Examples with DegraderTrackerClientImpl com.linkedin.d2.balancer.clients.DegraderTrackerClientImpl used on opensource projects

Example 6 with DegraderTrackerClientImpl

use of com.linkedin.d2.balancer.clients.DegraderTrackerClientImpl in project rest.li by linkedin.

the class DegraderLoadBalancerTest method createTrackerClient.

/**
 * create multiple trackerClients using the same clock
 * @return
 */
private List<DegraderTrackerClient> createTrackerClient(int n, TestClock clock, DegraderImpl.Config config) {
    String baseUri = "http://test.linkedin.com:10010/abc";
    List<DegraderTrackerClient> result = new LinkedList<>();
    for (int i = 0; i < n; i++) {
        URI uri = URI.create(baseUri + i);
        DegraderTrackerClient client = new DegraderTrackerClientImpl(uri, getDefaultPartitionData(1d), new TestLoadBalancerClient(uri), clock, config);
        result.add(client);
    }
    return result;
}

Example 7 with DegraderTrackerClientImpl

use of com.linkedin.d2.balancer.clients.DegraderTrackerClientImpl in project rest.li by linkedin.

the class DegraderLoadBalancerTest method testWeightedAndLatencyDegradationBalancingRingWithPartitions.

@Test(groups = { "small", "back-end" }, dataProvider = "consistentHashAlgorithms")
public void testWeightedAndLatencyDegradationBalancingRingWithPartitions(String consistentHashAlgorithm) throws URISyntaxException {
    DegraderLoadBalancerStrategyV3 strategy = getStrategy(consistentHashAlgorithm);
    Map<URI, TrackerClient> clientsForPartition0 = new HashMap<>();
    Map<URI, TrackerClient> clientsForPartition1 = new HashMap<>();
    URI uri1 = URI.create("http://someTestService/someTestUrl");
    URI uri2 = URI.create("http://abcxfweuoeueoueoueoukeueoueoueoueoueouo/2354");
    URI uri3 = URI.create("http://slashdot/blah");
    URI uri4 = URI.create("http://idle/server");
    TestClock clock1 = new TestClock();
    TestClock clock2 = new TestClock();
    TestClock clock3 = new TestClock();
    @SuppressWarnings("serial") DegraderTrackerClient client1 = new DegraderTrackerClientImpl(uri1, new HashMap<Integer, PartitionData>() {

        {
            put(0, new PartitionData(1d));
        }
    }, new TestLoadBalancerClient(uri1), clock1, null);
    @SuppressWarnings("serial") DegraderTrackerClient client2 = new DegraderTrackerClientImpl(uri2, new HashMap<Integer, PartitionData>() {

        {
            put(0, new PartitionData(0.5d));
            put(1, new PartitionData(0.5d));
        }
    }, new TestLoadBalancerClient(uri2), clock2, null);
    @SuppressWarnings("serial") DegraderTrackerClient client3 = new DegraderTrackerClientImpl(uri3, new HashMap<Integer, PartitionData>() {

