Search in sources :

Example 1 with ServiceUnavailableException

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

the class SimpleLoadBalancer method chooseTrackerClient.

private TrackerClient chooseTrackerClient(Request request, RequestContext requestContext, String serviceName, String clusterName, ClusterProperties cluster, LoadBalancerStateItem<UriProperties> uriItem, UriProperties uris, List<LoadBalancerState.SchemeStrategyPair> orderedStrategies, ServiceProperties serviceProperties) throws ServiceUnavailableException {
    // now try and find a tracker client for the uri
    TrackerClient trackerClient = null;
    URI targetHost = KeyMapper.TargetHostHints.getRequestContextTargetHost(requestContext);
    int partitionId = -1;
    URI requestUri = request.getURI();
    if (targetHost == null) {
        PartitionAccessor accessor = getPartitionAccessor(serviceName, clusterName);
        try {
            partitionId = accessor.getPartitionId(requestUri);
        } catch (PartitionAccessException e) {
            die(serviceName, "Error in finding the partition for URI: " + requestUri + ", " + e.getMessage());
        }
    } else {
        // This is the case of scatter/gather or search, where the target host may be chosen to be responsible for
        // more than one partitions (The target host was picked from a consistent hash ring, so load balancing is already in effect).
        // we randomly pick one partition to check for the call dropping
        // This is done for two reasons:
        // 1. Currently there is no way to know for which subset of partitions the target host is chosen for
        //    if it is serving more than one partitions. This can be added, but it requires the change of public interfaces (KeyMapper) so that
        //    more hints can be added to the request context for the concerned the partitions
        // 2. More importantly, there is no good way to check for call dropping even if the above problem is solved.
        //    For example, if a target host is chosen for partition 1, 5, 7, with call drop rates of 0, 0.2, 0.4 respectively
        //    A reasonable way to proceed would be use the highest drop rate and do the check once for the target host,
        //    but currently the check can only be done for each partition and only with boolean result (no access to drop rate)
        // The partition to check is picked at random to be conservative.
        // E.g. in the above example, we don't want to always use the drop rate of partition 1.
        Map<Integer, PartitionData> partitionDataMap = uris.getPartitionDataMap(targetHost);
        if (partitionDataMap == null || partitionDataMap.isEmpty()) {
            die(serviceName, "There is no partition data for server host: " + targetHost + ". URI: " + requestUri);
        }
        Set<Integer> partitions = partitionDataMap.keySet();
        Iterator<Integer> iterator = partitions.iterator();
        int index = _random.nextInt(partitions.size());
        for (int i = 0; i <= index; i++) {
            partitionId = iterator.next();
        }
    }
    List<TrackerClient> clientsToLoadBalance = null;
    for (LoadBalancerState.SchemeStrategyPair pair : orderedStrategies) {
        LoadBalancerStrategy strategy = pair.getStrategy();
        String scheme = pair.getScheme();
        clientsToLoadBalance = getPotentialClients(serviceName, serviceProperties, uris, scheme, partitionId);
        trackerClient = strategy.getTrackerClient(request, requestContext, uriItem.getVersion(), partitionId, clientsToLoadBalance);
        debug(_log, "load balancer strategy for ", serviceName, " returned: ", trackerClient);
        // break as soon as we find an available cluster client
        if (trackerClient != null) {
            break;
        }
    }
    if (trackerClient == null) {
        if (clientsToLoadBalance == null || clientsToLoadBalance.isEmpty()) {
            die(serviceName, "Service: " + serviceName + " unable to find a host to route the request" + " in partition: " + partitionId + " cluster: " + clusterName + ". Check what cluster your servers are announcing to.");
        } else {
            die(serviceName, "Service: " + serviceName + " is in a bad state (high latency/high error). " + "Dropping request. Cluster: " + clusterName + ", partitionId:" + partitionId + " (" + clientsToLoadBalance.size() + " hosts)");
        }
    }
    return trackerClient;
}
Also used : LoadBalancerStrategy(com.linkedin.d2.balancer.strategies.LoadBalancerStrategy) URI(java.net.URI) LoadBalancerState(com.linkedin.d2.balancer.LoadBalancerState) PartitionAccessor(com.linkedin.d2.balancer.util.partitions.PartitionAccessor) TrackerClient(com.linkedin.d2.balancer.clients.TrackerClient) PartitionData(com.linkedin.d2.balancer.properties.PartitionData) PartitionAccessException(com.linkedin.d2.balancer.util.partitions.PartitionAccessException)

Example 2 with ServiceUnavailableException

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

the class SimpleLoadBalancer method getClient.

