ClickHouse/src/Dictionaries/SSDCacheDictionaryStorage.h

#pragma once

#if defined(__linux__) || defined(__FreeBSD__)

#include <chrono>

#include <pcg_random.hpp>
#include <filesystem>
#include <city.h>
#include <fcntl.h>
#include <absl/container/flat_hash_map.h>
#include <absl/container/flat_hash_set.h>

#include <common/unaligned.h>
#include <Common/Stopwatch.h>
#include <Common/randomSeed.h>
#include <Common/Arena.h>
#include <Common/ArenaWithFreeLists.h>
#include <Common/MemorySanitizer.h>
#include <Common/HashTable/HashMap.h>
#include <IO/AIO.h>
#include <Dictionaries/DictionaryStructure.h>
#include <Dictionaries/ICacheDictionaryStorage.h>
#include <Dictionaries/DictionaryHelpers.h>

namespace ProfileEvents
{
    extern const Event FileOpen;
    extern const Event WriteBufferAIOWrite;
    extern const Event WriteBufferAIOWriteBytes;
}

namespace DB
{

namespace ErrorCodes
{
    extern const int AIO_READ_ERROR;
    extern const int AIO_WRITE_ERROR;
    extern const int CANNOT_ALLOCATE_MEMORY;
    extern const int CANNOT_CREATE_DIRECTORY;
    extern const int CANNOT_FSYNC;
    extern const int CANNOT_IO_GETEVENTS;
    extern const int CANNOT_IO_SUBMIT;
    extern const int CANNOT_OPEN_FILE;
    extern const int CORRUPTED_DATA;
    extern const int FILE_DOESNT_EXIST;
    extern const int UNSUPPORTED_METHOD;
    extern const int NOT_IMPLEMENTED;
}

struct SSDCacheDictionaryStorageConfiguration
{
    const size_t strict_max_lifetime_seconds;
    const DictionaryLifetime lifetime;

    const std::string file_path;
    const size_t max_partitions_count;
    const size_t block_size;
    const size_t file_blocks_size;
    const size_t read_buffer_blocks_size;
    const size_t write_buffer_blocks_size;
};


/** Simple Key is serialized in block with following structure
    key     | data_size | data
    8 bytes | 8 bytes   | data_size bytes

    Complex Key is serialized in block with following structure
    key_size     | key_data       | data_size | data
    8 bytes      | key_size bytes | 8 bytes   | data_size bytes
*/
template <typename TKeyType>
struct SSDCacheKey final
{
    using KeyType = TKeyType;

    SSDCacheKey(KeyType key_, size_t size_, const char * data_)
        : key(key_)
        , size(size_)
        , data(data_)
    {}

    KeyType key;
    size_t size;
    const char * data;
};

using SSDCacheSimpleKey = SSDCacheKey<UInt64>;
using SSDCacheComplexKey = SSDCacheKey<StringRef>;

/** Block is serialized with following structure
    check_sum | keys_size | [keys]
    8 bytes   | 8 bytes   |
*/
class SSDCacheBlock final
{
    static constexpr size_t block_header_check_sum_size = sizeof(size_t);
    static constexpr size_t block_header_keys_size = sizeof(size_t);
public:

    /// Block header size
    static constexpr size_t block_header_size = block_header_check_sum_size + block_header_keys_size;

    explicit SSDCacheBlock(size_t block_size_)
        : block_size(block_size_)
    {}

    /// Checks if simple key can be written in empty block with block_size
    static bool canBeWrittenInEmptyBlock(SSDCacheSimpleKey & simple_key, size_t block_size)
    {
        static constexpr size_t simple_key_size = sizeof(simple_key.key);

        return (block_header_size + simple_key_size + sizeof(simple_key.size) + simple_key.size) <= block_size;
    }

    /// Checks if complex key can be written in empty block with block_size
    static bool canBeWrittenInEmptyBlock(SSDCacheComplexKey & complex_key, size_t block_size)
    {
        StringRef & key = complex_key.key;
        size_t complex_key_size = sizeof(key.size) + key.size;

        return (block_header_size + complex_key_size + sizeof(complex_key.size) + complex_key.size) <= block_size;
    }

    /// Reset block with new block_data
    /// block_data must be filled with zeroes if it is new block
    inline void reset(char * new_block_data)
    {
        block_data = new_block_data;
        current_block_offset = block_header_size;
        keys_size = unalignedLoad<size_t>(new_block_data + block_header_check_sum_size);
    }

    /// Check if it is enough place to write key in block
    inline bool enoughtPlaceToWriteKey(const SSDCacheSimpleKey & cache_key) const
    {
        return (current_block_offset + (sizeof(cache_key.key) + sizeof(cache_key.size) + cache_key.size)) <= block_size;
    }

    /// Check if it is enough place to write key in block
    inline bool enoughtPlaceToWriteKey(const SSDCacheComplexKey & cache_key) const
    {
        const StringRef & key = cache_key.key;
        size_t complex_key_size = sizeof(key.size) + key.size;

        return (current_block_offset + (complex_key_size + sizeof(cache_key.size) + cache_key.size)) <= block_size;
    }

    /// Write key and returns offset in ssd cache block where data is written
    /// It is client responsibility to check if there is enough place in block to write key
    /// Returns true if key was written and false if there was not enough place to write key
    inline bool writeKey(const SSDCacheSimpleKey & cache_key, size_t & offset_in_block)
    {
        assert(cache_key.size > 0);

        if (!enoughtPlaceToWriteKey(cache_key))
            return false;

        char * current_block_offset_data = block_data + current_block_offset;

        /// Write simple key
        memcpy(reinterpret_cast<void *>(current_block_offset_data), reinterpret_cast<const void *>(&cache_key.key), sizeof(cache_key.key));
        current_block_offset_data += sizeof(cache_key.key);
        current_block_offset += sizeof(cache_key.key);

