Examples with Ring com.linkedin.d2.balancer.util.hashing.Ring used on opensource projects

Example 16 with Ring

use of com.linkedin.d2.balancer.util.hashing.Ring in project rest.li by linkedin.

the class DegraderLoadBalancerTest method testClusterRecovery2TC.

@Test(groups = { "small", "back-end" })
public void testClusterRecovery2TC() {
    final int NUM_CHECKS = 5;
    final Long TIME_INTERVAL = 5000L;
    Map<String, Object> myMap = new HashMap<String, Object>();
    // 1,2,4,8,16,32,64,100% steps, given a 2x recovery step coefficient
    int localStepsToFullRecovery = 8;
    myMap.put(PropertyKeys.HTTP_LB_INITIAL_RECOVERY_LEVEL, 0.005);
    myMap.put(PropertyKeys.HTTP_LB_RING_RAMP_FACTOR, 2d);
    myMap.put(PropertyKeys.HTTP_LB_STRATEGY_PROPERTIES_UPDATE_INTERVAL_MS, TIME_INTERVAL);
    TestClock clock = new TestClock();
    myMap.put(PropertyKeys.CLOCK, clock);
    DegraderLoadBalancerStrategyConfig config = DegraderLoadBalancerStrategyConfig.createHttpConfigFromMap(myMap);
    DegraderLoadBalancerStrategyV3 strategy = new DegraderLoadBalancerStrategyV3(config, "DegraderLoadBalancerTest", null);
    List<TrackerClient> clients = new ArrayList<TrackerClient>();
    URI uri1 = URI.create("http://test.linkedin.com:3242/fdsaf");
    URI uri2 = URI.create("http://test.linkedin.com:3243/fdsaf");
    URIRequest request = new URIRequest(uri1);
    List<CallCompletion> ccList = new ArrayList<CallCompletion>();
    CallCompletion cc;
    TrackerClient client1 = new TrackerClient(uri1, getDefaultPartitionData(1d), new TestLoadBalancerClient(uri1), clock, null);
    TrackerClient client2 = new TrackerClient(uri2, getDefaultPartitionData(1d), new TestLoadBalancerClient(uri2), clock, null);
    clients.add(client1);
    clients.add(client2);
    // force client1 to be disabled if we encounter errors/high latency
    DegraderControl dcClient1Default = client1.getDegraderControl(DEFAULT_PARTITION_ID);
    dcClient1Default.setMinCallCount(5);
    dcClient1Default.setOverrideMinCallCount(5);
    dcClient1Default.setUpStep(1.0);
    // force client2 to be disabled if we encounter errors/high latency
    DegraderControl dcClient2Default = client2.getDegraderControl(DEFAULT_PARTITION_ID);
    dcClient2Default.setOverrideMinCallCount(5);
    dcClient2Default.setMinCallCount(5);
    dcClient2Default.setUpStep(0.4);
    // Have one cycle of successful calls to verify valid tracker clients returned.
    // try load balancing on this updateState, need to updateState before forcing the strategy.
    TrackerClient resultTC = getTrackerClient(strategy, request, new RequestContext(), 1, clients);
    strategy.setStrategy(DEFAULT_PARTITION_ID, DegraderLoadBalancerStrategyV3.PartitionDegraderLoadBalancerState.Strategy.LOAD_BALANCE);
    resultTC = getTrackerClient(strategy, request, new RequestContext(), 1, clients);
    assertNotNull(resultTC, "expected non-null trackerclient");
    for (int j = 0; j < NUM_CHECKS; j++) {
        ccList.add(client1.getCallTracker().startCall());
        ccList.add(client2.getCallTracker().startCall());
    }
    clock.addMs(1);
    for (Iterator<CallCompletion> iter = ccList.listIterator(); iter.hasNext(); ) {
        cc = iter.next();
        cc.endCall();
    }
    // bump to next interval, and get stats.
    clock.addMs(5000);
    // try Load balancing on this updateState
    strategy.setStrategy(DEFAULT_PARTITION_ID, DegraderLoadBalancerStrategyV3.PartitionDegraderLoadBalancerState.Strategy.LOAD_BALANCE);
    resultTC = getTrackerClient(strategy, request, new RequestContext(), 1, clients);
    assertNotNull(resultTC, "expected non-null trackerclient");
    Assert.assertEquals(dcClient1Default.getCurrentComputedDropRate(), 0.0);
    Assert.assertEquals(dcClient2Default.getCurrentComputedDropRate(), 0.0);
    // now simulate a bad cluster state with high error and high latency
    for (int j = 0; j < NUM_CHECKS; j++) {
        ccList.add(client1.getCallTracker().startCall());
        ccList.add(client2.getCallTracker().startCall());
    }
    clock.addMs(3500);
    for (Iterator<CallCompletion> iter = ccList.listIterator(); iter.hasNext(); ) {
        cc = iter.next();
        cc.endCallWithError();
    }
    // go to next interval
    clock.addMs(5000);
    Assert.assertEquals(dcClient1Default.getCurrentComputedDropRate(), 1.0);
    Assert.assertEquals(dcClient2Default.getCurrentComputedDropRate(), 0.4);
    // trigger a state update, the returned TrackerClient should be client2
    // because client 1 should have gone up to a 1.0 drop rate, and the cluster should
    // be unhealthy
    strategy.setStrategy(DEFAULT_PARTITION_ID, DegraderLoadBalancerStrategyV3.PartitionDegraderLoadBalancerState.Strategy.LOAD_BALANCE);
    resultTC = getTrackerClient(strategy, request, new RequestContext(), 1, clients);
    assertEquals(resultTC, client2);
    // it's current drop rate.
    do {
        // go to next time interval.
        clock.addMs(TIME_INTERVAL);
        // adjust the hash ring this time.
        strategy.setStrategy(DEFAULT_PARTITION_ID, DegraderLoadBalancerStrategyV3.PartitionDegraderLoadBalancerState.Strategy.LOAD_BALANCE);
        resultTC = getTrackerClient(strategy, request, new RequestContext(), 1, clients);
        localStepsToFullRecovery--;
    } while (localStepsToFullRecovery > 0);
    assertNotNull(resultTC, "expected non-null trackerclient");
    assertTrue(strategy.getState().getPartitionState(DEFAULT_PARTITION_ID).getPointsMap().get(client1.getUri()) == client1.getPartitionWeight(DEFAULT_PARTITION_ID) * config.getPointsPerWeight(), "client1 did not recover to full weight in hash map.");
    Assert.assertEquals(dcClient2Default.getCurrentComputedDropRate(), 0.4, "client2 drop rate not as expected");
    cc = client1.getCallTracker().startCall();
    clock.addMs(10);
    cc.endCall();
    clock.addMs(TIME_INTERVAL);
    resultTC = getTrackerClient(strategy, request, new RequestContext(), 1, clients);
    assertNotNull(resultTC, "expected non-null trackerclient");
}

