Contact Us 1-800-596-4880

Flow Processing Strategies

Mule Runtime Engine versions 3.5, 3.6, and 3.7 reached End of Life on or before January 25, 2020. For more information, contact your Customer Success Manager to determine how you can migrate to the latest Mule version.

A flow processing strategy determines how Mule implements message processing for a given flow.

  • Should the message be processed synchronously (on the same thread) or asynchronously (on a different thread)?

  • If asynchronously, what are the properties of the pool of threads used to process the messages?

  • If asynchronously, how will messages wait for their turn to be processed in the second thread?

All Mule flows have an implicit processing strategy which Mule applies automatically: either synchronous or queued-asynchronous. Each of these processing strategies is optimal for certain flows. This page introduces these processing strategies and the criteria Mule uses to determine the optimal processing strategy for each flow.

This page also explains the circumstances under which you may want to override Mule’s choice of processing strategy and covers the parameters you can change or fine-tune in the processing strategy used for a given flow.

Assumptions

This page assumes that you are familiar with Mule flows and basic application architecture.

Implicit Processing Strategies

Even if you do not configure a processing strategy for your flow, the flow is following a processing strategy automatically applied by Mule. The processing strategy is either:

For the majority of use cases, the implicit strategy selected by Mule is optimal. See Changing the Processing Strategy for more information about why and how you can override the implicit processing strategy, if your use case demands it.

How Mule Selects a Flow Processing Strategy

Each flow varies in the degree to which it can benefit from the transactional reliability of synchronous processing, as opposed to the high throughput of the queued-asynchronous alternative.

Mule selects a processing strategy for a flow based on two criteria:

  • The flow’s exchange pattern

  • Whether or not the flow is transactional

A flow employs one of two exchange patterns, determined by the exchange pattern of the flow’s inbound endpoint:

  • a request-response exchange pattern is used when the sender of the messages (that is, the flow’s inbound endpoint) expects a response or some kind of result to return to the caller.

  • a one-way exchange pattern is used when no response is expected, or when the result is sent to some other flow or system and not back to the caller.

If the exchange pattern is request-response, Mule applies a synchronous processing strategy.

In addition, some inbound endpoints can be configured to operate transactionally, regardless of their exchange pattern. If a flow is transactional, Mule applies a synchronous processing strategy.

The following table summarizes how Mule chooses a flow processing strategy:

Exchange Pattern Transactional? Flow Processing Strategy

Request-Response

Yes

Synchronous

Request-Response

No

Synchronous

One-way

Yes

Synchronous

One-way

No

Queued-Asynchronous

The Synchronous Flow Processing Strategy

The synchronous approach is used to process messages in the same thread that initially received the message. After the flow receives a message, all processing, including the processing of the response, is done in that same thread (with the exception of asynchronous scopes like Async and For Each.) The synchronous strategy is ideally suited to flows where:

  • The sender of the message expects a response. This is known as a "request-response" exchange pattern.

  • The flow needs to meet the requirements of transactional processing. In other words, all the steps in the flow are considered a single unit, which must succeed entirely or fail entirely. Additionally, appropriate parties (such as the sender of the message or the administrator of the business process encapsulated by the flow) must be notified of the result. This means that a transactional flow must not hand off processing to other threads, where errors can occur after the transaction is completed.

  • The flow’s inbound endpoint must be notified of all errors that occur during the processing of the message. This situation is discussed further in Reliability Patterns.

The Queued-Asynchronous Flow Processing Strategy

The queued-asynchronous approach uses a queue to decouple the flow’s receiver from the rest of the steps in the flow.

After the receiver places a message into a queue, it can immediately return and accept a new incoming message. Each message waiting in the queue can be assigned a different thread from a pool of threads. A component called a Work Manager assigns pending messages to available threads, so they can be processed in parallel. Such parallel processing is ideal for situations where the receiver can, at peak times, accept messages significantly faster than the rest of the flow can process those messages.

However, the increased throughput facilitated by the asynchronous approach comes at the cost of transactional reliability.

The specific type of queue implemented for the queued-asynchronous flow processing strategy is known as a SEDA queue.

Behavior in a Cluster

A flow with an queued-asynchronous processing strategy can execute on any node in a cluster.

A flow with a synchronous processing strategy executes on the same node of the cluster until processing in the flow is complete.