        /// Write serialized columns size
        memcpy(reinterpret_cast<void *>(current_block_offset_data), reinterpret_cast<const void *>(&cache_key.size), sizeof(cache_key.size));
        current_block_offset_data += sizeof(cache_key.size);
        current_block_offset += sizeof(cache_key.size);

        offset_in_block = current_block_offset;

        memcpy(reinterpret_cast<void *>(current_block_offset_data), reinterpret_cast<const void *>(cache_key.data), cache_key.size);
        current_block_offset += cache_key.size;

        ++keys_size;

        return true;
    }

    inline bool writeKey(const SSDCacheComplexKey & cache_key, size_t & offset_in_block)
    {
        assert(cache_key.size > 0);

        if (!enoughtPlaceToWriteKey(cache_key))
            return false;

        char * current_block_offset_data = block_data + current_block_offset;

        const StringRef & key = cache_key.key;

        /// Write complex key
        memcpy(reinterpret_cast<void *>(current_block_offset_data), reinterpret_cast<const void *>(&key.size), sizeof(key.size));
        current_block_offset_data += sizeof(key.size);
        current_block_offset += sizeof(key.size);

        memcpy(reinterpret_cast<void *>(current_block_offset_data), reinterpret_cast<const void *>(key.data), key.size);
        current_block_offset_data += key.size;
        current_block_offset += key.size;

        /// Write serialized columns size
        memcpy(reinterpret_cast<void *>(current_block_offset_data), reinterpret_cast<const void *>(&cache_key.size), sizeof(cache_key.size));
        current_block_offset_data += sizeof(cache_key.size);
        current_block_offset += sizeof(cache_key.size);

        offset_in_block = current_block_offset;

        memcpy(reinterpret_cast<void *>(current_block_offset_data), reinterpret_cast<const void *>(cache_key.data), cache_key.size);
        current_block_offset += cache_key.size;

        ++keys_size;

        return true;
    }

    inline size_t getKeysSize() const { return keys_size; }

    /// Write keys size into block header
    inline void writeKeysSize()
    {
        char * keys_size_offset_data = block_data + block_header_check_sum_size;
        std::memcpy(keys_size_offset_data, &keys_size, sizeof(size_t));
    }

    /// Get check sum from block header
    inline size_t getCheckSum() const { return unalignedLoad<size_t>(block_data); }

    /// Calculate check sum in block
    inline size_t calculateCheckSum() const
    {
        size_t calculated_check_sum = static_cast<size_t>(CityHash_v1_0_2::CityHash64(block_data + block_header_check_sum_size, block_size - block_header_check_sum_size));

        return calculated_check_sum;
    }

    /// Check if check sum from block header matched calculated check sum in block
    inline bool checkCheckSum() const
    {
        size_t calculated_check_sum = calculateCheckSum();
        size_t check_sum = getCheckSum();

        return calculated_check_sum == check_sum;
    }

    /// Write check sum in block header
    inline void writeCheckSum()
    {
        size_t check_sum = static_cast<size_t>(CityHash_v1_0_2::CityHash64(block_data + block_header_check_sum_size, block_size - block_header_check_sum_size));
        std::memcpy(block_data, &check_sum, sizeof(size_t));
    }

    inline size_t getBlockSize() const { return block_size; }

    /// Returns block data
    inline char * getBlockData() const { return block_data; }

    /// Read keys that were serialized in block
    /// It is client responsibility to ensure that simple or complex keys were written in block
    void readSimpleKeys(PaddedPODArray<UInt64> & simple_keys) const
    {
        char * block_start = block_data + block_header_size;
        char * block_end = block_data + block_size;

        static constexpr size_t key_prefix_size = sizeof(UInt64) + sizeof(size_t);

        while (block_start + key_prefix_size < block_end)
        {
            UInt64 key = unalignedLoad<UInt64>(block_start);
            block_start += sizeof(UInt64);

            size_t allocated_size = unalignedLoad<size_t>(block_start);
            block_start += sizeof(size_t);

            /// If we read empty allocated size that means it is end of block
            if (allocated_size == 0)
                break;

            simple_keys.emplace_back(key);
            block_start += allocated_size;
        }
    }

    void readComplexKeys(PaddedPODArray<StringRef> & complex_keys) const
    {
        char * block_start = block_data + block_header_size;
        char * block_end = block_data + block_size;

        static constexpr size_t key_prefix_size = sizeof(size_t) + sizeof(size_t);

        while (block_start + key_prefix_size < block_end)
        {
            size_t key_size = unalignedLoad<size_t>(block_start);
            block_start += sizeof(key_size);

            StringRef complex_key (block_start, key_size);

            block_start += key_size;

            size_t allocated_size = unalignedLoad<size_t>(block_start);
            block_start += sizeof(size_t);

            /// If we read empty allocated size that means it is end of block
            if (allocated_size == 0)
                break;

            complex_keys.emplace_back(complex_key);
            block_start += allocated_size;
        }
    }

private:
    size_t block_size;
    char * block_data = nullptr;

    size_t current_block_offset = block_header_size;
    size_t keys_size = 0;
};

struct SSDCacheIndex
{
    SSDCacheIndex(size_t block_index_, size_t offset_in_block_)
        : block_index(block_index_)
        , offset_in_block(offset_in_block_)
    {}

    SSDCacheIndex() = default;

    size_t block_index = 0;
    size_t offset_in_block = 0;
};

inline bool operator==(const SSDCacheIndex & lhs, const SSDCacheIndex & rhs)
{
    return lhs.block_index == rhs.block_index && lhs.offset_in_block == rhs.offset_in_block;
}

/** Logically represents multiple memory_buffer_blocks_size SSDCacheBlocks and current write block.
  * If key cannot be written into current_write_block, current block keys size and check summ is written
  * and buffer increase index of current_write_block_index.
  * If current_write_block_index == memory_buffer_blocks_size write key will always returns true.
  * If reset is called current_write_block_index is set to 0.
  */
template <typename SSDCacheKeyType>
class SSDCacheMemoryBuffer
{
public:
    using KeyType = typename SSDCacheKeyType::KeyType;

    explicit SSDCacheMemoryBuffer(size_t block_size_, size_t memory_buffer_blocks_size_)
        : block_size(block_size_)
        , partition_blocks_size(memory_buffer_blocks_size_)
        , buffer(block_size * partition_blocks_size, 4096)
        , current_write_block(block_size)
    {
        current_write_block.reset(buffer.m_data);
    }

    bool writeKey(const SSDCacheKeyType & key, SSDCacheIndex & index)
    {
        if (current_block_index == partition_blocks_size)
            return false;

        size_t block_offset = 0;
        bool write_in_current_block = current_write_block.writeKey(key, block_offset);

        if (write_in_current_block)
        {
            index.block_index = current_block_index;
            index.offset_in_block = block_offset;
            return true;
        }

        current_write_block.writeKeysSize();
        current_write_block.writeCheckSum();

