001/* 002 * Licensed to the Apache Software Foundation (ASF) under one 003 * or more contributor license agreements. See the NOTICE file 004 * distributed with this work for additional information 005 * regarding copyright ownership. The ASF licenses this file 006 * to you under the Apache License, Version 2.0 (the 007 * "License"); you may not use this file except in compliance 008 * with the License. You may obtain a copy of the License at 009 * 010 * http://www.apache.org/licenses/LICENSE-2.0 011 * 012 * Unless required by applicable law or agreed to in writing, software 013 * distributed under the License is distributed on an "AS IS" BASIS, 014 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 015 * See the License for the specific language governing permissions and 016 * limitations under the License. 017 */ 018package org.apache.hadoop.hbase.wal; 019 020import java.io.IOException; 021import java.util.ArrayList; 022import java.util.List; 023import java.util.Map; 024import java.util.Set; 025import java.util.TreeSet; 026import org.apache.hadoop.hbase.Cell; 027import org.apache.hadoop.hbase.CellUtil; 028import org.apache.hadoop.hbase.HBaseInterfaceAudience; 029import org.apache.hadoop.hbase.KeyValue; 030import org.apache.hadoop.hbase.PrivateCellUtil; 031import org.apache.hadoop.hbase.client.RegionInfo; 032import org.apache.hadoop.hbase.codec.Codec; 033import org.apache.hadoop.hbase.io.HeapSize; 034import org.apache.hadoop.hbase.util.Bytes; 035import org.apache.hadoop.hbase.util.ClassSize; 036import org.apache.hadoop.hbase.util.EnvironmentEdgeManager; 037import org.apache.yetus.audience.InterfaceAudience; 038 039import org.apache.hadoop.hbase.shaded.protobuf.generated.WALProtos; 040import org.apache.hadoop.hbase.shaded.protobuf.generated.WALProtos.CompactionDescriptor; 041import org.apache.hadoop.hbase.shaded.protobuf.generated.WALProtos.FlushDescriptor; 042import org.apache.hadoop.hbase.shaded.protobuf.generated.WALProtos.RegionEventDescriptor; 043 044/** 045 * Used in HBase's transaction log (WAL) to represent a collection of edits (Cell/KeyValue objects) 046 * that came in as a single transaction. All the edits for a given transaction are written out as a 047 * single record, in PB format, followed (optionally) by Cells written via the WALCellEncoder. 048 * <p> 049 * This class is LimitedPrivate for CPs to read-only. The {@link #add} methods are classified as 050 * private methods, not for use by CPs. 051 * </p> 052 * <p> 053 * A particular WALEdit 'type' is the 'meta' type used to mark key operational events in the WAL 054 * such as compaction, flush, or region open. These meta types do not traverse hbase memstores. They 055 * are edits made by the hbase system rather than edit data submitted by clients. They only show in 056 * the WAL. These 'Meta' types have not been formally specified (or made into an explicit class 057 * type). They evolved organically. HBASE-8457 suggests codifying a WALEdit 'type' by adding a type 058 * field to WALEdit that gets serialized into the WAL. TODO. Would have to work on the 059 * consumption-side. Reading WALs on replay we seem to consume a Cell-at-a-time rather than by 060 * WALEdit. We are already in the below going out of our way to figure particular types -- e.g. if a 061 * compaction, replay, or close meta Marker -- during normal processing so would make sense to do 062 * this. Current system is an awkward marking of Cell columnfamily as {@link #METAFAMILY} and then 063 * setting qualifier based off meta edit type. For replay-time where we read Cell-at-a-time, there 064 * are utility methods below for figuring meta type. See also 065 * {@link #createBulkLoadEvent(RegionInfo, WALProtos.BulkLoadDescriptor)}, etc., for where we create 066 * meta WALEdit instances. 067 * </p> 068 * <p> 069 * WALEdit will accumulate a Set of all column family names referenced by the Cells 070 * {@link #add(Cell)}'d. This is an optimization. Usually when loading a WALEdit, we have the column 071 * family name to-hand.. just shove it into the WALEdit if available. Doing this, we can save on a 072 * parse of each Cell to figure column family down the line when we go to add the WALEdit to the WAL 073 * file. See the hand-off in FSWALEntry Constructor. 