CacheDictionary.cpp 36.4 KB
Newer Older
1
#include <functional>
2 3
#include <sstream>
#include <memory>
4
#include <Columns/ColumnsNumber.h>
5 6
#include <Columns/ColumnVector.h>
#include <Columns/ColumnString.h>
7 8 9 10 11
#include <Common/BitHelpers.h>
#include <Common/randomSeed.h>
#include <Common/HashTable/Hash.h>
#include <Common/Stopwatch.h>
#include <Common/ProfilingScopedRWLock.h>
12 13
#include <Common/ProfileEvents.h>
#include <Common/CurrentMetrics.h>
14
#include <Common/typeid_cast.h>
15 16 17
#include <DataTypes/DataTypesNumber.h>
#include <Dictionaries/CacheDictionary.h>
#include <Dictionaries/DictionaryBlockInputStream.h>
18 19 20
#include <ext/size.h>
#include <ext/range.h>
#include <ext/map.h>
21 22


23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42
namespace ProfileEvents
{
    extern const Event DictCacheKeysRequested;
    extern const Event DictCacheKeysRequestedMiss;
    extern const Event DictCacheKeysRequestedFound;
    extern const Event DictCacheKeysExpired;
    extern const Event DictCacheKeysNotFound;
    extern const Event DictCacheKeysHit;
    extern const Event DictCacheRequestTimeNs;
    extern const Event DictCacheRequests;
    extern const Event DictCacheLockWriteNs;
    extern const Event DictCacheLockReadNs;
}

namespace CurrentMetrics
{
    extern const Metric DictCacheRequests;
}


43 44 45 46 47
namespace DB
{

namespace ErrorCodes
{
48 49 50
    extern const int TYPE_MISMATCH;
    extern const int BAD_ARGUMENTS;
    extern const int UNSUPPORTED_METHOD;
51 52 53
}


V
Vladimir Smirnov 已提交
54
inline UInt64 CacheDictionary::getCellIdx(const Key id) const
55
{
56 57 58
    const auto hash = intHash64(id);
    const auto idx = hash & size_overlap_mask;
    return idx;
59 60 61 62
}


CacheDictionary::CacheDictionary(const std::string & name, const DictionaryStructure & dict_struct,
63 64 65 66 67 68 69 70
    DictionarySourcePtr source_ptr, const DictionaryLifetime dict_lifetime,
    const std::size_t size)
    : name{name}, dict_struct(dict_struct),
        source_ptr{std::move(source_ptr)}, dict_lifetime(dict_lifetime),
        size{roundUpToPowerOfTwoOrZero(std::max(size, size_t(max_collision_length)))},
        size_overlap_mask{this->size - 1},
        cells{this->size},
        rnd_engine{randomSeed()}
71
{
72 73 74 75
    if (!this->source_ptr->supportsSelectiveLoad())
        throw Exception{
            name + ": source cannot be used with CacheDictionary",
            ErrorCodes::UNSUPPORTED_METHOD};
76

77
    createAttributes();
78 79 80
}

CacheDictionary::CacheDictionary(const CacheDictionary & other)
81
    : CacheDictionary{other.name, other.dict_struct, other.source_ptr->clone(), other.dict_lifetime, other.size}
82 83 84
{}


A
Alexey Milovidov 已提交
85
void CacheDictionary::toParent(const PaddedPODArray<Key> & ids, PaddedPODArray<Key> & out) const
86
{
87
    const auto null_value = std::get<UInt64>(hierarchical_attribute->null_values);
88

89
    getItemsNumber<UInt64>(*hierarchical_attribute, ids, out, [&] (const std::size_t) { return null_value; });
90 91 92
}


93 94 95 96 97 98 99
/// Allow to use single value in same way as array.
static inline CacheDictionary::Key getAt(const PaddedPODArray<CacheDictionary::Key> & arr, const size_t idx) { return arr[idx]; }
static inline CacheDictionary::Key getAt(const CacheDictionary::Key & value, const size_t idx) { return value; }


template <typename AncestorType>
void CacheDictionary::isInImpl(
100 101 102
    const PaddedPODArray<Key> & child_ids,
    const AncestorType & ancestor_ids,
    PaddedPODArray<UInt8> & out) const
103
{
104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158
    /// Transform all children to parents until ancestor id or null_value will be reached.

    size_t size = out.size();
    memset(out.data(), 0xFF, size);        /// 0xFF means "not calculated"

    const auto null_value = std::get<UInt64>(hierarchical_attribute->null_values);

    PaddedPODArray<Key> children(size);
    PaddedPODArray<Key> parents(child_ids.begin(), child_ids.end());

    while (true)
    {
        size_t out_idx = 0;
        size_t parents_idx = 0;
        size_t new_children_idx = 0;

        while (out_idx < size)
        {
            /// Already calculated
            if (out[out_idx] != 0xFF)
            {
                ++out_idx;
                continue;
            }

            /// No parent
            if (parents[parents_idx] == null_value)
            {
                out[out_idx] = 0;
            }
            /// Found ancestor
            else if (parents[parents_idx] == getAt(ancestor_ids, parents_idx))
            {
                out[out_idx] = 1;
            }
            /// Found intermediate parent, add this value to search at next loop iteration
            else
            {
                children[new_children_idx] = parents[parents_idx];
                ++new_children_idx;
            }