        ++current_block_index;

        if (current_block_index == partition_blocks_size)
            return false;

        current_write_block.reset(buffer.m_data + (block_size * current_block_index));

        write_in_current_block = current_write_block.writeKey(key, block_offset);
        assert(write_in_current_block);

        index.block_index = current_block_index;
        index.offset_in_block = block_offset;

        return write_in_current_block;
    }

    void writeKeysSizeAndCheckSumForCurrentWriteBlock()
    {
        current_write_block.writeKeysSize();
        current_write_block.writeCheckSum();
    }

    inline char * getPlace(SSDCacheIndex index) const
    {
        return buffer.m_data + index.block_index * block_size + index.offset_in_block;
    }

    inline size_t getCurrentBlockIndex() const { return current_block_index; }

    inline const char * getData() const { return buffer.m_data; }

    inline size_t getSizeInBytes() const { return block_size * partition_blocks_size; }

    void readKeys(PaddedPODArray<KeyType> & keys) const
    {
        SSDCacheBlock block(block_size);

        for (size_t block_index = 0; block_index < partition_blocks_size; ++block_index)
        {
            block.reset(buffer.m_data + (block_index * block_size));

            if constexpr (std::is_same_v<KeyType, UInt64>)
                block.readSimpleKeys(keys);
            else
                block.readComplexKeys(keys);
        }
    }

    inline void reset()
    {
        current_block_index = 0;
        current_write_block.reset(buffer.m_data);
    }

    const size_t block_size;

    const size_t partition_blocks_size;

private:
    Memory<Allocator<true>> buffer;

    SSDCacheBlock current_write_block;

    size_t current_block_index = 0;
};

/// Logically represents multiple memory_buffer_blocks_size SSDCacheBlocks on file system
template <typename SSDCacheKeyType>
class SSDCacheFileBuffer : private boost::noncopyable
{
    static constexpr auto BIN_FILE_EXT = ".bin";

public:

    using KeyType = typename SSDCacheKeyType::KeyType;

    explicit SSDCacheFileBuffer(
        const std::string & file_path_,
        size_t block_size_,
        size_t file_blocks_size_)
        : file_path(file_path_ + BIN_FILE_EXT)
        , block_size(block_size_)
        , file_blocks_size(file_blocks_size_)
    {
        auto path = std::filesystem::path{file_path};
        auto parent_path_directory = path.parent_path();

        /// If cache file is in directory that does not exists create it
        if (!std::filesystem::exists(parent_path_directory))
            if (!std::filesystem::create_directories(parent_path_directory))
                throw Exception{"Failed to create directories.", ErrorCodes::CANNOT_CREATE_DIRECTORY};

        ProfileEvents::increment(ProfileEvents::FileOpen);

        file.fd = ::open(file_path.c_str(), O_RDWR | O_CREAT | O_TRUNC | O_DIRECT, 0666);
        if (file.fd == -1)
        {
            auto error_code = (errno == ENOENT) ? ErrorCodes::FILE_DOESNT_EXIST : ErrorCodes::CANNOT_OPEN_FILE;
            throwFromErrnoWithPath("Cannot open file " + file_path, file_path, error_code);
        }

        allocateSizeForNextPartition();
    }

    void allocateSizeForNextPartition()
    {
        if (preallocateDiskSpace(file.fd, current_blocks_size * block_size, block_size * file_blocks_size) < 0)
            throwFromErrnoWithPath("Cannot preallocate space for the file " + file_path, file_path, ErrorCodes::CANNOT_ALLOCATE_MEMORY);

        current_blocks_size += file_blocks_size;
    }

    bool writeBuffer(const char * buffer, size_t buffer_size_in_blocks)
    {
        if (current_block_index + buffer_size_in_blocks > current_blocks_size)
            return false;

        AIOContext aio_context{1};

        iocb write_request{};
        iocb * write_request_ptr{&write_request};

        #if defined(__FreeBSD__)
        write_request.aio.aio_lio_opcode = LIO_WRITE;
        write_request.aio.aio_fildes = file.fd;
        write_request.aio.aio_buf = reinterpret_cast<volatile void *>(const_cast<char *>(buffer));
        write_request.aio.aio_nbytes = block_size * buffer_size_in_blocks;
        write_request.aio.aio_offset = current_block_index * block_size;
        #else
        write_request.aio_lio_opcode = IOCB_CMD_PWRITE;
        write_request.aio_fildes = file.fd;
        write_request.aio_buf = reinterpret_cast<UInt64>(buffer);
        write_request.aio_nbytes = block_size * buffer_size_in_blocks;
        write_request.aio_offset = current_block_index * block_size;
        #endif

        while (io_submit(aio_context.ctx, 1, &write_request_ptr) < 0)
        {
            if (errno != EINTR)
                throw Exception("Cannot submit request for asynchronous IO on file " + file_path, ErrorCodes::CANNOT_IO_SUBMIT);
        }

        // CurrentMetrics::Increment metric_increment_write{CurrentMetrics::Write};

        io_event event;

        while (io_getevents(aio_context.ctx, 1, 1, &event, nullptr) < 0)
        {
            if (errno != EINTR)
                throw Exception("Failed to wait for asynchronous IO completion on file " + file_path, ErrorCodes::CANNOT_IO_GETEVENTS);
        }

        // Unpoison the memory returned from an uninstrumented system function.
        __msan_unpoison(&event, sizeof(event));

        auto bytes_written = eventResult(event);