074 * @see WALKey 075 */ 076// TODO: Do not expose this class to Coprocessors. It has set methods. A CP might meddle. 077@InterfaceAudience.LimitedPrivate({ HBaseInterfaceAudience.REPLICATION, 078 HBaseInterfaceAudience.COPROC }) 079public class WALEdit implements HeapSize { 080 // Below defines are for writing WALEdit 'meta' Cells.. 081 // TODO: Get rid of this system of special 'meta' Cells. See HBASE-8457. It suggests 082 // adding a type to WALEdit itself for use denoting meta Edits and their types. 083 public static final byte[] METAFAMILY = Bytes.toBytes("METAFAMILY"); 084 085 /** 086 * @deprecated Since 2.3.0. Not used. 087 */ 088 @Deprecated 089 public static final byte[] METAROW = Bytes.toBytes("METAROW"); 090 091 /** 092 * @deprecated Since 2.3.0. Make it protected, internal-use only. Use 093 * {@link #isCompactionMarker(Cell)} 094 */ 095 @Deprecated 096 @InterfaceAudience.Private 097 public static final byte[] COMPACTION = Bytes.toBytes("HBASE::COMPACTION"); 098 099 /** 100 * @deprecated Since 2.3.0. Make it protected, internal-use only. 101 */ 102 @Deprecated 103 @InterfaceAudience.Private 104 public static final byte[] FLUSH = Bytes.toBytes("HBASE::FLUSH"); 105 106 /** 107 * Qualifier for region event meta 'Marker' WALEdits start with the {@link #REGION_EVENT_PREFIX} 108 * prefix ('HBASE::REGION_EVENT::'). After the prefix, we note the type of the event which we get 109 * from the RegionEventDescriptor protobuf instance type (A RegionEventDescriptor protobuf 110 * instance is written as the meta Marker Cell value). Adding a type suffix means we do not have 111 * to deserialize the protobuf to figure out what type of event this is.. .just read the qualifier 112 * suffix. For example, a close region event descriptor will have a qualifier of 113 * HBASE::REGION_EVENT::REGION_CLOSE. See WAL.proto and the EventType in RegionEventDescriptor 114 * protos for all possible event types. 115 */ 116 private static final String REGION_EVENT_STR = "HBASE::REGION_EVENT"; 117 private static final String REGION_EVENT_PREFIX_STR = REGION_EVENT_STR + "::"; 118 private static final byte[] REGION_EVENT_PREFIX = Bytes.toBytes(REGION_EVENT_PREFIX_STR); 119 120 /** 121 * @deprecated Since 2.3.0. Remove. Not for external use. Not used. 122 */ 123 @Deprecated 124 public static final byte[] REGION_EVENT = Bytes.toBytes(REGION_EVENT_STR); 125 126 /** 127 * We use this define figuring if we are carrying a close event. 128 */ 129 private static final byte[] REGION_EVENT_CLOSE = 130 createRegionEventDescriptorQualifier(RegionEventDescriptor.EventType.REGION_CLOSE); 131 132 @InterfaceAudience.Private 133 public static final byte[] BULK_LOAD = Bytes.toBytes("HBASE::BULK_LOAD"); 134 135 /** 136 * Periodically {@link org.apache.hadoop.hbase.replication.regionserver.ReplicationMarkerChore} 137 * will create marker edits with family as {@link WALEdit#METAFAMILY} and 138 * {@link WALEdit#REPLICATION_MARKER} as qualifier and an empty value. 139 * org.apache.hadoop.hbase.replication.regionserver.ReplicationSourceWALReader will populate the 140 * Replication Marker edit with region_server_name, wal_name and wal_offset encoded in 141 * {@link org.apache.hadoop.hbase.shaded.protobuf.generated.WALProtos.ReplicationMarkerDescriptor} 142 * object. {@link org.apache.hadoop.hbase.replication.regionserver.Replication} will change the 143 * REPLICATION_SCOPE for this edit to GLOBAL so that it can replicate. On the sink cluster, 144 * {@link org.apache.hadoop.hbase.replication.regionserver.ReplicationSink} will convert the 145 * ReplicationMarkerDescriptor into a Put mutation to REPLICATION_SINK_TRACKER_TABLE_NAME_STR 146 * table. 