/**
   * Given a Request, returns a TransportClient that can handle requests for the Request.
   *
   *
   * @param request
   *          A request whose URI is a URL of the format "d2://&gt;servicename&lt;/optional/path".
   * @param requestContext context for this request
   * @return A client that can be called to retrieve data for the URN.
   * @throws ServiceUnavailableException
   *           If the load balancer can't figure out how to reach a service for the given
   *           URN, an ServiceUnavailableException will be thrown.
   */
@Override
public TransportClient getClient(Request request, RequestContext requestContext) throws ServiceUnavailableException {
    TransportClient client;
    URI uri = request.getURI();
    debug(_log, "get client for uri: ", uri);
    ServiceProperties service = listenToServiceAndCluster(uri);
    String serviceName = service.getServiceName();
    String clusterName = service.getClusterName();
    ClusterProperties cluster = getClusterProperties(serviceName, clusterName);
    // Check if we want to override the service URL and bypass choosing among the existing
    // tracker clients. This is useful when the service we want is not announcing itself to
    // the cluster, ie a private service for a set of clients.
    URI targetService = LoadBalancerUtil.TargetHints.getRequestContextTargetService(requestContext);
    if (targetService == null) {
        LoadBalancerStateItem<UriProperties> uriItem = getUriItem(serviceName, clusterName, cluster);
        UriProperties uris = uriItem.getProperty();
        List<LoadBalancerState.SchemeStrategyPair> orderedStrategies = _state.getStrategiesForService(serviceName, service.getPrioritizedSchemes());
        TrackerClient trackerClient = chooseTrackerClient(request, requestContext, serviceName, clusterName, cluster, uriItem, uris, orderedStrategies, service);
        String clusterAndServiceUriString = trackerClient.getUri() + service.getPath();
        client = new RewriteClient(serviceName, URI.create(clusterAndServiceUriString), trackerClient);
        _serviceAvailableStats.inc();
    } else {
        _log.debug("service hint found, using generic client for target: {}", targetService);
        TransportClient transportClient = _state.getClient(serviceName, targetService.getScheme());
        client = new RewriteClient(serviceName, targetService, transportClient);
    }
    return client;
}
Also used : ServiceProperties(com.linkedin.d2.balancer.properties.ServiceProperties) TransportClient(com.linkedin.r2.transport.common.bridge.client.TransportClient) TrackerClient(com.linkedin.d2.balancer.clients.TrackerClient) UriProperties(com.linkedin.d2.balancer.properties.UriProperties) ClusterProperties(com.linkedin.d2.balancer.properties.ClusterProperties) URI(java.net.URI) RewriteClient(com.linkedin.d2.balancer.clients.RewriteClient)

Example 3 with ServiceUnavailableException

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

the class SimpleLoadBalancer method getPartitionInformation.