            ++out_idx;
            ++parents_idx;
        }

        if (new_children_idx == 0)
            break;

        /// Transform all children to its parents.
        children.resize(new_children_idx);
        parents.resize(new_children_idx);

        toParent(children, parents);
    }
159 160 161
}

void CacheDictionary::isInVectorVector(
162 163 164
    const PaddedPODArray<Key> & child_ids,
    const PaddedPODArray<Key> & ancestor_ids,
    PaddedPODArray<UInt8> & out) const
165
{
166
    isInImpl(child_ids, ancestor_ids, out);
167
}
168

169
void CacheDictionary::isInVectorConstant(
170 171 172
    const PaddedPODArray<Key> & child_ids,
    const Key ancestor_id,
    PaddedPODArray<UInt8> & out) const
173
{
174
    isInImpl(child_ids, ancestor_id, out);
175
}
176

177
void CacheDictionary::isInConstantVector(
178 179 180
    const Key child_id,
    const PaddedPODArray<Key> & ancestor_ids,
    PaddedPODArray<UInt8> & out) const
181
{
182
    /// Special case with single child value.
183

184
    const auto null_value = std::get<UInt64>(hierarchical_attribute->null_values);
185

186 187 188
    PaddedPODArray<Key> child(1, child_id);
    PaddedPODArray<Key> parent(1);
    std::vector<Key> ancestors(1, child_id);
189

190 191 192 193
    /// Iteratively find all ancestors for child.
    while (true)
    {
        toParent(child, parent);
194

195 196
        if (parent[0] == null_value)
            break;
197

198 199 200
        child[0] = parent[0];
        ancestors.push_back(parent[0]);
    }
201

202 203 204
    /// Assuming short hierarchy, so linear search is Ok.
    for (size_t i = 0, size = out.size(); i < size; ++i)
        out[i] = std::find(ancestors.begin(), ancestors.end(), ancestor_ids[i]) != ancestors.end();
205
}
206 207


208
#define DECLARE(TYPE)\
A
Alexey Milovidov 已提交
209
void CacheDictionary::get##TYPE(const std::string & attribute_name, const PaddedPODArray<Key> & ids, PaddedPODArray<TYPE> & out) const\
210
{\
211 212 213 214 215 216 217 218 219
    auto & attribute = getAttribute(attribute_name);\
    if (!isAttributeTypeConvertibleTo(attribute.type, AttributeUnderlyingType::TYPE))\
        throw Exception{\
            name + ": type mismatch: attribute " + attribute_name + " has type " + toString(attribute.type),\
            ErrorCodes::TYPE_MISMATCH};\
    \
    const auto null_value = std::get<TYPE>(attribute.null_values);\
    \
    getItemsNumber<TYPE>(attribute, ids, out, [&] (const std::size_t) { return null_value; });\
220 221 222 223 224 225 226 227 228 229 230 231 232
}
DECLARE(UInt8)
DECLARE(UInt16)
DECLARE(UInt32)
DECLARE(UInt64)
DECLARE(Int8)
DECLARE(Int16)
DECLARE(Int32)
DECLARE(Int64)
DECLARE(Float32)
DECLARE(Float64)
#undef DECLARE

A
Alexey Milovidov 已提交
233
void CacheDictionary::getString(const std::string & attribute_name, const PaddedPODArray<Key> & ids, ColumnString * out) const
234
{
235 236 237 238 239
    auto & attribute = getAttribute(attribute_name);
    if (!isAttributeTypeConvertibleTo(attribute.type, AttributeUnderlyingType::String))
        throw Exception{
            name + ": type mismatch: attribute " + attribute_name + " has type " + toString(attribute.type),
            ErrorCodes::TYPE_MISMATCH};
240

241
    const auto null_value = StringRef{std::get<String>(attribute.null_values)};
242

243
    getItemsString(attribute, ids, out, [&] (const std::size_t) { return null_value; });
244 245 246 247
}

#define DECLARE(TYPE)\
void CacheDictionary::get##TYPE(\
248 249
    const std::string & attribute_name, const PaddedPODArray<Key> & ids, const PaddedPODArray<TYPE> & def,\
    PaddedPODArray<TYPE> & out) const\
250
{\
251 252 253 254 255 256 257
    auto & attribute = getAttribute(attribute_name);\
    if (!isAttributeTypeConvertibleTo(attribute.type, AttributeUnderlyingType::TYPE))\
        throw Exception{\
            name + ": type mismatch: attribute " + attribute_name + " has type " + toString(attribute.type),\
            ErrorCodes::TYPE_MISMATCH};\
    \
    getItemsNumber<TYPE>(attribute, ids, out, [&] (const std::size_t row) { return def[row]; });\
258 259 260 261 262 263 264 265 266 267 268 269 270 271
}
DECLARE(UInt8)
DECLARE(UInt16)
DECLARE(UInt32)
DECLARE(UInt64)
DECLARE(Int8)
DECLARE(Int16)
DECLARE(Int32)
DECLARE(Int64)
DECLARE(Float32)
DECLARE(Float64)
#undef DECLARE

void CacheDictionary::getString(
272 273
    const std::string & attribute_name, const PaddedPODArray<Key> & ids, const ColumnString * const def,
    ColumnString * const out) const
274
{
275 276 277 278 279
    auto & attribute = getAttribute(attribute_name);
    if (!isAttributeTypeConvertibleTo(attribute.type, AttributeUnderlyingType::String))
        throw Exception{
            name + ": type mismatch: attribute " + attribute_name + " has type " + toString(attribute.type),
            ErrorCodes::TYPE_MISMATCH};
280

