+
+    virtual ~ResourceRequest() = default;
+
+    /// Callback to trigger resource consumption.
+    /// IMPORTANT: is called from scheduler thread and must be fast,
+    /// just triggering start of a consumption, not doing the consumption itself
+    /// (e.g. setting an std::promise or creating a job in a thread pool)
+    virtual void execute() = 0;
+};
+
+}

src/IO/SchedulerNodeFactory.h

@@ -0,0 +1,57 @@
+#pragma once
+
+#include <Common/ErrorCodes.h>
+#include <Common/Exception.h>
+
+#include <IO/ISchedulerNode.h>
+
+#include <Poco/Util/AbstractConfiguration.h>
+
+#include <boost/noncopyable.hpp>
+
+#include <memory>
+#include <mutex>
+#include <unordered_map>
+
+namespace DB
+{
+
+namespace ErrorCodes
+{
+    extern const int INVALID_SCHEDULER_NODE;
+}
+
+class SchedulerNodeFactory : private boost::noncopyable
+{
+public:
+    static SchedulerNodeFactory & instance()
+    {
+        static SchedulerNodeFactory ret;
+        return ret;
+    }
+
+    SchedulerNodePtr get(const String & name, EventQueue * event_queue, const Poco::Util::AbstractConfiguration & config, const String & config_prefix)
+    {
+        std::lock_guard lock{mutex};
+        if (auto iter = methods.find(name); iter != methods.end())
+            return iter->second(event_queue, config, config_prefix);
+        throw Exception(ErrorCodes::INVALID_SCHEDULER_NODE, "Unknown scheduler node type: {}", name);
+    }
+
+    template <class TDerived>
+    void registerMethod(const String & name)
+    {
+        std::lock_guard lock{mutex};
+        methods[name] = [] (EventQueue * event_queue, const Poco::Util::AbstractConfiguration & config, const String & config_prefix)
+        {
+            return std::make_shared<TDerived>(event_queue, config, config_prefix);
+        };
+    }
+
+private:
+    std::mutex mutex;
+    using Method = std::function<SchedulerNodePtr(EventQueue * event_queue, const Poco::Util::AbstractConfiguration & config, const String & config_prefix)>;
+    std::unordered_map<String, Method> methods;
+};
+
+}

src/IO/SchedulerRoot.h

@@ -0,0 +1,250 @@
+#pragma once
+
+#include <base/defines.h>
+
+#include <Common/Stopwatch.h>
+#include <Common/ThreadPool.h>
+
+#include <IO/ISchedulerNode.h>
+#include <IO/ISchedulerConstraint.h>
+
+#include <Poco/Util/XMLConfiguration.h>
+
+#include <unordered_map>
+#include <map>
+#include <memory>
+#include <atomic>
+
+
+namespace DB
+{
+
+namespace ErrorCodes
+{
+    extern const int INVALID_SCHEDULER_NODE;
+}
+
+/*
+ * Resource scheduler root node with a dedicated thread.
+ * Immediate children correspond to different resources.
+ */
+class SchedulerRoot : public ISchedulerNode
+{
+private:
+    struct TResource
+    {
+        SchedulerNodePtr root;
+
+        // Intrusive cyclic list of active resources
+        TResource * next = nullptr;
+        TResource * prev = nullptr;
+
+        explicit TResource(const SchedulerNodePtr & root_)
+            : root(root_)
+        {
+            root->info.parent.ptr = this;
+        }
+
+        // Get pointer stored by ctor in info
+        static TResource * get(SchedulerNodeInfo & info)
+        {
+            return reinterpret_cast<TResource *>(info.parent.ptr);
+        }
+    };
+
+public:
+    SchedulerRoot()
+        : ISchedulerNode(&events)
+    {}
+
+    ~SchedulerRoot() override
+    {
+        stop();
+    }
+
+    /// Runs separate scheduler thread
+    void start()
+    {
+        if (!scheduler.joinable())
+            scheduler = ThreadFromGlobalPool([this] { schedulerThread(); });
+    }
+
+    /// Joins scheduler threads and execute every pending request iff graceful
+    void stop(bool graceful = true)
+    {
+        if (scheduler.joinable())
+        {
+            stop_flag.store(true);
+            events.enqueue([]{}); // just to wake up thread
+            scheduler.join();
+            if (graceful)
+            {
+                // Do the same cycle as schedulerThread() but never block, just exit instead
+                bool has_work = true;
+                while (has_work)
+                {
+                    auto [request, _] = dequeueRequest();
+                    if (request)
+                        execute(request);
+                    else
+                        has_work = false;
+                    while (events.tryProcess())
+                        has_work = true;
+                }
+            }
+        }
+    }
+
+    bool equals(ISchedulerNode * other) override
+    {
+        if (auto * o = dynamic_cast<SchedulerRoot *>(other))
+            return true;
+        return false;
+    }
+
+    void attachChild(const SchedulerNodePtr & child) override
+    {
+        // Take ownership
+        assert(child->parent == nullptr);
+        if (auto [it, inserted] = children.emplace(child.get(), child); !inserted)
+            throw Exception(
+                ErrorCodes::INVALID_SCHEDULER_NODE,
+                "Can't add the same scheduler node twice");
+
+        // Attach
+        child->setParent(this);
+
+        // Activate child if required
+        if (child->isActive())
+            activateChild(child.get());
+    }
+
+    void removeChild(ISchedulerNode * child) override
+    {
+        if (auto iter = children.find(child); iter != children.end())
+        {
+            SchedulerNodePtr removed = iter->second.root;
+
+            // Deactivate if required
+            deactivate(&iter->second);
+
+            // Detach
+            removed->setParent(nullptr);
+
+            // Remove ownership
+            children.erase(iter);
+        }
+    }
+
+    ISchedulerNode * getChild(const String &) override
+    {
+        abort(); // scheduler is allowed to have multiple children with the same name
+    }
+
+    std::pair<ResourceRequest *, bool> dequeueRequest() override
+    {
+        if (current == nullptr) // No active resources
+            return {nullptr, false};
+
+        // Dequeue request from current resource
+        auto [request, resource_active] = current->root->dequeueRequest();
+        assert(request != nullptr);
+
+        // Deactivate resource if required
+        if (!resource_active)
+            deactivate(current);
+        else
+            current = current->next; // Just move round-robin pointer
+
+        return {request, current != nullptr};
+    }
+
+    bool isActive() override
+    {
+        return current != nullptr;
+    }
+
+    void activateChild(ISchedulerNode * child) override
+    {
+        activate(TResource::get(child->info));
+    }
+
+    void setParent(ISchedulerNode *) override
+    {
+        abort(); // scheduler must be the root and this function should not be called
+    }
+
+private:
+    void activate(TResource * value)
+    {
+        assert(value->next == nullptr && value->prev == nullptr);
+        if (current == nullptr) // No active children
+        {
+            current = value;
+            value->prev = value;
+            value->next = value;
+        }
+        else
+        {
+            current->prev->next = value;
+            value->prev = current->prev;
+            current->prev = value;
+            value->next = current;
+        }
+    }
+
+    void deactivate(TResource * value)
+    {
+        if (value->next == nullptr)
+            return; // Already deactivated
+        assert(current != nullptr);
+        if (current == value)
+        {
+            if (current->next == current) // We are going to remove the last active child
+            {
+                value->next = nullptr;
+                value->prev = nullptr;
+                current = nullptr;
+                return;
+            }
+            else // Just move current to next to avoid invalidation
+                current = current->next;
+        }
+        value->prev->next = value->next;
+        value->next->prev = value->prev;
+        value->prev = nullptr;
+        value->next = nullptr;
+    }
+
+private:
+    void schedulerThread()
+    {
+        while (!stop_flag.load())
+        {
+            // Dequeue and execute single request
+            auto [request, _] = dequeueRequest();
+            if (request)
+                execute(request);
+            else // No more requests -- block until any event happens
+                events.process();
+
+            // Process all events before dequeuing to ensure fair competition
+            while (events.tryProcess()) {}
+        }
+    }
+
+    void execute(ResourceRequest * request)
+    {
+        request->execute_ns = clock_gettime_ns();
+        request->execute();
+    }
+
+private:
+    TResource * current = nullptr; // round-robin pointer
+    std::unordered_map<ISchedulerNode *, TResource> children; // resources by pointer
+    std::atomic<bool> stop_flag = false;
+    EventQueue events;
+    ThreadFromGlobalPool scheduler;
+};
+
+}

