| 1 | #include "CacheDictionary.h" |
| 2 | |
| 3 | #include <Columns/ColumnsNumber.h> |
| 4 | #include <Common/ProfilingScopedRWLock.h> |
| 5 | #include <Common/typeid_cast.h> |
| 6 | #include <DataStreams/IBlockInputStream.h> |
| 7 | #include <ext/chrono_io.h> |
| 8 | #include <ext/map.h> |
| 9 | #include <ext/range.h> |
| 10 | #include <ext/size.h> |
| 11 | |
| 12 | namespace ProfileEvents |
| 13 | { |
| 14 | extern const Event DictCacheKeysRequested; |
| 15 | extern const Event DictCacheKeysRequestedMiss; |
| 16 | extern const Event DictCacheKeysRequestedFound; |
| 17 | extern const Event DictCacheKeysExpired; |
| 18 | extern const Event DictCacheKeysNotFound; |
| 19 | extern const Event DictCacheKeysHit; |
| 20 | extern const Event DictCacheRequestTimeNs; |
| 21 | extern const Event DictCacheRequests; |
| 22 | extern const Event DictCacheLockWriteNs; |
| 23 | extern const Event DictCacheLockReadNs; |
| 24 | } |
| 25 | |
| 26 | namespace CurrentMetrics |
| 27 | { |
| 28 | extern const Metric DictCacheRequests; |
| 29 | } |
| 30 | |
| 31 | namespace DB |
| 32 | { |
| 33 | namespace ErrorCodes |
| 34 | { |
| 35 | extern const int TYPE_MISMATCH; |
| 36 | } |
| 37 | |
| 38 | template <typename AttributeType, typename OutputType, typename DefaultGetter> |
| 39 | void CacheDictionary::getItemsNumberImpl( |
| 40 | Attribute & attribute, const PaddedPODArray<Key> & ids, ResultArrayType<OutputType> & out, DefaultGetter && get_default) const |
| 41 | { |
| 42 | /// Mapping: <id> -> { all indices `i` of `ids` such that `ids[i]` = <id> } |
| 43 | std::unordered_map<Key, std::vector<size_t>> outdated_ids; |
| 44 | auto & attribute_array = std::get<ContainerPtrType<AttributeType>>(attribute.arrays); |
| 45 | const auto rows = ext::size(ids); |
| 46 | |
| 47 | size_t cache_expired = 0, cache_not_found = 0, cache_hit = 0; |
| 48 | |
| 49 | { |
| 50 | const ProfilingScopedReadRWLock read_lock{rw_lock, ProfileEvents::DictCacheLockReadNs}; |
| 51 | |
| 52 | const auto now = std::chrono::system_clock::now(); |
| 53 | /// fetch up-to-date values, decide which ones require update |
| 54 | for (const auto row : ext::range(0, rows)) |
| 55 | { |
| 56 | const auto id = ids[row]; |
| 57 | |
| 58 | /** cell should be updated if either: |
| 59 | * 1. ids do not match, |
| 60 | * 2. cell has expired, |
| 61 | * 3. explicit defaults were specified and cell was set default. */ |
| 62 | |
| 63 | const auto find_result = findCellIdx(id, now); |
| 64 | if (!find_result.valid) |
| 65 | { |
| 66 | outdated_ids[id].push_back(row); |
| 67 | if (find_result.outdated) |
| 68 | ++cache_expired; |
| 69 | else |
| 70 | ++cache_not_found; |
| 71 | } |
| 72 | else |
| 73 | { |
| 74 | ++cache_hit; |
| 75 | const auto & cell_idx = find_result.cell_idx; |
| 76 | const auto & cell = cells[cell_idx]; |
| 77 | out[row] = cell.isDefault() ? get_default(row) : static_cast<OutputType>(attribute_array[cell_idx]); |
| 78 | } |
| 79 | } |
| 80 | } |
| 81 | |
| 82 | ProfileEvents::increment(ProfileEvents::DictCacheKeysExpired, cache_expired); |
| 83 | ProfileEvents::increment(ProfileEvents::DictCacheKeysNotFound, cache_not_found); |