281
    getItemsString(attribute, ids, out, [&] (const std::size_t row) { return def->getDataAt(row); });
282 283 284 285
}

#define DECLARE(TYPE)\
void CacheDictionary::get##TYPE(\
286
    const std::string & attribute_name, const PaddedPODArray<Key> & ids, const TYPE def, PaddedPODArray<TYPE> & out) const\
287
{\
288 289 290 291 292 293 294
    auto & attribute = getAttribute(attribute_name);\
    if (!isAttributeTypeConvertibleTo(attribute.type, AttributeUnderlyingType::TYPE))\
        throw Exception{\
            name + ": type mismatch: attribute " + attribute_name + " has type " + toString(attribute.type),\
            ErrorCodes::TYPE_MISMATCH};\
    \
    getItemsNumber<TYPE>(attribute, ids, out, [&] (const std::size_t) { return def; });\
295 296 297 298 299 300 301 302 303 304 305 306 307 308
}
DECLARE(UInt8)
DECLARE(UInt16)
DECLARE(UInt32)
DECLARE(UInt64)
DECLARE(Int8)
DECLARE(Int16)
DECLARE(Int32)
DECLARE(Int64)
DECLARE(Float32)
DECLARE(Float64)
#undef DECLARE

void CacheDictionary::getString(
309 310
    const std::string & attribute_name, const PaddedPODArray<Key> & ids, const String & def,
    ColumnString * const out) const
311
{
312 313 314 315 316
    auto & attribute = getAttribute(attribute_name);
    if (!isAttributeTypeConvertibleTo(attribute.type, AttributeUnderlyingType::String))
        throw Exception{
            name + ": type mismatch: attribute " + attribute_name + " has type " + toString(attribute.type),
            ErrorCodes::TYPE_MISMATCH};
317

318
    getItemsString(attribute, ids, out, [&] (const std::size_t) { return StringRef{def}; });
319 320 321
}


322
/// returns cell_idx (always valid for replacing), 'cell is valid' flag, 'cell is outdated' flag
323 324 325 326 327 328 329 330
/// true  false   found and valid
/// false true    not found (something outdated, maybe our cell)
/// false false   not found (other id stored with valid data)
/// true  true    impossible
///
/// todo: split this func to two: find_for_get and find_for_set
CacheDictionary::FindResult CacheDictionary::findCellIdx(const Key & id, const CellMetadata::time_point_t now) const
{
331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359
    auto pos = getCellIdx(id);
    auto oldest_id = pos;
    auto oldest_time = CellMetadata::time_point_t::max();
    const auto stop = pos + max_collision_length;
    for (; pos < stop; ++pos)
    {
        const auto cell_idx = pos & size_overlap_mask;
        const auto & cell = cells[cell_idx];

        if (cell.id != id)
        {
            /// maybe we already found nearest expired cell (try minimize collision_length on insert)
            if (oldest_time > now && oldest_time > cell.expiresAt())
            {
                oldest_time = cell.expiresAt();
                oldest_id = cell_idx;
            }
            continue;
        }

        if (cell.expiresAt() < now)
        {
            return {cell_idx, false, true};
        }

        return {cell_idx, true, false};
    }

    return {oldest_id, false, false};
360 361
}

A
Alexey Milovidov 已提交
362
void CacheDictionary::has(const PaddedPODArray<Key> & ids, PaddedPODArray<UInt8> & out) const
363
{
364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411
    /// Mapping: <id> -> { all indices `i` of `ids` such that `ids[i]` = <id> }
    std::unordered_map<Key, std::vector<std::size_t>> outdated_ids;

    size_t cache_expired = 0, cache_not_found = 0, cache_hit = 0;

    const auto rows = ext::size(ids);
    {
        const ProfilingScopedReadRWLock read_lock{rw_lock, ProfileEvents::DictCacheLockReadNs};

        const auto now = std::chrono::system_clock::now();
        /// fetch up-to-date values, decide which ones require update
        for (const auto row : ext::range(0, rows))
        {
            const auto id = ids[row];
            const auto find_result = findCellIdx(id, now);
            const auto & cell_idx = find_result.cell_idx;
            if (!find_result.valid)
            {
                outdated_ids[id].push_back(row);
                if (find_result.outdated)
                    ++cache_expired;
                else
                    ++cache_not_found;
            }
            else
            {
                ++cache_hit;
                const auto & cell = cells[cell_idx];
                out[row] = !cell.isDefault();
            }
        }
    }

