HBASE-28596 Optimise BucketCache usage upon regions splits/merges. (#…

…5906)

Signed-off-by: Tak Lon (Stephen) Wu <taklwu@apache.org>
Reviewed0by: Duo Zhang <zhangduo@apache.org>

hbase-server/src/main/java/org/apache/hadoop/hbase/io/HalfStoreFileReader.java

@@ -20,6 +20,7 @@
 import java.io.IOException;
 import java.nio.ByteBuffer;
 import java.util.Optional;
+import java.util.concurrent.atomic.AtomicBoolean;
 import java.util.function.IntConsumer;
 import org.apache.hadoop.conf.Configuration;
 import org.apache.hadoop.hbase.Cell;
@@ -29,6 +30,7 @@
 import org.apache.hadoop.hbase.client.Scan;
 import org.apache.hadoop.hbase.io.hfile.CacheConfig;
 import org.apache.hadoop.hbase.io.hfile.HFileInfo;
+import org.apache.hadoop.hbase.io.hfile.HFileReaderImpl;
 import org.apache.hadoop.hbase.io.hfile.HFileScanner;
 import org.apache.hadoop.hbase.io.hfile.ReaderContext;
 import org.apache.hadoop.hbase.regionserver.StoreFileInfo;
@@ -64,6 +66,8 @@ public class HalfStoreFileReader extends StoreFileReader {
 
   private boolean firstKeySeeked = false;
 
+  private AtomicBoolean closed = new AtomicBoolean(false);
+
   /**
    * Creates a half file reader for a hfile referred to by an hfilelink.
    * @param context   Reader context info
@@ -335,4 +339,42 @@ public long getFilterEntries() {
     // Estimate the number of entries as half the original file; this may be wildly inaccurate.
     return super.getFilterEntries() / 2;
   }
+
+  /**
+   * Overrides close method to handle cache evictions for the referred file. If evictionOnClose is
+   * true, we will seek to the block containing the splitCell and evict all blocks from offset 0 up
+   * to that block offset if this is a bottom half reader, or the from the split block offset up to
+   * the end of the file if this is a top half reader.
+   * @param evictOnClose true if it should evict the file blocks from the cache.
+   */
+  @Override
+  public void close(boolean evictOnClose) throws IOException {
+    if (closed.compareAndSet(false, true)) {
+      if (evictOnClose) {
+        final HFileReaderImpl.HFileScannerImpl s =
+          (HFileReaderImpl.HFileScannerImpl) super.getScanner(false, true, false);
+        final String reference = this.reader.getHFileInfo().getHFileContext().getHFileName();
+        final String referred = StoreFileInfo.getReferredToRegionAndFile(reference).getSecond();
+        s.seekTo(splitCell);
+        if (s.getCurBlock() != null) {
+          long offset = s.getCurBlock().getOffset();
+          LOG.trace("Seeking to split cell in reader: {} for file: {} top: {}, split offset: {}",
+            this, reference, top, offset);
+          ((HFileReaderImpl) reader).getCacheConf().getBlockCache().ifPresent(cache -> {
+            int numEvictedReferred = top
+              ? cache.evictBlocksRangeByHfileName(referred, offset, Long.MAX_VALUE)
+              : cache.evictBlocksRangeByHfileName(referred, 0, offset);
+            int numEvictedReference = cache.evictBlocksByHfileName(reference);
+            LOG.trace(
+              "Closing reference: {}; referred file: {}; was top? {}; evicted for referred: {};"
+                + "evicted for reference: {}",
+              reference, referred, top, numEvictedReferred, numEvictedReference);
+          });
+        }
+        reader.close(false);
+      } else {
+        reader.close(evictOnClose);
+      }
+    }
+  }
 }

hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/BlockCache.java

@@ -250,4 +250,15 @@ default Optional<Map<String, Pair<String, Long>>> getFullyCachedFiles() {
   default Optional<Map<String, Long>> getRegionCachedInfo() {
     return Optional.empty();
   }
+
+  /**
+   * Evict all blocks for the given file name between the passed offset values.
+   * @param hfileName  The file for which blocks should be evicted.
+   * @param initOffset the initial offset for the range of blocks to be evicted.
+   * @param endOffset  the end offset for the range of blocks to be evicted.
+   * @return number of blocks evicted.
+   */
+  default int evictBlocksRangeByHfileName(String hfileName, long initOffset, long endOffset) {
+    return 0;
+  }
 }

hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/BlockCacheUtil.java

@@ -17,6 +17,8 @@
  */
 package org.apache.hadoop.hbase.io.hfile;
 
