I am using confluent parallel-consumer in order to acheive fast writes into different Data stores. I implemented my code and everything worked just fine locally with dockers.
Once I started several hosts with several consumers (with the same group id) I noticed that only one of the nodes (processes) is really consuming data. The topic I am reading from has 24 partitions, and I have 3 different nodes, I expected that kafka will split the work between them.
Here are parts of my code:
fun buildConsumer(config: KafkaConsumerConfig): KafkaConsumer<String, JsonObject> {
val props = Properties()
props[ConsumerConfig.BOOTSTRAP_SERVERS_CONFIG] = config.kafkaBootstrapServers
props[ConsumerConfig.AUTO_OFFSET_RESET_CONFIG] = "earliest"
props[ConsumerConfig.GROUP_ID_CONFIG] = "myGroup"
// Auto commit must be false in parallel consumer
props[ConsumerConfig.ENABLE_AUTO_COMMIT_CONFIG] = false
props[ConsumerConfig.KEY_DESERIALIZER_CLASS_CONFIG] = StringDeserializer::class.java.name
props[ConsumerConfig.VALUE_DESERIALIZER_CLASS_CONFIG] = JsonObjectDeSerializer::class.java.name
val consumer = KafkaConsumer<String, JsonObject>(props)
return consumer
}
private fun createReactParallelConsumer(): ReactorProcessor<String, JsonObject> {
val options = ParallelConsumerOptions.builder<String, JsonObject>()
.ordering(ParallelConsumerOptions.ProcessingOrder.KEY)
.maxConcurrency(10)
.batchSize(1)
.consumer(buildConsumer(kafkaConsumerConfig))
.build()
return ReactorProcessor(options)
}
And my main code:
pConsumer = createReactParallelConsumer()
pConsumer.subscribe(UniLists.of(kafkaConsumerConfig.kafkaTopic))
pConsumer.react { context ->
batchProcessor.processBatch(context)
}
Would appreciate any advice
We hit an issue that was closed in version 0.5.2.4 https://github.com/confluentinc/parallel-consumer/issues/409
The Parallel client kept old unfinished offsets, since our consumer was slow (many different reasons) we got to the end of the retention (earliest strategy), so every time we restarted the consumer, it was scanning all those incompatible offsets (which it did not truncate them - AKA the bug). Fix was just updating version from 0.5.2.3 to 0.5.2.4
Related
I have a piece of code that is essentially executing the following with Infinispan in embedded mode, using version 13.0.0 of the -core and -clustered-lock modules:
#Inject
lateinit var lockManager: ClusteredLockManager
private fun getLock(lockName: String): ClusteredLock {
lockManager.defineLock(lockName)
return lockManager.get(lockName)
}
fun createSession(sessionId: String) {
tryLockCounter.increment()
logger.debugf("Trying to start session %s. trying to acquire lock", sessionId)
Future.fromCompletionStage(getLock(sessionId).lock()).map {
acquiredLockCounter.increment()
logger.debugf("Starting session %s. Got lock", sessionId)
}.onFailure {
logger.errorf(it, "Failed to start session %s", sessionId)
}
}
I take this piece of code and deploy it to kubernetes. I then run it in six pods distributed over six nodes in the same region. The code exposes createSession with random Guids through an API. This API is called and creates sessions in chunks of 500, using a k8s service in front of the pods which means the load gets balanced over the pods. I notice that the execution time to acquire a lock grows linearly with the amount of sessions. In the beginning it's around 10ms, when there's about 20_000 sessions it takes about 100ms and the trend continues in a stable fashion.
I then take the same code and run it, but this time with twelve pods on twelve nodes. To my surprise I see that the performance characteristics are almost identical to when I had six pods. I've been digging in to the code but still haven't figured out why this is, I'm wondering if there's a good reason why infinispan here doesn't seem to perform better with more nodes?
For completeness the configuration of the locks are as follows:
val global = GlobalConfigurationBuilder.defaultClusteredBuilder()
global.addModule(ClusteredLockManagerConfigurationBuilder::class.java)
.reliability(Reliability.AVAILABLE)
.numOwner(1)
and looking at the code the clustered locks is using DIST_SYNC which should spread out the load of the cache onto the different nodes.
