Source code

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;
019
020import static org.junit.Assert.assertEquals;
021import static org.junit.Assert.assertNotNull;
022import static org.junit.Assert.assertTrue;
023import static org.junit.Assert.fail;
024import static org.mockito.Mockito.mock;
025import static org.mockito.Mockito.when;
026
027import java.io.IOException;
028import java.util.ArrayList;
029import java.util.Collections;
030import java.util.List;
031import java.util.Random;
032import org.apache.hadoop.conf.Configuration;
033import org.apache.hadoop.fs.FileSystem;
034import org.apache.hadoop.fs.Path;
035import org.apache.hadoop.hbase.CellComparatorImpl;
036import org.apache.hadoop.hbase.CellUtil;
037import org.apache.hadoop.hbase.HBaseClassTestRule;
038import org.apache.hadoop.hbase.HBaseTestingUtility;
039import org.apache.hadoop.hbase.KeyValue;
040import org.apache.hadoop.hbase.KeyValueUtil;
041import org.apache.hadoop.hbase.client.ColumnFamilyDescriptorBuilder;
042import org.apache.hadoop.hbase.client.Scan;
043import org.apache.hadoop.hbase.io.hfile.BlockCache;
044import org.apache.hadoop.hbase.io.hfile.CacheConfig;
045import org.apache.hadoop.hbase.io.hfile.CompoundBloomFilter;
046import org.apache.hadoop.hbase.io.hfile.CompoundBloomFilterWriter;
047import org.apache.hadoop.hbase.io.hfile.HFile;
048import org.apache.hadoop.hbase.io.hfile.HFileContext;
049import org.apache.hadoop.hbase.io.hfile.HFileContextBuilder;
050import org.apache.hadoop.hbase.io.hfile.RandomKeyValueUtil;
051import org.apache.hadoop.hbase.testclassification.MediumTests;
052import org.apache.hadoop.hbase.testclassification.RegionServerTests;
053import org.apache.hadoop.hbase.util.BloomFilterFactory;
054import org.apache.hadoop.hbase.util.BloomFilterUtil;
055import org.apache.hadoop.hbase.util.Bytes;
056import org.junit.Before;
057import org.junit.ClassRule;
058import org.junit.Test;
059import org.junit.experimental.categories.Category;
060import org.slf4j.Logger;
061import org.slf4j.LoggerFactory;
062
063/**
064 * Tests writing Bloom filter blocks in the same part of the file as data
065 * blocks.
066 */
067@Category({RegionServerTests.class, MediumTests.class})
068public class TestCompoundBloomFilter {
069
070  @ClassRule
071  public static final HBaseClassTestRule CLASS_RULE =
072      HBaseClassTestRule.forClass(TestCompoundBloomFilter.class);
073
074  private static final HBaseTestingUtility TEST_UTIL =
075      new HBaseTestingUtility();
076
077  private static final Logger LOG = LoggerFactory.getLogger(
078      TestCompoundBloomFilter.class);
079
080  private static final int NUM_TESTS = 9;
081  private static final BloomType BLOOM_TYPES[] = { BloomType.ROW,
082      BloomType.ROW, BloomType.ROWCOL, BloomType.ROWCOL, BloomType.ROW,
083      BloomType.ROWCOL, BloomType.ROWCOL, BloomType.ROWCOL, BloomType.ROW };
084
085  private static final int NUM_KV[];
086  static {
087    final int N = 10000; // Only used in initialization.
088    NUM_KV = new int[] { 21870, N, N, N, N, 1000, N, 7500, 7500};
089    assert NUM_KV.length == NUM_TESTS;
090  }
091
092  private static final int BLOCK_SIZES[];
093  static {
094    final int blkSize = 65536;
095    BLOCK_SIZES = new int[] { 512, 1000, blkSize, blkSize, blkSize, 128, 300,
096        blkSize, blkSize };
097    assert BLOCK_SIZES.length == NUM_TESTS;
098  }
099
100  /**
101   * Be careful not to specify too high a Bloom filter block size, otherwise
102   * there will only be one oversized chunk and the observed false positive
103   * rate will be too low.
