| 1 |
/* |
| 2 |
Copyright 2022 The Kubernetes Authors. |
| 3 |
|
| 4 |
Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 |
you may not use this file except in compliance with the License. |
| 6 |
You may obtain a copy of the License at |
| 7 |
|
| 8 |
http://www.apache.org/licenses/LICENSE-2.0 |
| 9 |
|
| 10 |
Unless required by applicable law or agreed to in writing, software |
| 11 |
distributed under the License is distributed on an "AS IS" BASIS, |
| 12 |
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 |
See the License for the specific language governing permissions and |
| 14 |
limitations under the License. |
| 15 |
*/ |
| 16 |
|
| 17 |
package cacher |
| 18 |
|
| 19 |
import ( |
| 20 |
"fmt" |
| 21 |
"strings" |
| 22 |
"sync" |
| 23 |
|
| 24 |
"github.com/google/btree" |
| 25 |
"k8s.io/client-go/tools/cache" |
| 26 |
) |
| 27 |
|
| 28 |
// newThreadedBtreeStoreIndexer returns a storage for cacher by adding locking over the two 2 data structures: |
| 29 |
// * btree based storage for efficient LIST operation on prefix |
| 30 |
// * map based indexer for retrieving values by index. |
| 31 |
// This separation is used to allow independent snapshotting those two storages in the future. |
| 32 |
// Intention is to utilize btree for its cheap snapshots that don't require locking if don't mutate data. |
| 33 |
func newThreadedBtreeStoreIndexer(indexers cache.Indexers, degree int) *threadedStoreIndexer { |
| 34 |
return &threadedStoreIndexer{ |
| 35 |
store: newBtreeStore(degree), |
| 36 |
indexer: newIndexer(indexers), |
| 37 |
} |
| 38 |
} |
| 39 |
|
| 40 |
type threadedStoreIndexer struct { |
| 41 |
lock sync.RWMutex |
| 42 |
store btreeStore |
| 43 |
indexer indexer |
| 44 |
} |
| 45 |
|
| 46 |
var _ orderedLister = (*threadedStoreIndexer)(nil) |
| 47 |
|
| 48 |
func (si *threadedStoreIndexer) Count(prefix, continueKey string) (count int) { |
| 49 |
si.lock.RLock() |
| 50 |
defer si.lock.RUnlock() |
| 51 |
return si.store.Count(prefix, continueKey) |
| 52 |
} |
| 53 |
|
| 54 |
func (si *threadedStoreIndexer) Clone() orderedLister { |
| 55 |
si.lock.RLock() |
| 56 |
defer si.lock.RUnlock() |
| 57 |
return si.store.Clone() |
| 58 |
} |
| 59 |
|
| 60 |
func (si *threadedStoreIndexer) Add(obj interface{}) error { |
| 61 |
return si.addOrUpdate(obj) |
| 62 |
} |
| 63 |
|
| 64 |
func (si *threadedStoreIndexer) Update(obj interface{}) error { |
| 65 |
return si.addOrUpdate(obj) |
| 66 |
} |
| 67 |
|
| 68 |
func (si *threadedStoreIndexer) addOrUpdate(obj interface{}) error { |
| 69 |
if obj == nil { |
| 70 |
return fmt.Errorf("obj cannot be nil") |
| 71 |
} |
| 72 |
newElem, ok := obj.(*storeElement) |
| 73 |
if !ok { |
| 74 |
return fmt.Errorf("obj not a storeElement: %#v", obj) |
| 75 |
} |
| 76 |
si.lock.Lock() |
| 77 |
defer si.lock.Unlock() |
| 78 |
oldElem := si.store.addOrUpdateElem(newElem) |
| 79 |
return si.indexer.updateElem(newElem.Key, oldElem, newElem) |
| 80 |
} |
| 81 |
|
| 82 |
func (si *threadedStoreIndexer) Delete(obj interface{}) error { |
| 83 |
storeElem, ok := obj.(*storeElement) |
| 84 |
if !ok { |
| 85 |
return fmt.Errorf("obj not a storeElement: %#v", obj) |