Example 17 with Ring

use of com.linkedin.d2.balancer.util.hashing.Ring in project rest.li by linkedin.

the class DegraderLoadBalancerTest method testClusterRecoverySlow1TC.

@Test(groups = { "small", "back-end" })
public void testClusterRecoverySlow1TC() {
    Map<String, Object> myMap = new HashMap<String, Object>();
    Long timeInterval = 5000L;
    TestClock clock = new TestClock();
    myMap.put(PropertyKeys.CLOCK, clock);
    // We want the degrader to have one cooling off period, and then re-enter the ring. This
    myMap.put(PropertyKeys.HTTP_LB_INITIAL_RECOVERY_LEVEL, 0.005);
    myMap.put(PropertyKeys.HTTP_LB_RING_RAMP_FACTOR, 2.0);
    myMap.put(PropertyKeys.HTTP_LB_STRATEGY_PROPERTIES_UPDATE_INTERVAL_MS, timeInterval);
    // it will take two intervals for the TC to be reintroduced into the hash ring.
    int stepsToFullRecovery = 1;
    //test Strategy V3
    DegraderLoadBalancerStrategyConfig config = DegraderLoadBalancerStrategyConfig.createHttpConfigFromMap(myMap);
    DegraderLoadBalancerStrategyV3 strategyV3 = new DegraderLoadBalancerStrategyV3(config, "DegraderLoadBalancerTest", null);
    DegraderLoadBalancerStrategyAdapter strategy = new DegraderLoadBalancerStrategyAdapter(strategyV3);
    clusterRecovery1TC(myMap, clock, stepsToFullRecovery, timeInterval, strategy, null, DegraderLoadBalancerStrategyV3.PartitionDegraderLoadBalancerState.Strategy.LOAD_BALANCE);
    //test Strategy V2
    config = DegraderLoadBalancerStrategyConfig.createHttpConfigFromMap(myMap);
    DegraderLoadBalancerStrategyV2_1 strategyV2 = new DegraderLoadBalancerStrategyV2_1(config, "DegraderLoadBalancerTest", null);
    strategy = new DegraderLoadBalancerStrategyAdapter(strategyV2);
    clusterRecovery1TC(myMap, clock, stepsToFullRecovery, timeInterval, strategy, DegraderLoadBalancerStrategyV2_1.DegraderLoadBalancerState.Strategy.LOAD_BALANCE, null);
}

Example 18 with Ring

use of com.linkedin.d2.balancer.util.hashing.Ring in project rest.li by linkedin.

the class DegraderLoadBalancerTest method testClusterRecoveryAfter100PercentDropCall.