        ProfileEvents::increment(ProfileEvents::WriteBufferAIOWrite);
        ProfileEvents::increment(ProfileEvents::WriteBufferAIOWriteBytes, bytes_written);

        if (bytes_written != static_cast<decltype(bytes_written)>(block_size * buffer_size_in_blocks))
            throw Exception("Not all data was written for asynchronous IO on file " + file_path + ". returned: " + std::to_string(bytes_written), ErrorCodes::AIO_WRITE_ERROR);

        if (::fsync(file.fd) < 0)
            throwFromErrnoWithPath("Cannot fsync " + file_path, file_path, ErrorCodes::CANNOT_FSYNC);

        current_block_index += buffer_size_in_blocks;

        return true;
    }

    bool readKeys(size_t block_start, size_t blocks_length, PaddedPODArray<KeyType> & out) const
    {
        if (block_start + blocks_length > current_blocks_size)
            return false;

        size_t buffer_size_in_bytes = blocks_length * block_size;

        Memory read_buffer_memory(block_size * blocks_length, block_size);

        iocb request{};
        iocb * request_ptr = &request;

        #if defined(__FreeBSD__)
        request.aio.aio_lio_opcode = LIO_READ;
        request.aio.aio_fildes = file.fd;
        request.aio.aio_buf = reinterpret_cast<volatile void *>(reinterpret_cast<UInt64>(read_buffer_memory.data()));
        request.aio.aio_nbytes = buffer_size_in_bytes;
        request.aio.aio_offset = block_start * block_size;
        request.aio_data = 0;
        #else
        request.aio_lio_opcode = IOCB_CMD_PREAD;
        request.aio_fildes = file.fd;
        request.aio_buf = reinterpret_cast<UInt64>(read_buffer_memory.data());
        request.aio_nbytes = buffer_size_in_bytes;
        request.aio_offset = block_start * block_size;
        request.aio_data = 0;
        #endif

        io_event event{};
        AIOContext aio_context(1);

        while (io_submit(aio_context.ctx, 1, &request_ptr) != 1)
        {
            if (errno != EINTR)
                throwFromErrno("io_submit: Failed to submit a request for asynchronous IO", ErrorCodes::CANNOT_IO_SUBMIT);
        }

        while (io_getevents(aio_context.ctx, 1, 1, &event, nullptr) != 1)
        {
            if (errno != EINTR)
                throwFromErrno("io_getevents: Failed to get an event for asynchronous IO", ErrorCodes::CANNOT_IO_GETEVENTS);
        }

        auto read_bytes = eventResult(event);

        if (read_bytes != static_cast<ssize_t>(buffer_size_in_bytes))
            throw Exception(ErrorCodes::AIO_READ_ERROR,
                "GC: AIO failed to read file ({}). Expected bytes ({}). Actual bytes ({})", file_path, buffer_size_in_bytes, read_bytes);

        SSDCacheBlock block(block_size);

        for (size_t i = 0; i < blocks_length; ++i)
        {
            block.reset(read_buffer_memory.data() + (i * block_size));

            if constexpr (std::is_same_v<SSDCacheKeyType, SSDCacheSimpleKey>)
                block.readSimpleKeys(out);
            else
                block.readComplexKeys(out);
        }

        return true;
    }

    template <typename FetchBlockFunc>
    void fetchBlocks(size_t read_from_file_buffer_blocks_size, const PaddedPODArray<size_t> & blocks_to_fetch, FetchBlockFunc && func) const
    {
        if (blocks_to_fetch.empty())
            return;

        Memory<Allocator<true>> read_buffer(read_from_file_buffer_blocks_size * block_size, 4096);

        size_t blocks_to_fetch_size = blocks_to_fetch.size();

        PaddedPODArray<iocb> requests;
        PaddedPODArray<iocb *> pointers;

        requests.reserve(blocks_to_fetch_size);
        pointers.reserve(blocks_to_fetch_size);

        for (size_t block_to_fetch_index = 0; block_to_fetch_index < blocks_to_fetch_size; ++block_to_fetch_index)
        {
            iocb request{};

            char * buffer_place = read_buffer.data() + block_size * (block_to_fetch_index % read_from_file_buffer_blocks_size);

            #if defined(__FreeBSD__)
            request.aio.aio_lio_opcode = LIO_READ;
            request.aio.aio_fildes = file.fd;
            request.aio.aio_buf = reinterpret_cast<volatile void *>(reinterpret_cast<UInt64>(buffer_place));
            request.aio.aio_nbytes = block_size;
            request.aio.aio_offset = block_size * blocks_to_fetch[block_to_fetch_index];
            request.aio_data = block_to_fetch_index;
            #else
            request.aio_lio_opcode = IOCB_CMD_PREAD;
            request.aio_fildes = file.fd;
            request.aio_buf = reinterpret_cast<UInt64>(buffer_place);
            request.aio_nbytes = block_size;
            request.aio_offset = block_size * blocks_to_fetch[block_to_fetch_index];
            request.aio_data = block_to_fetch_index;
            #endif

            requests.push_back(request);
            pointers.push_back(&requests.back());
        }

        AIOContext aio_context(read_from_file_buffer_blocks_size);

        PaddedPODArray<bool> processed(requests.size(), false);
        PaddedPODArray<io_event> events;
        events.resize_fill(requests.size());

        size_t to_push = 0;
        size_t to_pop = 0;

        while (to_pop < requests.size())
        {
            int popped = 0;

            while (to_pop < to_push && (popped = io_getevents(aio_context.ctx, to_push - to_pop, to_push - to_pop, &events[to_pop], nullptr)) <= 0)
            {
                if (errno != EINTR)
                    throwFromErrno("io_getevents: Failed to get an event for asynchronous IO", ErrorCodes::CANNOT_IO_GETEVENTS);
            }

            for (size_t i = to_pop; i < to_pop + popped; ++i)
            {
                size_t block_to_fetch_index = events[i].data;
                const auto & request = requests[block_to_fetch_index];

                const ssize_t read_bytes = eventResult(events[i]);

                if (read_bytes != static_cast<ssize_t>(block_size))
                    throw Exception(ErrorCodes::AIO_READ_ERROR,
                        "GC: AIO failed to read file ({}). Expected bytes ({}). Actual bytes ({})", file_path, block_size, read_bytes);

                char * request_buffer = getRequestBuffer(request);