| 86 |
} |
| 87 |
si.lock.Lock() |
| 88 |
defer si.lock.Unlock() |
| 89 |
oldObj, existed := si.store.deleteElem(storeElem) |
| 90 |
if !existed { |
| 91 |
return nil |
| 92 |
} |
| 93 |
return si.indexer.updateElem(storeElem.Key, oldObj, nil) |
| 94 |
} |
| 95 |
|
| 96 |
func (si *threadedStoreIndexer) List() []interface{} { |
| 97 |
si.lock.RLock() |
| 98 |
defer si.lock.RUnlock() |
| 99 |
return si.store.List() |
| 100 |
} |
| 101 |
|
| 102 |
func (si *threadedStoreIndexer) ListPrefix(prefix, continueKey string) []interface{} { |
| 103 |
si.lock.RLock() |
| 104 |
defer si.lock.RUnlock() |
| 105 |
return si.store.ListPrefix(prefix, continueKey) |
| 106 |
} |
| 107 |
|
| 108 |
func (si *threadedStoreIndexer) ListKeys() []string { |
| 109 |
si.lock.RLock() |
| 110 |
defer si.lock.RUnlock() |
| 111 |
return si.store.ListKeys() |
| 112 |
} |
| 113 |
|
| 114 |
func (si *threadedStoreIndexer) Get(obj interface{}) (item interface{}, exists bool, err error) { |
| 115 |
si.lock.RLock() |
| 116 |
defer si.lock.RUnlock() |
| 117 |
return si.store.Get(obj) |
| 118 |
} |
| 119 |
|
| 120 |
func (si *threadedStoreIndexer) GetByKey(key string) (item interface{}, exists bool, err error) { |
| 121 |
si.lock.RLock() |
| 122 |
defer si.lock.RUnlock() |
| 123 |
return si.store.GetByKey(key) |
| 124 |
} |
| 125 |
|
| 126 |
func (si *threadedStoreIndexer) Replace(objs []interface{}, resourceVersion string) error { |
| 127 |
si.lock.Lock() |
| 128 |
defer si.lock.Unlock() |
| 129 |
err := si.store.Replace(objs, resourceVersion) |
| 130 |
if err != nil { |
| 131 |
return err |
| 132 |
} |
| 133 |
return si.indexer.Replace(objs, resourceVersion) |
| 134 |
} |
| 135 |
|
| 136 |
func (si *threadedStoreIndexer) ByIndex(indexName, indexValue string) ([]interface{}, error) { |
| 137 |
si.lock.RLock() |
| 138 |
defer si.lock.RUnlock() |
| 139 |
return si.indexer.ByIndex(indexName, indexValue) |
| 140 |
} |
| 141 |
|
| 142 |
func newBtreeStore(degree int) btreeStore { |
| 143 |
return btreeStore{ |
| 144 |
tree: btree.NewG(degree, func(a, b *storeElement) bool { |
| 145 |
return a.Key < b.Key |
| 146 |
}), |
| 147 |
} |
| 148 |
} |
| 149 |
|
| 150 |
type btreeStore struct { |
| 151 |
tree *btree.BTreeG[*storeElement] |
| 152 |
} |
| 153 |
|
| 154 |
func (s *btreeStore) Clone() orderedLister { |
| 155 |
return &btreeStore{ |
| 156 |
tree: s.tree.Clone(), |
| 157 |
} |
| 158 |
} |
| 159 |
|
| 160 |
func (s *btreeStore) Add(obj interface{}) error { |
| 161 |
if obj == nil { |
| 162 |
return fmt.Errorf("obj cannot be nil") |
| 163 |
} |
| 164 |
storeElem, ok := obj.(*storeElement) |
| 165 |
if !ok { |
| 166 |
return fmt.Errorf("obj not a storeElement: %#v", obj) |
| 167 |
} |
| 168 |
s.addOrUpdateElem(storeElem) |
| 169 |
return nil |
| 170 |
} |
| 171 |
|
| 172 |
func (s *btreeStore) Update(obj interface{}) error { |
| 173 |
if obj == nil { |
| 174 |
return fmt.Errorf("obj cannot be nil") |
| 175 |
} |
| 176 |
storeElem, ok := obj.(*storeElement) |
| 177 |
if !ok { |
| 178 |
return fmt.Errorf("obj not a storeElement: %#v", obj) |
| 179 |
} |
| 180 |
s.addOrUpdateElem(storeElem) |
| 181 |
