上一篇说到实际的核心逻辑是这2行代码
LookupOptions options = LookupOptions.builder()
.authoritative(authoritative)
.advertisedListenerName(advertisedListenerName)
.loadTopicsInBundle(true) // 这里这个条件是true
.build();
pulsarService.getNamespaceService().getBrokerServiceUrlAsync(topicName, options)
这里传递的参数将loadTopicsInBundle 设置了成true。我们看下在处理lookup请求过程中是否有loadtopic的逻辑。
NamespaceService.findBrokerServiceUrl
这个函数我们注意到有ownershipCache.getOwnerAsync 和searchForCandidateBroker 这2个地方没有细说
我们先看一下ownershipCache。
private CompletableFuture> findBrokerServiceUrl(
NamespaceBundle bundle, LookupOptions options) {
....
return targetMap.computeIfAbsent(bundle, (k) -> {
...
ownershipCache.getOwnerAsync(bundle)
.thenAccept(nsData -> {
// nsData : Optional
if (!nsData.isPresent()) {
...
// 目前还没有人负责这个bundle 尝试查找这个bundle的owner
pulsar.getExecutor().execute(() -> {
searchForCandidateBroker(bundle, future, options);
});
...
}
...
});
}
OwnerShipCache类
从javadoc 里面可以知道这个类的主要功能。
- cache zk里面关于 service unit 的ownership信息
- 提供zk的读写功能
- 可以用来查找owner信息
- 可以用来获取一个 service unit 的ownership
getOwnerAsync 这个方法主要是查看zk cache里面是否有信息，如果没有信息，则尝试读取zk节点，
如果节点有信息则说明有人拿到了这个bundle的ownership
如果这个节点就是当前机器，则会通知bundle load的信息给listener
如果这个节点没有信息，说明当前还没有人负责这个bundle。
// org.apache.pulsar.broker.namespace.OwnerShipCache
public
CompletableFuture>
getOwnerAsync(NamespaceBundle suName)
{
// 这里的路径是 /namespace/{namespace}/0x{lowerEndpoint}_0x{upperEndpoint}
String path = ServiceUnitZkUtils.path(suName);
// ownedBundleFuture 还是一个 AsyncLoadingCache
// 这里不会尝试去加载这个cache信息，因为调用的getIfPresent
CompletableFuture ownedBundleFuture = ownedBundlesCache.getIfPresent(path);
// 如果之前有内容的话就说明当前broker是owner（这部分逻辑在cache的加载代码里面，后面会说）
if (ownedBundleFuture != null) {
// Either we're the owners or we're trying to become the owner.
return ownedBundleFuture.thenApply(serviceUnit -> {
// We are the owner of the service unit
return Optional.of(serviceUnit.isActive() ? selfOwnerInfo : selfOwnerInfoDisabled);
});
}
// 如果cache里面没有，我们确认下当前的owner是谁。
// If we're not the owner, we need to check if anybody else is
return resolveOwnership(path)
.thenApply(optional -> optional.map(Map.Entry::getKey));
}
private CompletableFuture>> resolveOwnership(String path) {
return ownershipReadOnlyCache.getWithStatAsync(path) // 这个逻辑是从zk里面读取这个bundle路径下的内容
.thenApply(optionalOwnerDataWithStat -> {
// 如果这个路径下有数据，则说明有人已经成功获取了这个bundle的ownership信息
if (optionalOwnerDataWithStat.isPresent()) {
Map.Entry ownerDataWithStat = optionalOwnerDataWithStat.get();
Stat stat = ownerDataWithStat.getValue();
// 如果这个zk临时节点的owner就是当前的broker
if (stat.getEphemeralOwner() == localZkCache.getZooKeeper().getSessionId()) {
LOG.info("Successfully reestablish ownership of {}", path);
// 这里是更新缓存的逻辑
OwnedBundle ownedBundle = new OwnedBundle(ServiceUnitZkUtils.suBundleFromPath(path, bundleFactory));
if (selfOwnerInfo.getNativeUrl().equals(ownerDataWithStat.getKey().getNativeUrl())) {
ownedBundlesCache.put(path, CompletableFuture.completedFuture(ownedBundle));
}
ownershipReadOnlyCache.invalidate(path);
// 这里会通知callback（和主要逻辑无关）
namespaceService.onNamespaceBundleOwned(ownedBundle.getNamespaceBundle());
}
}
// 这里返回的是一个Optional对象，如果这个节点不存在的话返回的实际是一个Empty
// 说明这个时候没有人负责这个bundle
// 也可能返回带有信息的optional，这时候负责这个节点的broker可能是当前机器也可能是其他机器。
return optionalOwnerDataWithStat;
});
}
我们看一下如果没有任何人负责这个bundle的情况。
NamespaceService.searchForCandidateBroker
这个方法的逻辑是选出当前这个bundle的owner是哪个broker
主要依靠LeaderElectionService 和LoadManager 选出。
如果选出来的broker是本机的话，则会尝试获取这个bundle的ownership。
如果是其他机器的话则会把这个请求转发给其他机器，请求其他机器来获取ownership。
private void searchForCandidateBroker(NamespaceBundle bundle,
CompletableFuture> lookupFuture,
LookupOptions options) {
...