                // Unpoison the memory returned from an uninstrumented system function.
                __msan_unpoison(request_buffer, block_size);

                SSDCacheBlock block(block_size);
                block.reset(request_buffer);

                if (!block.checkCheckSum())
                {
                    std::string calculated_check_sum = std::to_string(block.calculateCheckSum());
                    std::string check_sum = std::to_string(block.getCheckSum());
                    throw Exception("Cache data corrupted. Checksum validation failed. Calculated " +  calculated_check_sum + " in block " + check_sum, ErrorCodes::CORRUPTED_DATA);
                }

                std::forward<FetchBlockFunc>(func)(blocks_to_fetch[block_to_fetch_index], block.getBlockData());

                processed[block_to_fetch_index] = true;
            }

            while (to_pop < requests.size() && processed[to_pop])
                ++to_pop;

            /// add new io tasks
            const int new_tasks_count = std::min(read_from_file_buffer_blocks_size - (to_push - to_pop), requests.size() - to_push);

            int pushed = 0;
            while (new_tasks_count > 0 && (pushed = io_submit(aio_context.ctx, new_tasks_count, &pointers[to_push])) <= 0)
            {
                if (errno != EINTR)
                    throwFromErrno("io_submit: Failed to submit a request for asynchronous IO", ErrorCodes::CANNOT_IO_SUBMIT);
            }

            to_push += pushed;
        }
    }

    inline size_t getCurrentBlockIndex() const { return current_block_index; }

    inline void reset()
    {
        current_block_index = 0;
    }
private:
    struct FileDescriptor : private boost::noncopyable
    {

        FileDescriptor() = default;

        FileDescriptor(FileDescriptor && rhs) : fd(rhs.fd) { rhs.fd = -1; }

        FileDescriptor & operator=(FileDescriptor && rhs)
        {
            close(fd);

            fd = rhs.fd;
            rhs.fd = -1;
        }

        ~FileDescriptor()
        {
            if (fd != -1)
                close(fd);
        }

        int fd = -1;
    };

    inline static int preallocateDiskSpace(int fd, size_t offset, size_t len)
    {
        #if defined(__FreeBSD__)
            return posix_fallocate(fd, offset, len);
        #else
            return fallocate(fd, 0, offset, len);
        #endif
    }

    inline static char * getRequestBuffer(const iocb & request)
    {
        char * result = nullptr;

        #if defined(__FreeBSD__)
            result = reinterpret_cast<char *>(reinterpret_cast<UInt64>(request.aio.aio_buf));
        #else
            result = reinterpret_cast<char *>(request.aio_buf);
        #endif

        return result;
    }

    inline static ssize_t eventResult(io_event & event)
    {
        ssize_t  bytes_written;

        #if defined(__FreeBSD__)
            bytes_written = aio_return(reinterpret_cast<struct aiocb *>(event.udata));
        #else
            bytes_written = event.res;
        #endif

        return bytes_written;
    }

    String file_path;
    size_t block_size;
    size_t file_blocks_size;
    FileDescriptor file;

    size_t current_block_index = 0;
    size_t current_blocks_size = 0;
};

/** ICacheDictionaryStorage implementation that keeps column data serialized in memory index and in disk partitions.
  * Data is first written in memory buffer.
  * If memory buffer is full then buffer is flushed to disk partition.
  * If memory buffer cannot be flushed to associated disk partition, then if partition
  * can be allocated (current partition index < max_partitions_size) storage allocates new partition, if not old partitions are reused.
  * Index maps key to partition block and offset.
  */
template <DictionaryKeyType dictionary_key_type>
class SSDCacheDictionaryStorage final : public ICacheDictionaryStorage
{
public:
    using SSDCacheKeyType = std::conditional_t<dictionary_key_type == DictionaryKeyType::simple, SSDCacheSimpleKey, SSDCacheComplexKey>;
    using KeyType = std::conditional_t<dictionary_key_type == DictionaryKeyType::simple, UInt64, StringRef>;

    explicit SSDCacheDictionaryStorage(const SSDCacheDictionaryStorageConfiguration & configuration_)
        : configuration(configuration_)
        , file_buffer(configuration_.file_path, configuration.block_size, configuration.file_blocks_size)
        , rnd_engine(randomSeed())
    {
        memory_buffer_partitions.emplace_back(configuration.block_size, configuration.write_buffer_blocks_size);
    }

    bool returnsFetchedColumnsInOrderOfRequestedKeys() const override { return false; }

    String getName() const override
    {
        if (dictionary_key_type == DictionaryKeyType::simple)
            return "SSDCache";
        else
            return "SSDComplexKeyCache";
    }

    bool supportsSimpleKeys() const override { return dictionary_key_type == DictionaryKeyType::simple; }

    SimpleKeysStorageFetchResult fetchColumnsForKeys(
        const PaddedPODArray<UInt64> & keys,
        const DictionaryStorageFetchRequest & fetch_request) override
    {
        if constexpr (dictionary_key_type == DictionaryKeyType::simple)
            return fetchColumnsForKeysImpl<SimpleKeysStorageFetchResult>(keys, fetch_request);
        else
            throw Exception("Method insertColumnsForKeys is not supported for complex key storage", ErrorCodes::NOT_IMPLEMENTED);
    }

    void insertColumnsForKeys(const PaddedPODArray<UInt64> & keys, Columns columns) override
    {
        if constexpr (dictionary_key_type == DictionaryKeyType::simple)
            insertColumnsForKeysImpl(keys, columns);
        else
            throw Exception("Method insertColumnsForKeys is not supported for complex key storage", ErrorCodes::NOT_IMPLEMENTED);
    }

    void insertDefaultKeys(const PaddedPODArray<UInt64> & keys) override
    {
        if constexpr (dictionary_key_type == DictionaryKeyType::simple)
            insertDefaultKeysImpl(keys);
        else
            throw Exception("Method insertDefaultKeysImpl is not supported for complex key storage", ErrorCodes::NOT_IMPLEMENTED);
    }

