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.regionserver.wal;
019
020import static org.apache.hadoop.hbase.util.ConcurrentMapUtils.computeIfAbsent;
021
022import java.util.ArrayList;
023import java.util.Collections;
024import java.util.HashMap;
025import java.util.List;
026import java.util.Map;
027import java.util.Set;
028import java.util.TreeMap;
029import java.util.concurrent.ConcurrentHashMap;
030import java.util.concurrent.ConcurrentMap;
031import java.util.stream.Collectors;
032import org.apache.hadoop.hbase.HConstants;
033import org.apache.hadoop.hbase.util.Bytes;
034import org.apache.hadoop.hbase.util.ImmutableByteArray;
035import org.apache.yetus.audience.InterfaceAudience;
036import org.slf4j.Logger;
037import org.slf4j.LoggerFactory;
038
039/**
040 * Accounting of sequence ids per region and then by column family. So we can keep our accounting
041 * current, call startCacheFlush and then finishedCacheFlush or abortCacheFlush so this instance can
042 * keep abreast of the state of sequence id persistence. Also call update per append.
043 * <p>
044 * For the implementation, we assume that all the {@code encodedRegionName} passed in are gotten by
045 * {@link org.apache.hadoop.hbase.client.RegionInfo#getEncodedNameAsBytes()}. So it is safe to use
046 * it as a hash key. And for family name, we use {@link ImmutableByteArray} as key. This is because
047 * hash based map is much faster than RBTree or CSLM and here we are on the critical write path. See
048 * HBASE-16278 for more details.
049 * </p>
050 */
051@InterfaceAudience.Private
052class SequenceIdAccounting {
053  private static final Logger LOG = LoggerFactory.getLogger(SequenceIdAccounting.class);
054
055  /**
056   * This lock ties all operations on {@link SequenceIdAccounting#flushingSequenceIds} and
057   * {@link #lowestUnflushedSequenceIds} Maps. {@link #lowestUnflushedSequenceIds} has the lowest
058   * outstanding sequence ids EXCEPT when flushing. When we flush, the current lowest set for the
059   * region/column family are moved (atomically because of this lock) to
060   * {@link #flushingSequenceIds}.
061   * <p>
062   * The two Maps are tied by this locking object EXCEPT when we go to update the lowest entry; see
063   * {@link #lowestUnflushedSequenceIds}. In here is a putIfAbsent call on
064   * {@link #lowestUnflushedSequenceIds}. In this latter case, we will add this lowest sequence id
065   * if we find that there is no entry for the current column family. There will be no entry only if
066   * we just came up OR we have moved aside current set of lowest sequence ids because the current
067   * set are being flushed (by putting them into {@link #flushingSequenceIds}). This is how we pick
068   * up the next 'lowest' sequence id per region per column family to be used figuring what is in
069   * the next flush.
070   */
071  private final Object tieLock = new Object();
072
073  /**
074   * Map of encoded region names and family names to their OLDEST -- i.e. their first, the
075   * longest-lived, their 'earliest', the 'lowest' -- sequence id.
076   * <p>
077   * When we flush, the current lowest sequence ids get cleared and added to
078   * {@link #flushingSequenceIds}. The next append that comes in, is then added here to
079   * {@link #lowestUnflushedSequenceIds} as the next lowest sequenceid.
080   * <p>
081   * If flush fails, currently server is aborted so no need to restore previous sequence ids.
082   * <p>
083   * Needs to be concurrent Maps because we use putIfAbsent updating oldest.
084   */
085  private final ConcurrentMap<byte[],
086    ConcurrentMap<ImmutableByteArray, Long>> lowestUnflushedSequenceIds = new ConcurrentHashMap<>();
087
088  /**
089   * Map of encoded region names and family names to their lowest or OLDEST sequence/edit id
090   * currently being flushed out to hfiles. Entries are moved here from
091   * {@link #lowestUnflushedSequenceIds} while the lock {@link #tieLock} is held (so movement
092   * between the Maps is atomic).
093   */
094  private final Map<byte[], Map<ImmutableByteArray, Long>> flushingSequenceIds = new HashMap<>();
095
096  /**
097   * <p>
098   * Map of region encoded names to the latest/highest region sequence id. Updated on each call to
099   * append.
100   * </p>
101   * <p>
102   * This map uses byte[] as the key, and uses reference equality. It works in our use case as we
103   * use {@link org.apache.hadoop.hbase.client.RegionInfo#getEncodedNameAsBytes()} as keys. For a
104   * given region, it always returns the same array.
105   * </p>
106   */
107  private Map<byte[], Long> highestSequenceIds = new HashMap<>();
108
109  /**
110   * Returns the lowest unflushed sequence id for the region.