147 */ 148 @InterfaceAudience.Private 149 public static final byte[] REPLICATION_MARKER = Bytes.toBytes("HBASE::REPLICATION_MARKER"); 150 151 private final transient boolean replay; 152 153 private ArrayList<Cell> cells; 154 155 /** 156 * All the Cell families in <code>cells</code>. Updated by {@link #add(Cell)} and 157 * {@link #add(Map)}. This Set is passed to the FSWALEntry so it does not have to recalculate the 158 * Set of families in a transaction; makes for a bunch of CPU savings. 159 */ 160 private Set<byte[]> families = null; 161 162 public WALEdit() { 163 this(1, false); 164 } 165 166 /** 167 * @deprecated since 2.0.1 and will be removed in 4.0.0. Use {@link #WALEdit(int, boolean)} 168 * instead. 169 * @see #WALEdit(int, boolean) 170 * @see <a href="https://issues.apache.org/jira/browse/HBASE-20781">HBASE-20781</a> 171 */ 172 @Deprecated 173 public WALEdit(boolean replay) { 174 this(1, replay); 175 } 176 177 /** 178 * @deprecated since 2.0.1 and will be removed in 4.0.0. Use {@link #WALEdit(int, boolean)} 179 * instead. 180 * @see #WALEdit(int, boolean) 181 * @see <a href="https://issues.apache.org/jira/browse/HBASE-20781">HBASE-20781</a> 182 */ 183 @Deprecated 184 public WALEdit(int cellCount) { 185 this(cellCount, false); 186 } 187 188 /** 189 * @param cellCount Pass so can pre-size the WALEdit. Optimization. 190 */ 191 public WALEdit(int cellCount, boolean isReplay) { 192 this.replay = isReplay; 193 cells = new ArrayList<>(cellCount); 194 } 195 196 private Set<byte[]> getOrCreateFamilies() { 197 if (this.families == null) { 198 this.families = new TreeSet<>(Bytes.BYTES_COMPARATOR); 199 } 200 return this.families; 201 } 202 203 /** 204 * For use by FSWALEntry ONLY. An optimization. 205 * @return All families in {@link #getCells()}; may be null. 206 */ 207 public Set<byte[]> getFamilies() { 208 return this.families; 209 } 210 211 /** 212 * @return True is <code>f</code> is {@link #METAFAMILY} 213 * @deprecated Since 2.3.0. Do not expose. Make protected. 214 */ 215 @Deprecated 216 public static boolean isMetaEditFamily(final byte[] f) { 217 return Bytes.equals(METAFAMILY, f); 218 } 219 220 /** 221 * Replaying WALs can read Cell-at-a-time so need this method in those cases. 222 */ 223 public static boolean isMetaEditFamily(Cell cell) { 224 return CellUtil.matchingFamily(cell, METAFAMILY); 225 } 226 227 /** 228 * @return True if this is a meta edit; has one edit only and its columnfamily is 229 * {@link #METAFAMILY}. 230 */ 231 public boolean isMetaEdit() { 232 return this.families != null && this.families.size() == 1 && this.families.contains(METAFAMILY); 233 } 234 235 /** 236 * @return True when current WALEdit is created by log replay. Replication skips WALEdits from 237 * replay. 238 */ 239 public boolean isReplay() { 240 return this.replay; 241 } 242 243 @InterfaceAudience.Private 244 public WALEdit add(Cell cell, byte[] family) { 245 getOrCreateFamilies().add(family); 246 return addCell(cell); 247 } 248 249 @InterfaceAudience.Private 250 public WALEdit add(Cell cell) { 251 // We clone Family each time we add a Cell. Expensive but safe. For CPU savings, use 252 // add(Map) or add(Cell, family). 253 return add(cell, CellUtil.cloneFamily(cell)); 254 } 255 256 public boolean isEmpty() { 257 return cells.isEmpty(); 258 } 259 260 public int size() { 261 return cells.size(); 262 } 263 264 public ArrayList<Cell> getCells() { 265 return cells; 266 } 267 268 /** 269 * This is not thread safe. This will change the WALEdit and shouldn't be used unless you are sure 270 * that nothing else depends on the contents being immutable. 271 * @param cells the list of cells that this WALEdit now contains. 272 */ 273 @InterfaceAudience.Private 274 // Used by replay. 275 public void setCells(ArrayList<Cell> cells) { 276 this.cells = cells; 277 this.families = null; 278 } 279 280 /** 281 * Reads WALEdit from cells. 282 * @param cellDecoder Cell decoder. 283 * @param expectedCount Expected cell count. 