    PaddedPODArray<UInt64> getCachedSimpleKeys() const override
    {
        if constexpr (dictionary_key_type == DictionaryKeyType::simple)
            return getCachedKeysImpl();
        else
            throw Exception("Method getCachedSimpleKeys is not supported for complex key storage", ErrorCodes::NOT_IMPLEMENTED);
    }

    bool supportsComplexKeys() const override { return dictionary_key_type == DictionaryKeyType::complex; }

    ComplexKeysStorageFetchResult fetchColumnsForKeys(
        const PaddedPODArray<StringRef> & keys,
        const DictionaryStorageFetchRequest & fetch_request) override
    {
        if constexpr (dictionary_key_type == DictionaryKeyType::complex)
            return fetchColumnsForKeysImpl<ComplexKeysStorageFetchResult>(keys, fetch_request);
        else
            throw Exception("Method fetchColumnsForKeys is not supported for simple key storage", ErrorCodes::NOT_IMPLEMENTED);
    }

    void insertColumnsForKeys(const PaddedPODArray<StringRef> & keys, Columns columns) override
    {
        if constexpr (dictionary_key_type == DictionaryKeyType::complex)
            insertColumnsForKeysImpl(keys, columns);
        else
            throw Exception("Method insertColumnsForKeys is not supported for simple key storage", ErrorCodes::NOT_IMPLEMENTED);
    }

    void insertDefaultKeys(const PaddedPODArray<StringRef> & keys) override
    {
        if constexpr (dictionary_key_type == DictionaryKeyType::complex)
            insertDefaultKeysImpl(keys);
        else
            throw Exception("Method insertDefaultKeysImpl is not supported for simple key storage", ErrorCodes::NOT_IMPLEMENTED);
    }

    PaddedPODArray<StringRef> getCachedComplexKeys() const override
    {
        if constexpr (dictionary_key_type == DictionaryKeyType::complex)
            return getCachedKeysImpl();
        else
            throw Exception("Method getCachedSimpleKeys is not supported for simple key storage", ErrorCodes::NOT_IMPLEMENTED);
    }

    size_t getSize() const override { return index.size(); }

    double getLoadFactor() const override
    {
        size_t partitions_size = memory_buffer_partitions.size();

        if (partitions_size == configuration.max_partitions_count)
            return 1.0;

        auto & current_memory_partition = memory_buffer_partitions[current_partition_index];

        size_t full_partitions = partitions_size - 1;
        size_t blocks_in_memory = (full_partitions * configuration.write_buffer_blocks_size) + current_memory_partition.getCurrentBlockIndex();
        size_t blocks_on_disk = file_buffer.getCurrentBlockIndex();

        size_t max_blocks_size = (configuration.file_blocks_size + configuration.write_buffer_blocks_size) * configuration.max_partitions_count;

        double load_factor = static_cast<double>(blocks_in_memory + blocks_on_disk) / max_blocks_size;
        return load_factor;
    }

    size_t getBytesAllocated() const override
    {
        size_t memory_partitions_bytes_size = memory_buffer_partitions.size() * configuration.write_buffer_blocks_size * configuration.block_size;
        size_t file_partitions_bytes_size = memory_buffer_partitions.size() * configuration.file_blocks_size * configuration.block_size;

        return index.getBufferSizeInBytes() + memory_partitions_bytes_size + file_partitions_bytes_size;
    }

private:

    using TimePoint = std::chrono::system_clock::time_point;

    struct Cell
    {
        enum CellState
        {
            in_memory,
            on_disk,
            default_value
        };

        time_t deadline;
        SSDCacheIndex index;
        size_t in_memory_partition_index;
        CellState state;

        inline bool isInMemory() const { return state == in_memory; }
        inline bool isOnDisk() const { return state == on_disk; }
        inline bool isDefaultValue() const { return state == default_value; }
    };

    struct KeyToBlockOffset
    {
        KeyToBlockOffset(size_t key_index_, size_t offset_in_block_)
            : key_index(key_index_), offset_in_block(offset_in_block_)
        {}

        size_t key_index = 0;
        size_t offset_in_block = 0;
    };

    template <typename Result>
    Result fetchColumnsForKeysImpl(
        const PaddedPODArray<KeyType> & keys,
        const DictionaryStorageFetchRequest & fetch_request) const
    {
        Result result;

        result.fetched_columns = fetch_request.makeAttributesResultColumns();
        result.key_index_to_state.resize_fill(keys.size());

        const time_t now = std::chrono::system_clock::to_time_t(std::chrono::system_clock::now());

        size_t fetched_columns_index = 0;

        using BlockIndexToKeysMap = absl::flat_hash_map<size_t, PaddedPODArray<KeyToBlockOffset>, DefaultHash<size_t>>;
        BlockIndexToKeysMap block_to_keys_map;
        absl::flat_hash_set<size_t, DefaultHash<size_t>> unique_blocks_to_request;
        PaddedPODArray<size_t> blocks_to_request;

        time_t strict_max_lifetime_seconds = static_cast<time_t>(configuration.strict_max_lifetime_seconds);
        size_t keys_size = keys.size();

        for (size_t attribute_size = 0; attribute_size < fetch_request.attributesSize(); ++attribute_size)
            if (fetch_request.shouldFillResultColumnWithIndex(attribute_size))
                result.fetched_columns[attribute_size]->reserve(keys_size);

        for (size_t key_index = 0; key_index < keys_size; ++key_index)
        {
            auto key = keys[key_index];

            const auto * it = index.find(key);

            if (!it)
            {
                ++result.not_found_keys_size;
                continue;
            }

            const auto & cell = it->getMapped();

            if (unlikely(now > cell.deadline + strict_max_lifetime_seconds))
            {
                ++result.not_found_keys_size;
                continue;
            }

            bool cell_is_expired = false;
            KeyState::State key_state = KeyState::found;

            if (now > cell.deadline)
            {
                cell_is_expired = true;
                key_state = KeyState::expired;
            }

            result.expired_keys_size += static_cast<size_t>(cell_is_expired);
            result.found_keys_size += static_cast<size_t>(!cell_is_expired);

            switch (cell.state)
            {
                case Cell::in_memory:
                {
                    result.key_index_to_state[key_index] = {key_state, fetched_columns_index};
                    ++fetched_columns_index;

                    const auto & partition = memory_buffer_partitions[cell.in_memory_partition_index];
                    char * serialized_columns_place = partition.getPlace(cell.index);
                    deserializeAndInsertIntoColumns(result.fetched_columns, fetch_request, serialized_columns_place);
                    break;
                }
                case Cell::on_disk:
                {
                    PaddedPODArray<KeyToBlockOffset> & keys_block = block_to_keys_map[cell.index.block_index];
                    keys_block.emplace_back(key_index, cell.index.offset_in_block);