| 84 | ProfileEvents::increment(ProfileEvents::DictCacheKeysHit, cache_hit); |
| 85 | |
| 86 | query_count.fetch_add(rows, std::memory_order_relaxed); |
| 87 | hit_count.fetch_add(rows - outdated_ids.size(), std::memory_order_release); |
| 88 | |
| 89 | if (outdated_ids.empty()) |
| 90 | return; |
| 91 | |
| 92 | std::vector<Key> required_ids(outdated_ids.size()); |
| 93 | std::transform(std::begin(outdated_ids), std::end(outdated_ids), std::begin(required_ids), [](auto & pair) { return pair.first; }); |
| 94 | |
| 95 | /// request new values |
| 96 | update( |
| 97 | required_ids, |
| 98 | [&](const auto id, const auto cell_idx) |
| 99 | { |
| 100 | const auto attribute_value = attribute_array[cell_idx]; |
| 101 | |
| 102 | for (const size_t row : outdated_ids[id]) |
| 103 | out[row] = static_cast<OutputType>(attribute_value); |
| 104 | }, |
| 105 | [&](const auto id, const auto) |
| 106 | { |
| 107 | for (const size_t row : outdated_ids[id]) |
| 108 | out[row] = get_default(row); |
| 109 | }); |
| 110 | } |
| 111 | |
| 112 | template <typename DefaultGetter> |
| 113 | void CacheDictionary::getItemsString( |
| 114 | Attribute & attribute, const PaddedPODArray<Key> & ids, ColumnString * out, DefaultGetter && get_default) const |
| 115 | { |
| 116 | const auto rows = ext::size(ids); |
| 117 | |
| 118 | /// save on some allocations |
| 119 | out->getOffsets().reserve(rows); |
| 120 | |
| 121 | auto & attribute_array = std::get<ContainerPtrType<StringRef>>(attribute.arrays); |
| 122 | |
| 123 | auto found_outdated_values = false; |
| 124 | |
| 125 | /// perform optimistic version, fallback to pessimistic if failed |
| 126 | { |
| 127 | const ProfilingScopedReadRWLock read_lock{rw_lock, ProfileEvents::DictCacheLockReadNs}; |
| 128 | |
| 129 | const auto now = std::chrono::system_clock::now(); |
| 130 | /// fetch up-to-date values, discard on fail |
| 131 | for (const auto row : ext::range(0, rows)) |
| 132 | { |
| 133 | const auto id = ids[row]; |
| 134 | |
| 135 | const auto find_result = findCellIdx(id, now); |
| 136 | if (!find_result.valid) |
| 137 | { |
| 138 | found_outdated_values = true; |
| 139 | break; |
| 140 | } |
| 141 | else |
| 142 | { |
| 143 | const auto & cell_idx = find_result.cell_idx; |
| 144 | const auto & cell = cells[cell_idx]; |
| 145 | const auto string_ref = cell.isDefault() ? get_default(row) : attribute_array[cell_idx]; |
| 146 | out->insertData(string_ref.data, string_ref.size); |
| 147 | } |
| 148 | } |
| 149 | } |
| 150 | |
| 151 | /// optimistic code completed successfully |
| 152 | if (!found_outdated_values) |
| 153 | { |
| 154 | query_count.fetch_add(rows, std::memory_order_relaxed); |
| 155 | hit_count.fetch_add(rows, std::memory_order_release); |
| 156 | return; |
| 157 | } |
| 158 | |
| 159 | /// now onto the pessimistic one, discard possible partial results from the optimistic path |
| 160 | out->getChars().resize_assume_reserved(0); |
| 161 | out->getOffsets().resize_assume_reserved(0); |
| 162 | |
| 163 | /// Mapping: <id> -> { all indices `i` of `ids` such that `ids[i]` = <id> } |
| 164 | std::unordered_map<Key, std::vector<size_t>> outdated_ids; |