    ProfileEvents::increment(ProfileEvents::DictCacheKeysExpired, cache_expired);
    ProfileEvents::increment(ProfileEvents::DictCacheKeysNotFound, cache_not_found);
    ProfileEvents::increment(ProfileEvents::DictCacheKeysHit, cache_hit);

    query_count.fetch_add(rows, std::memory_order_relaxed);
    hit_count.fetch_add(rows - outdated_ids.size(), std::memory_order_release);

    if (outdated_ids.empty())
        return;

    std::vector<Key> required_ids(outdated_ids.size());
    std::transform(std::begin(outdated_ids), std::end(outdated_ids), std::begin(required_ids),
        [] (auto & pair) { return pair.first; });

    /// request new values
A
Alexey Milovidov 已提交
412 413 414
    update(required_ids,
    [&] (const auto id, const auto)
    {
415 416
        for (const auto row : outdated_ids[id])
            out[row] = true;
A
Alexey Milovidov 已提交
417 418 419
    },
    [&] (const auto id, const auto)
    {
420 421 422
        for (const auto row : outdated_ids[id])
            out[row] = false;
    });
423 424 425 426 427
}


void CacheDictionary::createAttributes()
{
428 429
    const auto attributes_size = dict_struct.attributes.size();
    attributes.reserve(attributes_size);
430 431

    bytes_allocated += size * sizeof(CellMetadata);
432
    bytes_allocated += attributes_size * sizeof(attributes.front());
433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448

    for (const auto & attribute : dict_struct.attributes)
    {
        attribute_index_by_name.emplace(attribute.name, attributes.size());
        attributes.push_back(createAttributeWithType(attribute.underlying_type, attribute.null_value));

        if (attribute.hierarchical)
        {
            hierarchical_attribute = &attributes.back();

            if (hierarchical_attribute->type != AttributeUnderlyingType::UInt64)
                throw Exception{
                    name + ": hierarchical attribute must be UInt64.",
                    ErrorCodes::TYPE_MISMATCH};
        }
    }
449 450
}

A
Alexey Milovidov 已提交
451
CacheDictionary::Attribute CacheDictionary::createAttributeWithType(const AttributeUnderlyingType type, const Field & null_value)
452
{
453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516
    Attribute attr{type};

    switch (type)
    {
        case AttributeUnderlyingType::UInt8:
            std::get<UInt8>(attr.null_values) = null_value.get<UInt64>();
            std::get<ContainerPtrType<UInt8>>(attr.arrays) = std::make_unique<ContainerType<UInt8>>(size);
            bytes_allocated += size * sizeof(UInt8);
            break;
        case AttributeUnderlyingType::UInt16:
            std::get<UInt16>(attr.null_values) = null_value.get<UInt64>();
            std::get<ContainerPtrType<UInt16>>(attr.arrays) = std::make_unique<ContainerType<UInt16>>(size);
            bytes_allocated += size * sizeof(UInt16);
            break;
        case AttributeUnderlyingType::UInt32:
            std::get<UInt32>(attr.null_values) = null_value.get<UInt64>();
            std::get<ContainerPtrType<UInt32>>(attr.arrays) = std::make_unique<ContainerType<UInt32>>(size);
            bytes_allocated += size * sizeof(UInt32);
            break;
        case AttributeUnderlyingType::UInt64:
            std::get<UInt64>(attr.null_values) = null_value.get<UInt64>();
            std::get<ContainerPtrType<UInt64>>(attr.arrays) = std::make_unique<ContainerType<UInt64>>(size);
            bytes_allocated += size * sizeof(UInt64);
            break;
        case AttributeUnderlyingType::Int8:
            std::get<Int8>(attr.null_values) = null_value.get<Int64>();
            std::get<ContainerPtrType<Int8>>(attr.arrays) = std::make_unique<ContainerType<Int8>>(size);
            bytes_allocated += size * sizeof(Int8);
            break;
        case AttributeUnderlyingType::Int16:
            std::get<Int16>(attr.null_values) = null_value.get<Int64>();
            std::get<ContainerPtrType<Int16>>(attr.arrays) = std::make_unique<ContainerType<Int16>>(size);
            bytes_allocated += size * sizeof(Int16);
            break;
        case AttributeUnderlyingType::Int32:
            std::get<Int32>(attr.null_values) = null_value.get<Int64>();
            std::get<ContainerPtrType<Int32>>(attr.arrays) = std::make_unique<ContainerType<Int32>>(size);
            bytes_allocated += size * sizeof(Int32);
            break;
        case AttributeUnderlyingType::Int64:
            std::get<Int64>(attr.null_values) = null_value.get<Int64>();
            std::get<ContainerPtrType<Int64>>(attr.arrays) = std::make_unique<ContainerType<Int64>>(size);
            bytes_allocated += size * sizeof(Int64);
            break;
        case AttributeUnderlyingType::Float32:
            std::get<Float32>(attr.null_values) = null_value.get<Float64>();
            std::get<ContainerPtrType<Float32>>(attr.arrays) = std::make_unique<ContainerType<Float32>>(size);
            bytes_allocated += size * sizeof(Float32);
            break;
        case AttributeUnderlyingType::Float64:
            std::get<Float64>(attr.null_values) = null_value.get<Float64>();
            std::get<ContainerPtrType<Float64>>(attr.arrays) = std::make_unique<ContainerType<Float64>>(size);
            bytes_allocated += size * sizeof(Float64);
            break;
        case AttributeUnderlyingType::String:
            std::get<String>(attr.null_values) = null_value.get<String>();
            std::get<ContainerPtrType<StringRef>>(attr.arrays) = std::make_unique<ContainerType<StringRef>>(size);
            bytes_allocated += size * sizeof(StringRef);
            if (!string_arena)
                string_arena = std::make_unique<ArenaWithFreeLists>();
            break;
    }