        {
            put(1, new PartitionData(1d));
        }
    }, new TestLoadBalancerClient(uri3), clock3, null);
    final int partitionId0 = 0;
    clientsForPartition0.put(client1.getUri(), client1);
    clientsForPartition0.put(client2.getUri(), client2);
    final int partitionId1 = 1;
    clientsForPartition1.put(client2.getUri(), client2);
    clientsForPartition1.put(client3.getUri(), client3);
    // force client2 to be disabled
    DegraderControl dcClient2Partition0 = client2.getDegraderControl(0);
    DegraderControl dcClient2Partition1 = client2.getDegraderControl(1);
    dcClient2Partition0.setOverrideMinCallCount(1);
    dcClient2Partition0.setMinCallCount(1);
    dcClient2Partition0.setMaxDropRate(1d);
    dcClient2Partition0.setUpStep(0.4d);
    dcClient2Partition0.setHighErrorRate(0);
    dcClient2Partition1.setOverrideMinCallCount(1);
    dcClient2Partition1.setMinCallCount(1);
    dcClient2Partition1.setMaxDropRate(1d);
    dcClient2Partition1.setUpStep(0.4d);
    dcClient2Partition1.setHighErrorRate(0);
    CallCompletion cc = client2.getCallTracker().startCall();
    clock2.addMs(1);
    cc.endCallWithError();
    // force client3 to be disabled
    DegraderControl dcClient3Partition1 = client3.getDegraderControl(1);
    dcClient3Partition1.setOverrideMinCallCount(1);
    dcClient3Partition1.setMinCallCount(1);
    dcClient3Partition1.setMaxDropRate(1d);
    dcClient3Partition1.setHighErrorRate(0);
    dcClient3Partition1.setUpStep(0.2d);
    CallCompletion cc3 = client3.getCallTracker().startCall();
    clock3.addMs(1);
    cc3.endCallWithError();
    clock1.addMs(15000);
    clock2.addMs(5000);
    clock3.addMs(5000);
    // trigger a state update
    assertNotNull(strategy.getTrackerClient(null, new RequestContext(), 1, partitionId0, clientsForPartition0));
    assertNotNull(strategy.getTrackerClient(null, new RequestContext(), 1, partitionId1, clientsForPartition1));
    assertNotNull(strategy.getRing(1, partitionId0, clientsForPartition0));
    assertNotNull(strategy.getRing(1, partitionId1, clientsForPartition1));
    // now do a basic verification to verify getTrackerClient is properly weighting things
    int calls = 10000;
    int client1Count = 0;
    int client2Count = 0;
    double tolerance = 0.05d;
    for (int i = 0; i < calls; ++i) {
        TrackerClient client = strategy.getTrackerClient(null, new RequestContext(), 1, partitionId0, clientsForPartition0);
        assertNotNull(client);
        if (client.getUri().equals(uri1)) {
            ++client1Count;
        } else {
            ++client2Count;
        }
    }
    assertTrue(Math.abs((client1Count / (double) calls) - (100 / 130d)) < tolerance);
    assertTrue(Math.abs((client2Count / (double) calls) - (30 / 130d)) < tolerance);
    client2Count = 0;
    int client3Count = 0;
    int client4Count = 0;
    for (int i = 0; i < calls; ++i) {
        TrackerClient client = strategy.getTrackerClient(null, new RequestContext(), 1, partitionId1, clientsForPartition1);
        assertNotNull(client);
        if (client.getUri().equals(uri3)) {
            ++client3Count;
        } else if (client.getUri().equals(uri2)) {
            ++client2Count;
        } else {
            ++client4Count;
        }
    }
    assertTrue(Math.abs((client3Count / (double) calls) - (80 / 110d)) < tolerance);
    assertTrue(Math.abs((client2Count / (double) calls) - (30 / 110d)) < tolerance);
    assertTrue(client4Count == 0);
}

Example 8 with DegraderTrackerClientImpl

use of com.linkedin.d2.balancer.clients.DegraderTrackerClientImpl in project rest.li by linkedin.

the class DegraderLoadBalancerTest method stressTest.

@Test(groups = { "small", "back-end" })
public void stressTest() {
    final DegraderLoadBalancerStrategyV3 strategyV3 = getStrategy();
    TestClock testClock = new TestClock();
    String baseUri = "http://linkedin.com:9999";
    int numberOfPartitions = 10;
    Map<String, String> degraderProperties = new HashMap<>();
    degraderProperties.put(PropertyKeys.DEGRADER_HIGH_ERROR_RATE, "0.5");
    degraderProperties.put(PropertyKeys.DEGRADER_LOW_ERROR_RATE, "0.2");
    DegraderImpl.Config degraderConfig = DegraderConfigFactory.toDegraderConfig(degraderProperties);
    final List<DegraderTrackerClient> clients = new ArrayList<>();
    for (int i = 0; i < numberOfPartitions; i++) {
        URI uri = URI.create(baseUri + i);
        DegraderTrackerClient client = new DegraderTrackerClientImpl(uri, getDefaultPartitionData(1, numberOfPartitions), new TestLoadBalancerClient(uri), testClock, degraderConfig);
        clients.add(client);
    }
    final ExecutorService executor = Executors.newFixedThreadPool(100);
    final CountDownLatch startLatch = new CountDownLatch(1);
    final CountDownLatch finishLatch = new CountDownLatch(100);
    try {
        for (int i = 0; i < numberOfPartitions; i++) {
            Assert.assertFalse(strategyV3.getState().getPartitionState(i).isInitialized());
        }
        for (int i = 0; i < 100; i++) {
            final int partitionId = i % numberOfPartitions;
            executor.submit(new Runnable() {