104   */
105  private static final int BLOOM_BLOCK_SIZES[] = { 1000, 4096, 4096, 4096,
106      8192, 128, 1024, 600, 600 };
107  static { assert BLOOM_BLOCK_SIZES.length == NUM_TESTS; }
108
109  private static final double TARGET_ERROR_RATES[] = { 0.025, 0.01, 0.015,
110      0.01, 0.03, 0.01, 0.01, 0.07, 0.07 };
111  static { assert TARGET_ERROR_RATES.length == NUM_TESTS; }
112
113  /** A false positive rate that is obviously too high. */
114  private static final double TOO_HIGH_ERROR_RATE;
115  static {
116    double m = 0;
117    for (double errorRate : TARGET_ERROR_RATES)
118      m = Math.max(m, errorRate);
119    TOO_HIGH_ERROR_RATE = m + 0.03;
120  }
121
122  private static Configuration conf;
123  private static CacheConfig cacheConf;
124  private FileSystem fs;
125  private BlockCache blockCache;
126
127  /**
128   * A message of the form <code>in test#&lt;number&gt;</code>" to include in logging.
129   */
130  private String testIdMsg;
131
132  private static final int GENERATION_SEED = 2319;
133  private static final int EVALUATION_SEED = 135;
134
135  @Before
136  public void setUp() throws IOException {
137    conf = TEST_UTIL.getConfiguration();
138
139    // This test requires the most recent HFile format (i.e. v2).
140    conf.setInt(HFile.FORMAT_VERSION_KEY, HFile.MAX_FORMAT_VERSION);
141
142    fs = FileSystem.get(conf);
143
144    cacheConf = new CacheConfig(conf);
145    blockCache = cacheConf.getBlockCache();
146    assertNotNull(blockCache);
147  }
148
149  private List<KeyValue> createSortedKeyValues(Random rand, int n) {
150    List<KeyValue> kvList = new ArrayList<>(n);
151    for (int i = 0; i < n; ++i)
152      kvList.add(RandomKeyValueUtil.randomKeyValue(rand));
153    Collections.sort(kvList, CellComparatorImpl.COMPARATOR);
154    return kvList;
155  }
156
157  @Test
158  public void testCompoundBloomFilter() throws IOException {
159    conf.setBoolean(BloomFilterFactory.IO_STOREFILE_BLOOM_ENABLED, true);
160    for (int t = 0; t < NUM_TESTS; ++t) {
161      conf.setFloat(BloomFilterFactory.IO_STOREFILE_BLOOM_ERROR_RATE,
162          (float) TARGET_ERROR_RATES[t]);
163
164      testIdMsg = "in test #" + t + ":";
165      Random generationRand = new Random(GENERATION_SEED);
166      List<KeyValue> kvs = createSortedKeyValues(generationRand, NUM_KV[t]);
167      BloomType bt = BLOOM_TYPES[t];
168      Path sfPath = writeStoreFile(t, bt, kvs);
169      readStoreFile(t, bt, kvs, sfPath);
170    }
171  }
172
173  /**
174   * Validates the false positive ratio by computing its z-value and comparing
175   * it to the provided threshold.
176   *
177   * @param falsePosRate experimental positive rate
178   * @param nTrials the number of Bloom filter checks
179   * @param zValueBoundary z-value boundary, positive for an upper bound and
180   *          negative for a lower bound
181   * @param cbf the compound Bloom filter we are using
182   * @param additionalMsg additional message to include in log output and
183   *          assertion failures
184   */
185  private void validateFalsePosRate(double falsePosRate, int nTrials,
186      double zValueBoundary, CompoundBloomFilter cbf, String additionalMsg) {
187    double p = BloomFilterFactory.getErrorRate(conf);
188    double zValue = (falsePosRate - p) / Math.sqrt(p * (1 - p) / nTrials);
189
190    String assortedStatsStr = " (targetErrorRate=" + p + ", falsePosRate="
191        + falsePosRate + ", nTrials=" + nTrials + ")";
192    LOG.info("z-value is " + zValue + assortedStatsStr);
193
194    boolean isUpperBound = zValueBoundary > 0;
195
196    if (isUpperBound && zValue > zValueBoundary ||
197        !isUpperBound && zValue < zValueBoundary) {
198      String errorMsg = "False positive rate z-value " + zValue + " is "
199          + (isUpperBound ? "higher" : "lower") + " than " + zValueBoundary
200          + assortedStatsStr + ". Per-chunk stats:\n"
201          + cbf.formatTestingStats();
202      fail(errorMsg + additionalMsg);
203    }
204  }
205
206  private void readStoreFile(int t, BloomType bt, List<KeyValue> kvs,
207      Path sfPath) throws IOException {
208    HStoreFile sf = new HStoreFile(fs, sfPath, conf, cacheConf, bt, true);
209    sf.initReader();
210    StoreFileReader r = sf.getReader();
211    final boolean pread = true; // does not really matter
212    StoreFileScanner scanner = r.getStoreFileScanner(true, pread, false, 0, 0, false);
213
214    {
215      // Test for false negatives (not allowed).