    return attr;
517 518 519 520 521
}


template <typename OutputType, typename DefaultGetter>
void CacheDictionary::getItemsNumber(
522 523 524 525
    Attribute & attribute,
    const PaddedPODArray<Key> & ids,
    PaddedPODArray<OutputType> & out,
    DefaultGetter && get_default) const
526
{
527
    if (false) {}
528
#define DISPATCH(TYPE) \
529 530 531 532 533 534 535 536 537 538 539 540
    else if (attribute.type == AttributeUnderlyingType::TYPE) \
        getItemsNumberImpl<TYPE, OutputType>(attribute, ids, out, std::forward<DefaultGetter>(get_default));
    DISPATCH(UInt8)
    DISPATCH(UInt16)
    DISPATCH(UInt32)
    DISPATCH(UInt64)
    DISPATCH(Int8)
    DISPATCH(Int16)
    DISPATCH(Int32)
    DISPATCH(Int64)
    DISPATCH(Float32)
    DISPATCH(Float64)
541
#undef DISPATCH
542 543
    else
        throw Exception("Unexpected type of attribute: " + toString(attribute.type), ErrorCodes::LOGICAL_ERROR);
544 545 546 547
}

template <typename AttributeType, typename OutputType, typename DefaultGetter>
void CacheDictionary::getItemsNumberImpl(
548 549 550 551
    Attribute & attribute,
    const PaddedPODArray<Key> & ids,
    PaddedPODArray<OutputType> & out,
    DefaultGetter && get_default) const
552
{
553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607
    /// Mapping: <id> -> { all indices `i` of `ids` such that `ids[i]` = <id> }
    std::unordered_map<Key, std::vector<std::size_t>> outdated_ids;
    auto & attribute_array = std::get<ContainerPtrType<AttributeType>>(attribute.arrays);
    const auto rows = ext::size(ids);

    size_t cache_expired = 0, cache_not_found = 0, cache_hit = 0;

    {
        const ProfilingScopedReadRWLock read_lock{rw_lock, ProfileEvents::DictCacheLockReadNs};

        const auto now = std::chrono::system_clock::now();
        /// fetch up-to-date values, decide which ones require update
        for (const auto row : ext::range(0, rows))
        {
            const auto id = ids[row];

            /** cell should be updated if either:
                *    1. ids do not match,
                *    2. cell has expired,
                *    3. explicit defaults were specified and cell was set default. */

            const auto find_result = findCellIdx(id, now);
            if (!find_result.valid)
            {
                outdated_ids[id].push_back(row);
                if (find_result.outdated)
                    ++cache_expired;
                else
                    ++cache_not_found;
            }
            else
            {
                ++cache_hit;
                const auto & cell_idx = find_result.cell_idx;
                const auto & cell = cells[cell_idx];
                out[row] = cell.isDefault() ? get_default(row) : attribute_array[cell_idx];
            }
        }
    }

    ProfileEvents::increment(ProfileEvents::DictCacheKeysExpired, cache_expired);
    ProfileEvents::increment(ProfileEvents::DictCacheKeysNotFound, cache_not_found);
    ProfileEvents::increment(ProfileEvents::DictCacheKeysHit, cache_hit);

    query_count.fetch_add(rows, std::memory_order_relaxed);
    hit_count.fetch_add(rows - outdated_ids.size(), std::memory_order_release);

    if (outdated_ids.empty())
        return;

    std::vector<Key> required_ids(outdated_ids.size());
    std::transform(std::begin(outdated_ids), std::end(outdated_ids), std::begin(required_ids),
        [] (auto & pair) { return pair.first; });

    /// request new values
A
Alexey Milovidov 已提交
608 609 610 611
    update(required_ids,
    [&] (const auto id, const auto cell_idx)
    {
        const auto attribute_value = attribute_array[cell_idx];
612

A
Alexey Milovidov 已提交
613 614 615 616 617 618 619 620
        for (const auto row : outdated_ids[id])
            out[row] = attribute_value;
    },
    [&] (const auto id, const auto cell_idx)
    {
        for (const auto row : outdated_ids[id])
            out[row] = get_default(row);
    });
621 622 623 624
}

template <typename DefaultGetter>
void CacheDictionary::getItemsString(
625 626 627 628
    Attribute & attribute,
    const PaddedPODArray<Key> & ids,
    ColumnString * out,
    DefaultGetter && get_default) const
629
{
630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729
    const auto rows = ext::size(ids);

    /// save on some allocations
    out->getOffsets().reserve(rows);

    auto & attribute_array = std::get<ContainerPtrType<StringRef>>(attribute.arrays);

    auto found_outdated_values = false;

