What is causing Azure Event Hubs ReceiverDisconnectedException/LeaseLostException?

问题

I'm receiving events from an EventHub using EventProcessorHost and an IEventProcessor class (call it: MyEventProcessor). I scale this out to two servers by running my EPH on both servers, and having them connect to the Hub using the same ConsumerGroup, but unique hostName's (using the machine name).

The problem is: at random hours of the day/night, the app logs this:

Exception information: 
Exception type: ReceiverDisconnectedException 
Exception message: New receiver with higher epoch of '186' is created hence current receiver with epoch '186' is getting disconnected. If you are recreating the receiver, make sure a higher epoch is used.
  at Microsoft.ServiceBus.Common.ExceptionDispatcher.Throw(Exception exception)
  at Microsoft.ServiceBus.Common.Parallel.TaskHelpers.EndAsyncResult(IAsyncResult asyncResult)
  at Microsoft.ServiceBus.Messaging.IteratorAsyncResult`1.StepCallback(IAsyncResult result)

This Exception occurs at the same time as a LeaseLostException, thrown from MyEventProcessor's CloseAsync method when it tries to checkpoint. (Presumably Close is being called because of the ReceiverDisconnectedException?)

I think this is occurring due to Event Hubs' automatic lease management when scaling out to multiple machines. But I'm wondering if I need to do something different to make it work more cleanly and avoid these Exceptions? Eg: something with epochs?

回答1:

TLDR: This behavior is absolutely normal.

Why can't Lease Management be smooth & exception-free: To give more control on the situation to developer.

The really long story - all-the-way from Basics EventProcessorhost (hereby EPH - is very similar to what __consumer_offset topic does for Kafka Consumers - partition ownership & checkpoint store) is written by Microsoft Azure EventHubs team themselves - to translate all of the EventHubs partition receiver Gu into a simple onReceive(Events) callback.

EPH is used to address 2 general, major, well-known problems while reading out of a high-throughput partitioned streams like EventHubs:

fault tolerant receive pipe-line - for ex: a simpler version of the problem - if the host running a PartitionReceiver dies and comes back - it needs to resume processing from where it left. To remember the last successfully processed EventData, EPH uses the blob supplied to EPH constructor to store the checkpoints - when ever user invokes context.CheckpointAsync(). Eventually, when the host process dies (for ex: abruptly reboots or hits a hardware fault and never/comesback) - any EPH instance can pick up this task and resume from that Checkpoint.
Balance/distribute partitions across EPH instances - lets say, if there are 10 partitions and 2 EPH instances processing events from these 10 partitions - we need a way to divide partitions across the instances (PartitionManager component of EPH library does this). We use Azure Storage - Blob LeaseManagement-feature to implement this. As of version 2.2.10 - to simplify the problem, EPH assumes that all partitions are loaded equally.

With this, lets try to see what's going on: So, to start with, in the above example of 10 event hub partitions and 2 EPH instances processing events out of them:

lets say the first EPH instance - EPH1 started, at-first, alone and a part of start-up, it created receivers to all 10 partitions and is processing events. In the start up - EPH1 will announce that it owns all these 10 partitions by acquiring Leases on 10 storage blobs representing these 10 event hub partitions (with a standard nomenclature- which EPH internally creates in the Storage account - from the StorageConnectionString passed to the ctor). Leases will be acquired for a set time, after which the EPH instance will loose the ownership on this Partition.
EPH1 continually announces once in a while - that it is still owning those partitions - by renewing leases on the blob. Frequency of renewal, along with other useful tuning, can be performed using PartitionManagerOptions
now, lets say, EPH2 starts up - and you supplied the same AzureStorageAccount as EPH1 to the ctor of EPH2 as well. Right now, it has 0 partitions to process. So, to achieve balance of partitions across EPH instances, it will go ahead and download the list of all leaseblobs which has the mapping of owner to partitionId. From this, it will STEAL leases for its fair share of partitions - which is 5 in our example, and will announce that information on that lease blob. As part of this, EPH2 reads the latest checkpoint written by PartitionX it wants to steal the lease for and goes ahead and creates corresponding PartitionReceiver's with the EPOCH same as the one in the Checkpoint.
As a result, EPH1 will loose ownership of these 5 partitions and will run into different errors based on the exact state it is in.
- if EPH1 is actually invoking the PartitionReceiver.Receive() call - while EPH2 is creating the PartitionReceiver on the same receiver - EPH1 will experience ReceiverDisconnectedException. This will eventually, invoke IEventProcessor.Close(CloseReason=LeaseLost). Note that, probability of hitting this specific Exception is higher, if the messages being received are larger or the PrefetchCount is smaller - as in both cases the receiver would be performing more aggressive I/O.
- if EPH1 is in the state of checkpointing the lease or renewing the lease, while the EPH2 stole the lease, the EventProcessorOptions.ExceptionReceived eventHandler would be signaled with a leaselostException (with 409 conflict error on the leaseblob) - which also eventually invokes IEventProcess.Close(LeaseLost).

Why can't Lease Management be smooth & exception-free:

To keep the consumer simple and error-free, lease management related exceptions could have been swallowed by EPH and not notified to the user-code at all. However, we realized, throwing LeaseLostException could empower customers to find interesting bugs in IEventProcessor.ProcessEvents() callback - for which the symptom would be - frequent partition-moves

minor network outage on a specific machine - due to which EPH1 fails to renew leases and comes back up! - and imagine if the n/w of this machine stands flaky for a day - EPH instances are going to play ping-pong with Partitions! This machine will continuously try to steal the lease from other machine - which is legitimate from EPH point-of-view - but, is a total disaster for the user of EPH - as it completely interferes with the processing pipe. EPH - would exactly see a ReceiverDisconnectedException, when the n/w comes back up on this flaky m/c! We think the best and infact the only way is to enable the developer to smell this!
or a simple scenario like, having a bug in ProcessEvents logic - which throws unhandled exceptions which are fatal and brings down the whole process - ex: a poison event. This partition is going to move around a lot.
customers, performing write/delete operations on the same storage account which EPH is also using - by mistake (like an automated clean-up script) etc.
last but not the least - which we never wish could happen - say a 5 min outage on Azure d.c where a specific EventHub.Partition is located - say n/w incident. Partitions are going to move around across EPH instances.

Basically, in majority of situations, it would be tricky - for us to detect the diff. between these situations and a legitimate leaseLost due to balancing and we want to delegate control of these situations to the Developer.