216      int numChecked = 0;
217      for (KeyValue kv : kvs) {
218        byte[] row = CellUtil.cloneRow(kv);
219        boolean present = isInBloom(scanner, row, CellUtil.cloneQualifier(kv));
220        assertTrue(testIdMsg + " Bloom filter false negative on row "
221            + Bytes.toStringBinary(row) + " after " + numChecked
222            + " successful checks", present);
223        ++numChecked;
224      }
225    }
226
227    // Test for false positives (some percentage allowed). We test in two modes:
228    // "fake lookup" which ignores the key distribution, and production mode.
229    for (boolean fakeLookupEnabled : new boolean[] { true, false }) {
230      BloomFilterUtil.setFakeLookupMode(fakeLookupEnabled);
231      try {
232        String fakeLookupModeStr = ", fake lookup is " + (fakeLookupEnabled ?
233            "enabled" : "disabled");
234        CompoundBloomFilter cbf = (CompoundBloomFilter) r.getGeneralBloomFilter();
235        cbf.enableTestingStats();
236        int numFalsePos = 0;
237        Random rand = new Random(EVALUATION_SEED);
238        int nTrials = NUM_KV[t] * 10;
239        for (int i = 0; i < nTrials; ++i) {
240          byte[] query = RandomKeyValueUtil.randomRowOrQualifier(rand);
241          if (isInBloom(scanner, query, bt, rand)) {
242            numFalsePos += 1;
243          }
244        }
245        double falsePosRate = numFalsePos * 1.0 / nTrials;
246        LOG.debug(String.format(testIdMsg
247            + " False positives: %d out of %d (%f)",
248            numFalsePos, nTrials, falsePosRate) + fakeLookupModeStr);
249
250        // Check for obvious Bloom filter crashes.
251        assertTrue("False positive is too high: " + falsePosRate + " (greater "
252            + "than " + TOO_HIGH_ERROR_RATE + ")" + fakeLookupModeStr,
253            falsePosRate < TOO_HIGH_ERROR_RATE);
254
255        // Now a more precise check to see if the false positive rate is not
256        // too high. The reason we use a relaxed restriction for the real-world
257        // case as opposed to the "fake lookup" case is that our hash functions
258        // are not completely independent.
259
260        double maxZValue = fakeLookupEnabled ? 1.96 : 2.5;
261        validateFalsePosRate(falsePosRate, nTrials, maxZValue, cbf,
262            fakeLookupModeStr);
263
264        // For checking the lower bound we need to eliminate the last chunk,
265        // because it is frequently smaller and the false positive rate in it
266        // is too low. This does not help if there is only one under-sized
267        // chunk, though.