111   * @return Lowest outstanding unflushed sequenceid for <code>encodedRegionName</code>. Will return
112   *         {@link HConstants#NO_SEQNUM} when none.
113   */
114  long getLowestSequenceId(final byte[] encodedRegionName) {
115    synchronized (this.tieLock) {
116      Map<?, Long> m = this.flushingSequenceIds.get(encodedRegionName);
117      long flushingLowest = m != null ? getLowestSequenceId(m) : Long.MAX_VALUE;
118      m = this.lowestUnflushedSequenceIds.get(encodedRegionName);
119      long unflushedLowest = m != null ? getLowestSequenceId(m) : HConstants.NO_SEQNUM;
120      return Math.min(flushingLowest, unflushedLowest);
121    }
122  }
123
124  /**
125   * @return Lowest outstanding unflushed sequenceid for <code>encodedRegionname</code> and
126   *         <code>familyName</code>. Returned sequenceid may be for an edit currently being
127   *         flushed.
128   */
129  long getLowestSequenceId(final byte[] encodedRegionName, final byte[] familyName) {
130    ImmutableByteArray familyNameWrapper = ImmutableByteArray.wrap(familyName);
131    synchronized (this.tieLock) {
132      Map<ImmutableByteArray, Long> m = this.flushingSequenceIds.get(encodedRegionName);
133      if (m != null) {
134        Long lowest = m.get(familyNameWrapper);
135        if (lowest != null) {
136          return lowest;
137        }
138      }
139      m = this.lowestUnflushedSequenceIds.get(encodedRegionName);
140      if (m != null) {
141        Long lowest = m.get(familyNameWrapper);
142        if (lowest != null) {
143          return lowest;
144        }
145      }
146    }
147    return HConstants.NO_SEQNUM;
148  }
149
150  /**
151   * Reset the accounting of highest sequenceid by regionname.
152   * @return Return the previous accounting Map of regions to the last sequence id written into
153   *         each.
154   */
155  Map<byte[], Long> resetHighest() {
156    Map<byte[], Long> old = this.highestSequenceIds;
157    this.highestSequenceIds = new HashMap<>();
158    return old;
159  }
160
161  /**
162   * We've been passed a new sequenceid for the region. Set it as highest seen for this region and
163   * if we are to record oldest, or lowest sequenceids, save it as oldest seen if nothing currently
164   * older. nnn * @param lowest Whether to keep running account of oldest sequence id.
165   */
166  void update(byte[] encodedRegionName, Set<byte[]> families, long sequenceid,
167    final boolean lowest) {
168    Long l = Long.valueOf(sequenceid);
169    this.highestSequenceIds.put(encodedRegionName, l);
170    if (lowest) {
171      ConcurrentMap<ImmutableByteArray, Long> m = getOrCreateLowestSequenceIds(encodedRegionName);
172      for (byte[] familyName : families) {
173        m.putIfAbsent(ImmutableByteArray.wrap(familyName), l);
174      }
175    }
176  }
177
178  /**
179   * Clear all the records of the given region as it is going to be closed.
180   * <p/>
181   * We will call this once we get the region close marker. We need this because that, if we use
182   * Durability.ASYNC_WAL, after calling startCacheFlush, we may still get some ongoing wal entries
183   * that has not been processed yet, this will lead to orphan records in the
184   * lowestUnflushedSequenceIds and then cause too many WAL files.
185   * <p/>
186   * See HBASE-23157 for more details.
