Apache
/
hbase
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

HBASE-16288 HFile intermediate block level indexes might recurse forever creating multi TB files

This commit is contained in:
Enis Soztutar 2016-07-29 19:19:20 -07:00
parent 2c5a0fcf1f
commit aa0235f98c
3 changed files with 93 additions and 10 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

45
hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/HFileBlockIndex.java
View File

@ -76,6 +76,16 @@ public class HFileBlockIndex {
*/
public static final String MAX_CHUNK_SIZE_KEY = "hfile.index.block.max.size";
/**
* Minimum number of entries in a single index block. Even if we are above the
* hfile.index.block.max.size we will keep writing to the same block unless we have that many
* entries. We should have at least a few entries so that we don't have too many levels in the
* multi-level index. This should be at least 2 to make sure there is no infinite recursion.
*/
public static final String MIN_INDEX_NUM_ENTRIES_KEY = "hfile.index.block.min.entries";
static final int DEFAULT_MIN_INDEX_NUM_ENTRIES = 16;
/**
* The number of bytes stored in each "secondary index" entry in addition to
* key bytes in the non-root index block format. The first long is the file
@ -120,6 +130,7 @@ public class HFileBlockIndex {
searchTreeLevel = treeLevel;
}
@Override
protected long calculateHeapSizeForBlockKeys(long heapSize) {
// Calculating the size of blockKeys
if (blockKeys != null) {
@ -162,10 +173,12 @@ public class HFileBlockIndex {
return null;
}
@Override
protected void initialize(int numEntries) {
blockKeys = new byte[numEntries][];
}
@Override
protected void add(final byte[] key, final long offset, final int dataSize) {
blockOffsets[rootCount] = offset;
blockKeys[rootCount] = key;
@ -242,6 +255,7 @@ public class HFileBlockIndex {
searchTreeLevel = treeLevel;
}
@Override
protected long calculateHeapSizeForBlockKeys(long heapSize) {
if (blockKeys != null) {
heapSize += ClassSize.REFERENCE;
@ -430,6 +444,7 @@ public class HFileBlockIndex {
return targetMidKey;
}
@Override
protected void initialize(int numEntries) {
blockKeys = new Cell[numEntries];
}
@ -441,6 +456,7 @@ public class HFileBlockIndex {
* @param offset file offset where the block is stored
* @param dataSize the uncompressed data size
*/
@Override
protected void add(final byte[] key, final long offset, final int dataSize) {
blockOffsets[rootCount] = offset;
// Create the blockKeys as Cells once when the reader is opened
@ -985,6 +1001,9 @@ public class HFileBlockIndex {
/** The maximum size guideline of all multi-level index blocks. */
private int maxChunkSize;
/** The maximum level of multi-level index blocks */
private int minIndexNumEntries;
/** Whether we require this block index to always be single-level. */
private boolean singleLevelOnly;
@ -1017,15 +1036,23 @@ public class HFileBlockIndex {
this.cacheConf = cacheConf;
this.nameForCaching = nameForCaching;
this.maxChunkSize = HFileBlockIndex.DEFAULT_MAX_CHUNK_SIZE;
this.minIndexNumEntries = HFileBlockIndex.DEFAULT_MIN_INDEX_NUM_ENTRIES;
}
public void setMaxChunkSize(int maxChunkSize) {
if (maxChunkSize <= 0) {
throw new IllegalArgumentException("Invald maximum index block size");
throw new IllegalArgumentException("Invalid maximum index block size");
}
this.maxChunkSize = maxChunkSize;
}
public void setMinIndexNumEntries(int minIndexNumEntries) {
if (minIndexNumEntries <= 1) {
throw new IllegalArgumentException("Invalid maximum index level, should be >= 2");
}
this.minIndexNumEntries = minIndexNumEntries;
}
/**
* Writes the root level and intermediate levels of the block index into
* the output stream, generating the tree from bottom up. Assumes that the
@ -1057,7 +1084,11 @@ public class HFileBlockIndex {
: null;
if (curInlineChunk != null) {
while (rootChunk.getRootSize() > maxChunkSize) {
while (rootChunk.getRootSize() > maxChunkSize
// HBASE-16288: if firstKey is larger than maxChunkSize we will loop indefinitely
&& rootChunk.getNumEntries() > minIndexNumEntries
// Sanity check. We will not hit this (minIndexNumEntries ^ 16) blocks can be addressed
&& numLevels < 16) {
rootChunk = writeIntermediateLevel(out, rootChunk);
numLevels += 1;
}
@ -1153,9 +1184,13 @@ public class HFileBlockIndex {
curChunk.add(currentLevel.getBlockKey(i),
currentLevel.getBlockOffset(i), currentLevel.getOnDiskDataSize(i));
if (curChunk.getRootSize() >= maxChunkSize)
// HBASE-16288: We have to have at least minIndexNumEntries(16) items in the index so that
// we won't end up with too-many levels for a index with very large rowKeys. Also, if the
// first key is larger than maxChunkSize this will cause infinite recursion.
if (i >= minIndexNumEntries && curChunk.getRootSize() >= maxChunkSize) {
writeIntermediateBlock(out, parent, curChunk);
}
}
if (curChunk.getNumEntries() > 0) {
writeIntermediateBlock(out, parent, curChunk);
@ -1647,4 +1682,8 @@ public class HFileBlockIndex {
public static int getMaxChunkSize(Configuration conf) {
return conf.getInt(MAX_CHUNK_SIZE_KEY, DEFAULT_MAX_CHUNK_SIZE);
}
public static int getMinIndexNumEntries(Configuration conf) {
return conf.getInt(MIN_INDEX_NUM_ENTRIES_KEY, DEFAULT_MIN_INDEX_NUM_ENTRIES);
}
}

