Sql partition aware routing[API-1862] (#1167)

* sql aware draft

* sql_result update

* LRU Cache Interface

* std::priority_queue used

* sql service implementation

* revert unused file

* clang format

* LRU Test added

* LRU Test added

* basic type tests

* unit tests for complex types

* test case update

* test added for routing

* clang format

* 3 members for sql test

* cluster version check added

* test case failure fix

* member3_ reset

* additional test cases

* statement partition arg index set

* tdd review fix

* clang format

* comments added

* clang format

* jenkins fix

* compiler error fix for jenkins

* jenkins linux compiler error

* read_optimized_lru_cache removed from public API

* code review fixes

* initial value of lock

* rename variable

* clang format

* compiler error fixed

* windows bat file is updated

Author: akeles85

hazelcast/include/hazelcast/client/client_properties.h

@@ -99,6 +99,8 @@ class HAZELCAST_API client_properties
 
     const client_property& cloud_base_url() const;
 
+    const client_property& partition_arg_cache_size() const;
+
     /**
      * Client will be sending heartbeat messages to members and this is the
      * timeout. If there is no any message passing between client and member
@@ -264,6 +266,18 @@ class HAZELCAST_API client_properties
     static constexpr const char* CLOUD_URL_BASE_DEFAULT =
       "api.viridian.hazelcast.com";
 
+    /**
+     * Parametrized SQL queries touching only a single partition benefit from
+     * using the partition owner as the query coordinator, if the partition
+     * owner can be determined from one of the query parameters. When such a
+     * query is executed, the cluster sends the index of such argument to the
+     * client. This parameter configures the size of the cache the client uses
+     * for storing this information.
+     */
+    static constexpr const char* PARTITION_ARGUMENT_CACHE_SIZE =
+      "hazelcast.client.sql.partition.argument.cache.size";
+    static constexpr const char* PARTITION_ARGUMENT_CACHE_SIZE_DEFAULT = "1024";
+
     /**
      * Returns the configured boolean value of a {@link ClientProperty}.
      *
@@ -314,6 +328,7 @@ class HAZELCAST_API client_properties
     client_property backup_timeout_millis_;
     client_property fail_on_indeterminate_state_;
     client_property cloud_base_url_;
+    client_property partition_arg_cache_size_;
 
     std::unordered_map<std::string, std::string> properties_map_;
 };