187   */
188  void onRegionClose(byte[] encodedRegionName) {
189    synchronized (tieLock) {
190      this.lowestUnflushedSequenceIds.remove(encodedRegionName);
191      Map<ImmutableByteArray, Long> flushing = this.flushingSequenceIds.remove(encodedRegionName);
192      if (flushing != null) {
193        LOG.warn("Still have flushing records when closing {}, {}",
194          Bytes.toString(encodedRegionName),
195          flushing.entrySet().stream().map(e -> e.getKey().toString() + "->" + e.getValue())
196            .collect(Collectors.joining(",", "{", "}")));
197      }
198    }
199    this.highestSequenceIds.remove(encodedRegionName);
200  }
201
202  /**
203   * Update the store sequence id, e.g., upon executing in-memory compaction
204   */
205  void updateStore(byte[] encodedRegionName, byte[] familyName, Long sequenceId,
206    boolean onlyIfGreater) {
207    if (sequenceId == null) {
208      return;
209    }
210    Long highest = this.highestSequenceIds.get(encodedRegionName);
211    if (highest == null || sequenceId > highest) {
212      this.highestSequenceIds.put(encodedRegionName, sequenceId);
213    }
214    ImmutableByteArray familyNameWrapper = ImmutableByteArray.wrap(familyName);
215    synchronized (this.tieLock) {
216      ConcurrentMap<ImmutableByteArray, Long> m = getOrCreateLowestSequenceIds(encodedRegionName);
217      boolean replaced = false;
218      while (!replaced) {
219        Long oldSeqId = m.get(familyNameWrapper);
220        if (oldSeqId == null) {
221          m.put(familyNameWrapper, sequenceId);
222          replaced = true;
223        } else if (onlyIfGreater) {
224          if (sequenceId > oldSeqId) {
225            replaced = m.replace(familyNameWrapper, oldSeqId, sequenceId);
226          } else {
227            return;
228          }
229        } else { // replace even if sequence id is not greater than oldSeqId
230          m.put(familyNameWrapper, sequenceId);
231          return;
232        }
233      }
234    }
235  }
236
237  ConcurrentMap<ImmutableByteArray, Long> getOrCreateLowestSequenceIds(byte[] encodedRegionName) {
238    // Intentionally, this access is done outside of this.regionSequenceIdLock. Done per append.
239    return computeIfAbsent(this.lowestUnflushedSequenceIds, encodedRegionName,
240      ConcurrentHashMap::new);
241  }
242
243  /**
244   * @param sequenceids Map to search for lowest value.
245   * @return Lowest value found in <code>sequenceids</code>.
246   */
247  private static long getLowestSequenceId(Map<?, Long> sequenceids) {
248    long lowest = HConstants.NO_SEQNUM;
249    for (Map.Entry<?, Long> entry : sequenceids.entrySet()) {
250      if (entry.getKey().toString().equals("METAFAMILY")) {
251        continue;
252      }
253      Long sid = entry.getValue();
254      if (lowest == HConstants.NO_SEQNUM || sid.longValue() < lowest) {
255        lowest = sid.longValue();
256      }
257    }
258    return lowest;
259  }
260
261  /**
262   * n * @return New Map that has same keys as <code>src</code> but instead of a Map for a value, it
263   * instead has found the smallest sequence id and it returns that as the value instead.
264   */
265  private <T extends Map<?, Long>> Map<byte[], Long> flattenToLowestSequenceId(Map<byte[], T> src) {
266    if (src == null || src.isEmpty()) {
267      return null;
268    }
269    Map<byte[], Long> tgt = new HashMap<>();
270    for (Map.Entry<byte[], T> entry : src.entrySet()) {
271      long lowestSeqId = getLowestSequenceId(entry.getValue());
272      if (lowestSeqId != HConstants.NO_SEQNUM) {
273        tgt.put(entry.getKey(), lowestSeqId);
274      }
275    }
276    return tgt;
277  }
278
279  /**
280   * @param encodedRegionName Region to flush.
281   * @param families          Families to flush. May be a subset of all families in the region.
282   * @return Returns {@link HConstants#NO_SEQNUM} if we are flushing the whole region OR if we are
283   *         flushing a subset of all families but there are no edits in those families not being
284   *         flushed; in other words, this is effectively same as a flush of all of the region
285   *         though we were passed a subset of regions. Otherwise, it returns the sequence id of the
286   *         oldest/lowest outstanding edit.
287   */
288  Long startCacheFlush(final byte[] encodedRegionName, final Set<byte[]> families) {
289    Map<byte[], Long> familytoSeq = new HashMap<>();
290    for (byte[] familyName : families) {
291      familytoSeq.put(familyName, HConstants.NO_SEQNUM);
292    }
293    return startCacheFlush(encodedRegionName, familytoSeq);
294  }
295
296  Long startCacheFlush(final byte[] encodedRegionName, final Map<byte[], Long> familyToSeq) {
297    Map<ImmutableByteArray, Long> oldSequenceIds = null;
298    Long lowestUnflushedInRegion = HConstants.NO_SEQNUM;
299    synchronized (tieLock) {
300      Map<ImmutableByteArray, Long> m = this.lowestUnflushedSequenceIds.get(encodedRegionName);
301      if (m != null) {
302        // NOTE: Removal from this.lowestUnflushedSequenceIds must be done in controlled
303        // circumstance because another concurrent thread now may add sequenceids for this family
304        // (see above in getOrCreateLowestSequenceId). Make sure you are ok with this. Usually it
305        // is fine because updates are blocked when this method is called. Make sure!!!