284 * @return Number of KVs read. 285 */ 286 public int readFromCells(Codec.Decoder cellDecoder, int expectedCount) throws IOException { 287 cells.clear(); 288 cells.ensureCapacity(expectedCount); 289 while (cells.size() < expectedCount && cellDecoder.advance()) { 290 add(cellDecoder.current()); 291 } 292 return cells.size(); 293 } 294 295 @Override 296 public long heapSize() { 297 long ret = ClassSize.ARRAYLIST; 298 for (Cell cell : cells) { 299 ret += cell.heapSize(); 300 } 301 return ret; 302 } 303 304 public long estimatedSerializedSizeOf() { 305 long ret = 0; 306 for (Cell cell : cells) { 307 ret += PrivateCellUtil.estimatedSerializedSizeOf(cell); 308 } 309 return ret; 310 } 311 312 @Override 313 public String toString() { 314 StringBuilder sb = new StringBuilder(); 315 316 sb.append("[#edits: ").append(cells.size()).append(" = <"); 317 for (Cell cell : cells) { 318 sb.append(cell); 319 sb.append("; "); 320 } 321 sb.append(">]"); 322 return sb.toString(); 323 } 324 325 public static WALEdit createFlushWALEdit(RegionInfo hri, FlushDescriptor f) { 326 KeyValue kv = new KeyValue(getRowForRegion(hri), METAFAMILY, FLUSH, 327 EnvironmentEdgeManager.currentTime(), f.toByteArray()); 328 return new WALEdit().add(kv, METAFAMILY); 329 } 330 331 public static FlushDescriptor getFlushDescriptor(Cell cell) throws IOException { 332 return CellUtil.matchingColumn(cell, METAFAMILY, FLUSH) 333 ? FlushDescriptor.parseFrom(CellUtil.cloneValue(cell)) 334 : null; 335 } 336 337 /** 338 * @return A meta Marker WALEdit that has a single Cell whose value is the passed in 339 * <code>regionEventDesc</code> serialized and whose row is this region, columnfamily is 340 * {@link #METAFAMILY} and qualifier is {@link #REGION_EVENT_PREFIX} + 341 * {@link RegionEventDescriptor#getEventType()}; for example 342 * HBASE::REGION_EVENT::REGION_CLOSE. 343 */ 344 public static WALEdit createRegionEventWALEdit(RegionInfo hri, 345 RegionEventDescriptor regionEventDesc) { 346 return createRegionEventWALEdit(getRowForRegion(hri), regionEventDesc); 347 } 348 349 @InterfaceAudience.Private 350 public static WALEdit createRegionEventWALEdit(byte[] rowForRegion, 351 RegionEventDescriptor regionEventDesc) { 352 KeyValue kv = new KeyValue(rowForRegion, METAFAMILY, 353 createRegionEventDescriptorQualifier(regionEventDesc.getEventType()), 354 EnvironmentEdgeManager.currentTime(), regionEventDesc.toByteArray()); 355 return new WALEdit().add(kv, METAFAMILY); 356 } 357 358 /** 359 * @return Cell qualifier for the passed in RegionEventDescriptor Type; e.g. we'll return 360 * something like a byte array with HBASE::REGION_EVENT::REGION_OPEN in it. 361 */ 362 @InterfaceAudience.Private 363 public static byte[] createRegionEventDescriptorQualifier(RegionEventDescriptor.EventType t) { 364 return Bytes.toBytes(REGION_EVENT_PREFIX_STR + t.toString()); 365 } 366 367 /** 368 * Public so can be accessed from regionserver.wal package. 369 * @return True if this is a Marker Edit and it is a RegionClose type. 370 */ 371 public boolean isRegionCloseMarker() { 372 return isMetaEdit() && PrivateCellUtil.matchingQualifier(this.cells.get(0), REGION_EVENT_CLOSE, 373 0, REGION_EVENT_CLOSE.length); 374 } 375 376 /** 377 * @return Returns a RegionEventDescriptor made by deserializing the content of the passed in 378 * <code>cell</code>, IFF the <code>cell</code> is a RegionEventDescriptor type WALEdit. 379 */ 380 public static RegionEventDescriptor getRegionEventDescriptor(Cell cell) throws IOException { 381 return CellUtil.matchingColumnFamilyAndQualifierPrefix(cell, METAFAMILY, REGION_EVENT_PREFIX) 382 ? RegionEventDescriptor.parseFrom(CellUtil.cloneValue(cell)) 383 : null; 384 } 385 386 /** Returns A Marker WALEdit that has <code>c</code> serialized as its value */ 387 public static WALEdit createCompaction(final RegionInfo hri, final CompactionDescriptor c) { 388 byte[] pbbytes = c.toByteArray(); 389 KeyValue kv = new KeyValue(getRowForRegion(hri), METAFAMILY, COMPACTION, 390 EnvironmentEdgeManager.currentTime(), pbbytes); 391 return new WALEdit().add(kv, METAFAMILY); // replication scope null so this won't be replicated 392 } 393 394 public static byte[] getRowForRegion(RegionInfo hri) { 395 byte[] startKey = hri.getStartKey(); 396 if (startKey.length == 0) { 397 // empty row key is not allowed in mutations because it is both the start key and the end key 398 // we return the smallest byte[] that is bigger (in lex comparison) than byte[0]. 