+import static org.apache.hadoop.hbase.io.hfile.HFileBlock.FILL_HEADER;
+
 import java.io.IOException;
 import java.nio.ByteBuffer;
 import java.util.NavigableMap;
@@ -25,7 +27,9 @@
 import java.util.concurrent.ConcurrentSkipListSet;
 import org.apache.hadoop.conf.Configuration;
 import org.apache.hadoop.hbase.metrics.impl.FastLongHistogram;
+import org.apache.hadoop.hbase.nio.ByteBuff;
 import org.apache.hadoop.hbase.util.Bytes;
+import org.apache.hadoop.hbase.util.ChecksumType;
 import org.apache.hadoop.hbase.util.GsonUtil;
 import org.apache.yetus.audience.InterfaceAudience;
 import org.slf4j.Logger;
@@ -244,6 +248,44 @@ public static int getMaxCachedBlocksByFile(Configuration conf) {
     return conf == null ? DEFAULT_MAX : conf.getInt("hbase.ui.blockcache.by.file.max", DEFAULT_MAX);
   }
 
+  /**
+   * Similarly to HFileBlock.Writer.getBlockForCaching(), creates a HFileBlock instance without
+   * checksum for caching. This is needed for when we cache blocks via readers (either prefetch or
+   * client read), otherwise we may fail equality comparison when checking against same block that
+   * may already have been cached at write time.
+   * @param cacheConf the related CacheConfig object.
+   * @param block     the HFileBlock instance to be converted.
+   * @return the resulting HFileBlock instance without checksum.
+   */
+  public static HFileBlock getBlockForCaching(CacheConfig cacheConf, HFileBlock block) {
+    // Calculate how many bytes we need for checksum on the tail of the block.
+    int numBytes = cacheConf.shouldCacheCompressed(block.getBlockType().getCategory())
+      ? 0
+      : (int) ChecksumUtil.numBytes(block.getOnDiskDataSizeWithHeader(),
+        block.getHFileContext().getBytesPerChecksum());
+    ByteBuff buff = block.getBufferReadOnly();
+    HFileBlockBuilder builder = new HFileBlockBuilder();
+    return builder.withBlockType(block.getBlockType())
+      .withOnDiskSizeWithoutHeader(block.getOnDiskSizeWithoutHeader())
+      .withUncompressedSizeWithoutHeader(block.getUncompressedSizeWithoutHeader())
+      .withPrevBlockOffset(block.getPrevBlockOffset()).withByteBuff(buff)
+      .withFillHeader(FILL_HEADER).withOffset(block.getOffset()).withNextBlockOnDiskSize(-1)
+      .withOnDiskDataSizeWithHeader(block.getOnDiskDataSizeWithHeader() + numBytes)
+      .withHFileContext(cloneContext(block.getHFileContext()))
+      .withByteBuffAllocator(cacheConf.getByteBuffAllocator()).withShared(!buff.hasArray()).build();
+  }
+
+  public static HFileContext cloneContext(HFileContext context) {
+    HFileContext newContext = new HFileContextBuilder().withBlockSize(context.getBlocksize())
+      .withBytesPerCheckSum(0).withChecksumType(ChecksumType.NULL) // no checksums in cached data
+      .withCompression(context.getCompression())
+      .withDataBlockEncoding(context.getDataBlockEncoding())
+      .withHBaseCheckSum(context.isUseHBaseChecksum()).withCompressTags(context.isCompressTags())
+      .withIncludesMvcc(context.isIncludesMvcc()).withIncludesTags(context.isIncludesTags())
+      .withColumnFamily(context.getColumnFamily()).withTableName(context.getTableName()).build();
+    return newContext;
+  }
+
   /**
    * Use one of these to keep a running account of cached blocks by file. Throw it away when done.
    * This is different than metrics in that it is stats on current state of a cache. See

hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/CacheConfig.java

@@ -72,6 +72,8 @@ public class CacheConfig implements ConfigurationObserver {
    */
   public static final String EVICT_BLOCKS_ON_CLOSE_KEY = "hbase.rs.evictblocksonclose";
 