src/Interpreters/Context.cpp

@@ -54,6 +54,7 @@
 #include <Access/SettingsConstraintsAndProfileIDs.h>
 #include <Access/ExternalAuthenticators.h>
 #include <Access/GSSAcceptor.h>
+#include <IO/ResourceManagerFactory.h>
 #include <Backups/BackupsWorker.h>
 #include <Dictionaries/Embedded/GeoDictionariesLoader.h>
 #include <Interpreters/EmbeddedDictionaries.h>
@@ -221,6 +222,7 @@ struct ContextSharedPart : boost::noncopyable
     String system_profile_name;                             /// Profile used by system processes
     String buffer_profile_name;                             /// Profile used by Buffer engine for flushing to the underlying
     std::unique_ptr<AccessControl> access_control;
+    mutable ResourceManagerPtr resource_manager;
     mutable UncompressedCachePtr uncompressed_cache;        /// The cache of decompressed blocks.
     mutable MarkCachePtr mark_cache;                        /// Cache of marks in compressed files.
     mutable std::unique_ptr<ThreadPool> load_marks_threadpool; /// Threadpool for loading marks cache.
@@ -1065,6 +1067,21 @@ std::vector<UUID> Context::getEnabledProfiles() const
 }
 
 
+ResourceManagerPtr Context::getResourceManager() const
+{
+    auto lock = getLock();
+    if (!shared->resource_manager)
+        shared->resource_manager = ResourceManagerFactory::instance().get(getConfigRef().getString("resource_manager", "static"));
+    return shared->resource_manager;
+}
+
+ClassifierPtr Context::getClassifier() const
+{
+    auto lock = getLock();
+    return getResourceManager()->acquire(getSettingsRef().workload);
+}
+
+
 const Scalars & Context::getScalars() const
 {
     return scalars;

src/Interpreters/Context.h

@@ -19,6 +19,7 @@
 #include <base/types.h>
 #include <Storages/MergeTree/ParallelReplicasReadingCoordinator.h>
 #include <Storages/ColumnsDescription.h>
+#include <IO/IResourceManager.h>
 
 #include <Server/HTTP/HTTPContext.h>
 
@@ -540,6 +541,10 @@ public:
     std::shared_ptr<const EnabledQuota> getQuota() const;
     std::optional<QuotaUsage> getQuotaUsage() const;
 
+    /// Resource management related
+    ResourceManagerPtr getResourceManager() const;
+    ClassifierPtr getClassifier() const;
+
     /// We have to copy external tables inside executeQuery() to track limits. Therefore, set callback for it. Must set once.
     void setExternalTablesInitializer(ExternalTablesInitializer && initializer);
     /// This method is called in executeQuery() and will call the external tables initializer.