399 return new byte[] { 0 }; 400 } 401 return startKey; 402 } 403 404 /** 405 * Deserialized and returns a CompactionDescriptor is the KeyValue contains one. 406 * @param kv the key value 407 * @return deserialized CompactionDescriptor or null. 408 */ 409 public static CompactionDescriptor getCompaction(Cell kv) throws IOException { 410 return isCompactionMarker(kv) ? CompactionDescriptor.parseFrom(CellUtil.cloneValue(kv)) : null; 411 } 412 413 /** 414 * Returns true if the given cell is a serialized {@link CompactionDescriptor} 415 * @see #getCompaction(Cell) 416 */ 417 public static boolean isCompactionMarker(Cell cell) { 418 return CellUtil.matchingColumn(cell, METAFAMILY, COMPACTION); 419 } 420 421 /** 422 * Create a bulk loader WALEdit 423 * @param hri The RegionInfo for the region in which we are bulk loading 424 * @param bulkLoadDescriptor The descriptor for the Bulk Loader 425 * @return The WALEdit for the BulkLoad 426 */ 427 public static WALEdit createBulkLoadEvent(RegionInfo hri, 428 WALProtos.BulkLoadDescriptor bulkLoadDescriptor) { 429 KeyValue kv = new KeyValue(getRowForRegion(hri), METAFAMILY, BULK_LOAD, 430 EnvironmentEdgeManager.currentTime(), bulkLoadDescriptor.toByteArray()); 431 return new WALEdit().add(kv, METAFAMILY); 432 } 433 434 /** 435 * Deserialized and returns a BulkLoadDescriptor from the passed in Cell 436 * @param cell the key value 437 * @return deserialized BulkLoadDescriptor or null. 438 */ 439 public static WALProtos.BulkLoadDescriptor getBulkLoadDescriptor(Cell cell) throws IOException { 440 return CellUtil.matchingColumn(cell, METAFAMILY, BULK_LOAD) 441 ? WALProtos.BulkLoadDescriptor.parseFrom(CellUtil.cloneValue(cell)) 442 : null; 443 } 444 445 /** 446 * Append the given map of family->edits to a WALEdit data structure. This does not write to the 447 * WAL itself. Note that as an optimization, we will stamp the Set of column families into the 448 * WALEdit to save on our having to calculate column families subsequently down in the actual WAL 449 * writing. 450 * @param familyMap map of family->edits 451 */ 452 public void add(Map<byte[], List<Cell>> familyMap) { 453 for (Map.Entry<byte[], List<Cell>> e : familyMap.entrySet()) { 454 // 'foreach' loop NOT used. See HBASE-12023 "...creates too many iterator objects." 455 int listSize = e.getValue().size(); 456 // Add all Cells first and then at end, add the family rather than call {@link #add(Cell)} 457 // and have it clone family each time. Optimization! 458 for (int i = 0; i < listSize; i++) { 459 addCell(e.getValue().get(i)); 460 } 461 addFamily(e.getKey()); 462 } 463 } 464 465 private void addFamily(byte[] family) { 466 getOrCreateFamilies().add(family); 467 } 468 469 private WALEdit addCell(Cell cell) { 470 this.cells.add(cell); 471 return this; 472 } 473 474 /** 475 * Creates a replication tracker edit with {@link #METAFAMILY} family and 476 * {@link #REPLICATION_MARKER} qualifier and has null value. 477 * @param rowKey rowkey 478 * @param timestamp timestamp 479 */ 480 public static WALEdit createReplicationMarkerEdit(byte[] rowKey, long timestamp) { 481 KeyValue kv = 482 new KeyValue(rowKey, METAFAMILY, REPLICATION_MARKER, timestamp, KeyValue.Type.Put); 483 return new WALEdit().add(kv); 484 } 485 486 /** 487 * Checks whether this edit is a replication marker edit. 488 * @param edit edit 489 * @return true if the cell within an edit has column = METAFAMILY and qualifier = 490 * REPLICATION_MARKER, false otherwise 491 */ 492 public static boolean isReplicationMarkerEdit(WALEdit edit) { 493 // Check just the first cell from the edit. ReplicationMarker edit will have only 1 cell. 494 return edit.getCells().size() == 1 495 && CellUtil.matchingColumn(edit.getCells().get(0), METAFAMILY, REPLICATION_MARKER); 496 } 497}