| 165 | /// we are going to store every string separately |
| 166 | std::unordered_map<Key, String> map; |
| 167 | |
| 168 | size_t total_length = 0; |
| 169 | size_t cache_expired = 0, cache_not_found = 0, cache_hit = 0; |
| 170 | { |
| 171 | const ProfilingScopedReadRWLock read_lock{rw_lock, ProfileEvents::DictCacheLockReadNs}; |
| 172 | |
| 173 | const auto now = std::chrono::system_clock::now(); |
| 174 | for (const auto row : ext::range(0, ids.size())) |
| 175 | { |
| 176 | const auto id = ids[row]; |
| 177 | |
| 178 | const auto find_result = findCellIdx(id, now); |
| 179 | if (!find_result.valid) |
| 180 | { |
| 181 | outdated_ids[id].push_back(row); |
| 182 | if (find_result.outdated) |
| 183 | ++cache_expired; |
| 184 | else |
| 185 | ++cache_not_found; |
| 186 | } |
| 187 | else |
| 188 | { |
| 189 | ++cache_hit; |
| 190 | const auto & cell_idx = find_result.cell_idx; |
| 191 | const auto & cell = cells[cell_idx]; |
| 192 | const auto string_ref = cell.isDefault() ? get_default(row) : attribute_array[cell_idx]; |
| 193 | |
| 194 | if (!cell.isDefault()) |
| 195 | map[id] = String{string_ref}; |
| 196 | |
| 197 | total_length += string_ref.size + 1; |
| 198 | } |
| 199 | } |
| 200 | } |
| 201 | |
| 202 | ProfileEvents::increment(ProfileEvents::DictCacheKeysExpired, cache_expired); |
| 203 | ProfileEvents::increment(ProfileEvents::DictCacheKeysNotFound, cache_not_found); |
| 204 | ProfileEvents::increment(ProfileEvents::DictCacheKeysHit, cache_hit); |
| 205 | |
| 206 | query_count.fetch_add(rows, std::memory_order_relaxed); |
| 207 | hit_count.fetch_add(rows - outdated_ids.size(), std::memory_order_release); |
| 208 | |
| 209 | /// request new values |
| 210 | if (!outdated_ids.empty()) |
| 211 | { |
| 212 | std::vector<Key> required_ids(outdated_ids.size()); |
| 213 | std::transform(std::begin(outdated_ids), std::end(outdated_ids), std::begin(required_ids), [](auto & pair) { return pair.first; }); |
| 214 | |
| 215 | update( |
| 216 | required_ids, |
| 217 | [&](const auto id, const auto cell_idx) |
| 218 | { |
| 219 | const auto attribute_value = attribute_array[cell_idx]; |
| 220 | |
| 221 | map[id] = String{attribute_value}; |
| 222 | total_length += (attribute_value.size + 1) * outdated_ids[id].size(); |
| 223 | }, |
| 224 | [&](const auto id, const auto) |
| 225 | { |
| 226 | for (const auto row : outdated_ids[id]) |
| 227 | total_length += get_default(row).size + 1; |
| 228 | }); |
| 229 | } |
| 230 | |
| 231 | out->getChars().reserve(total_length); |
| 232 | |
| 233 | for (const auto row : ext::range(0, ext::size(ids))) |
| 234 | { |
| 235 | const auto id = ids[row]; |
| 236 | const auto it = map.find(id); |
| 237 | |
| 238 | const auto string_ref = it != std::end(map) ? StringRef{it->second} : get_default(row); |
| 239 | out->insertData(string_ref.data, string_ref.size); |
| 240 | } |
| 241 | } |
| 242 | |
| 243 | template <typename PresentIdHandler, typename AbsentIdHandler> |
| 244 | void CacheDictionary::update( |
| 245 | const std::vector<Key> & requested_ids, PresentIdHandler && on_cell_updated, AbsentIdHandler && on_id_not_found) const |
| 246 | { |