                    KeyState::State state = cell_is_expired ? KeyState::expired : KeyState::found;

                    /// Fetched column index will be set later during fetch blocks
                    result.key_index_to_state[key_index] = {state, 0};

                    auto insert_result = unique_blocks_to_request.insert(cell.index.block_index);
                    bool was_inserted = insert_result.second;

                    if (was_inserted)
                        blocks_to_request.emplace_back(cell.index.block_index);

                    break;
                }
                case Cell::default_value:
                {
                    result.key_index_to_state[key_index] = {key_state, fetched_columns_index};
                    result.key_index_to_state[key_index].setDefault();
                    ++fetched_columns_index;
                    ++result.default_keys_size;

                    insertDefaultValuesIntoColumns(result.fetched_columns, fetch_request, key_index);
                    break;
                }
            }
        }

        /// Sort blocks by offset before start async io requests
        std::sort(blocks_to_request.begin(), blocks_to_request.end());

        file_buffer.fetchBlocks(configuration.read_buffer_blocks_size, blocks_to_request, [&](size_t block_index, char * block_data)
        {
            auto & keys_in_block = block_to_keys_map[block_index];

            for (auto & key_in_block : keys_in_block)
            {
                char * key_data = block_data + key_in_block.offset_in_block;
                deserializeAndInsertIntoColumns(result.fetched_columns, fetch_request, key_data);

                result.key_index_to_state[key_in_block.key_index].setFetchedColumnIndex(fetched_columns_index);

                ++fetched_columns_index;
            }
        });

        return result;
    }

    void insertColumnsForKeysImpl(const PaddedPODArray<KeyType> & keys, Columns columns)
    {
        size_t columns_to_serialize_size = columns.size();
        PaddedPODArray<StringRef> temporary_column_data(columns_to_serialize_size);

        Arena temporary_values_pool;

        const auto now = std::chrono::system_clock::now();

        for (size_t key_index = 0; key_index < keys.size(); ++key_index)
        {
            size_t allocated_size_for_columns = 0;
            const char * block_start = nullptr;

            auto key = keys[key_index];

            for (size_t column_index = 0; column_index < columns_to_serialize_size; ++column_index)
            {
                auto & column = columns[column_index];
                temporary_column_data[column_index] = column->serializeValueIntoArena(key_index, temporary_values_pool, block_start);
                allocated_size_for_columns += temporary_column_data[column_index].size;
            }

            SSDCacheKeyType ssd_cache_key { key, allocated_size_for_columns, block_start };

            if (!SSDCacheBlock::canBeWrittenInEmptyBlock(ssd_cache_key, configuration.block_size))
                throw Exception("Serialized columns size is greater than allowed block size and metadata", ErrorCodes::UNSUPPORTED_METHOD);

            /// We cannot reuse place that is already allocated in file or memory cache so we erase key from index
            eraseKeyFromIndex(key);

            Cell cell;
            setCellDeadline(cell, now);

            if constexpr (dictionary_key_type == DictionaryKeyType::complex)
            {
                /// Copy complex key into arena and put in cache
                size_t key_size = key.size;
                char * place_for_key = complex_key_arena.alloc(key_size);
                memcpy(reinterpret_cast<void *>(place_for_key), reinterpret_cast<const void *>(key.data), key_size);
                KeyType updated_key{place_for_key, key_size};
                ssd_cache_key.key = updated_key;
            }

            insertCell(ssd_cache_key, cell);

            temporary_values_pool.rollback(allocated_size_for_columns);
        }
    }

    void insertDefaultKeysImpl(const PaddedPODArray<KeyType> & keys)
    {
        const auto now = std::chrono::system_clock::now();

        for (auto key : keys)
        {
            eraseKeyFromIndex(key);

            Cell cell;

            setCellDeadline(cell, now);
            cell.index = {0, 0};
            cell.in_memory_partition_index = 0;
            cell.state = Cell::default_value;


            if constexpr (dictionary_key_type == DictionaryKeyType::complex)
            {
                /// Copy complex key into arena and put in cache
                size_t key_size = key.size;
                char * place_for_key = complex_key_arena.alloc(key_size);
                memcpy(reinterpret_cast<void *>(place_for_key), reinterpret_cast<const void *>(key.data), key_size);
                KeyType updated_key{place_for_key, key_size};
                key = updated_key;
            }

            index[key] = cell;
        }
    }

    PaddedPODArray<KeyType> getCachedKeysImpl() const
    {
        PaddedPODArray<KeyType> result;
        result.reserve(index.size());

        for (auto & node : index)
        {
            auto & cell = node.getMapped();

            if (cell.state == Cell::default_value)
                continue;

            result.emplace_back(node.getKey());
        }

        return result;
    }

    void insertCell(SSDCacheKeyType & ssd_cache_key, Cell & cell)
    {
        /** InsertCell has following flow

            1. We try to write key into current memory buffer, if write succeeded then return.
            2. Then if we does not write key into current memory buffer, we try to flush current memory buffer
            to disk.

            If flush succeeded then reset current memory buffer, write key into it and return.
            If flush failed that means that current partition on disk is full, need to allocate new partition
            or start reusing old ones.