@Test(groups = { "small", "back-end" })
public void testClusterRecoveryAfter100PercentDropCall() {
    Map<String, Object> myMap = new HashMap<String, Object>();
    Long timeInterval = 5000L;
    TestClock clock = new TestClock();
    myMap.put(PropertyKeys.CLOCK, clock);
    // We want the degrader to have one cooling off period, and then re-enter the ring. This
    myMap.put(PropertyKeys.HTTP_LB_INITIAL_RECOVERY_LEVEL, 0.005);
    myMap.put(PropertyKeys.HTTP_LB_RING_RAMP_FACTOR, 2.0);
    myMap.put(PropertyKeys.HTTP_LB_STRATEGY_PROPERTIES_UPDATE_INTERVAL_MS, timeInterval);
    myMap.put(PropertyKeys.HTTP_LB_GLOBAL_STEP_UP, 1.0);
    myMap.put(PropertyKeys.HTTP_LB_GLOBAL_STEP_DOWN, 0.8);
    myMap.put(PropertyKeys.HTTP_LB_CLUSTER_MIN_CALL_COUNT_HIGH_WATER_MARK, 1l);
    myMap.put(PropertyKeys.HTTP_LB_CLUSTER_MIN_CALL_COUNT_LOW_WATER_MARK, 1l);
    //test Strategy V3
    DegraderLoadBalancerStrategyConfig config = DegraderLoadBalancerStrategyConfig.createHttpConfigFromMap(myMap);
    DegraderLoadBalancerStrategyV3 strategyV3 = new DegraderLoadBalancerStrategyV3(config, "DegraderLoadBalancerTest", null);
    DegraderLoadBalancerStrategyAdapter strategy = new DegraderLoadBalancerStrategyAdapter(strategyV3);
    clusterTotalRecovery1TC(myMap, clock, timeInterval, strategy);
    //test Strategy V2
    config = DegraderLoadBalancerStrategyConfig.createHttpConfigFromMap(myMap);
    DegraderLoadBalancerStrategyV2_1 strategyV2 = new DegraderLoadBalancerStrategyV2_1(config, "DegraderLoadBalancerTest", null);
    strategy = new DegraderLoadBalancerStrategyAdapter(strategyV2);
    clusterTotalRecovery1TC(myMap, clock, timeInterval, strategy);
}

Example 19 with Ring

use of com.linkedin.d2.balancer.util.hashing.Ring in project rest.li by linkedin.

the class DegraderLoadBalancerTest method testDegraderLoadBalancerSimulator.

private void testDegraderLoadBalancerSimulator(DegraderLoadBalancerStrategyAdapter adapter, TestClock clock, long timeInterval, List<TrackerClient> 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<URI>();
    HashSet<URI> removedUris = new HashSet<URI>();
    for (TrackerClient client : clients) {
        uris.add(client.getUri());
    }
    LinkedList<TrackerClient> removedClients = new LinkedList<TrackerClient>();
    //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
        TrackerClient newClient = new TrackerClient(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 20 with Ring

use of com.linkedin.d2.balancer.util.hashing.Ring in project rest.li by linkedin.

the class LoadBalancerSimulator method getPoints.

/**
   * Given a serviceName and partition number, return the hashring points for each URI
   * @param serviceName
   * @param partition
   * @return
   * @throws ServiceUnavailableException
   */
public Map<URI, Integer> getPoints(String serviceName, int partition) throws ServiceUnavailableException {
    URI serviceUri = URI.create("d2://" + serviceName);
    Ring<URI> ring = _loadBalancer.getRings(serviceUri).get(partition);
    Map<URI, Integer> pointsMap = new HashMap<>();
    Iterator<URI> iter = ring.getIterator(0);
    iter.forEachRemaining(uri -> pointsMap.compute(uri, (k, v) -> v == null ? 1 : v + 1));
    return pointsMap;
}