/**
   * If given a collection of keys, the method will maps keys to partitions and
   * return the servers that belongs to that partition up to limitHostPerPartition.
   *
   * If no keys are specified, the method will return hosts in all partitions
   *
   * @param serviceUri for example d2://articles
   * @param keys all the keys we want to find the partition for
   * @param limitHostPerPartition the number of hosts that we should return for this partition. Must be larger than 0.
   * @param hash this will be used to create Iterator for the hosts in the hash ring
   * @return Number of hosts in requested partitions. See {@link com.linkedin.d2.balancer.util.HostToKeyMapper} for more details.
   * @throws ServiceUnavailableException
   */
@Override
public <K> HostToKeyMapper<K> getPartitionInformation(URI serviceUri, Collection<K> keys, int limitHostPerPartition, int hash) throws ServiceUnavailableException {
    if (limitHostPerPartition <= 0) {
        throw new IllegalArgumentException("limitHostPartition cannot be 0 or less");
    }
    ServiceProperties service = listenToServiceAndCluster(serviceUri);
    String serviceName = service.getServiceName();
    String clusterName = service.getClusterName();
    ClusterProperties cluster = getClusterProperties(serviceName, clusterName);
    LoadBalancerStateItem<UriProperties> uriItem = getUriItem(serviceName, clusterName, cluster);
    UriProperties uris = uriItem.getProperty();
    List<LoadBalancerState.SchemeStrategyPair> orderedStrategies = _state.getStrategiesForService(serviceName, service.getPrioritizedSchemes());
    Map<Integer, Integer> partitionWithoutEnoughHost = new HashMap<Integer, Integer>();
    if (!orderedStrategies.isEmpty()) {
        // get the partitionId -> keys mapping
        final PartitionAccessor accessor = getPartitionAccessor(serviceName, clusterName);
        int maxPartitionId = accessor.getMaxPartitionId();
        List<K> unmappedKeys = new ArrayList<K>();
        Map<Integer, Set<K>> partitionSet = getPartitionSet(keys, accessor, unmappedKeys);
        final LoadBalancerState.SchemeStrategyPair pair = orderedStrategies.get(0);
        //get the partitionId -> host URIs list
        Map<Integer, KeysAndHosts<K>> partitionDataMap = new HashMap<Integer, KeysAndHosts<K>>();
        for (Integer partitionId : partitionSet.keySet()) {
            Set<URI> possibleUris = uris.getUriBySchemeAndPartition(pair.getScheme(), partitionId);
            List<TrackerClient> trackerClients = getPotentialClients(serviceName, service, possibleUris);
            int size = trackerClients.size() <= limitHostPerPartition ? trackerClients.size() : limitHostPerPartition;
            List<URI> rankedUri = new ArrayList<URI>(size);
            Ring<URI> ring = pair.getStrategy().getRing(uriItem.getVersion(), partitionId, trackerClients);
            Iterator<URI> iterator = ring.getIterator(hash);
            while (iterator.hasNext() && rankedUri.size() < size) {
                URI uri = iterator.next();
                if (!rankedUri.contains(uri)) {
                    rankedUri.add(uri);
                }
            }
            if (rankedUri.size() < limitHostPerPartition) {
                partitionWithoutEnoughHost.put(partitionId, limitHostPerPartition - rankedUri.size());
            }
            KeysAndHosts<K> keysAndHosts = new KeysAndHosts<K>(partitionSet.get(partitionId), rankedUri);
            partitionDataMap.put(partitionId, keysAndHosts);
        }
        return new HostToKeyMapper<K>(unmappedKeys, partitionDataMap, limitHostPerPartition, maxPartitionId + 1, partitionWithoutEnoughHost);
    } else {
        throw new ServiceUnavailableException(serviceName, "Unable to find a load balancer strategy");
    }
}
Also used : Set(java.util.Set) HashSet(java.util.HashSet) IdentityHashMap(java.util.IdentityHashMap) HashMap(java.util.HashMap) ArrayList(java.util.ArrayList) KeysAndHosts(com.linkedin.d2.balancer.util.KeysAndHosts) ServiceUnavailableException(com.linkedin.d2.balancer.ServiceUnavailableException) URI(java.net.URI) LoadBalancerState(com.linkedin.d2.balancer.LoadBalancerState) TrackerClient(com.linkedin.d2.balancer.clients.TrackerClient) UriProperties(com.linkedin.d2.balancer.properties.UriProperties) HostToKeyMapper(com.linkedin.d2.balancer.util.HostToKeyMapper) ServiceProperties(com.linkedin.d2.balancer.properties.ServiceProperties) PartitionAccessor(com.linkedin.d2.balancer.util.partitions.PartitionAccessor) ClusterProperties(com.linkedin.d2.balancer.properties.ClusterProperties)

Example 4 with ServiceUnavailableException

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

the class SimpleLoadBalancer method getRings.

@Override
public <K> MapKeyResult<Ring<URI>, K> getRings(URI serviceUri, Iterable<K> keys) throws ServiceUnavailableException {
    ServiceProperties service = listenToServiceAndCluster(serviceUri);
    String serviceName = service.getServiceName();
    String clusterName = service.getClusterName();
    ClusterProperties cluster = getClusterProperties(serviceName, clusterName);
    LoadBalancerStateItem<UriProperties> uriItem = getUriItem(serviceName, clusterName, cluster);
    UriProperties uris = uriItem.getProperty();
    List<LoadBalancerState.SchemeStrategyPair> orderedStrategies = _state.getStrategiesForService(serviceName, service.getPrioritizedSchemes());
    if (!orderedStrategies.isEmpty()) {
        LoadBalancerState.SchemeStrategyPair pair = orderedStrategies.get(0);
        PartitionAccessor accessor = getPartitionAccessor(serviceName, clusterName);
        // first distribute keys to partitions
        Map<Integer, Set<K>> partitionSet = new HashMap<Integer, Set<K>>();
        List<MapKeyResult.UnmappedKey<K>> unmappedKeys = new ArrayList<MapKeyResult.UnmappedKey<K>>();
        for (final K key : keys) {
            int partitionId;
            try {
                partitionId = accessor.getPartitionId(key.toString());
            } catch (PartitionAccessException e) {
                unmappedKeys.add(new MapKeyResult.UnmappedKey<K>(key, MapKeyResult.ErrorType.FAIL_TO_FIND_PARTITION));
                continue;
            }
            Set<K> set = partitionSet.get(partitionId);
            if (set == null) {
                set = new HashSet<K>();
                partitionSet.put(partitionId, set);
            }
            set.add(key);
        }
        // then we find the ring for each partition and create a map of Ring<URI> to Set<K>
        final Map<Ring<URI>, Collection<K>> ringMap = new IdentityHashMap<Ring<URI>, Collection<K>>(partitionSet.size() * 2);
        for (Map.Entry<Integer, Set<K>> entry : partitionSet.entrySet()) {
            int partitionId = entry.getKey();
            List<TrackerClient> clients = getPotentialClients(serviceName, service, uris, pair.getScheme(), partitionId);
            Ring<URI> ring = pair.getStrategy().getRing(uriItem.getVersion(), partitionId, clients);
            // make sure the same ring is not used in other partition
            Object oldValue = ringMap.put(ring, entry.getValue());
            assert (oldValue == null);
        }
        return new MapKeyResult<Ring<URI>, K>(ringMap, unmappedKeys);
    } else {
        throw new ServiceUnavailableException(serviceName, "Unable to find a load balancer strategy");
    }
}
Also used : Set(java.util.Set) HashSet(java.util.HashSet) IdentityHashMap(java.util.IdentityHashMap) HashMap(java.util.HashMap) IdentityHashMap(java.util.IdentityHashMap) ArrayList(java.util.ArrayList) ServiceUnavailableException(com.linkedin.d2.balancer.ServiceUnavailableException) URI(java.net.URI) LoadBalancerState(com.linkedin.d2.balancer.LoadBalancerState) TrackerClient(com.linkedin.d2.balancer.clients.TrackerClient) UriProperties(com.linkedin.d2.balancer.properties.UriProperties) MapKeyResult(com.linkedin.d2.balancer.util.MapKeyResult) ServiceProperties(com.linkedin.d2.balancer.properties.ServiceProperties) PartitionAccessor(com.linkedin.d2.balancer.util.partitions.PartitionAccessor) Ring(com.linkedin.d2.balancer.util.hashing.Ring) ClusterProperties(com.linkedin.d2.balancer.properties.ClusterProperties) Collection(java.util.Collection) PartitionAccessException(com.linkedin.d2.balancer.util.partitions.PartitionAccessException) Map(java.util.Map) IdentityHashMap(java.util.IdentityHashMap) HashMap(java.util.HashMap) TreeMap(java.util.TreeMap)