                @Override
                public void run() {
                    try {
                        startLatch.await();
                    } catch (InterruptedException ex) {
                    }
                    strategyV3.getRing(1, partitionId, toMap(clients));
                    finishLatch.countDown();
                }
            });
        }
        // all threads would try to getRing simultanously
        startLatch.countDown();
        if (!finishLatch.await(10, TimeUnit.SECONDS)) {
            fail("Stress test failed to finish within 10 seconds");
        }
        for (int i = 0; i < numberOfPartitions; i++) {
            Assert.assertTrue(strategyV3.getState().getPartitionState(i).isInitialized());
        }
    } catch (InterruptedException ex) {
    } finally {
        executor.shutdownNow();
    }
}

Example 9 with DegraderTrackerClientImpl

use of com.linkedin.d2.balancer.clients.DegraderTrackerClientImpl in project rest.li by linkedin.

the class DegraderLoadBalancerTest method testWeightedAndLatencyDegradationBalancingRing.

@Test(groups = { "small", "back-end" }, dataProvider = "consistentHashAlgorithms")
public void testWeightedAndLatencyDegradationBalancingRing(String consistentHashAlgorithm) throws URISyntaxException {
    DegraderLoadBalancerStrategyV3 strategy = getStrategy(consistentHashAlgorithm);
    List<DegraderTrackerClient> clients = new ArrayList<>();
    URI uri1 = URI.create("http://test.linkedin.com:3242/fdsaf");
    URI uri2 = URI.create("http://test.linkedin.com:3243/fdsaf");
    TestClock clock1 = new TestClock();
    TestClock clock2 = new TestClock();
    DegraderTrackerClient client1 = new DegraderTrackerClientImpl(uri1, getDefaultPartitionData(1d), new TestLoadBalancerClient(uri1), clock1, null);
    DegraderTrackerClient client2 = new DegraderTrackerClientImpl(uri2, getDefaultPartitionData(0.8d), new TestLoadBalancerClient(uri2), clock2, null);
    clients.add(client1);
    clients.add(client2);
    DegraderControl dcClient2Default = client2.getDegraderControl(DEFAULT_PARTITION_ID);
    dcClient2Default.setOverrideMinCallCount(1);
    dcClient2Default.setMinCallCount(1);
    dcClient2Default.setMaxDropRate(1d);
    dcClient2Default.setUpStep(0.4d);
    dcClient2Default.setHighErrorRate(0);
    CallCompletion cc = client2.getCallTracker().startCall();
    clock2.addMs(1);
    cc.endCallWithError();
    clock1.addMs(15000);
    clock2.addMs(5000);
    // trigger a state update
    assertNotNull(getTrackerClient(strategy, null, new RequestContext(), 1, clients));
    // now do a basic verification to verify getTrackerClient is properly weighting things
    double calls = 10000d;
    int client1Count = 0;
    int client2Count = 0;
    double tolerance = 0.05d;
    for (int i = 0; i < calls; ++i) {
        TrackerClient client = getTrackerClient(strategy, null, new RequestContext(), 1, clients);
        assertNotNull(client);
        if (client.getUri().equals(uri1)) {
            ++client1Count;
        } else {
            ++client2Count;
        }
    }
    assertTrue(Math.abs((client1Count / calls) - (100 / 148d)) < tolerance);
    assertTrue(Math.abs((client2Count / calls) - (48 / 148d)) < tolerance);
}

Example 10 with DegraderTrackerClientImpl

use of com.linkedin.d2.balancer.clients.DegraderTrackerClientImpl in project rest.li by linkedin.