return nil |
| 182 |
} |
| 183 |
|
| 184 |
func (s *btreeStore) Delete(obj interface{}) error { |
| 185 |
if obj == nil { |
| 186 |
return fmt.Errorf("obj cannot be nil") |
| 187 |
} |
| 188 |
storeElem, ok := obj.(*storeElement) |
| 189 |
if !ok { |
| 190 |
return fmt.Errorf("obj not a storeElement: %#v", obj) |
| 191 |
} |
| 192 |
s.deleteElem(storeElem) |
| 193 |
return nil |
| 194 |
} |
| 195 |
|
| 196 |
func (s *btreeStore) deleteElem(storeElem *storeElement) (*storeElement, bool) { |
| 197 |
return s.tree.Delete(storeElem) |
| 198 |
} |
| 199 |
|
| 200 |
func (s *btreeStore) List() []interface{} { |
| 201 |
items := make([]interface{}, 0, s.tree.Len()) |
| 202 |
s.tree.Ascend(func(item *storeElement) bool { |
| 203 |
items = append(items, item) |
| 204 |
return true |
| 205 |
}) |
| 206 |
return items |
| 207 |
} |
| 208 |
|
| 209 |
func (s *btreeStore) ListKeys() []string { |
| 210 |
items := make([]string, 0, s.tree.Len()) |
| 211 |
s.tree.Ascend(func(item *storeElement) bool { |
| 212 |
items = append(items, item.Key) |
| 213 |
return true |
| 214 |
}) |
| 215 |
return items |
| 216 |
} |
| 217 |
|
| 218 |
func (s *btreeStore) Get(obj interface{}) (item interface{}, exists bool, err error) { |
| 219 |
storeElem, ok := obj.(*storeElement) |
| 220 |
if !ok { |
| 221 |
return nil, false, fmt.Errorf("obj is not a storeElement") |
| 222 |
} |
| 223 |
item, exists = s.tree.Get(storeElem) |
| 224 |
return item, exists, nil |
| 225 |
} |
| 226 |
|
| 227 |
func (s *btreeStore) GetByKey(key string) (item interface{}, exists bool, err error) { |
| 228 |
return s.getByKey(key) |
| 229 |
} |
| 230 |
|
| 231 |
func (s *btreeStore) Replace(objs []interface{}, _ string) error { |
| 232 |
s.tree.Clear(false) |
| 233 |
for _, obj := range objs { |
| 234 |
storeElem, ok := obj.(*storeElement) |
| 235 |
if !ok { |
| 236 |
return fmt.Errorf("obj not a storeElement: %#v", obj) |
| 237 |
} |
| 238 |
s.addOrUpdateElem(storeElem) |
| 239 |
} |
| 240 |
return nil |
| 241 |
} |
| 242 |
|
| 243 |
// addOrUpdateLocked assumes a lock is held and is used for Add |
| 244 |
// and Update operations. |
| 245 |
func (s *btreeStore) addOrUpdateElem(storeElem *storeElement) *storeElement { |
| 246 |
oldObj, _ := s.tree.ReplaceOrInsert(storeElem) |
| 247 |
return oldObj |
| 248 |
} |
| 249 |
|
| 250 |
func (s *btreeStore) getByKey(key string) (item interface{}, exists bool, err error) { |
| 251 |
keyElement := &storeElement{Key: key} |
| 252 |
item, exists = s.tree.Get(keyElement) |
| 253 |
return item, exists, nil |
| 254 |
} |
| 255 |
|
| 256 |
func (s *btreeStore) ListPrefix(prefix, continueKey string) []interface{} { |
| 257 |
if continueKey == "" { |
| 258 |
continueKey = prefix |
| 259 |
} |
| 260 |
var result []interface{} |
| 261 |
s.tree.AscendGreaterOrEqual(&storeElement{Key: continueKey}, func(item *storeElement) bool { |
| 262 |
if !strings.HasPrefix(item.Key, prefix) { |
| 263 |
return false |
| 264 |
} |
| 265 |
result = append(result, item) |
| 266 |
return true |
| 267 |
}) |
| 268 |
return result |
| 269 |
} |
| 270 |
|
| 271 |
func (s *btreeStore) Count(prefix, continueKey string) (count int) { |