+  public static final String EVICT_BLOCKS_ON_SPLIT_KEY = "hbase.rs.evictblocksonsplit";
+
   /**
    * Configuration key to prefetch all blocks of a given file into the block cache when the file is
    * opened.
@@ -113,6 +115,7 @@ public class CacheConfig implements ConfigurationObserver {
   public static final boolean DEFAULT_CACHE_INDEXES_ON_WRITE = false;
   public static final boolean DEFAULT_CACHE_BLOOMS_ON_WRITE = false;
   public static final boolean DEFAULT_EVICT_ON_CLOSE = false;
+  public static final boolean DEFAULT_EVICT_ON_SPLIT = true;
   public static final boolean DEFAULT_CACHE_DATA_COMPRESSED = false;
   public static final boolean DEFAULT_PREFETCH_ON_OPEN = false;
   public static final boolean DEFAULT_CACHE_COMPACTED_BLOCKS_ON_WRITE = false;

hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/CombinedBlockCache.java

@@ -109,8 +109,6 @@ public Cacheable getBlock(BlockCacheKey cacheKey, boolean caching, boolean repea
         }
       } else {
         if (existInL1) {
-          LOG.warn("Cache key {} had block type {}, but was found in L1 cache.", cacheKey,
-            cacheKey.getBlockType());
           updateBlockMetrics(block, cacheKey, l1Cache, caching);
         } else {
           updateBlockMetrics(block, cacheKey, l2Cache, caching);
@@ -504,4 +502,9 @@ public Optional<Integer> getBlockSize(BlockCacheKey key) {
     return l1Result.isPresent() ? l1Result : l2Cache.getBlockSize(key);
   }
 
+  @Override
+  public int evictBlocksRangeByHfileName(String hfileName, long initOffset, long endOffset) {
+    return l1Cache.evictBlocksRangeByHfileName(hfileName, initOffset, endOffset)
+      + l2Cache.evictBlocksRangeByHfileName(hfileName, initOffset, endOffset);
+  }
 }

hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/HFileBlock.java

@@ -697,7 +697,7 @@ public boolean isUnpacked() {
    * when block is returned to the cache.
    * @return the offset of this block in the file it was read from
    */
-  long getOffset() {
+  public long getOffset() {
     if (offset < 0) {
       throw new IllegalStateException("HFile block offset not initialized properly");
     }
@@ -1205,16 +1205,7 @@ void writeBlock(BlockWritable bw, FSDataOutputStream out) throws IOException {
      * being wholesome (ECC memory or if file-backed, it does checksumming).
      */
     HFileBlock getBlockForCaching(CacheConfig cacheConf) {
-      HFileContext newContext = new HFileContextBuilder().withBlockSize(fileContext.getBlocksize())
-        .withBytesPerCheckSum(0).withChecksumType(ChecksumType.NULL) // no checksums in cached data
-        .withCompression(fileContext.getCompression())
-        .withDataBlockEncoding(fileContext.getDataBlockEncoding())
-        .withHBaseCheckSum(fileContext.isUseHBaseChecksum())
-        .withCompressTags(fileContext.isCompressTags())
-        .withIncludesMvcc(fileContext.isIncludesMvcc())
-        .withIncludesTags(fileContext.isIncludesTags())
-        .withColumnFamily(fileContext.getColumnFamily()).withTableName(fileContext.getTableName())
-        .build();
+      HFileContext newContext = BlockCacheUtil.cloneContext(fileContext);
       // Build the HFileBlock.
       HFileBlockBuilder builder = new HFileBlockBuilder();
       ByteBuff buff;

hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/HFilePreadReader.java

@@ -45,7 +45,7 @@ public HFilePreadReader(ReaderContext context, HFileInfo fileInfo, CacheConfig c
     });
 