UPDATE:
The two counters in the code above are simply micrometer counters. It is through them and prometheus that I can see how the lock creation starts to slow down.
It's correctly observed that there's one lock created per session id, this is per design what we'd like. Our use case is that we want to ensure that a session is running in at least one place. Without going to deep into detail this can be achieved by ensuring that we at least have two pods that are trying to acquire the same lock. The Infinispan library is great in that it tells us directly when the lock holder dies without any additional extra chattiness between pods, which means that we have a "cheap" way of ensuring that execution of the session continues when one pod is removed.
After digging deeper into the code I found the following in CacheNotifierImpl in the core library:
private CompletionStage<Void> doNotifyModified(K key, V value, Metadata metadata, V previousValue,
Metadata previousMetadata, boolean pre, InvocationContext ctx, FlagAffectedCommand command) {
if (clusteringDependentLogic.running().commitType(command, ctx, extractSegment(command, key), false).isLocal()
&& (command == null || !command.hasAnyFlag(FlagBitSets.PUT_FOR_STATE_TRANSFER))) {
EventImpl<K, V> e = EventImpl.createEvent(cache.wired(), CACHE_ENTRY_MODIFIED);
boolean isLocalNodePrimaryOwner = isLocalNodePrimaryOwner(key);
Object batchIdentifier = ctx.isInTxScope() ? null : Thread.currentThread();
try {
AggregateCompletionStage<Void> aggregateCompletionStage = null;
for (CacheEntryListenerInvocation<K, V> listener : cacheEntryModifiedListeners) {
// Need a wrapper per invocation since converter could modify the entry in it
configureEvent(listener, e, key, value, metadata, pre, ctx, command, previousValue, previousMetadata);
aggregateCompletionStage = composeStageIfNeeded(aggregateCompletionStage,
listener.invoke(new EventWrapper<>(key, e), isLocalNodePrimaryOwner));
}
The lock library uses a clustered Listener on the entry modified event, and this one uses a filter to only notify when the key for the lock is modified. It seems to me the core library still has to check this condition on every registered listener, which of course becomes a very big list as the number of sessions grow. I suspect this to be the reason and if it is it would be really really awesome if the core library supported a kind of key filter so that it could use a hashmap for these listeners instead of going through a whole list with all listeners.
I believe you are creating a clustered lock per session id. Is this what you need ? what is the acquiredLockCounter? We are about to deprecate the "lock" method in favour of "tryLock" with timeout since the lock method will block forever if the clustered lock is never acquired. Do you ever unlock the clustered lock in another piece of code? If you shared a complete reproducer of the code will be very helpful for us. Thanks!
I make some performance testing of my PoC. What I saw is my actor is not receiving all messages that are sent to him and the performance is very low. I sent around 150k messages to my app, and it causes a peak on my processor to reach 100% utilization. But when I stop sending requests 2/3 of messages are not delivered to the actor. Here is a simple metrics from app insights:
To prove I have almost the same number of event persistent in mongo that my actor received messages.
Secondly, performance of processing messages is very disappointing. I get around 300 messages per second.
I know Akka.NET message delivery is at most once by default but I don't get any error saying that message were dropped.
Here is code:
Cluster shard registration:
services.AddSingleton<ValueCoordinatorProvider>(provider =>
{
var shardRegion = ClusterSharding.Get(_actorSystem).Start(
typeName: "values-actor",
entityProps: _actorSystem.DI().Props<ValueActor>(),
settings: ClusterShardingSettings.Create(_actorSystem),
messageExtractor: new ValueShardMsgRouter());
return () => shardRegion;
});
Controller:
[ApiController]
[Route("api/[controller]")]
public class ValueController : ControllerBase
{
private readonly IActorRef _valueCoordinator;
public ValueController(ValueCoordinatorProvider valueCoordinatorProvider)
{
_valueCoordinator = valuenCoordinatorProvider();
}
[HttpPost]
public Task<IActionResult> PostAsync(Message message)
{
_valueCoordinator.Tell(message);
return Task.FromResult((IActionResult)Ok());
}
}
Actor:
public class ValueActor : ReceivePersistentActor
{
public override string PersistenceId { get; }
private decimal _currentValue;
public ValueActor()
{
PersistenceId = Context.Self.Path.Name;
Command<Message>(Handle);
}
private void Handle(Message message)
{
Context.IncrementMessagesReceived();
var accepted = new ValueAccepted(message.ValueId, message.Value);
Persist(accepted, valueAccepted =>
{
_currentValue = valueAccepted.BidValue;
});
}
}
Message router.