Changing the Processing Strategy

Mule’s choice of processing strategy is almost always optimal for the flow to which it is applied. However, the following options exist:

  • In cases where Mule has selected queued-asynchronous processing for a flow, you can specify a synchronous flow instead. You might want to specify a synchronous flow to achieve reliability. Remember that the synchronous strategy is ideally suited for flows where the flow’s inbound endpoint must be notified of all errors that occur during the processing of the message. The Reliability Patterns document discusses this situation further. Although this is not applicable to most use cases, if Mule has selected queued-asynchronous processing for a flow, you can specify another kind of asynchronous processing strategy instead. You cannot force a queued-asynchronous processing strategy on a synchronous flow.

  • You can accept Mule’s choice of the queued-asynchronous flow processing strategy, but then fine-tune the processing strategy. Note that you cannot do any fine-tuning for a synchronous flow.

  • You can wrap one or more message processors within your synchronous flow in an asynchronous scope to define a processing block that Mule processes asynchronously with the main flow. You can also create an asynchronous flow by wrapping a flow-ref in an asynchronous scope.

  • You can call out to a subflow (which always processes synchronously) or a synchronous flow from any flow, regardless of what processing strategy is applied to the triggering flow. The triggering flow waits until this synchronous processing block returns a message before proceeding with processing. See Flows and Subflows for more information about subflows and synchronous flows.

  • You can create a custom flow processing strategy to fit your exact needs. For instance, you might prefer a queued-asynchronous flow that uses an increased number of threads to handle high peak loads.

Summary

Flow Processing Strategy implictly applied by Mule Can you specify a different
processing strategy?
Can you fine-tune
the processing strategy?
Can you create a processing block that executes using a different processing strategy from the main flow? Can you apply a custom processing strategy using Spring?

Synchronous

No. You cannot force a flow with a request-response exchange pattern and/or transactionality to be asynchronous.

No. You cannot fine-tune the synchronous processing strategy. You can, however, customize the inbound endpoint connector receiver threading profile.

Yes. You can use an Async Scope or an asynchronous flow to cause Mule to process a selected block of message processors asynchronously. Creating an Asynchronous Processing Block.

Yes. Creating a Custom Processing Strategy.

Queued-Asynchronous

Yes. You can override an implicitly selected queued-asynchronous processing strategy by explicitly declaring a synchronous processing strategy (or, in rare use cases, a different kind of asynchronous processing strategy) instead. Specifying a Processing Strategy.

Yes. You can fine-tune the queued-asynchronous processing strategy to meet your needs. Fine-Tuning a Queued-Asynchronous Processing Strategy.

Yes. A synchronous flow or subflow processes a selected block of message processors synchronously, regardless of the processing strategy of the main flow.

Yes. Creating a Custom Processing Strategy.

Specifying a Processing Strategy

The procedure to change a processing strategy for an individual flow is straightforward. You may only specify a processing strategy for flows to which Mule would implicitly apply a queued-asynchronous processing strategy. Thus, only flows with one-way exchange patterns and which are not transactional can have a processing strategy specified to override Mule’s selection of the queued-asynchronous processing strategy.

The most frequent use case for specifying a processing strategy is to force a flow that would otherwise be queued-asynchronous to be synchronous instead. To force a flow to be synchronous, add the processingStrategy attribute to the flow that you want to change and set it to synchronous. This is illustrated in the code example below.

<flow name="asynchronousToSynchronous" processingStrategy="synchronous">
  <vm:inbound-endpoint path="anyUniqueEndpointName" exchange-pattern="one-way"/>
  <vm:outbound-endpoint path="output" exchange-pattern="one-way"/>
</flow>

In rare use cases, you might want to specify another kind of asynchronous processing strategy on a flow that would otherwise follow a queued-asynchronous processing strategy. You can either declare it directly in the flow configuration, just as in the example immediately above which declares a synchronous processing strategy, or you can create a global element and further fine tune the processing strategy. The table below lists the processing strategy names, which can each be declared as values for the processingStrategy attribute, or as a global element.

Additional Asynchronous Processing Strategy Element/Global Element Names Description

asynchronous-processing-strategy

Not applicable to most use cases. Same as queued-asynchronous processing strategy except that it doesn’t use a queue. Use this only if for some reason you do not want your processing to be distributed across nodes.

queued-thread-per-processor-processing-strategy

Not applicable to most use cases. Writes messages to a queue, then every processor in the scope runs sequentially in a different thread.

thread-per-processor-processing-strategy

Not applicable to most use cases. Every processor in the scope runs sequentially in a different thread.