| 272 |
if continueKey == "" { |
| 273 |
continueKey = prefix |
| 274 |
} |
| 275 |
s.tree.AscendGreaterOrEqual(&storeElement{Key: continueKey}, func(item *storeElement) bool { |
| 276 |
if !strings.HasPrefix(item.Key, prefix) { |
| 277 |
return false |
| 278 |
} |
| 279 |
count++ |
| 280 |
return true |
| 281 |
}) |
| 282 |
return count |
| 283 |
} |
| 284 |
|
| 285 |
// newIndexer returns a indexer similar to storeIndex from client-go/tools/cache. |
| 286 |
// TODO: Unify the indexer code with client-go/cache package. |
| 287 |
// Major differences is type of values stored and their mutability: |
| 288 |
// * Indexer in client-go stores object keys, that are not mutable. |
| 289 |
// * Indexer in cacher stores whole objects, which is mutable. |
| 290 |
// Indexer in client-go uses keys as it is used in conjunction with map[key]value |
| 291 |
// allowing for fast value retrieval, while btree used in cacher would provide additional overhead. |
| 292 |
// Difference in mutability of stored values is used for optimizing some operations in client-go Indexer. |
| 293 |
func newIndexer(indexers cache.Indexers) indexer { |
| 294 |
return indexer{ |
| 295 |
indices: map[string]map[string]map[string]*storeElement{}, |
| 296 |
indexers: indexers, |
| 297 |
} |
| 298 |
} |
| 299 |
|
| 300 |
type indexer struct { |
| 301 |
indices map[string]map[string]map[string]*storeElement |
| 302 |
indexers cache.Indexers |
| 303 |
} |
| 304 |
|
| 305 |
func (i *indexer) ByIndex(indexName, indexValue string) ([]interface{}, error) { |
| 306 |
indexFunc := i.indexers[indexName] |
| 307 |
if indexFunc == nil { |
| 308 |
return nil, fmt.Errorf("index with name %s does not exist", indexName) |
| 309 |
} |
| 310 |
index := i.indices[indexName] |
| 311 |
set := index[indexValue] |
| 312 |
list := make([]interface{}, 0, len(set)) |
| 313 |
for _, obj := range set { |
| 314 |
list = append(list, obj) |
| 315 |
} |
| 316 |
return list, nil |
| 317 |
} |
| 318 |
|
| 319 |
func (i *indexer) Replace(objs []interface{}, resourceVersion string) error { |
| 320 |
i.indices = map[string]map[string]map[string]*storeElement{} |
| 321 |
for _, obj := range objs { |
| 322 |
storeElem, ok := obj.(*storeElement) |
| 323 |
if !ok { |
| 324 |
return fmt.Errorf("obj not a storeElement: %#v", obj) |
| 325 |
} |
| 326 |
err := i.updateElem(storeElem.Key, nil, storeElem) |
| 327 |
if err != nil { |
| 328 |
return err |
| 329 |
} |
| 330 |
} |
| 331 |
return nil |
| 332 |
} |
| 333 |
|
| 334 |
func (i *indexer) updateElem(key string, oldObj, newObj *storeElement) (err error) { |
| 335 |
var oldIndexValues, indexValues []string |
| 336 |
for name, indexFunc := range i.indexers { |
| 337 |
if oldObj != nil { |
| 338 |
oldIndexValues, err = indexFunc(oldObj) |
| 339 |
} else { |
| 340 |
oldIndexValues = oldIndexValues[:0] |
| 341 |
} |
| 342 |
if err != nil { |
| 343 |
return fmt.Errorf("unable to calculate an index entry for key %q on index %q: %w", key, name, err) |
| 344 |
} |
| 345 |
if newObj != nil { |
| 346 |
indexValues, err = indexFunc(newObj) |
| 347 |
} else { |
| 348 |
indexValues = indexValues[:0] |
| 349 |
