/**
* Updates metadata when a worker periodically heartbeats with the master.
*
* @param workerId the worker id
* @param usedBytesOnTiers a mapping from tier alias to the used bytes
* @param removedBlockIds a list of block ids removed from this worker
* @param addedBlocksOnTiers a mapping from tier alias to the added blocks
* @return an optional command for the worker to execute
*/
public Command workerHeartbeat(
long workerId,
Map<String, Long> usedBytesOnTiers,
List<Long> removedBlockIds,
Map<String, List<Long>> addedBlocksOnTiers) {
MasterWorkerInfo worker = mWorkers.getFirstByField(ID_INDEX, workerId);
if (worker == null) {
LOG.warn("Could not find worker id: {} for heartbeat.", workerId);
return new Command(CommandType.Register, new ArrayList<Long>());
}
synchronized (worker) {
// Technically, 'worker' should be confirmed to still be in the data structure. Lost worker
// detection can remove it. However, we are intentionally ignoring this race, since the worker
// will just re-register regardless.
processWorkerRemovedBlocks(worker, removedBlockIds);
processWorkerAddedBlocks(worker, addedBlocksOnTiers);
worker.updateUsedBytes(usedBytesOnTiers);
worker.updateLastUpdatedTimeMs();
List<Long> toRemoveBlocks = worker.getToRemoveBlocks();
if (toRemoveBlocks.isEmpty()) {
return new Command(CommandType.Nothing, new ArrayList<Long>());
}
return new Command(CommandType.Free, toRemoveBlocks);
}
}
// TODO(binfan): check the logic is correct or not when commitBlock is a retry
public void commitBlock(
long workerId, long usedBytesOnTier, String tierAlias, long blockId, long length)
throws NoWorkerException {
LOG.debug(
"Commit block from workerId: {}, usedBytesOnTier: {}, blockId: {}, length: {}",
workerId,
usedBytesOnTier,
blockId,
length);
long counter = AsyncJournalWriter.INVALID_FLUSH_COUNTER;
MasterWorkerInfo worker = mWorkers.getFirstByField(ID_INDEX, workerId);
// TODO(peis): Check lost workers as well.
if (worker == null) {
throw new NoWorkerException(ExceptionMessage.NO_WORKER_FOUND.getMessage(workerId));
}
// Lock the worker metadata first.
synchronized (worker) {
// Loop until block metadata is successfully locked.
for (; ; ) {
boolean newBlock = false;
MasterBlockInfo block = mBlocks.get(blockId);
if (block == null) {
// The block metadata doesn't exist yet.
block = new MasterBlockInfo(blockId, length);
newBlock = true;
}
// Lock the block metadata.
synchronized (block) {
boolean writeJournal = false;
if (newBlock) {
if (mBlocks.putIfAbsent(blockId, block) != null) {
// Another thread already inserted the metadata for this block, so start loop over.
continue;
}
// Successfully added the new block metadata. Append a journal entry for the new
// metadata.
writeJournal = true;
} else if (block.getLength() != length && block.getLength() == Constants.UNKNOWN_SIZE) {
// The block size was previously unknown. Update the block size with the committed
// size, and append a journal entry.
block.updateLength(length);
writeJournal = true;
}
if (writeJournal) {
BlockInfoEntry blockInfo =
BlockInfoEntry.newBuilder().setBlockId(blockId).setLength(length).build();
counter = appendJournalEntry(JournalEntry.newBuilder().setBlockInfo(blockInfo).build());
}
// At this point, both the worker and the block metadata are locked.
// Update the block metadata with the new worker location.
block.addWorker(workerId, tierAlias);
// This worker has this block, so it is no longer lost.
mLostBlocks.remove(blockId);
// Update the worker information for this new block.
// TODO(binfan): when retry commitBlock on master is expected, make sure metrics are not
// double counted.
worker.addBlock(blockId);
worker.updateUsedBytes(tierAlias, usedBytesOnTier);
worker.updateLastUpdatedTimeMs();
}
break;
}
}
waitForJournalFlush(counter);
}