Example 5 with ServiceUnavailableException

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

the class SimpleLoadBalancer method listenToService.

private void listenToService(String serviceName) throws ServiceUnavailableException {
    if (_timeout > 0) {
        CountDownLatch latch = new CountDownLatch(1);
        SimpleLoadBalancerCountDownCallback callback = new SimpleLoadBalancerCountDownCallback(latch) {

            @Override
            public void done(int type, String name) {
                super.done(type, name);
            }
        };
        _state.listenToService(serviceName, callback);
        try {
            if (!latch.await(_timeout, _unit)) {
                warn(_log, "timed out during wait while trying to add service: ", serviceName);
            }
        } catch (InterruptedException e) {
            _log.error("got interrupt while waiting for a service to be registered", e);
            die(serviceName, "got interrupt while waiting for a service to be registered");
        }
    } else {
        _state.listenToService(serviceName, new NullStateListenerCallback());
        _log.info("No timeout for service {}", serviceName);
    }
}
Also used : NullStateListenerCallback(com.linkedin.d2.balancer.LoadBalancerState.NullStateListenerCallback) CountDownLatch(java.util.concurrent.CountDownLatch)

Aggregations

URI (java.net.URI)16 ServiceProperties (com.linkedin.d2.balancer.properties.ServiceProperties)15 Test (org.testng.annotations.Test)14 ClusterProperties (com.linkedin.d2.balancer.properties.ClusterProperties)12 UriProperties (com.linkedin.d2.balancer.properties.UriProperties)12 ServiceUnavailableException (com.linkedin.d2.balancer.ServiceUnavailableException)11 HashMap (java.util.HashMap)11 ArrayList (java.util.ArrayList)10 CountDownLatch (java.util.concurrent.CountDownLatch)10 DegraderLoadBalancerStrategyFactoryV3 (com.linkedin.d2.balancer.strategies.degrader.DegraderLoadBalancerStrategyFactoryV3)9 RequestContext (com.linkedin.r2.message.RequestContext)9 URIRequest (com.linkedin.d2.balancer.util.URIRequest)7 PropertyEventShutdownCallback (com.linkedin.d2.discovery.event.PropertyEventThread.PropertyEventShutdownCallback)7 Map (java.util.Map)7 None (com.linkedin.common.util.None)6 SimpleLoadBalancerSimulation (com.linkedin.d2.balancer.simulator.SimpleLoadBalancerSimulation)6 LoadBalancerStrategy (com.linkedin.d2.balancer.strategies.LoadBalancerStrategy)6 RandomLoadBalancerStrategyFactory (com.linkedin.d2.balancer.strategies.random.RandomLoadBalancerStrategyFactory)6 TransportClientFactory (com.linkedin.r2.transport.common.TransportClientFactory)6 HashSet (java.util.HashSet)6