} |
| 350 |
if err != nil { |
| 351 |
return fmt.Errorf("unable to calculate an index entry for key %q on index %q: %w", key, name, err) |
| 352 |
} |
| 353 |
index := i.indices[name] |
| 354 |
if index == nil { |
| 355 |
index = map[string]map[string]*storeElement{} |
| 356 |
i.indices[name] = index |
| 357 |
} |
| 358 |
if len(indexValues) == 1 && len(oldIndexValues) == 1 && indexValues[0] == oldIndexValues[0] { |
| 359 |
// We optimize for the most common case where indexFunc returns a single value which has not been changed |
| 360 |
i.add(key, indexValues[0], newObj, index) |
| 361 |
continue |
| 362 |
} |
| 363 |
for _, value := range oldIndexValues { |
| 364 |
i.delete(key, value, index) |
| 365 |
} |
| 366 |
for _, value := range indexValues { |
| 367 |
i.add(key, value, newObj, index) |
| 368 |
} |
| 369 |
} |
| 370 |
return nil |
| 371 |
} |
| 372 |
|
| 373 |
func (i *indexer) add(key, value string, obj *storeElement, index map[string]map[string]*storeElement) { |
| 374 |
set := index[value] |
| 375 |
if set == nil { |
| 376 |
set = map[string]*storeElement{} |
| 377 |
index[value] = set |
| 378 |
} |
| 379 |
set[key] = obj |
| 380 |
} |
| 381 |
|
| 382 |
func (i *indexer) delete(key, value string, index map[string]map[string]*storeElement) { |
| 383 |
set := index[value] |
| 384 |
if set == nil { |
| 385 |
return |
| 386 |
} |
| 387 |
delete(set, key) |
| 388 |
// If we don's delete the set when zero, indices with high cardinality |
| 389 |
// short lived resources can cause memory to increase over time from |
| 390 |
// unused empty sets. See `kubernetes/kubernetes/issues/84959`. |
| 391 |
if len(set) == 0 { |
| 392 |
delete(index, value) |
| 393 |
} |
| 394 |
} |
| 395 |
|
| 396 |
// newStoreSnapshotter returns a storeSnapshotter that stores snapshots for |
| 397 |
// serving read requests with exact resource versions (RV) and pagination. |
| 398 |
// |
| 399 |
// Snapshots are created by calling Clone method on orderedLister, which is |
| 400 |
// expected to be fast and efficient thanks to usage of B-trees. |
| 401 |
// B-trees can create a lazy copy of the tree structure, minimizing overhead. |
| 402 |
// |
| 403 |
// Assuming the watch cache observes all events and snapshots cache after each of them, |
| 404 |
// requests for a specific resource version can be served by retrieving |
| 405 |
// the snapshot with the greatest RV less than or equal to the requested RV. |
| 406 |
// To make snapshot retrivial efficient we need an ordered data structure, such as tree. |
| 407 |
// |
| 408 |
// The initial implementation uses a B-tree to achieve the following performance characteristics (n - number of snapshots stored): |
| 409 |
// - `Add`: Adds a new snapshot. |
| 410 |
// Complexity: O(log n). |
| 411 |
// Executed for each watch event observed by the cache. |
| 412 |
// - `GetLessOrEqual`: Retrieves the snapshot with the greatest RV less than or equal to the requested RV. |
| 413 |
// Complexity: O(log n). |
| 414 |
// Executed for each LIST request with match=Exact or continuation. |
| 415 |
// - `RemoveLess`: Cleans up snapshots outside the watch history window. |