hazelcast/include/hazelcast/client/sql/impl/read_optimized_lru_cache.h

@@ -0,0 +1,203 @@
+/*
+ * Copyright (c) 2008-2023, Hazelcast, Inc. All Rights Reserved.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#pragma once
+
+#include <queue>
+#include "hazelcast/util/SynchronizedMap.h"
+#include "hazelcast/util/export.h"
+
+namespace hazelcast {
+namespace client {
+namespace sql {
+namespace impl {
+
+/**
+ * Implementation of an LRU cache optimized for read-heavy use cases.
+ * <p>
+ * It stores the entries in a {SynchronizedMap}, along with the last
+ * access time. It allows the size to grow beyond the capacity, up to
+ * `cleanup_threshold_`, at which point the inserting thread will remove a batch
+ * of the eldest items in two passes.
+ * <p>
+ * The cleanup process isn't synchronized to guarantee that the capacity is not
+ * exceeded. The cache is available during the cleanup for reads and writes. If
+ * there's a large number of writes by many threads, the one thread doing the
+ * cleanup might not be quick enough and there's no upper bound on the actual
+ * size of the cache. This is done to optimize the happy path when the keys fit
+ * into the cache.
+ */
+template<typename K, typename V>
+class read_optimized_lru_cache
+{
+public:
+    /**
+     * @param capacity Capacity of the cache
+     * @param cleanup_threshold The size at which the cache will clean up oldest
+     *    entries in batch. `cleanup_threshold - capacity` entries will be
+     * removed
+     * @throws exception::illegal_argument if capacity is smaller or equal to 0,
+     * or if the cleanup_threshold is smaller than capacity
+     */
+    explicit read_optimized_lru_cache(const uint32_t capacity,
+                                      const uint32_t cleanup_threshold)
+    {
+        if (capacity == 0) {
+            BOOST_THROW_EXCEPTION(
+              client::exception::illegal_argument("capacity == 0"));
+        }
+        if (cleanup_threshold <= capacity) {
+            BOOST_THROW_EXCEPTION(client::exception::illegal_argument(
+              "cleanupThreshold <= capacity"));
+        }
+
+        capacity_ = capacity;
+        cleanup_threshold_ = cleanup_threshold;
+    }
+
+    /**
+     * @param key the key of the cache entry
+     * @param default_value the default value if the key is not cached.
+     * @returns Returns the value to which the specified key is cached,
+     * or default value if this cache contains no mapping for the key.
+     */
+    std::shared_ptr<V> get_or_default(const K& key,
+                                      const std::shared_ptr<V>& default_value)
+    {
+        const auto existing_value = get(key);
+        return (existing_value != nullptr) ? existing_value : default_value;
+    }
+
+    /**
+     * @param key the key of the cache entry
+     * Returns the value to which the specified key is cached,
+     * or {@code null} if this cache contains no mapping for the key.
+     * @returns Returns the value to which the specified key is cached
+     */
+    std::shared_ptr<V> get(const K& key)
+    {
+        auto value_from_cache = cache_.get(key);
+        if (value_from_cache == nullptr) {
+            return nullptr;
+        }
+        value_from_cache->touch();
+        return std::make_shared<int32_t>(value_from_cache->value_);
+    }
+
+    /**
+     * @param key the key of the cache entry
+     * @param value the value of the cache entry
+     * @throws exception::illegal_argument if the value equals to nullptr
+     */
+    void put(const K& key, const std::shared_ptr<V>& value)
+    {
+        if (value == nullptr) {
+            BOOST_THROW_EXCEPTION(client::exception::illegal_argument(
+              "Null values are disallowed"));
+        }
+
+        auto old_value =
+          cache_.put(key, std::make_shared<value_and_timestamp<V>>(*value));
+        if (old_value == nullptr && cache_.size() > cleanup_threshold_) {
+            do_cleanup();
+        }
+    }
+
+    /**
+     * @param key the key of the cache entry
+     * Removes the cached value for the given key
+     */
+    void remove(const K& key) { cache_.remove(key); }
+
+protected:
+    /**
+     * Helper class to hold the value with timestamp.
+     */
+    template<typename T>
+    class value_and_timestamp
+    {
+    public:
+        const T value_;
+        int64_t timestamp_;
+
+        value_and_timestamp(T value)
+          : value_(value)
+        {
+            touch();
+        }
+
+        void touch() { timestamp_ = util::current_time_nanos(); }
+    };
+
+    util::SynchronizedMap<K, value_and_timestamp<V>> cache_;
+
+private:
+    /**
+     * Cleans the cache
+     */
+    void do_cleanup()
+    {
+        bool expected = false;
+        // if no thread is cleaning up, we'll do it
+        if (!cleanup_lock_.compare_exchange_strong(expected, true)) {
+            return;
+        }
+
+        util::finally release_lock(
+          [this]() { this->cleanup_lock_.store(false); });
+
+        if (capacity_ >= cache_.size()) {
+            // this can happen if the cache is concurrently modified
+            return;
+        }
+        auto entries_to_remove = cache_.size() - capacity_;
+
+        /*max heap*/
+        std::priority_queue<int64_t> oldest_timestamps;
+
+        // 1st pass
+        const auto values = cache_.values();
+        for (const auto& value_and_timestamp : values) {
+            oldest_timestamps.push(value_and_timestamp->timestamp_);
+            if (oldest_timestamps.size() > entries_to_remove) {
+                oldest_timestamps.pop();
+            }
+        }
+
+        // find out the highest value in the queue - the value, below which
+        // entries will be removed
+        if (oldest_timestamps.empty()) {
+            // this can happen if the cache is concurrently modified
+            return;
+        }
+        int64_t remove_threshold = oldest_timestamps.top();
+        oldest_timestamps.pop();
+
+        // 2nd pass
+        cache_.remove_values_if(
+          [remove_threshold](const value_and_timestamp<V>& v) -> bool {
+              return (v.timestamp_ <= remove_threshold);
+          });
+    }
+
+    std::atomic<bool> cleanup_lock_{ false };
+    uint32_t capacity_;
+    uint32_t cleanup_threshold_;
+};
+
+} // namespace impl
+} // namespace sql
+} // namespace client
+} // namespace hazelcast

hazelcast/include/hazelcast/client/sql/sql_service.h

@@ -19,6 +19,7 @@
 #include "hazelcast/client/sql/sql_result.h"
 #include "hazelcast/client/sql/sql_statement.h"
 #include "hazelcast/client/sql/hazelcast_sql_exception.h"
+#include "hazelcast/client/sql/impl/read_optimized_lru_cache.h"
 
 #if defined(WIN32) || defined(_WIN32) || defined(WIN64) || defined(_WIN64)
 #pragma warning(push)
@@ -127,12 +128,17 @@ class HAZELCAST_API sql_service
     boost::future<std::shared_ptr<sql_result>> execute(
       const sql_statement& statement);
 
+    std::shared_ptr<impl::read_optimized_lru_cache<std::string, int32_t>>
+      partition_argument_index_cache_;
+
 private:
     friend client::impl::hazelcast_client_instance_impl;
     friend sql_result;
 
     client::spi::ClientContext& client_context_;
 