268        int nChunks = cbf.getNumChunks();
269        if (nChunks > 1) {
270          numFalsePos -= cbf.getNumPositivesForTesting(nChunks - 1);
271          nTrials -= cbf.getNumQueriesForTesting(nChunks - 1);
272          falsePosRate = numFalsePos * 1.0 / nTrials;
273          LOG.info(testIdMsg + " False positive rate without last chunk is " +
274              falsePosRate + fakeLookupModeStr);
275        }
276
277        validateFalsePosRate(falsePosRate, nTrials, -2.58, cbf,
278            fakeLookupModeStr);
279      } finally {
280        BloomFilterUtil.setFakeLookupMode(false);
281      }
282    }
283
284    r.close(true); // end of test so evictOnClose
285  }
286
287  private boolean isInBloom(StoreFileScanner scanner, byte[] row, BloomType bt,
288      Random rand) {
289    return isInBloom(scanner, row, RandomKeyValueUtil.randomRowOrQualifier(rand));
290  }
291
292  private boolean isInBloom(StoreFileScanner scanner, byte[] row,
293      byte[] qualifier) {
294    Scan scan = new Scan().withStartRow(row).withStopRow(row, true);
295    scan.addColumn(Bytes.toBytes(RandomKeyValueUtil.COLUMN_FAMILY_NAME), qualifier);
296    HStore store = mock(HStore.class);
297    when(store.getColumnFamilyDescriptor())
298        .thenReturn(ColumnFamilyDescriptorBuilder.of(RandomKeyValueUtil.COLUMN_FAMILY_NAME));
299    return scanner.shouldUseScanner(scan, store, Long.MIN_VALUE);
300  }
301
302  private Path writeStoreFile(int t, BloomType bt, List<KeyValue> kvs)
303      throws IOException {
304    conf.setInt(BloomFilterFactory.IO_STOREFILE_BLOOM_BLOCK_SIZE,
305        BLOOM_BLOCK_SIZES[t]);
306    conf.setBoolean(CacheConfig.CACHE_BLOCKS_ON_WRITE_KEY, true);
307    cacheConf = new CacheConfig(conf);
308    HFileContext meta = new HFileContextBuilder().withBlockSize(BLOCK_SIZES[t]).build();
309    StoreFileWriter w = new StoreFileWriter.Builder(conf, cacheConf, fs)
310            .withOutputDir(TEST_UTIL.getDataTestDir())
311            .withBloomType(bt)
312            .withFileContext(meta)
313            .build();
314
315    assertTrue(w.hasGeneralBloom());
316    assertTrue(w.getGeneralBloomWriter() instanceof CompoundBloomFilterWriter);
317    CompoundBloomFilterWriter cbbf =
318        (CompoundBloomFilterWriter) w.getGeneralBloomWriter();
319
320    int keyCount = 0;
321    KeyValue prev = null;
322    LOG.debug("Total keys/values to insert: " + kvs.size());
323    for (KeyValue kv : kvs) {
324      w.append(kv);
325
326      // Validate the key count in the Bloom filter.
327      boolean newKey = true;
328      if (prev != null) {
329        newKey = !(bt == BloomType.ROW ? CellUtil.matchingRows(kv,
330            prev) : CellUtil.matchingRowColumn(kv, prev));
331      }
332      if (newKey)
333        ++keyCount;
334      assertEquals(keyCount, cbbf.getKeyCount());
335
336      prev = kv;
337    }
338    w.close();
339
340    return w.getPath();
341  }
342
343  @Test
344  public void testCompoundBloomSizing() {
345    int bloomBlockByteSize = 4096;
346    int bloomBlockBitSize = bloomBlockByteSize * 8;
347    double targetErrorRate = 0.01;
348    long maxKeysPerChunk = BloomFilterUtil.idealMaxKeys(bloomBlockBitSize,
349        targetErrorRate);
350
351    long bloomSize1 = bloomBlockByteSize * 8;
352    long bloomSize2 = BloomFilterUtil.computeBitSize(maxKeysPerChunk,
353        targetErrorRate);
354
355    double bloomSizeRatio = (bloomSize2 * 1.0 / bloomSize1);
356    assertTrue(Math.abs(bloomSizeRatio - 0.9999) < 0.0001);
357  }
358
359  @Test
360  public void testCreateKey() {
361    byte[] row = "myRow".getBytes();
362    byte[] qualifier = "myQualifier".getBytes();
363    // Mimic what Storefile.createBloomKeyValue() does
364    byte[] rowKey = KeyValueUtil.createFirstOnRow(row, 0, row.length, new byte[0], 0, 0, row, 0, 0).getKey();
365    byte[] rowColKey = KeyValueUtil.createFirstOnRow(row, 0, row.length,
366        new byte[0], 0, 0, qualifier, 0, qualifier.length).getKey();
367    KeyValue rowKV = KeyValueUtil.createKeyValueFromKey(rowKey);
368    KeyValue rowColKV = KeyValueUtil.createKeyValueFromKey(rowColKey);
369    assertEquals(rowKV.getTimestamp(), rowColKV.getTimestamp());
370    assertEquals(Bytes.toStringBinary(rowKV.getRowArray(), rowKV.getRowOffset(),
371      rowKV.getRowLength()), Bytes.toStringBinary(rowColKV.getRowArray(), rowColKV.getRowOffset(),
372      rowColKV.getRowLength()));
373    assertEquals(0, rowKV.getQualifierLength());
374  }
375
376
377}
378