            Retry to step 1.
         */

        SSDCacheIndex cache_index {0, 0};

        while (true)
        {
            bool started_reusing_old_partitions = memory_buffer_partitions.size() == configuration.max_partitions_count;

            auto & current_memory_buffer_partition = memory_buffer_partitions[current_partition_index];

            bool write_into_memory_buffer_result = current_memory_buffer_partition.writeKey(ssd_cache_key, cache_index);

            if (write_into_memory_buffer_result)
            {
                cell.state = Cell::in_memory;
                cell.index = cache_index;
                cell.in_memory_partition_index = current_partition_index;

                index[ssd_cache_key.key] = cell;
                break;
            }
            else
            {
                /// Partition memory write buffer if full flush it to disk and retry
                size_t block_index_in_file_before_write = file_buffer.getCurrentBlockIndex();

                if (started_reusing_old_partitions)
                {
                    /// If we start reusing old partitions we need to remove old keys on disk from index before writing buffer
                    PaddedPODArray<KeyType> old_keys;
                    file_buffer.readKeys(block_index_in_file_before_write, configuration.write_buffer_blocks_size, old_keys);

                    size_t file_read_end_block_index = block_index_in_file_before_write + configuration.write_buffer_blocks_size;

                    for (auto old_key : old_keys)
                    {
                        auto * it = index.find(old_key);

                        if (it)
                        {
                            const Cell & old_key_cell = it->getMapped();

                            size_t old_key_block = old_key_cell.index.block_index;

                            /// Check if key in index is key from old partition blocks
                            if (old_key_cell.isOnDisk() &&
                                old_key_block >= block_index_in_file_before_write &&
                                old_key_block < file_read_end_block_index)
                                eraseKeyFromIndex(old_key);
                        }
                    }
                }

                const char * partition_data = current_memory_buffer_partition.getData();

                bool flush_to_file_result = file_buffer.writeBuffer(partition_data, configuration.write_buffer_blocks_size);

                if (flush_to_file_result)
                {
                    /// Update index cells keys offset and block index
                    PaddedPODArray<KeyType> keys_to_update;
                    current_memory_buffer_partition.readKeys(keys_to_update);

                    absl::flat_hash_set<KeyType, DefaultHash<KeyType>> updated_keys;

                    Int64 keys_to_update_size = static_cast<Int64>(keys_to_update.size());

                    /// Start from last to first because there can be multiple keys in same partition.
                    /// The valid key is the latest.
                    for (Int64 i = keys_to_update_size - 1; i >= 0; --i)
                    {
                        auto key_to_update = keys_to_update[i];
                        auto * it = index.find(key_to_update);

                        /// If there are no key to update or key to update not in memory
                        if (!it || it->getMapped().state != Cell::in_memory)
                            continue;

                        /// If there were duplicated keys in memory buffer partition
                        if (updated_keys.contains(it->getKey()))
                            continue;

                        updated_keys.insert(key_to_update);

                        Cell & cell_to_update = it->getMapped();

                        cell_to_update.state = Cell::on_disk;
                        cell_to_update.index.block_index += block_index_in_file_before_write;
                    }

                    /// Memory buffer partition flushed to disk start reusing it
                    current_memory_buffer_partition.reset();
                    memset(const_cast<char *>(current_memory_buffer_partition.getData()), 0, current_memory_buffer_partition.getSizeInBytes());

                    write_into_memory_buffer_result = current_memory_buffer_partition.writeKey(ssd_cache_key, cache_index);
                    assert(write_into_memory_buffer_result);

                    cell.state = Cell::in_memory;
                    cell.index = cache_index;
                    cell.in_memory_partition_index = current_partition_index;

                    index[ssd_cache_key.key] = cell;
                    break;
                }
                else
                {
                    /// Partition is full need to try next partition

                    if (memory_buffer_partitions.size() < configuration.max_partitions_count)
                    {
                        /// Try tro create next partition without reusing old partitions
                        ++current_partition_index;
                        file_buffer.allocateSizeForNextPartition();
                        memory_buffer_partitions.emplace_back(configuration.block_size, configuration.write_buffer_blocks_size);
                    }
                    else
                    {
                        /// Start reusing old partitions
                        current_partition_index = (current_partition_index + 1) % memory_buffer_partitions.size();
                        file_buffer.reset();
                    }
                }
            }
        }
    }

    inline void setCellDeadline(Cell & cell, TimePoint now)
    {
        if (configuration.lifetime.min_sec == 0 && configuration.lifetime.max_sec == 0)
        {
            auto deadline = std::chrono::time_point<std::chrono::system_clock>::max() - 2 * std::chrono::seconds(configuration.strict_max_lifetime_seconds);
            cell.deadline = std::chrono::system_clock::to_time_t(deadline);
            return;
        }

        size_t min_sec_lifetime = configuration.lifetime.min_sec;
        size_t max_sec_lifetime = configuration.lifetime.max_sec;

        std::uniform_int_distribution<UInt64> distribution{min_sec_lifetime, max_sec_lifetime};
        auto deadline = now + std::chrono::seconds(distribution(rnd_engine));
        cell.deadline = std::chrono::system_clock::to_time_t(deadline);
    }

    inline void eraseKeyFromIndex(KeyType key)
    {
        auto it = index.find(key);

        if (it == nullptr)
            return;

        /// In case of complex key in arena key is serialized from hash table
        KeyType key_copy = it->getKey();

        index.erase(key);

        if constexpr (std::is_same_v<KeyType, StringRef>)
            complex_key_arena.free(const_cast<char *>(key_copy.data), key_copy.size);
    }

    SSDCacheDictionaryStorageConfiguration configuration;

    SSDCacheFileBuffer<SSDCacheKeyType> file_buffer;

    std::vector<SSDCacheMemoryBuffer<SSDCacheKeyType>> memory_buffer_partitions;

    pcg64 rnd_engine;

    using SimpleKeyHashMap = HashMap<UInt64, Cell>;
    using ComplexKeyHashMap = HashMapWithSavedHash<StringRef, Cell>;

    using CacheMap = std::conditional_t<
        dictionary_key_type == DictionaryKeyType::simple,
        SimpleKeyHashMap,
        ComplexKeyHashMap>;

    ArenaWithFreeLists complex_key_arena;

    CacheMap index;

    size_t current_partition_index = 0;

};

}

#endif