+    bool is_smart_routing_;
+
     struct sql_execute_response_parameters
     {
         int64_t update_count;
@@ -141,6 +147,8 @@ class HAZELCAST_API sql_service
         boost::optional<impl::sql_error> error;
         bool is_infinite_rows = false;
         bool is_infinite_rows_exist = false;
+        int32_t partition_argument_index = -1;
+        bool is_partition_argument_index_exists = false;
     };
 
     struct sql_fetch_response_parameters
@@ -152,6 +160,12 @@ class HAZELCAST_API sql_service
     explicit sql_service(client::spi::ClientContext& context);
 
     std::shared_ptr<connection::Connection> query_connection();
+    std::shared_ptr<connection::Connection> query_connection(
+      int32_t partition_id);
+
+    boost::optional<int32_t> extract_partition_id(
+      const sql_statement& statement,
+      const int32_t arg_index) const;
 
     void rethrow(const std::exception& exc_ptr);
     void rethrow(const std::exception& cause_ptr,
@@ -160,10 +174,13 @@ class HAZELCAST_API sql_service
     boost::uuids::uuid client_id();
 
     std::shared_ptr<sql_result> handle_execute_response(
+      const std::string& sql_query,
+      const int32_t original_partition_argument_index,
       protocol::ClientMessage& msg,
       std::shared_ptr<connection::Connection> connection,
       impl::query_id id,
-      int32_t cursor_buffer_size);
+      int32_t cursor_buffer_size,
+      std::weak_ptr<std::atomic<int32_t>> statement_par_arg_index_ptr);
 
     static sql_execute_response_parameters decode_execute_response(
       protocol::ClientMessage& msg);

hazelcast/include/hazelcast/client/sql/sql_statement.h

@@ -194,6 +194,13 @@ class HAZELCAST_API sql_statement
      */
     sql_statement& expected_result_type(sql_expected_result_type type);
 
+    /**
+     * Get the partition argument index value
+     *
+     * @return partition argument index, -1 if not set.
+     */
+    std::shared_ptr<std::atomic<int32_t>> partition_argument_index() const;
+
     /**
      * Gets the schema name.
      *
@@ -224,12 +231,26 @@ class HAZELCAST_API sql_statement
 
     sql_statement(spi::ClientContext& client_context, std::string query);
 
+    /**
+     * Set the partition argument index. If there's no such argument, use -1.
+     * <p>
+     * Setting a wrong argument index will not cause incorrect query results,
+     * but might cause performance degradation due to more network
+     * communication. Setting a value higher than the actual number of arguments
+     * will have no effect.
+     *
+     * @param partition_argument_index index of the partition-determining
+     * argument of the statement
+     */
+    sql_statement& partition_argument_index(int32_t partition_argument_index);
+
     std::string sql_;
     std::vector<data> serialized_parameters_;
     int32_t cursor_buffer_size_;
     std::chrono::milliseconds timeout_;
     sql::sql_expected_result_type expected_result_type_;
     boost::optional<std::string> schema_;
+    std::shared_ptr<std::atomic<int32_t>> partition_argument_index_;
 
     serialization_service& serialization_service_;
 

hazelcast/include/hazelcast/util/SynchronizedMap.h

@@ -242,6 +242,24 @@ class SynchronizedMap
         return internal_map_.empty();
     }
 
+    /**
+     * @param comp Map is iterated from beginning to the end and removed if the
+     * lambda comp return true.
+     */
+    template<typename Comparator>
+    void remove_values_if(Comparator comp)
+    {
+        std::lock_guard<std::mutex> lg(map_lock_);
+
+        for (auto iter = internal_map_.begin(); iter != internal_map_.end();) {
+            if (iter->second != nullptr && comp(*(iter->second))) {
+                iter = internal_map_.erase(iter);
+            } else {
+                ++iter;
+            }
+        }
+    }
+
 private:
     std::unordered_map<K, std::shared_ptr<V>, Hash> internal_map_;
     mutable std::mutex map_lock_;

hazelcast/src/hazelcast/client/client_impl.cpp

@@ -1074,6 +1074,8 @@ client_properties::client_properties(
   , fail_on_indeterminate_state_(FAIL_ON_INDETERMINATE_OPERATION_STATE,
                                  FAIL_ON_INDETERMINATE_OPERATION_STATE_DEFAULT)
   , cloud_base_url_(CLOUD_URL_BASE, CLOUD_URL_BASE_DEFAULT)
+  , partition_arg_cache_size_(PARTITION_ARGUMENT_CACHE_SIZE,
+                              PARTITION_ARGUMENT_CACHE_SIZE_DEFAULT)
   , properties_map_(properties)
 {}
 
@@ -1202,6 +1204,12 @@ client_properties::cloud_base_url() const
     return cloud_base_url_;
 }
 
+const client_property&
+client_properties::partition_arg_cache_size() const
+{
+    return partition_arg_cache_size_;
+}
+
 namespace exception {
 iexception::iexception(std::string exception_name,
                        std::string source,