| 247 | CurrentMetrics::Increment metric_increment{CurrentMetrics::DictCacheRequests}; |
| 248 | ProfileEvents::increment(ProfileEvents::DictCacheKeysRequested, requested_ids.size()); |
| 249 | |
| 250 | std::unordered_map<Key, UInt8> remaining_ids{requested_ids.size()}; |
| 251 | for (const auto id : requested_ids) |
| 252 | remaining_ids.insert({id, 0}); |
| 253 | |
| 254 | const auto now = std::chrono::system_clock::now(); |
| 255 | |
| 256 | const ProfilingScopedWriteRWLock write_lock{rw_lock, ProfileEvents::DictCacheLockWriteNs}; |
| 257 | |
| 258 | if (now > backoff_end_time) |
| 259 | { |
| 260 | try |
| 261 | { |
| 262 | if (error_count) |
| 263 | { |
| 264 | /// Recover after error: we have to clone the source here because |
| 265 | /// it could keep connections which should be reset after error. |
| 266 | source_ptr = source_ptr->clone(); |
| 267 | } |
| 268 | |
| 269 | Stopwatch watch; |
| 270 | auto stream = source_ptr->loadIds(requested_ids); |
| 271 | stream->readPrefix(); |
| 272 | |
| 273 | while (const auto block = stream->read()) |
| 274 | { |
| 275 | const auto id_column = typeid_cast<const ColumnUInt64 *>(block.safeGetByPosition(0).column.get()); |
| 276 | if (!id_column) |
| 277 | throw Exception{name + ": id column has type different from UInt64." , ErrorCodes::TYPE_MISMATCH}; |
| 278 | |
| 279 | const auto & ids = id_column->getData(); |
| 280 | |
| 281 | /// cache column pointers |
| 282 | const auto column_ptrs = ext::map<std::vector>( |
| 283 | ext::range(0, attributes.size()), [&block](size_t i) { return block.safeGetByPosition(i + 1).column.get(); }); |
| 284 | |
| 285 | for (const auto i : ext::range(0, ids.size())) |
| 286 | { |
| 287 | const auto id = ids[i]; |
| 288 | |
| 289 | const auto find_result = findCellIdx(id, now); |
| 290 | const auto & cell_idx = find_result.cell_idx; |
| 291 | |
| 292 | auto & cell = cells[cell_idx]; |
| 293 | |
| 294 | for (const auto attribute_idx : ext::range(0, attributes.size())) |
| 295 | { |
| 296 | const auto & attribute_column = *column_ptrs[attribute_idx]; |
| 297 | auto & attribute = attributes[attribute_idx]; |
| 298 | |
| 299 | setAttributeValue(attribute, cell_idx, attribute_column[i]); |
| 300 | } |
| 301 | |
| 302 | /// if cell id is zero and zero does not map to this cell, then the cell is unused |
| 303 | if (cell.id == 0 && cell_idx != zero_cell_idx) |
| 304 | element_count.fetch_add(1, std::memory_order_relaxed); |
| 305 | |
| 306 | cell.id = id; |
| 307 | if (dict_lifetime.min_sec != 0 && dict_lifetime.max_sec != 0) |
| 308 | { |
| 309 | std::uniform_int_distribution<UInt64> distribution{dict_lifetime.min_sec, dict_lifetime.max_sec}; |
| 310 | cell.setExpiresAt(now + std::chrono::seconds{distribution(rnd_engine)}); |
| 311 | } |
| 312 | else |
| 313 | cell.setExpiresAt(std::chrono::time_point<std::chrono::system_clock>::max()); |
| 314 | |
| 315 | /// inform caller |
| 316 | on_cell_updated(id, cell_idx); |
| 317 | /// mark corresponding id as found |
| 318 | remaining_ids[id] = 1; |
| 319 | } |
| 320 | } |
| 321 | |
| 322 | stream->readSuffix(); |
| 323 | |
| 324 | error_count = 0; |
| 325 | last_exception = std::exception_ptr{}; |