    /// perform optimistic version, fallback to pessimistic if failed
    {
        const ProfilingScopedReadRWLock read_lock{rw_lock, ProfileEvents::DictCacheLockReadNs};

        const auto now = std::chrono::system_clock::now();
        /// fetch up-to-date values, discard on fail
        for (const auto row : ext::range(0, rows))
        {
            const auto id = ids[row];

            const auto find_result = findCellIdx(id, now);
            if (!find_result.valid)
            {
                found_outdated_values = true;
                break;
            }
            else
            {
                const auto & cell_idx = find_result.cell_idx;
                const auto & cell = cells[cell_idx];
                const auto string_ref = cell.isDefault() ? get_default(row) : attribute_array[cell_idx];
                out->insertData(string_ref.data, string_ref.size);
            }
        }
    }

    /// optimistic code completed successfully
    if (!found_outdated_values)
    {
        query_count.fetch_add(rows, std::memory_order_relaxed);
        hit_count.fetch_add(rows, std::memory_order_release);
        return;
    }

    /// now onto the pessimistic one, discard possible partial results from the optimistic path
    out->getChars().resize_assume_reserved(0);
    out->getOffsets().resize_assume_reserved(0);

    /// Mapping: <id> -> { all indices `i` of `ids` such that `ids[i]` = <id> }
    std::unordered_map<Key, std::vector<std::size_t>> outdated_ids;
    /// we are going to store every string separately
    std::unordered_map<Key, String> map;

    std::size_t total_length = 0;
    size_t cache_expired = 0, cache_not_found = 0, cache_hit = 0;
    {
        const ProfilingScopedReadRWLock read_lock{rw_lock, ProfileEvents::DictCacheLockReadNs};

        const auto now = std::chrono::system_clock::now();
        for (const auto row : ext::range(0, ids.size()))
        {
            const auto id = ids[row];

            const auto find_result = findCellIdx(id, now);
            if (!find_result.valid)
            {
                outdated_ids[id].push_back(row);
                if (find_result.outdated)
                    ++cache_expired;
                else
                    ++cache_not_found;
            }
            else
            {
                ++cache_hit;
                const auto & cell_idx = find_result.cell_idx;
                const auto & cell = cells[cell_idx];
                const auto string_ref = cell.isDefault() ? get_default(row) : attribute_array[cell_idx];

                if (!cell.isDefault())
                    map[id] = String{string_ref};

                total_length += string_ref.size + 1;
            }
        }
    }

    ProfileEvents::increment(ProfileEvents::DictCacheKeysExpired, cache_expired);
    ProfileEvents::increment(ProfileEvents::DictCacheKeysNotFound, cache_not_found);
    ProfileEvents::increment(ProfileEvents::DictCacheKeysHit, cache_hit);

    query_count.fetch_add(rows, std::memory_order_relaxed);
    hit_count.fetch_add(rows - outdated_ids.size(), std::memory_order_release);

    /// request new values
    if (!outdated_ids.empty())
    {
        std::vector<Key> required_ids(outdated_ids.size());
        std::transform(std::begin(outdated_ids), std::end(outdated_ids), std::begin(required_ids),
            [] (auto & pair) { return pair.first; });

A
Alexey Milovidov 已提交
730 731 732
        update(required_ids,
        [&] (const auto id, const auto cell_idx)
        {
733 734 735 736
            const auto attribute_value = attribute_array[cell_idx];

            map[id] = String{attribute_value};
            total_length += (attribute_value.size + 1) * outdated_ids[id].size();
A
Alexey Milovidov 已提交
737 738 739
        },
        [&] (const auto id, const auto cell_idx)
        {
740 741 742 743 744 745 746 747 748 749 750 751 752 753 754
            for (const auto row : outdated_ids[id])
                total_length += get_default(row).size + 1;
        });
    }

    out->getChars().reserve(total_length);

    for (const auto row : ext::range(0, ext::size(ids)))
    {
        const auto id = ids[row];
        const auto it = map.find(id);

        const auto string_ref = it != std::end(map) ? StringRef{it->second} : get_default(row);
        out->insertData(string_ref.data, string_ref.size);
    }
755 756 757 758
}

template <typename PresentIdHandler, typename AbsentIdHandler>
void CacheDictionary::update(
759 760
    const std::vector<Key> & requested_ids,
    PresentIdHandler && on_cell_updated,
761
    AbsentIdHandler && on_id_not_found) const
762
{
763 764 765 766 767 768 769 770 771 772
    std::unordered_map<Key, UInt8> remaining_ids{requested_ids.size()};
    for (const auto id : requested_ids)
        remaining_ids.insert({ id, 0 });

    std::uniform_int_distribution<UInt64> distribution{
        dict_lifetime.min_sec,
        dict_lifetime.max_sec
    };

    const ProfilingScopedWriteRWLock write_lock{rw_lock, ProfileEvents::DictCacheLockWriteNs};
773

774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792
    {
        CurrentMetrics::Increment metric_increment{CurrentMetrics::DictCacheRequests};
        Stopwatch watch;
        auto stream = source_ptr->loadIds(requested_ids);
        stream->readPrefix();

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

        while (const auto block = stream->read())
        {
            const auto id_column = typeid_cast<const ColumnUInt64 *>(block.safeGetByPosition(0).column.get());
            if (!id_column)
                throw Exception{
                    name + ": id column has type different from UInt64.",
                    ErrorCodes::TYPE_MISMATCH};

            const auto & ids = id_column->getData();