2
hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/HFileWriterImpl.java
View File

@ -287,6 +287,8 @@ public class HFileWriterImpl implements HFile.Writer {
cacheIndexesOnWrite ? name : null);
dataBlockIndexWriter.setMaxChunkSize(
HFileBlockIndex.getMaxChunkSize(conf));
dataBlockIndexWriter.setMinIndexNumEntries(
HFileBlockIndex.getMinIndexNumEntries(conf));
inlineBlockWriters.add(dataBlockIndexWriter);
// Meta data block index writer

44
hbase-server/src/test/java/org/apache/hadoop/hbase/io/hfile/TestHFileBlockIndex.java
View File

@ -671,6 +671,48 @@ public class TestHFileBlockIndex {
valueRead);
}
@Test(timeout=10000)
public void testIntermediateLevelIndicesWithLargeKeys() throws IOException {
testIntermediateLevelIndicesWithLargeKeys(16);
}
@Test(timeout=10000)
public void testIntermediateLevelIndicesWithLargeKeysWithMinNumEntries() throws IOException {
// because of the large rowKeys, we will end up with a 50-level block index without sanity check
testIntermediateLevelIndicesWithLargeKeys(2);
}
public void testIntermediateLevelIndicesWithLargeKeys(int minNumEntries) throws IOException {
Path hfPath = new Path(TEST_UTIL.getDataTestDir(),
"testIntermediateLevelIndicesWithLargeKeys.hfile");
int maxChunkSize = 1024;
FileSystem fs = FileSystem.get(conf);
CacheConfig cacheConf = new CacheConfig(conf);
conf.setInt(HFileBlockIndex.MAX_CHUNK_SIZE_KEY, maxChunkSize);
conf.setInt(HFileBlockIndex.MIN_INDEX_NUM_ENTRIES_KEY, minNumEntries);
HFileContext context = new HFileContextBuilder().withBlockSize(16).build();
HFile.Writer hfw = new HFile.WriterFactory(conf, cacheConf)
.withFileContext(context)
.withPath(fs, hfPath).create();
List<byte[]> keys = new ArrayList<byte[]>();
// This should result in leaf-level indices and a root level index
for (int i=0; i < 100; i++) {
byte[] rowkey = new byte[maxChunkSize + 1];
byte[] b = Bytes.toBytes(i);
System.arraycopy(b, 0, rowkey, rowkey.length - b.length, b.length);
keys.add(rowkey);
hfw.append(CellUtil.createCell(rowkey));
}
hfw.close();
HFile.Reader reader = HFile.createReader(fs, hfPath, cacheConf, conf);
// Scanner doesn't do Cells yet. Fix.
HFileScanner scanner = reader.getScanner(true, true);
for (int i = 0; i < keys.size(); ++i) {
scanner.seekTo(CellUtil.createCell(keys.get(i)));
}
reader.close();
}
}