public sealed class ValueShardMsgRouter : HashCodeMessageExtractor
{
public const int DefaultShardCount = 1_000_000_000;
public ValueShardMsgRouter() : this(DefaultShardCount)
{
}
public ValueShardMsgRouter(int maxNumberOfShards) : base(maxNumberOfShards)
{
}
public override string EntityId(object message)
{
return message switch
{
IWithValueId valueMsg => valueMsg.ValueId,
_ => null
};
}
}
akka.conf
akka {
stdout-loglevel = ERROR
loglevel = ERROR
actor {
debug {
unhandled = on
}
provider = cluster
serializers {
hyperion = "Akka.Serialization.HyperionSerializer, Akka.Serialization.Hyperion"
}
serialization-bindings {
"System.Object" = hyperion
}
deployment {
/valuesRouter {
router = consistent-hashing-group
routees.paths = ["/values"]
cluster {
enabled = on
}
}
}
}
remote {
dot-netty.tcp {
hostname = "desktop-j45ou76"
port = 5054
}
}
cluster {
seed-nodes = ["akka.tcp://valuessystem#desktop-j45ou76:5054"]
}
persistence {
journal {
plugin = "akka.persistence.journal.mongodb"
mongodb {
class = "Akka.Persistence.MongoDb.Journal.MongoDbJournal, Akka.Persistence.MongoDb"
connection-string = "mongodb://localhost:27017/akkanet"
auto-initialize = off
plugin-dispatcher = "akka.actor.default-dispatcher"
collection = "EventJournal"
metadata-collection = "Metadata"
legacy-serialization = off
}
}
snapshot-store {
plugin = "akka.persistence.snapshot-store.mongodb"
mongodb {
class = "Akka.Persistence.MongoDb.Snapshot.MongoDbSnapshotStore, Akka.Persistence.MongoDb"
connection-string = "mongodb://localhost:27017/akkanet"
auto-initialize = off
plugin-dispatcher = "akka.actor.default-dispatcher"
collection = "SnapshotStore"
legacy-serialization = off
}
}
}
}
So there are two issues going on here: actor performance and missing messages.
It's not clear from your writeup, but I'm going to make an assumption: 100% of these messages are going to a single actor.
Actor Performance
The end-to-end throughput of a single actor depends on:
The amount of work it takes to route the message to the actor (i.e. through the sharding system, hierarchy, over the network, etc)
The amount of time it takes the actor to process a single message, as this determines the rate at which a mailbox can be emptied; and
Any flow control that affects which messages can be processed when - i.e. if an actor uses stashing and behavior switching, the amount of time an actor spends stashing messages while waiting for its state to change will have a cumulative impact on the end-to-end processing time for all stashed messages.
You will have poor performance due to item 3 on this list. The design that you are implementing calls Persist and blocks the actor from doing any additional processing until the message is successfully persisted. All other messages sent to the actor are stashed internally until the previous one is successfully persisted.
Akka.Persistence offers four options for persisting messages from the point of view of a single actor:
Persist - highest consistency (no other messages can be processed until persistence is confirmed), lowest performance;
PersistAsync - lower consistency, much higher performance. Doesn't wait for the message to be persisted before processing the next message in the mailbox. Allows multiple messages from a single persistent actor to be processed concurrently in-flight - the order in which those events are persisted will be preserved (because they're sent to the internal Akka.Persistence journal IActorRef in that order) but the actor will continue to process additional messages before the persisted ones are confirmed. This means you probably have to modify your actor's in-memory state before you call PersistAsync and not after the fact.
PersistAll - high consistency, but batches multiple persistent events at once. Same ordering and control flow semantics as Persist - but you're just persisting an array of messages together.