            /// cache column pointers
793 794
            const auto column_ptrs = ext::map<std::vector>(ext::range(0, attributes.size()), [&block] (size_t i)
            {
795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870
                return block.safeGetByPosition(i + 1).column.get();
            });

            for (const auto i : ext::range(0, ids.size()))
            {
                const auto id = ids[i];

                const auto find_result = findCellIdx(id, now);
                const auto & cell_idx = find_result.cell_idx;

                auto & cell = cells[cell_idx];

                for (const auto attribute_idx : ext::range(0, attributes.size()))
                {
                    const auto & attribute_column = *column_ptrs[attribute_idx];
                    auto & attribute = attributes[attribute_idx];

                    setAttributeValue(attribute, cell_idx, attribute_column[i]);
                }

                /// if cell id is zero and zero does not map to this cell, then the cell is unused
                if (cell.id == 0 && cell_idx != zero_cell_idx)
                    element_count.fetch_add(1, std::memory_order_relaxed);

                cell.id = id;
                if (dict_lifetime.min_sec != 0 && dict_lifetime.max_sec != 0)
                    cell.setExpiresAt(std::chrono::system_clock::now() + std::chrono::seconds{distribution(rnd_engine)});
                else
                    cell.setExpiresAt(std::chrono::time_point<std::chrono::system_clock>::max());

                /// inform caller
                on_cell_updated(id, cell_idx);
                /// mark corresponding id as found
                remaining_ids[id] = 1;
            }
        }

        stream->readSuffix();

        ProfileEvents::increment(ProfileEvents::DictCacheKeysRequested, requested_ids.size());
        ProfileEvents::increment(ProfileEvents::DictCacheRequestTimeNs, watch.elapsed());
    }

    size_t not_found_num = 0, found_num = 0;

    const auto now = std::chrono::system_clock::now();
    /// Check which ids have not been found and require setting null_value
    for (const auto id_found_pair : remaining_ids)
    {
        if (id_found_pair.second)
        {
            ++found_num;
            continue;
        }
        ++not_found_num;

        const auto id = id_found_pair.first;

        const auto find_result = findCellIdx(id, now);
        const auto & cell_idx = find_result.cell_idx;

        auto & cell = cells[cell_idx];

        /// Set null_value for each attribute
        for (auto & attribute : attributes)
            setDefaultAttributeValue(attribute, cell_idx);

        /// Check if cell had not been occupied before and increment element counter if it hadn't
        if (cell.id == 0 && cell_idx != zero_cell_idx)
            element_count.fetch_add(1, std::memory_order_relaxed);

        cell.id = id;
        if (dict_lifetime.min_sec != 0 && dict_lifetime.max_sec != 0)
            cell.setExpiresAt(std::chrono::system_clock::now() + std::chrono::seconds{distribution(rnd_engine)});
        else
            cell.setExpiresAt(std::chrono::time_point<std::chrono::system_clock>::max());
871

872 873 874 875 876 877 878 879 880
        cell.setDefault();

        /// inform caller that the cell has not been found
        on_id_not_found(id, cell_idx);
    }

    ProfileEvents::increment(ProfileEvents::DictCacheKeysRequestedMiss, not_found_num);
    ProfileEvents::increment(ProfileEvents::DictCacheKeysRequestedFound, found_num);
    ProfileEvents::increment(ProfileEvents::DictCacheRequests);
881 882 883
}


A
Alexey Milovidov 已提交
884
void CacheDictionary::setDefaultAttributeValue(Attribute & attribute, const Key idx) const
885
{
886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913
    switch (attribute.type)
    {
        case AttributeUnderlyingType::UInt8: std::get<ContainerPtrType<UInt8>>(attribute.arrays)[idx] = std::get<UInt8>(attribute.null_values); break;
        case AttributeUnderlyingType::UInt16: std::get<ContainerPtrType<UInt16>>(attribute.arrays)[idx] = std::get<UInt16>(attribute.null_values); break;
        case AttributeUnderlyingType::UInt32: std::get<ContainerPtrType<UInt32>>(attribute.arrays)[idx] = std::get<UInt32>(attribute.null_values); break;
        case AttributeUnderlyingType::UInt64: std::get<ContainerPtrType<UInt64>>(attribute.arrays)[idx] = std::get<UInt64>(attribute.null_values); break;
        case AttributeUnderlyingType::Int8: std::get<ContainerPtrType<Int8>>(attribute.arrays)[idx] = std::get<Int8>(attribute.null_values); break;
        case AttributeUnderlyingType::Int16: std::get<ContainerPtrType<Int16>>(attribute.arrays)[idx] = std::get<Int16>(attribute.null_values); break;
        case AttributeUnderlyingType::Int32: std::get<ContainerPtrType<Int32>>(attribute.arrays)[idx] = std::get<Int32>(attribute.null_values); break;
        case AttributeUnderlyingType::Int64: std::get<ContainerPtrType<Int64>>(attribute.arrays)[idx] = std::get<Int64>(attribute.null_values); break;
        case AttributeUnderlyingType::Float32: std::get<ContainerPtrType<Float32>>(attribute.arrays)[idx] = std::get<Float32>(attribute.null_values); break;
        case AttributeUnderlyingType::Float64: std::get<ContainerPtrType<Float64>>(attribute.arrays)[idx] = std::get<Float64>(attribute.null_values); break;
        case AttributeUnderlyingType::String:
        {
            const auto & null_value_ref = std::get<String>(attribute.null_values);
            auto & string_ref = std::get<ContainerPtrType<StringRef>>(attribute.arrays)[idx];