the class DegraderLoadBalancerTest method testDegraderLoadBalancerSimulator.

private void testDegraderLoadBalancerSimulator(DegraderLoadBalancerStrategyAdapter adapter, TestClock clock, long timeInterval, List<DegraderTrackerClient> clients, double qps, DegraderImpl.Config degraderConfig) {
    long clusterGenerationId = 1;
    double overrideDropRate = 0.0;
    // simulate latency 4000 ms
    // 1st round we use LOAD_BALANCING strategy. Since we have a high latency we will decrease the number of points
    // from 100 to 80 (transmissionRate * points per weight).
    TrackerClient resultTC = simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 80, true, 0.0, 4000, false, false);
    assertNotNull(resultTC);
    // 2nd round drop rate should be increased by DegraderLoadBalancerStrategyConfig.DEFAULT_GLOBAL_STEP_UP
    overrideDropRate += DegraderLoadBalancerStrategyConfig.DEFAULT_GLOBAL_STEP_UP;
    resultTC = simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 80, false, overrideDropRate, 4000, false, false);
    // 3rd round. We alternate back to LOAD_BALANCING strategy and we drop the points even more
    resultTC = simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 39, true, overrideDropRate, 4000, false, false);
    // 4th round. The drop rate should be increased again like 2nd round
    overrideDropRate += DegraderLoadBalancerStrategyConfig.DEFAULT_GLOBAL_STEP_UP;
    resultTC = simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 39, false, overrideDropRate, 4000, false, false);
    // 5th round. Alternate to changing hash ring again.
    resultTC = simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 1, true, overrideDropRate, 4000, false, false);
    // 6th round. Same as 5th round, we'll increase the drop rate
    overrideDropRate += DegraderLoadBalancerStrategyConfig.DEFAULT_GLOBAL_STEP_UP;
    resultTC = simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 1, false, overrideDropRate, 4000, false, false);
    // 7th round. The # of point in hashring is at the minimum so we can't decrease it further. At this point the client
    // is in recovery mode. But since we can't change the hashring anymore, we'll always in CALL_DROPPING mode
    // so the next strategy is expected to be LOAD_BALANCING mode.
    overrideDropRate += DegraderLoadBalancerStrategyConfig.DEFAULT_GLOBAL_STEP_UP;
    resultTC = simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 1, false, overrideDropRate, 4000, false, false);
    // 8th round. We'll increase the drop rate to the max.
    overrideDropRate += DegraderLoadBalancerStrategyConfig.DEFAULT_GLOBAL_STEP_UP;
    resultTC = simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 1, false, overrideDropRate, 4000, false, false);
    // 9th round, now we'll simulate as if there still a call even though we drop 100% of all request to get
    // tracker client. The assumption is there's some thread that still holds tracker client and we want
    // to make sure we can handle the request and we can't degrade the cluster even further.
    resultTC = simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 1, false, overrideDropRate, 4000, false, false);
    // 10th round, now we'll simulate as if there's no call because we dropped all request
    // even though we are in LOAD_BALANCING mode and this tracker client is in recovery mode and there's no call
    // so the hashring doesn't change so we go back to reducing the drop rate to 0.8 and that means the next
    // strategy is LOAD_BALANCE
    overrideDropRate -= DegraderLoadBalancerStrategyConfig.DEFAULT_GLOBAL_STEP_DOWN;
    resultTC = simulateAndTestOneInterval(timeInterval, clock, 0.0, clients, adapter, clusterGenerationId, 1, false, overrideDropRate, 4000, false, false);
    // 11th round, this time we'll simulate the latency is now 1000 ms (so it's within low and high watermark). Drop rate
    // should stay the same and everything else should stay the same
    resultTC = simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 1, false, overrideDropRate, 1000, false, false);
    // we'll simulate the client dying one by one until all the clients are gone
    int numberOfClients = clients.size();
    HashSet<URI> uris = new HashSet<>();
    HashSet<URI> removedUris = new HashSet<>();
    for (TrackerClient client : clients) {
        uris.add(client.getUri());
    }
    LinkedList<TrackerClient> removedClients = new LinkedList<>();
    // loadBalancing strategy will always be picked because there is no hash ring changes
    boolean isLoadBalancingStrategyTurn = true;