PersistAllAsync - highest performance. Same semantics as PersistAsync but it's an atomic batch of messages in an array being persisted together.
To get an idea as to how the performance characteristics of Akka.Persistence changes with each of these methods, take a look at the detailed benchmark data the Akka.NET organization has put together around Akka.Persistence.Linq2Db, the new high performance RDBMS Akka.Persistence library: https://github.com/akkadotnet/Akka.Persistence.Linq2Db#performance - it's a difference between 15,000 per second and 250 per second on SQL; the write performance is likely even higher in a system like MongoDB.
One of the key properties of Akka.Persistence is that it intentionally routes all of the persistence commands through a set of centralized "journal" and "snapshot" actors on each node in a cluster - so messages from multiple persistent actors can be batched together across a small number of concurrent database connections. There are many users running hundreds of thousands of persistent actors simultaneously - if each actor had their own unique connection to the database it would melt even the most robustly vertically scaled database instances on Earth. This connection pooling / sharing is why the individual persistent actors rely on flow control.
You'll see similar performance using any persistent actor framework (i.e. Orleans, Service Fabric) because they all employ a similar design for the same reasons Akka.NET does.
To improve your performance, you will need to either batch received messages together and persist them in a group with PersistAll (think of this as de-bouncing) or use asynchronous persistence semantics using PersistAsync.
You'll also see better aggregate performance if you spread your workload out across many concurrent actors with different entity ids - that way you can benefit from actor concurrency and parallelism.
Missing Messages
There could be any number of reasons why this might occur - most often it's going to be the result of:
Actors being terminated (not the same as restarting) and dumping all of their messages into the DeadLetter collection;
Network disruptions resulting in dropped connections - this can happen when nodes are sitting at 100% CPU - messages that are queued for delivery at the time can be dropped; and
The Akka.Persistence journal receiving timeouts back from the database will result in persistent actors terminating themselves due to loss of consistency.
You should look for the following in your logs:
DeadLetter warnings / counts
OpenCircuitBreakerExceptions coming from Akka.Persistence
You'll usually see both of those appear together - I suspect that's what is happening to your system. The other possibility could be Akka.Remote throwing DisassociationExceptions, which I would also look for.
You can fix the Akka.Remote issues by changing the heartbeat values for the Akka.Cluster failure-detector in configuration https://getakka.net/articles/configuration/akka.cluster.html:
akka.cluster.failure-detector {
# FQCN of the failure detector implementation.
# It must implement akka.remote.FailureDetector and have
# a public constructor with a com.typesafe.config.Config and
# akka.actor.EventStream parameter.
implementation-class = "Akka.Remote.PhiAccrualFailureDetector, Akka.Remote"
# How often keep-alive heartbeat messages should be sent to each connection.
heartbeat-interval = 1 s
# Defines the failure detector threshold.
# A low threshold is prone to generate many wrong suspicions but ensures
# a quick detection in the event of a real crash. Conversely, a high
# threshold generates fewer mistakes but needs more time to detect
# actual crashes.
threshold = 8.0
# Number of the samples of inter-heartbeat arrival times to adaptively
# calculate the failure timeout for connections.
max-sample-size = 1000
# Minimum standard deviation to use for the normal distribution in
# AccrualFailureDetector. Too low standard deviation might result in
# too much sensitivity for sudden, but normal, deviations in heartbeat
# inter arrival times.
min-std-deviation = 100 ms
# Number of potentially lost/delayed heartbeats that will be
# accepted before considering it to be an anomaly.
# This margin is important to be able to survive sudden, occasional,
# pauses in heartbeat arrivals, due to for example garbage collect or
# network drop.
acceptable-heartbeat-pause = 3 s
# Number of member nodes that each member will send heartbeat messages to,
# i.e. each node will be monitored by this number of other nodes.
monitored-by-nr-of-members = 9
# After the heartbeat request has been sent the first failure detection
# will start after this period, even though no heartbeat mesage has
# been received.
expected-response-after = 1 s
}
Bump the acceptable-heartbeat-pause = 3 s value to something larger like 10,20,30 if needed.