// 首先会按照一定逻辑来选出这个bundle的可能的broker节点
String candidateBroker = null;
...
boolean authoritativeRedirect = les.isLeader();
try {
// check if this is Heartbeat or SLAMonitor namespace
...
if (candidateBroker == null) {
if (options.isAuthoritative()) {
// leader broker already assigned the current broker as owner
candidateBroker = pulsar.getSafeWebServiceAddress();
} else
// 如果这个LeaderElectionService 是leader ||
// 不是中心化的loadManager（这个是均衡负载用的）||
// 如果当前这个leader的broker还不是active的
if (!this.loadManager.get().isCentralized()
|| pulsar.getLeaderElectionService().isLeader()
// If leader is not active, fallback to pick the least loaded from current broker loadmanager
|| !isBrokerActive(pulsar.getLeaderElectionService().getCurrentLeader().getServiceUrl())
) {
// 从loadManager选一个负载最轻的broker出来
Optional availableBroker = getLeastLoadedFromLoadManager(bundle);
if (!availableBroker.isPresent()) {
lookupFuture.complete(Optional.empty());
return;
}
candidateBroker = availableBroker.get();
authoritativeRedirect = true;
} else {
// forward to leader broker to make assignment
candidateBroker = pulsar.getLeaderElectionService().getCurrentLeader().getServiceUrl();
}
}
} catch (Exception e) {
...
}
// 到这里就选出一个候选的broker地址了
try {
checkNotNull(candidateBroker);
// 如果这个候选broker就是当前机器
if (candidateBroker.equals(pulsar.getSafeWebServiceAddress())) {
...
// 这里使用ownerShipCache尝试获取这个bundle的ownership
ownershipCache.tryAcquiringOwnership(bundle)
.thenAccept(ownerInfo -> {
...
// 这里就是文章开始的时候说的是否需要load 所有在bundle里面的topic
if (options.isLoadTopicsInBundle()) {
// Schedule the task to pre-load topics
pulsar.loadNamespaceTopics(bundle);
}
// find the target
// 走到这里说明已经把当前的broker作为这个bundle的owner了，直接返回本机的信息给请求者
lookupFuture.complete(Optional.of(new LookupResult(ownerInfo)));
return;
}
}).exceptionally(exception -> {
...
});
} else {
...
// 这里是把这个lookup 请求转发给其他broker
// Load managed decider some other broker should try to acquire ownership
// Now setting the redirect url
createLookupResult(candidateBroker, authoritativeRedirect, options.getAdvertisedListenerName())
.thenAccept(lookupResult -> lookupFuture.complete(Optional.of(lookupResult)))
.exceptionally(ex -> {
lookupFuture.completeExceptionally(ex);
return null;
});
}
} catch (Exception e) {
...
}
}
OwnershipCache.tryAcquiringOwnership
这里就是尝试获取这个bundle的ownership的逻辑了。
只需要在zk上记录当前节点的信息就可以了。
（也会有维护这个cache的逻辑）
public CompletableFuture
tryAcquiringOwnership(NamespaceBundle bundle) throws Exception {
String path = ServiceUnitZkUtils.path(bundle);
CompletableFuture future = new CompletableFuture<>();
...