     // Prefetch file blocks upon open if requested
-    if (cacheConf.shouldPrefetchOnOpen() && cacheIfCompactionsOff() && shouldCache.booleanValue()) {
+    if (cacheConf.shouldPrefetchOnOpen() && shouldCache.booleanValue()) {
       PrefetchExecutor.request(path, new Runnable() {
         @Override
         public void run() {

hbase-server/src/main/java/org/apache/hadoop/hbase/io/hfile/HFileReaderImpl.java

@@ -17,7 +17,6 @@
  */
 package org.apache.hadoop.hbase.io.hfile;
 
-import static org.apache.hadoop.hbase.regionserver.CompactSplit.HBASE_REGION_SERVER_ENABLE_COMPACTION;
 import static org.apache.hadoop.hbase.trace.HBaseSemanticAttributes.BLOCK_CACHE_KEY_KEY;
 
 import io.opentelemetry.api.common.Attributes;
@@ -42,14 +41,12 @@
 import org.apache.hadoop.hbase.SizeCachedKeyValue;
 import org.apache.hadoop.hbase.SizeCachedNoTagsByteBufferKeyValue;
 import org.apache.hadoop.hbase.SizeCachedNoTagsKeyValue;
-import org.apache.hadoop.hbase.io.HFileLink;
 import org.apache.hadoop.hbase.io.compress.Compression;
 import org.apache.hadoop.hbase.io.encoding.DataBlockEncoder;
 import org.apache.hadoop.hbase.io.encoding.DataBlockEncoding;
 import org.apache.hadoop.hbase.io.encoding.HFileBlockDecodingContext;
 import org.apache.hadoop.hbase.nio.ByteBuff;
 import org.apache.hadoop.hbase.regionserver.KeyValueScanner;
-import org.apache.hadoop.hbase.regionserver.StoreFileInfo;
 import org.apache.hadoop.hbase.util.ByteBufferUtils;
 import org.apache.hadoop.hbase.util.Bytes;
 import org.apache.hadoop.hbase.util.IdLock;
@@ -159,6 +156,10 @@ public BlockIndexNotLoadedException(Path path) {
     }
   }
 
+  public CacheConfig getCacheConf() {
+    return cacheConf;
+  }
+
   private Optional<String> toStringFirstKey() {
     return getFirstKey().map(CellUtil::getCellKeyAsString);
   }
@@ -307,7 +308,7 @@ public NotSeekedException(Path path) {
     }
   }
 
-  protected static class HFileScannerImpl implements HFileScanner {
+  public static class HFileScannerImpl implements HFileScanner {
     private ByteBuff blockBuffer;
     protected final boolean cacheBlocks;
     protected final boolean pread;
@@ -331,6 +332,7 @@ protected static class HFileScannerImpl implements HFileScanner {
      * loaded yet.
      */
     protected Cell nextIndexedKey;
+
     // Current block being used. NOTICE: DON't release curBlock separately except in shipped() or
     // close() methods. Because the shipped() or close() will do the release finally, even if any
     // exception occur the curBlock will be released by the close() method (see
@@ -340,6 +342,11 @@ protected static class HFileScannerImpl implements HFileScanner {
     // Whether we returned a result for curBlock's size in recordBlockSize().
     // gets reset whenever curBlock is changed.
     private boolean providedCurrentBlockSize = false;
+
+    public HFileBlock getCurBlock() {
+      return curBlock;
+    }
+
     // Previous blocks that were used in the course of the read
     protected final ArrayList<HFileBlock> prevBlocks = new ArrayList<>();
 
@@ -1283,8 +1290,6 @@ public HFileBlock readBlock(long dataBlockOffset, long onDiskBlockSize, final bo
       new BlockCacheKey(path, dataBlockOffset, this.isPrimaryReplicaReader(), expectedBlockType);
     Attributes attributes = Attributes.of(BLOCK_CACHE_KEY_KEY, cacheKey.toString());
 
