Examples with MasterWorkerInfo alluxio.master.block.meta.MasterWorkerInfo used on opensource projects

Example 31 with MasterWorkerInfo

use of alluxio.master.block.meta.MasterWorkerInfo in project alluxio by Alluxio.

the class BlockMasterRegisterStreamIntegrationTest method verifyWorkerCanReregister.

// Verify a worker can reregister and have the correct final blocks
private void verifyWorkerCanReregister(long workerId, List<RegisterWorkerPRequest> requestChunks, int expectedBlockCount) throws Exception {
    Queue<Throwable> errorQueue = new ConcurrentLinkedQueue<>();
    sendStreamToMaster(requestChunks, RegisterStreamTestUtils.getErrorCapturingResponseObserver(errorQueue));
    assertEquals(0, errorQueue.size());
    MasterWorkerInfo worker = mBlockMaster.getWorker(workerId);
    assertEquals(expectedBlockCount, worker.getBlockCount());
    assertEquals(0, worker.getToRemoveBlockCount());
    assertEquals(1, mBlockMaster.getWorkerCount());
}

Example 32 with MasterWorkerInfo

use of alluxio.master.block.meta.MasterWorkerInfo in project alluxio by Alluxio.

the class BlockMasterRegisterStreamIntegrationTest method registerLostWorker.

@Test
public // The master has marked the worker as lost.
void registerLostWorker() throws Exception {
    long workerId = mBlockMaster.getWorkerId(NET_ADDRESS_1);
    // The worker registers to the master
    List<RegisterWorkerPRequest> requestChunks = RegisterStreamTestUtils.generateRegisterStreamForWorker(workerId);
    prepareBlocksOnMaster(requestChunks);
    Queue<Throwable> errorQueue = new ConcurrentLinkedQueue<>();
    sendStreamToMaster(requestChunks, RegisterStreamTestUtils.getErrorCapturingResponseObserver(errorQueue));
    // Verify the worker has been registered
    assertEquals(0, errorQueue.size());
    assertEquals(1, mBlockMaster.getWorkerCount());
    // The worker has lost heartbeat and been forgotten
    MasterWorkerInfo worker = mBlockMaster.getWorker(workerId);
    long newTimeMs = worker.getLastUpdatedTimeMs() + MASTER_WORKER_TIMEOUT + 1;
    mClock.setTimeMs(newTimeMs);
    DefaultBlockMaster.LostWorkerDetectionHeartbeatExecutor lostWorkerDetector = ((DefaultBlockMaster) mBlockMaster).new LostWorkerDetectionHeartbeatExecutor();
    lostWorkerDetector.heartbeat();
    // Verify the worker has been forgotten
    assertEquals(0, mBlockMaster.getWorkerCount());
    // Register again
    Queue<Throwable> newErrorQueue = new ConcurrentLinkedQueue<>();
    sendStreamToMaster(requestChunks, RegisterStreamTestUtils.getErrorCapturingResponseObserver(newErrorQueue));
    // Verify the worker is registered again
    assertEquals(0, errorQueue.size());
    MasterWorkerInfo updatedWorker = mBlockMaster.getWorker(workerId);
    assertEquals(TIER_BLOCK_TOTAL, updatedWorker.getBlockCount());
    assertEquals(0, updatedWorker.getToRemoveBlockCount());
    assertEquals(1, mBlockMaster.getWorkerCount());
    // Verify the worker is readable and writable
    verifyWorkerWritable(workerId);
}

Example 33 with MasterWorkerInfo

use of alluxio.master.block.meta.MasterWorkerInfo in project alluxio by Alluxio.

the class BlockMasterRegisterStreamIntegrationTest method reregisterWithDelete.