| 416 |
// Complexity: O(k log n), k - number of snapshots to remove, usually only one if watch capacity was not reduced. |
| 417 |
// Executed per watch event observed when the cache is full. |
| 418 |
// - `Reset`: Cleans up all snapshots. |
| 419 |
// Complexity: O(1). |
| 420 |
// Executed when the watch cache is reinitialized. |
| 421 |
// |
| 422 |
// Further optimization is possible by leveraging the property that adds always |
| 423 |
// increase the maximum RV and deletes only increase the minimum RV. |
| 424 |
// For example, a binary search on a cyclic buffer of (RV, snapshot) |
| 425 |
// should reduce number of allocations and improve removal complexity. |
| 426 |
// However, this solution is more complex and is deferred for future implementation. |
| 427 |
// |
| 428 |
// TODO: Rewrite to use a cyclic buffer |
| 429 |
func newStoreSnapshotter() *storeSnapshotter { |
| 430 |
s := &storeSnapshotter{ |
| 431 |
snapshots: btree.NewG[rvSnapshot](btreeDegree, func(a, b rvSnapshot) bool { |
| 432 |
return a.resourceVersion < b.resourceVersion |
| 433 |
}), |
| 434 |
} |
| 435 |
return s |
| 436 |
} |
| 437 |
|
| 438 |
var _ Snapshotter = (*storeSnapshotter)(nil) |
| 439 |
|
| 440 |
type Snapshotter interface { |
| 441 |
Reset() |
| 442 |
GetLessOrEqual(rv uint64) (orderedLister, bool) |
| 443 |
Add(rv uint64, indexer orderedLister) |
| 444 |
RemoveLess(rv uint64) |
| 445 |
Len() int |
| 446 |
} |
| 447 |
|
| 448 |
type storeSnapshotter struct { |
| 449 |
mux sync.RWMutex |
| 450 |
snapshots *btree.BTreeG[rvSnapshot] |
| 451 |
} |
| 452 |
|
| 453 |
type rvSnapshot struct { |
| 454 |
resourceVersion uint64 |
| 455 |
snapshot orderedLister |
| 456 |
} |
| 457 |
|
| 458 |
func (s *storeSnapshotter) Reset() { |
| 459 |
s.mux.Lock() |
| 460 |
defer s.mux.Unlock() |
| 461 |
s.snapshots.Clear(false) |
| 462 |
} |
| 463 |
|
| 464 |
func (s *storeSnapshotter) GetLessOrEqual(rv uint64) (orderedLister, bool) { |
| 465 |
s.mux.RLock() |
| 466 |
defer s.mux.RUnlock() |
| 467 |
|
| 468 |
var result *rvSnapshot |
| 469 |
s.snapshots.DescendLessOrEqual(rvSnapshot{resourceVersion: rv}, func(rvs rvSnapshot) bool { |
| 470 |
result = &rvs |
| 471 |
return false |
| 472 |
}) |
| 473 |
if result == nil { |
| 474 |
return nil, false |
| 475 |
} |
| 476 |
return result.snapshot, true |
| 477 |
} |
| 478 |
|
| 479 |
func (s *storeSnapshotter) Add(rv uint64, indexer orderedLister) { |
| 480 |
s.mux.Lock() |
| 481 |
defer s.mux.Unlock() |
| 482 |
s.snapshots.ReplaceOrInsert(rvSnapshot{resourceVersion: rv, snapshot: indexer.Clone()}) |
| 483 |
} |
| 484 |
|
| 485 |
func (s *storeSnapshotter) RemoveLess(rv uint64) { |
| 486 |
s.mux.Lock() |
| 487 |
defer s.mux.Unlock() |
| 488 |
for s.snapshots.Len() > 0 { |
| 489 |
oldest, ok := s.snapshots.Min() |
| 490 |
if !ok { |
| 491 |
break |
| 492 |
} |
| 493 |
if rv <= oldest.resourceVersion { |
| 494 |
break |
| 495 |
} |
| 496 |
s.snapshots.DeleteMin() |
| 497 |
} |
| 498 |
} |
| 499 |
|
| 500 |
func (s *storeSnapshotter) Len() int { |
| 501 |
s.mux.RLock() |
| 502 |
defer s.mux.RUnlock() |
| 503 |
|
| 504 |
return s.snapshots.Len() |
| 505 |
} |