306        for (Map.Entry<byte[], Long> entry : familyToSeq.entrySet()) {
307          ImmutableByteArray familyNameWrapper = ImmutableByteArray.wrap((byte[]) entry.getKey());
308          Long seqId = null;
309          if (entry.getValue() == HConstants.NO_SEQNUM) {
310            seqId = m.remove(familyNameWrapper);
311          } else {
312            seqId = m.replace(familyNameWrapper, entry.getValue());
313          }
314          if (seqId != null) {
315            if (oldSequenceIds == null) {
316              oldSequenceIds = new HashMap<>();
317            }
318            oldSequenceIds.put(familyNameWrapper, seqId);
319          }
320        }
321        if (oldSequenceIds != null && !oldSequenceIds.isEmpty()) {
322          if (this.flushingSequenceIds.put(encodedRegionName, oldSequenceIds) != null) {
323            LOG.warn("Flushing Map not cleaned up for " + Bytes.toString(encodedRegionName)
324              + ", sequenceid=" + oldSequenceIds);
325          }
326        }
327        if (m.isEmpty()) {
328          // Remove it otherwise it will be in oldestUnflushedStoreSequenceIds for ever
329          // even if the region is already moved to other server.
330          // Do not worry about data racing, we held write lock of region when calling
331          // startCacheFlush, so no one can add value to the map we removed.
332          this.lowestUnflushedSequenceIds.remove(encodedRegionName);
333        } else {
334          // Flushing a subset of the region families. Return the sequence id of the oldest entry.
335          lowestUnflushedInRegion = Collections.min(m.values());
336        }
337      }
338    }
339    // Do this check outside lock.
340    if (oldSequenceIds != null && oldSequenceIds.isEmpty()) {
341      // TODO: if we have no oldStoreSeqNum, and WAL is not disabled, presumably either
342      // the region is already flushing (which would make this call invalid), or there
343      // were no appends after last flush, so why are we starting flush? Maybe we should
344      // assert not empty. Less rigorous, but safer, alternative is telling the caller to stop.
345      // For now preserve old logic.
346      LOG.warn("Couldn't find oldest sequenceid for " + Bytes.toString(encodedRegionName));
347    }
348    return lowestUnflushedInRegion;
349  }
350
351  void completeCacheFlush(byte[] encodedRegionName, long maxFlushedSeqId) {
352    // This is a simple hack to avoid maxFlushedSeqId go backwards.
353    // The system works fine normally, but if we make use of Durability.ASYNC_WAL and we are going
354    // to flush all the stores, the maxFlushedSeqId will be next seq id of the region, but we may
355    // still have some unsynced WAL entries in the ringbuffer after we call startCacheFlush, and
356    // then it will be recorded as the lowestUnflushedSeqId by the above update method, which is
357    // less than the current maxFlushedSeqId. And if next time we only flush the family with this
358    // unusual lowestUnflushedSeqId, the maxFlushedSeqId will go backwards.
359    // This is an unexpected behavior so we should fix it, otherwise it may cause unexpected
360    // behavior in other area.
361    // The solution here is a bit hack but fine. Just replace the lowestUnflushedSeqId with
362    // maxFlushedSeqId + 1 if it is lesser. The meaning of maxFlushedSeqId is that, all edits less
363    // than or equal to it have been flushed, i.e, persistent to HFile, so set
364    // lowestUnflushedSequenceId to maxFlushedSeqId + 1 will not cause data loss.
365    // And technically, using +1 is fine here. If the maxFlushesSeqId is just the flushOpSeqId, it
366    // means we have flushed all the stores so the seq id for actual data should be at least plus 1.
367    // And if we do not flush all the stores, then the maxFlushedSeqId is calculated by
368    // lowestUnflushedSeqId - 1, so here let's plus the 1 back.
369    Long wrappedSeqId = Long.valueOf(maxFlushedSeqId + 1);
370    synchronized (tieLock) {
371      this.flushingSequenceIds.remove(encodedRegionName);
372      Map<ImmutableByteArray, Long> unflushed = lowestUnflushedSequenceIds.get(encodedRegionName);
373      if (unflushed == null) {
374        return;
375      }
376      for (Map.Entry<ImmutableByteArray, Long> e : unflushed.entrySet()) {
377        if (e.getValue().longValue() <= maxFlushedSeqId) {
378          e.setValue(wrappedSeqId);
379        }
380      }
381    }
382  }
383
384  void abortCacheFlush(final byte[] encodedRegionName) {
385    // Method is called when we are crashing down because failed write flush AND it is called
386    // if we fail prepare. The below is for the fail prepare case; we restore the old sequence ids.