/**
 * Tests below cover the race conditions during concurrent executions.
 *
 * When the worker registers for the 1st time, no clients should know this worker.
 * Therefore there is no concurrent client-incurred write operations on this worker.
 * The races happen typically when the worker re-registers with the master,
 * where some clients already know this worker and can direct invoke writes on the worker.
 *
 * Tests here verify the integrity of the master-side metadata.
 * In other words, we assume those writers succeed on the worker, and the subsequent
 * update on the master-side metadata should also succeed and be correct.
 */
@Test
public void reregisterWithDelete() throws Exception {
    // Register the worker so the worker is marked active in master
    long workerId = mBlockMaster.getWorkerId(NET_ADDRESS_1);
    List<RegisterWorkerPRequest> requestChunks = RegisterStreamTestUtils.generateRegisterStreamForWorker(workerId);
    prepareBlocksOnMaster(requestChunks);
    Queue<Throwable> errorQueue = new ConcurrentLinkedQueue<>();
    sendStreamToMaster(requestChunks, RegisterStreamTestUtils.getErrorCapturingResponseObserver(errorQueue));
    assertEquals(0, errorQueue.size());
    assertEquals(1, mBlockMaster.getWorkerCount());
    // Find a block to remove
    long blockToRemove = RegisterStreamTestUtils.findFirstBlock(requestChunks);
    // Register again
    CountDownLatch latch = new CountDownLatch(1);
    Queue<Throwable> newErrorQueue = new ConcurrentLinkedQueue<>();
    Future f = mExecutorService.submit(() -> {
        sendStreamToMasterAndSignal(requestChunks, RegisterStreamTestUtils.getErrorCapturingResponseObserver(newErrorQueue), latch);
    });
    // During the register stream, trigger a delete on worker
    latch.await();
    mBlockMaster.removeBlocks(ImmutableList.of(blockToRemove), true);
    // Wait for the register to finish
    f.get();
    assertThrows(BlockInfoException.class, () -> {
        mBlockMaster.getBlockInfo(blockToRemove);
    });
    MasterWorkerInfo worker = mBlockMaster.getWorker(workerId);
    assertEquals(1, mBlockMaster.getWorkerCount());
    assertEquals(TIER_BLOCK_TOTAL - 1, worker.getBlockCount());
    // BlockMaster.removeBlocks() will first remove the block from master metadata
    // (with block lock) then update the block locations (with worker lock).
    // The worker lock is being held by the registering worker, but the 1st part
    // will likely succeed.
    // So during registration when checking on the block, the block is not recognized
    // any more and will remain in MasterWorkerInfo.mToRemoveBlocks.
    // In the next heartbeat the master will issue a command to remove the block
    // from the worker.
    // Even if the block is already removed on the worker it is fine,
    // because deletion of a not-found block is a noop.
    Command command = sendHeartbeatToMaster(workerId);
    assertEquals(Command.newBuilder().addData(blockToRemove).setCommandType(CommandType.Free).build(), command);
}

Example 34 with MasterWorkerInfo

use of alluxio.master.block.meta.MasterWorkerInfo in project alluxio by Alluxio.

the class BlockMasterRegisterStreamIntegrationTest method registerExistingWorkerBlocksLost.

@Test
public void registerExistingWorkerBlocksLost() throws Exception {
    long workerId = mBlockMaster.getWorkerId(NET_ADDRESS_1);
    // Register the worker for the 1st time
    List<RegisterWorkerPRequest> requestChunks = RegisterStreamTestUtils.generateRegisterStreamForWorker(workerId);
    prepareBlocksOnMaster(requestChunks);
    Queue<Throwable> errorQueue = new ConcurrentLinkedQueue<>();
    sendStreamToMaster(requestChunks, RegisterStreamTestUtils.getErrorCapturingResponseObserver(errorQueue));
    assertEquals(0, errorQueue.size());
    // Verify the worker has registered
    assertEquals(1, mBlockMaster.getWorkerCount());
    MasterWorkerInfo worker = mBlockMaster.getWorker(workerId);
    assertEquals(TIER_BLOCK_TOTAL, worker.getBlockCount());
    assertEquals(0, worker.getToRemoveBlockCount());
    // Manually generate the blocks again and remove some
    List<String> tierAliases = getTierAliases(parseTierConfig(TIER_CONFIG));
    Map<BlockStoreLocation, List<Long>> blockMap = RpcBenchPreparationUtils.generateBlockIdOnTiers(parseTierConfig(TIER_CONFIG));
    Set<Long> lostBlocks = removeSomeBlocks(blockMap);
    // Regenerate the requests
    RegisterStreamer newRegisterStreamer = new RegisterStreamer(null, workerId, tierAliases, CAPACITY_MAP, USAGE_MAP, blockMap, LOST_STORAGE, EMPTY_CONFIG);
    List<RegisterWorkerPRequest> newRequestChunks = ImmutableList.copyOf(newRegisterStreamer);
    int newExpectedBatchCount = (int) Math.ceil((TIER_BLOCK_TOTAL - lostBlocks.size()) / (double) BATCH_SIZE);
    assertEquals(newExpectedBatchCount, newRequestChunks.size());
    // Register again with the updated stream
    Queue<Throwable> newErrorQueue = new ConcurrentLinkedQueue<>();
    sendStreamToMaster(newRequestChunks, RegisterStreamTestUtils.getErrorCapturingResponseObserver(newErrorQueue));
    assertEquals(0, newErrorQueue.size());
    // Verify the worker is registered
    assertEquals(1, mBlockMaster.getWorkerCount());
    MasterWorkerInfo updatedWorker = mBlockMaster.getWorker(workerId);
    assertEquals(TIER_BLOCK_TOTAL - lostBlocks.size(), updatedWorker.getBlockCount());
    // The master will mark the lost blocks as to be removed
    // This is to ensure the unrecognized blocks do no live on the worker anymore
    assertEquals(lostBlocks.size(), updatedWorker.getToRemoveBlockCount());
    // The update is received during the registration so no command to send to the worker
    Command command = sendHeartbeatToMaster(workerId);
    assertEquals(CommandType.Free, command.getCommandType());
    assertEquals(lostBlocks, new HashSet<>(command.getDataList()));
    // Verify the worker is readable and writable
    verifyWorkerWritable(workerId);
}