    for (int i = numberOfClients; i > 0; i--) {
        TrackerClient removed = clients.remove(0);
        uris.remove(removed.getUri());
        removedClients.addLast(removed);
        removedUris.add(removed.getUri());
        clusterGenerationId++;
        resultTC = simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 1, isLoadBalancingStrategyTurn, overrideDropRate, 1000, false, false);
        if (i == 1) {
            assertNull(resultTC);
        } else {
            // the override drop rate is 0.8)
            if (resultTC != null) {
                assertTrue(uris.contains(resultTC.getUri()));
                assertFalse(removedUris.contains(resultTC.getUri()));
            }
        }
    }
    assertTrue(uris.isEmpty());
    assertTrue(clients.isEmpty());
    assertEquals(removedUris.size(), numberOfClients);
    assertEquals(removedClients.size(), numberOfClients);
    // we'll simulate the client start reviving one by one until all clients are back up again
    for (int i = numberOfClients; i > 0; i--) {
        TrackerClient added = removedClients.remove(0);
        // we have to create a new client. The old client has a degraded DegraderImpl. And in production enviroment
        // when a new client join a cluster, it should be in good state. This means there should be 100 points
        // in the hash ring for this client
        DegraderTrackerClient newClient = new DegraderTrackerClientImpl(added.getUri(), getDefaultPartitionData(1d), new TestLoadBalancerClient(added.getUri()), clock, degraderConfig);
        clients.add(newClient);
        uris.add(added.getUri());
        removedUris.remove(added.getUri());
        clusterGenerationId++;
        resultTC = simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 100, isLoadBalancingStrategyTurn, overrideDropRate, 1000, false, false);
        if (resultTC != null) {
            assertTrue(uris.contains(resultTC.getUri()));
            assertFalse(removedUris.contains(resultTC.getUri()));
        }
    }
    // the number of points because there is no hash ring changes
    for (overrideDropRate -= DegraderLoadBalancerStrategyConfig.DEFAULT_GLOBAL_STEP_DOWN; overrideDropRate >= 0; overrideDropRate -= DegraderLoadBalancerStrategyConfig.DEFAULT_GLOBAL_STEP_DOWN) {
        resultTC = simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 100, false, overrideDropRate, 300, false, false);
    }
    // we should have recovered fully by this time
    overrideDropRate = 0.0;
    resultTC = simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 100, false, overrideDropRate, 300, false, false);
    assertNotNull(resultTC);
    clusterGenerationId++;
    // simulate the increase of certain error (connect exception, closedChannelException) rate will cause degradation.
    simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 80, true, 0.0, 300, false, true);
    // switching to call dropping strategy
    simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 80, false, 0.0, 300, false, true);
    // continue the degradation
    simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 39, true, 0.0, 300, false, true);
    simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 39, false, 0.0, 300, false, true);
    simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 1, true, 0.0, 300, false, true);
    // now let's remove all the error and see how the cluster recover but we have to wait until next round because
    // this round is CALL_DROP strategy
    simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 1, false, 0.0, 300, false, false);
    simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 39, true, 0.0, 300, false, false);
    simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 39, false, 0.0, 300, false, false);
    simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 80, true, 0.0, 300, false, false);
    simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 80, false, 0.0, 300, false, false);
    simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 100, true, 0.0, 300, false, false);
    // make sure if we have error that is not from CONNECT_EXCEPTION or CLOSED_CHANNEL_EXCEPTION we don't degrade
    simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 100, false, 0.0, 300, true, false);
    // since there's no change in hash ring due to error NOT of CONNECT_EXCEPTION or CLOSED_CHANNEL_EXCEPTION,
    // the strategy won't change to CALL_DROPPING
    simulateAndTestOneInterval(timeInterval, clock, qps, clients, adapter, clusterGenerationId, 100, false, 0.0, 300, true, false);
}