-    boolean cacheable = cacheBlock && cacheIfCompactionsOff();
-
     boolean useLock = false;
     IdLock.Entry lockEntry = null;
     final Span span = Span.current();
@@ -1326,7 +1331,7 @@ public HFileBlock readBlock(long dataBlockOffset, long onDiskBlockSize, final bo
             return cachedBlock;
           }
 
-          if (!useLock && cacheable && cacheConf.shouldLockOnCacheMiss(expectedBlockType)) {
+          if (!useLock && cacheBlock && cacheConf.shouldLockOnCacheMiss(expectedBlockType)) {
             // check cache again with lock
             useLock = true;
             continue;
@@ -1337,7 +1342,7 @@ public HFileBlock readBlock(long dataBlockOffset, long onDiskBlockSize, final bo
         span.addEvent("block cache miss", attributes);
         // Load block from filesystem.
         HFileBlock hfileBlock = fsBlockReader.readBlockData(dataBlockOffset, onDiskBlockSize, pread,
-          !isCompaction, shouldUseHeap(expectedBlockType, cacheable));
+          !isCompaction, shouldUseHeap(expectedBlockType, cacheBlock));
         try {
           validateBlockType(hfileBlock, expectedBlockType);
         } catch (IOException e) {
@@ -1350,25 +1355,30 @@ public HFileBlock readBlock(long dataBlockOffset, long onDiskBlockSize, final bo
 
         // Don't need the unpacked block back and we're storing the block in the cache compressed
         if (cacheOnly && cacheCompressed && cacheOnRead) {
+          HFileBlock blockNoChecksum = BlockCacheUtil.getBlockForCaching(cacheConf, hfileBlock);
           cacheConf.getBlockCache().ifPresent(cache -> {
             LOG.debug("Skipping decompression of block {} in prefetch", cacheKey);
             // Cache the block if necessary
-            if (cacheable && cacheConf.shouldCacheBlockOnRead(category)) {
-              cache.cacheBlock(cacheKey, hfileBlock, cacheConf.isInMemory(), cacheOnly);
+            if (cacheBlock && cacheConf.shouldCacheBlockOnRead(category)) {
+              cache.cacheBlock(cacheKey, blockNoChecksum, cacheConf.isInMemory(), cacheOnly);
             }
           });
 
           if (updateCacheMetrics && hfileBlock.getBlockType().isData()) {
             HFile.DATABLOCK_READ_COUNT.increment();
           }
-          return hfileBlock;
+          return blockNoChecksum;
         }
         HFileBlock unpacked = hfileBlock.unpack(hfileContext, fsBlockReader);
+        HFileBlock unpackedNoChecksum = BlockCacheUtil.getBlockForCaching(cacheConf, unpacked);
         // Cache the block if necessary
         cacheConf.getBlockCache().ifPresent(cache -> {
-          if (cacheable && cacheConf.shouldCacheBlockOnRead(category)) {
+          if (cacheBlock && cacheConf.shouldCacheBlockOnRead(category)) {
             // Using the wait on cache during compaction and prefetching.
-            cache.cacheBlock(cacheKey, cacheCompressed ? hfileBlock : unpacked,
+            cache.cacheBlock(cacheKey,
+              cacheCompressed
+                ? BlockCacheUtil.getBlockForCaching(cacheConf, hfileBlock)
+                : unpackedNoChecksum,
               cacheConf.isInMemory(), cacheOnly);
           }
         });
@@ -1380,7 +1390,7 @@ public HFileBlock readBlock(long dataBlockOffset, long onDiskBlockSize, final bo
           HFile.DATABLOCK_READ_COUNT.increment();
         }
 
-        return unpacked;
+        return unpackedNoChecksum;
       }
     } finally {
       if (lockEntry != null) {
@@ -1691,9 +1701,4 @@ public int getMajorVersion() {
   public void unbufferStream() {
     fsBlockReader.unbufferStream();
   }
-
-  protected boolean cacheIfCompactionsOff() {
-    return (!StoreFileInfo.isReference(name) && !HFileLink.isHFileLink(name))
-      || !conf.getBoolean(HBASE_REGION_SERVER_ENABLE_COMPACTION, true);
-  }
 }