            if (string_ref.data != null_value_ref.data())
            {
                if (string_ref.data)
                    string_arena->free(const_cast<char *>(string_ref.data), string_ref.size);

                string_ref = StringRef{null_value_ref};
            }

            break;
        }
    }
914 915
}

A
Alexey Milovidov 已提交
916
void CacheDictionary::setAttributeValue(Attribute & attribute, const Key idx, const Field & value) const
917
{
918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952
    switch (attribute.type)
    {
        case AttributeUnderlyingType::UInt8: std::get<ContainerPtrType<UInt8>>(attribute.arrays)[idx] = value.get<UInt64>(); break;
        case AttributeUnderlyingType::UInt16: std::get<ContainerPtrType<UInt16>>(attribute.arrays)[idx] = value.get<UInt64>(); break;
        case AttributeUnderlyingType::UInt32: std::get<ContainerPtrType<UInt32>>(attribute.arrays)[idx] = value.get<UInt64>(); break;
        case AttributeUnderlyingType::UInt64: std::get<ContainerPtrType<UInt64>>(attribute.arrays)[idx] = value.get<UInt64>(); break;
        case AttributeUnderlyingType::Int8: std::get<ContainerPtrType<Int8>>(attribute.arrays)[idx] = value.get<Int64>(); break;
        case AttributeUnderlyingType::Int16: std::get<ContainerPtrType<Int16>>(attribute.arrays)[idx] = value.get<Int64>(); break;
        case AttributeUnderlyingType::Int32: std::get<ContainerPtrType<Int32>>(attribute.arrays)[idx] = value.get<Int64>(); break;
        case AttributeUnderlyingType::Int64: std::get<ContainerPtrType<Int64>>(attribute.arrays)[idx] = value.get<Int64>(); break;
        case AttributeUnderlyingType::Float32: std::get<ContainerPtrType<Float32>>(attribute.arrays)[idx] = value.get<Float64>(); break;
        case AttributeUnderlyingType::Float64: std::get<ContainerPtrType<Float64>>(attribute.arrays)[idx] = value.get<Float64>(); break;
        case AttributeUnderlyingType::String:
        {
            const auto & string = value.get<String>();
            auto & string_ref = std::get<ContainerPtrType<StringRef>>(attribute.arrays)[idx];
            const auto & null_value_ref = std::get<String>(attribute.null_values);

            /// free memory unless it points to a null_value
            if (string_ref.data && string_ref.data != null_value_ref.data())
                string_arena->free(const_cast<char *>(string_ref.data), string_ref.size);

            const auto size = string.size();
            if (size != 0)
            {
                auto string_ptr = string_arena->alloc(size + 1);
                std::copy(string.data(), string.data() + size + 1, string_ptr);
                string_ref = StringRef{string_ptr, size};
            }
            else
                string_ref = {};

            break;
        }
    }
953 954
}

A
Alexey Milovidov 已提交
955
CacheDictionary::Attribute & CacheDictionary::getAttribute(const std::string & attribute_name) const
956
{
957 958 959 960 961 962 963 964
    const auto it = attribute_index_by_name.find(attribute_name);
    if (it == std::end(attribute_index_by_name))
        throw Exception{
            name + ": no such attribute '" + attribute_name + "'",
            ErrorCodes::BAD_ARGUMENTS
        };

    return attributes[it->second];
965 966
}

967 968
bool CacheDictionary::isEmptyCell(const UInt64 idx) const
{
969
    return (idx != zero_cell_idx && cells[idx].id == 0) || (cells[idx].data
970 971 972 973 974
        == ext::safe_bit_cast<CellMetadata::time_point_urep_t>(CellMetadata::time_point_t()));
}

PaddedPODArray<CacheDictionary::Key> CacheDictionary::getCachedIds() const
{
975 976
    const ProfilingScopedReadRWLock read_lock{rw_lock, ProfileEvents::DictCacheLockReadNs};

977 978 979
    PaddedPODArray<Key> array;
    for (size_t idx = 0; idx < cells.size(); ++idx)
    {
N
Nikolai Kochetov 已提交
980
        auto & cell = cells[idx];
981
        if (!isEmptyCell(idx) && !cells[idx].isDefault())
982 983 984 985 986 987 988
        {
            array.push_back(cell.id);
        }
    }
    return array;
}

989
BlockInputStreamPtr CacheDictionary::getBlockInputStream(const Names & column_names, size_t max_block_size) const
990
{
991
    using BlockInputStreamType = DictionaryBlockInputStream<CacheDictionary, Key>;
992
    return std::make_shared<BlockInputStreamType>(shared_from_this(), max_block_size, getCachedIds(), column_names);
993 994 995
}


996
}