LOG.info("Trying to acquire ownership of {}", bundle);
// 这里调用的是get，这个方法会触发cache加载的逻辑。
// Doing a get() on the ownedBundlesCache will trigger an async ZK write to acquire the lock over the
// service unit
ownedBundlesCache.get(path)
.thenAccept(namespaceBundle -> {
// 到这里说明已经获得了这个bundle的ownership了,直接返回。
LOG.info("Successfully acquired ownership of {}", path);
namespaceService.onNamespaceBundleOwned(bundle);
future.complete(selfOwnerInfo);
}).exceptionally(exception -> {
// 这里如果加载过程中出现问题（可能是其他人成为了leader）
// Failed to acquire ownership
if (exception instanceof CompletionException
&& exception.getCause() instanceof KeeperException.NodeExistsException) {
// 确认当前的leader是谁
resolveOwnership(path)
.thenAccept(optionalOwnerDataWithStat -> {
// 这里会拿到之前成功获得ownership的节点信息
if (optionalOwnerDataWithStat.isPresent()) {
Map.Entry ownerDataWithStat = optionalOwnerDataWithStat.get();
NamespaceEphemeralData ownerData = ownerDataWithStat.getKey();
Stat stat = ownerDataWithStat.getValue();
if (stat.getEphemeralOwner() != localZkCache.getZooKeeper().getSessionId()) {
LOG.info("Failed to acquire ownership of {} -- Already owned by broker {}",
path, ownerData);
}
// 直接返回即可
future.complete(ownerData);
} else {
...
}{
}).exceptionally(ex -> {
....
});
} else {
...
}
return null;
});
return future;
}
OwnershipCache 加载逻辑
这里逻辑比较简单，序列化本机的连接信息，写入到这个bundle的path下面就行了
private class OwnedServiceUnitCacheLoader implements AsyncCacheLoader {
@SuppressWarnings("deprecation")
@Override
public CompletableFuture asyncLoad(String namespaceBundleZNode, Executor executor) {
if (LOG.isDebugEnabled()) {
LOG.debug("Acquiring zk lock on namespace {}", namespaceBundleZNode);
}
byte[] znodeContent;
try {
znodeContent = jsonMapper.writeValueAsBytes(selfOwnerInfo);
} catch (JsonProcessingException e) {
// Failed to serialize to JSON
return FutureUtil.failedFuture(e);
}
CompletableFuture future = new CompletableFuture<>();
ZkUtils.asyncCreateFullPathOptimistic(localZkCache.getZooKeeper(), namespaceBundleZNode, znodeContent,
Ids.OPEN_ACL_UNSAFE, CreateMode.EPHEMERAL, (rc, path, ctx, name) -> {
if (rc == KeeperException.Code.OK.intValue()) {
if (LOG.isDebugEnabled()) {
LOG.debug("Successfully acquired zk lock on {}", namespaceBundleZNode);
}
ownershipReadOnlyCache.invalidate(namespaceBundleZNode);
future.complete(new OwnedBundle(
ServiceUnitZkUtils.suBundleFromPath(namespaceBundleZNode, bundleFactory)));
} else {
// Failed to acquire lock
future.completeExceptionally(KeeperException.create(rc));
}
}, null);
return future;
}
}
加载bundle下所有topic
到这里我们已经可以拿到bundle的ownership了。我们看一下之前加载所有topic的逻辑。
PulsarService.loadNamespaceTopics
public void loadNamespaceTopics(NamespaceBundle bundle) {
executor.submit(() -> {
NamespaceName nsName = bundle.getNamespaceObject();
List> persistentTopics = Lists.newArrayList();
long topicLoadStart = System.nanoTime();
for (String topic : getNamespaceService().getListOfPersistentTopics(nsName).join()) {
try {
TopicName topicName = TopicName.get(topic);
if (bundle.includes(topicName)) {
// 到这里会创建一个Topic对象保存在BrokerService里面
// 这部分后面会说，涉及到 ManagedLedger 里面的初始化
CompletableFuture future = brokerService.getOrCreateTopic(topic);
if (future != null) {
persistentTopics.add(future);
}
}
}
...
}
...
return null;
});
}
NamespaceService.getListOfPersistentTopics
这里就比较容易了
读取zk的/managed-ledgers/%s/persistent所有子节点即可。
public CompletableFuture> getListOfPersistentTopics(NamespaceName namespaceName) {
// For every topic there will be a managed ledger created.
String path = String.format("/managed-ledgers/%s/persistent", namespaceName);
if (LOG.isDebugEnabled()) {
LOG.debug("Getting children from managed-ledgers now: {}", path);
}
return pulsar.getLocalZkCacheService().managedLedgerListCache().getAsync(path)
.thenApply(znodes -> {
List topics = Lists.newArrayList();
for (String znode : znodes) {
topics.add(String.format("persistent://%s/%s", namespaceName, Codec.decode(znode)));
}
topics.sort(null);
return topics;
});
}
具体加载一个topic的逻辑我们后面再说。