387    Map<ImmutableByteArray, Long> flushing = null;
388    Map<ImmutableByteArray, Long> tmpMap = new HashMap<>();
389    // Here we are moving sequenceids from flushing back to unflushed; doing opposite of what
390    // happened in startCacheFlush. During prepare phase, we have update lock on the region so
391    // no edits should be coming in via append.
392    synchronized (tieLock) {
393      flushing = this.flushingSequenceIds.remove(encodedRegionName);
394      if (flushing != null) {
395        Map<ImmutableByteArray, Long> unflushed = getOrCreateLowestSequenceIds(encodedRegionName);
396        for (Map.Entry<ImmutableByteArray, Long> e : flushing.entrySet()) {
397          // Set into unflushed the 'old' oldest sequenceid and if any value in flushed with this
398          // value, it will now be in tmpMap.
399          tmpMap.put(e.getKey(), unflushed.put(e.getKey(), e.getValue()));
400        }
401      }
402    }
403
404    // Here we are doing some 'test' to see if edits are going in out of order. What is it for?
405    // Carried over from old code.
406    if (flushing != null) {
407      for (Map.Entry<ImmutableByteArray, Long> e : flushing.entrySet()) {
408        Long currentId = tmpMap.get(e.getKey());
409        if (currentId != null && currentId.longValue() < e.getValue().longValue()) {
410          String errorStr = Bytes.toString(encodedRegionName) + " family " + e.getKey().toString()
411            + " acquired edits out of order current memstore seq=" + currentId
412            + ", previous oldest unflushed id=" + e.getValue();
413          LOG.error(errorStr);
414          Runtime.getRuntime().halt(1);
415        }
416      }
417    }
418  }
419
420  /**
421   * See if passed <code>sequenceids</code> are lower -- i.e. earlier -- than any outstanding
422   * sequenceids, sequenceids we are holding on to in this accounting instance.
423   * @param sequenceids Keyed by encoded region name. Cannot be null (doesn't make sense for it to
424   *                    be null).
425   * @return true if all sequenceids are lower, older than, the old sequenceids in this instance.
426   */
427  boolean areAllLower(Map<byte[], Long> sequenceids) {
428    Map<byte[], Long> flushing = null;
429    Map<byte[], Long> unflushed = null;
430    synchronized (this.tieLock) {
431      // Get a flattened -- only the oldest sequenceid -- copy of current flushing and unflushed
432      // data structures to use in tests below.
433      flushing = flattenToLowestSequenceId(this.flushingSequenceIds);
434      unflushed = flattenToLowestSequenceId(this.lowestUnflushedSequenceIds);
435    }
436    for (Map.Entry<byte[], Long> e : sequenceids.entrySet()) {
437      long oldestFlushing = Long.MAX_VALUE;
438      long oldestUnflushed = Long.MAX_VALUE;
439      if (flushing != null && flushing.containsKey(e.getKey())) {
440        oldestFlushing = flushing.get(e.getKey());
441      }
442      if (unflushed != null && unflushed.containsKey(e.getKey())) {
443        oldestUnflushed = unflushed.get(e.getKey());
444      }
445      long min = Math.min(oldestFlushing, oldestUnflushed);
446      if (min <= e.getValue()) {
447        return false;
448      }
449    }
450    return true;
451  }
452
453  /**
454   * Iterates over the given Map and compares sequence ids with corresponding entries in
455   * {@link #lowestUnflushedSequenceIds}. If a region in {@link #lowestUnflushedSequenceIds} has a
456   * sequence id less than that passed in <code>sequenceids</code> then return it.
457   * @param sequenceids Sequenceids keyed by encoded region name.
458   * @return stores of regions found in this instance with sequence ids less than those passed in.
459   */
460  Map<byte[], List<byte[]>> findLower(Map<byte[], Long> sequenceids) {
461    Map<byte[], List<byte[]>> toFlush = null;
462    // Keeping the old behavior of iterating unflushedSeqNums under oldestSeqNumsLock.
463    synchronized (tieLock) {
464      for (Map.Entry<byte[], Long> e : sequenceids.entrySet()) {
465        Map<ImmutableByteArray, Long> m = this.lowestUnflushedSequenceIds.get(e.getKey());
466        if (m == null) {
467          continue;
468        }
469        for (Map.Entry<ImmutableByteArray, Long> me : m.entrySet()) {
470          if (me.getValue() <= e.getValue()) {
471            if (toFlush == null) {
472              toFlush = new TreeMap(Bytes.BYTES_COMPARATOR);
473            }
474            toFlush.computeIfAbsent(e.getKey(), k -> new ArrayList<>())
475              .add(Bytes.toBytes(me.getKey().toString()));
476          }
477        }
478      }
479    }
480    return toFlush;
481  }
482}