| 326 | backoff_end_time = std::chrono::system_clock::time_point{}; |
| 327 | |
| 328 | ProfileEvents::increment(ProfileEvents::DictCacheRequestTimeNs, watch.elapsed()); |
| 329 | } |
| 330 | catch (...) |
| 331 | { |
| 332 | ++error_count; |
| 333 | last_exception = std::current_exception(); |
| 334 | backoff_end_time = now + std::chrono::seconds(calculateDurationWithBackoff(rnd_engine, error_count)); |
| 335 | |
| 336 | tryLogException(last_exception, log, "Could not update cache dictionary '" + getFullName() + |
| 337 | "', next update is scheduled at " + ext::to_string(backoff_end_time)); |
| 338 | } |
| 339 | } |
| 340 | |
| 341 | size_t not_found_num = 0, found_num = 0; |
| 342 | |
| 343 | /// Check which ids have not been found and require setting null_value |
| 344 | for (const auto & id_found_pair : remaining_ids) |
| 345 | { |
| 346 | if (id_found_pair.second) |
| 347 | { |
| 348 | ++found_num; |
| 349 | continue; |
| 350 | } |
| 351 | ++not_found_num; |
| 352 | |
| 353 | const auto id = id_found_pair.first; |
| 354 | |
| 355 | const auto find_result = findCellIdx(id, now); |
| 356 | const auto & cell_idx = find_result.cell_idx; |
| 357 | auto & cell = cells[cell_idx]; |
| 358 | |
| 359 | if (error_count) |
| 360 | { |
| 361 | if (find_result.outdated) |
| 362 | { |
| 363 | /// We have expired data for that `id` so we can continue using it. |
| 364 | bool was_default = cell.isDefault(); |
| 365 | cell.setExpiresAt(backoff_end_time); |
| 366 | if (was_default) |
| 367 | cell.setDefault(); |
| 368 | if (was_default) |
| 369 | on_id_not_found(id, cell_idx); |
| 370 | else |
| 371 | on_cell_updated(id, cell_idx); |
| 372 | continue; |
| 373 | } |
| 374 | /// We don't have expired data for that `id` so all we can do is to rethrow `last_exception`. |
| 375 | std::rethrow_exception(last_exception); |
| 376 | } |
| 377 | |
| 378 | /// Check if cell had not been occupied before and increment element counter if it hadn't |
| 379 | if (cell.id == 0 && cell_idx != zero_cell_idx) |
| 380 | element_count.fetch_add(1, std::memory_order_relaxed); |
| 381 | |
| 382 | cell.id = id; |
| 383 | |
| 384 | if (dict_lifetime.min_sec != 0 && dict_lifetime.max_sec != 0) |
| 385 | { |
| 386 | std::uniform_int_distribution<UInt64> distribution{dict_lifetime.min_sec, dict_lifetime.max_sec}; |
| 387 | cell.setExpiresAt(now + std::chrono::seconds{distribution(rnd_engine)}); |
| 388 | } |
| 389 | else |
| 390 | cell.setExpiresAt(std::chrono::time_point<std::chrono::system_clock>::max()); |
| 391 | |
| 392 | /// Set null_value for each attribute |
| 393 | cell.setDefault(); |
| 394 | for (auto & attribute : attributes) |
| 395 | setDefaultAttributeValue(attribute, cell_idx); |
| 396 | |
| 397 | /// inform caller that the cell has not been found |
| 398 | on_id_not_found(id, cell_idx); |
| 399 | } |
| 400 | |
| 401 | ProfileEvents::increment(ProfileEvents::DictCacheKeysRequestedMiss, not_found_num); |
| 402 | ProfileEvents::increment(ProfileEvents::DictCacheKeysRequestedFound, found_num); |
| 403 | ProfileEvents::increment(ProfileEvents::DictCacheRequests); |
| 404 | } |
| 405 | |
| 406 | } |
| 407 | |