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

Example 46 with TrackerClient

use of com.linkedin.d2.balancer.clients.TrackerClient 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 47 with TrackerClient

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

the class DegraderLoadBalancerTest method testRandom.

@Test(groups = { "small", "back-end" })
public void testRandom() throws URISyntaxException {
    DegraderLoadBalancerStrategyV3 strategy = getStrategy();
    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");
    clients.add(getClient(uri1, new TestClock()));
    clients.add(getClient(uri2, new TestClock()));
    // since cluster call count is 0, we will default to random
    for (int i = 0; i < 1000; ++i) {
        TrackerClient client = getTrackerClient(strategy, null, new RequestContext(), 0, clients);
        assertNotNull(client);
        assertTrue(clients.contains(client));
    }
}

Example 48 with TrackerClient

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

the class DegraderLoadBalancerTest method testInconsistentHashAndTrackerclients.

@Test(groups = { "small", "back-end" }, dataProvider = "consistentHashAlgorithms")
public void testInconsistentHashAndTrackerclients(String consistentHashAlgorithm) throws URISyntaxException, InterruptedException {
    // check if the inconsistent Hash ring and trackerlients can be handled
    TestClock clock = new TestClock();
    Map<String, Object> myMap = lbDefaultConfig();
    myMap.put(PropertyKeys.CLOCK, clock);
    myMap.put(PropertyKeys.HTTP_LB_CONSISTENT_HASH_ALGORITHM, consistentHashAlgorithm);
    DegraderLoadBalancerStrategyConfig config = DegraderLoadBalancerStrategyConfig.createHttpConfigFromMap(myMap);
    DegraderLoadBalancerStrategyV3 strategy = new DegraderLoadBalancerStrategyV3(config, "DegraderLoadBalancerTest", null, DEGRADER_STATE_LISTENER_FACTORIES);
    List<DegraderTrackerClient> clients = new ArrayList<>();
    clients.add(getClient(URI.create("http://test.linkedin.com:3242/fdsaf"), clock));
    clients.add(getClient(URI.create("http://test.linkedin.com:3243/fdsaf"), clock));
    TrackerClient chosen = getTrackerClient(strategy, null, new RequestContext(), 0, clients);
    assertNotNull(chosen);
    // remove the client from the list, now the ring and the trackerClient list are inconsistent
    clients.remove(chosen);
    assertNotNull(getTrackerClient(strategy, null, new RequestContext(), 0, clients));
    // update the hash ring we should get the results as well
    assertNotNull(getTrackerClient(strategy, null, new RequestContext(), 1, clients));
}

Example 49 with TrackerClient

use of com.linkedin.d2.balancer.clients.TrackerClient 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);
}

Example 50 with TrackerClient

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

the class SimpleLoadBalancerStateTest method testUpdatePartitionDataMap.

@Test(groups = { "small", "back-end" })
public void testUpdatePartitionDataMap() {
    reset();
    URI uri = URI.create("http://cluster-1/test");
    List<String> schemes = new ArrayList<>();
    Map<Integer, PartitionData> partitionDataMap = new HashMap<>(1);
    Map<URI, Map<Integer, PartitionData>> uriData = new HashMap<>();
    uriData.put(uri, partitionDataMap);
    schemes.add("http");
    _state.listenToCluster("cluster-1", new NullStateListenerCallback());
    _state.listenToService("service-1", new NullStateListenerCallback());
    _serviceRegistry.put("service-1", new ServiceProperties("service-1", "cluster-1", "/test", Arrays.asList("random"), Collections.<String, Object>emptyMap(), null, null, schemes, null));
    // first we announce uri with empty partition data map
    _uriRegistry.put("cluster-1", new UriProperties("cluster-1", uriData));
    TrackerClient client = _state.getClient("service-1", uri);
    assertNotNull(client);
    assertEquals(client.getUri(), uri);
    // tracker client should see empty partition data map
    assertTrue(client.getPartitionDataMap().isEmpty());
    // then we update this uri to have a non-empty partition data map
    partitionDataMap.put(DefaultPartitionAccessor.DEFAULT_PARTITION_ID, new PartitionData(1d));
    _uriRegistry.put("cluster-1", new UriProperties("cluster-1", uriData));
    TrackerClient updatedClient = _state.getClient("service-1", uri);
    assertNotNull(updatedClient);
    // should have got a different tracker client
    assertNotSame(client, updatedClient);
    assertEquals(updatedClient.getUri(), uri);
    // this updated client should have updated partition data map
    assertFalse(updatedClient.getPartitionDataMap().isEmpty());
    assertEquals(updatedClient.getPartitionDataMap(), partitionDataMap);
}