Sharding Configuration
One last thing I want to point out with your code - the shard count is way too high. You should have about ~10 shards per node. Reduce it to something reasonable.
Situation is the following.
We have setup SSL + ACLs in Kafka Broker.
We are setting up stream, which reads messages from two topics:
KStream<String, String> stringInput
= kBuilder.stream( STRING_SERDE, STRING_SERDE, inTopicName );
stringInput
.filter( streamFilter::passOrFilterMessages )
.map( processor )
.to( outTopicName );
It is done like two times (in the loop).
Then we are setting general error handler:
streams.setUncaughtExceptionHandler( ( Thread t, Throwable e ) -> {
synchronized ( this ) {
LOG.fatal( ... );
this.stop();
}
}
);
Problem is the following. If for example in one topic certificate is no more valid. The stream is throwing exception Not authorized to access topics ...
So far so good.
But the exception is handled by general error handler, so the complete application stops even if the second topic has no problems.
The question is, how to handle this exception per topic?
How to avoid situation that at some moment complete application stops due to the problem that one single topic has problems with authorization?
I understand that if Broker is not available, then complete app may stop. But if only one topic is not available, then single stream shall stop, and not complete application, or?
By design, Kafka Streams treats the topology a one and cannot distinguish between both parts. For your specific case, as you loop and build to independent pipelines, you could run two KafkaStreams instances in parallel (within the same application/JVM) to isolate both from each other. Thus, if one fails, the other one is not affected. You would need to use two different application.id for both instances.
I'd like to to use an Ignite cluster to warm a PARTITIONED cache from an existing database. The existing database is not partitioned and expensive to scan, so I'd like to perform a single scan when the cache is created by the cluster. Once the job completes, the result would be a cache containing all data from the existing database partitioned and evenly distributed across the cluster.
How do you implement a job that runs when a cache is created by Ignite?
Ignite integrates with underlying stores via CacheStore [1] implementations. Refer to [2] for details about your particular use case.
[1] https://apacheignite.readme.io/docs/persistent-store
[2] https://apacheignite.readme.io/docs/data-loading
You can create a Service that runs once on cluster start and then cancels itself. It can use a cache to store state, so it will not run if it's deployed in the cluster a second time.
The following abstract Service runs executeOnce once per cluster the first time it's deployed after cluster start:
abstract class ExecuteOnceService extends Service {
val ExecuteOnceCacheName = "_execute_once_service"
val config = new CacheConfiguration[String, java.lang.Boolean](ExecuteOnceCacheName)
.setCacheMode(CacheMode.PARTITIONED)
.setAtomicityMode(CacheAtomicityMode.TRANSACTIONAL)
#IgniteInstanceResource
var ignite: Ignite = _
override def execute(ctx: ServiceContext): Unit = {
val cache = ignite.getOrCreateCache(config)
val executed = cache.getAndPutIfAbsent(ctx.name(), java.lang.Boolean.TRUE)
if (executed != java.lang.Boolean.TRUE) executeOnce(ctx)
ignite.services().cancel(ctx.name())
}
def executeOnce(ctx: ServiceContext): Unit
}
The code below is from source code of Alamofire
let queue = dispatch_queue_create(nil, DISPATCH_QUEUE_SERIAL)
public func request(URLRequest: URLRequestConvertible) -> Request {
var dataTask: NSURLSessionDataTask!
dispatch_sync(queue) { dataTask = self.session.dataTaskWithRequest(URLRequest.URLRequest) }
let request = Request(session: session, task: dataTask)
self.delegate[request.delegate.task] = request.delegate
if startRequestsImmediately {
request.resume()
}
return request
}
It seems like every time it creates a dataTask, it dispatch that creating process to a serial queue. Would this measure protect the program from any kind of multi-thread trap?
I can't figure out what's the difference without that queue.
The reason why we implemented that check is due to Alamofire Issue #393. We were seeing duplicate task identifiers without the serial queue when creating data and upload tasks in parallel from multiple threads. It appears that Apple has a thread safety issue when incrementing the task identifiers. Therefore in Alamofire, we eliminate the issue by creating the tasks on a serial queue.
Cheers. 🍻