/**
* Returns the local filename for the block if that file exists on the local file system. This is
* an alpha power-api feature for applications that want short-circuit-read files directly. There
* is no guarantee that the file still exists after this call returns, as Tachyon may evict blocks
* from memory at any time.
*
* @param blockIndex The index of the block in the file.
* @return filename on local file system or null if file not present on local file system.
* @throws IOException
*/
public String getLocalFilename(int blockIndex) throws IOException {
ClientBlockInfo blockInfo = getClientBlockInfo(blockIndex);
long blockId = blockInfo.getBlockId();
int blockLockId = mTachyonFS.getBlockLockId();
String filename = mTachyonFS.lockBlock(blockId, blockLockId);
if (filename != null) {
mTachyonFS.unlockBlock(blockId, blockLockId);
}
return filename;
}
@Override
public int read(byte[] b, int off, int len) throws IOException {
if (b == null) {
throw new NullPointerException();
} else if (off < 0 || len < 0 || len > b.length - off) {
throw new IndexOutOfBoundsException();
} else if (len == 0) {
return 0;
} else if (mBlockPos == mBlockInfo.length) {
return -1;
}
// We read at most len bytes, but if mBlockPos + len exceeds the length of the file, we only
// read up to the end of the file
len = (int) Math.min(len, mBlockInfo.length - mBlockPos);
int bytesLeft = len;
// Lazy initialization of the out stream for caching to avoid collisions with other caching
// attempts that are invalidated later due to seek/skips
if (bytesLeft > 0 && mBlockOutStream == null && mRecache) {
try {
mBlockOutStream = BlockOutStream.get(mFile, WriteType.TRY_CACHE, mBlockIndex, mTachyonConf);
// We should only cache when we are writing to a local worker
if (mBlockOutStream instanceof RemoteBlockOutStream) {
LOG.info("Cannot find a local worker to write to, recache attempt cancelled.");
cancelRecache();
}
} catch (IOException ioe) {
LOG.warn("Recache attempt failed.", ioe);
cancelRecache();
}
}
// While we still have bytes to read, make sure the buffer is set to read the byte at mBlockPos.
// If we fail to set mCurrentBuffer, we stream the rest from the underfs
while (bytesLeft > 0 && mAttemptReadFromWorkers && updateCurrentBuffer()) {
int bytesToRead = Math.min(bytesLeft, mCurrentBuffer.remaining());
mCurrentBuffer.get(b, off, bytesToRead);
if (mRecache) {
mBlockOutStream.write(b, off, bytesToRead);
}
off += bytesToRead;
bytesLeft -= bytesToRead;
mBlockPos += bytesToRead;
}
mBytesReadRemote += len - bytesLeft;
mTachyonFS.getClientMetrics().incBytesReadRemote(len - bytesLeft);
if (bytesLeft > 0) {
// Unable to read from worker memory, reading this block from underfs in the future.
mAttemptReadFromWorkers = false;
// We failed to read everything from mCurrentBuffer, so we need to stream the rest from the
// underfs
if (!setupStreamFromUnderFs()) {
LOG.error(
"Failed to read at position "
+ mBlockPos
+ " in block "
+ mBlockInfo.getBlockId()
+ " from workers or underfs");
// Return the number of bytes we managed to read
return len - bytesLeft;
}
while (bytesLeft > 0) {
int readBytes = mCheckpointInputStream.read(b, off, bytesLeft);
if (readBytes <= 0) {
LOG.error("Checkpoint stream read 0 bytes, which shouldn't ever happen");
return len - bytesLeft;
}
if (mRecache) {
mBlockOutStream.write(b, off, readBytes);
}
off += readBytes;
bytesLeft -= readBytes;
mBlockPos += readBytes;
mCheckpointPos += readBytes;
mTachyonFS.getClientMetrics().incBytesReadUfs(readBytes);
}
}
return len;
}
/**
* Promote block back to top layer after access
*
* @param blockIndex the index of the block
* @return true if success, false otherwise
* @throws IOException
*/
public boolean promoteBlock(int blockIndex) throws IOException {
ClientBlockInfo blockInfo = getClientBlockInfo(blockIndex);
return mTachyonFS.promoteBlock(blockInfo.getBlockId());
}