Fine-Tuning a Queued-Asynchronous Processing Strategy

If Mule has applied a queued-asynchronous flow processing strategy, you can fine-tune it to adjust how it behaves. Note that you can only fine-tune a queued-asynchronous strategy; you cannot do any fine-tuning for a synchronous flow.

You can fine-tune a queued-asynchronous processing strategy by:

  • Changing the number of threads available to the flow.

  • Limiting the number of messages that can be queued.

  • Specifying a queue store to persist data.

You achieve this fine-tuning by specifying parameters for a global processing strategy, then referencing the parameters within the flow or flows you wish to fine-tune. If you don’t specify a certain configuration parameter at either the global or local levels, Mule sets a default value for the parameter. The table below lists these default values.

In addition to fine-tuning the attributes of the queued-asynchronous processing strategy, you can also tune performance of your Mule application by analyzing and configuring thread profiles.

The following example defines a global processing strategy (queued-asynchronous-processing-strategy), which sets the maximum number of threads (maxThreads) to 500. The example also shows how a flow references the global processing strategy. This flow:

  • is asynchronous, because it refers to the queued-asynchronous processing strategy

  • allows up to 500 concurrent threads, because of the value set for maxThreads

    <queued-asynchronous-processing-strategy name="allow500Threads" maxThreads="500"/>
    
    <flow name="manyThreads" processingStrategy="allow500Threads">
      <vm:inbound-endpoint path="manyThreads" exchange-pattern="one-way"/>
      <vm:outbound-endpoint path="output" exchange-pattern="one-way"/>
    </flow>

The following table lists the configuration parameters you can fine-tune for asynchronous processing strategies. (The synchronous processing strategy cannot be configured.) All of these attributes can be configured on the global element.

Attribute Type Queued only? Default value Description Optional?

maxBufferSize

integer

no

1

Determines how many requests are queued when the pool reaches maximum capacity and the pool exhausted action is WAIT. The buffer is used as an overflow.

yes

maxQueueSize

integer

yes

n/a

The maximum number of messages that can be queued.

yes

maxThreads

integer

no

16

The maximum number of threads that can be used.

yes

minThreads

integer

no

n/a

The number of idle threads kept in the pool when there is no load.

yes

poolExhaustedAction

enum

no

WHEN_EXHAUSTED_RUN

When the maximum pool size or queue size is bounded, this value determines how to handle incoming tasks.

yes

queueTimeout

integer

yes

n/a

The timeout used when taking events from the queue.

yes

threadTTL

integer

no

60000

Determines how long an inactive thread is kept in the pool before being discarded.

yes

threadWaitTimeout

integer

no

30000

How long to wait in milliseconds when the pool exhausted action is WAIT. If the value is negative, the wait is infinite.

yes

doThreading

boolean

no

True

Whether threading should be used.

yes

In addition, you can define a queue store using one of the following nested elements:

Queue Store nested element Description

simple-in-memory-queue-store

A simple in-memory queue store.

default-in-memory-queue-store

This is the default queue store used for non-persistent queues.

default-persistent-queue-store

This is the default queue store used for persistent queues.

file-queue-store

A simple file queue store.

queue-store

A reference to a queue store defined elsewhere.

custom-queue-store

A custom queue store defined with Spring properties.

Creating an Asynchronous Processing Block

If Mule has applied a synchronous processing strategy to your flow, you can separate out a processing block that executes simultaneously with the main flow and does not return messages back to the main flow. Achieve this in one of two ways:

  1. wrap one or more processors in an async scope

  2. create an asynchronous flow by wrapping a flow-ref element in an async scope so that the contents of the flow will be processed asynchronously with the triggering flow

Both of these methods allow you to block off a set of processing steps that may be very time consuming to execute. Because this asynchronous processing block is one-way, the main flow does not wait for a response and is free to continue processing in the main thread.

If no processing strategy is configured for the async scope, Mule applies a queued-asynchronous processing strategy. However, you can specify or fine-tune the asynchronous processing strategy of the scope by defining a global element and referencing it from the async element.

The following global elements are available for configuring the processing strategy of an asynchronous scope.

Global Element Description

asynchronous-processing-strategy

Not applicable to most use cases. Same as queued-asynchronous processing strategy (which is what Mule applies if no other processing strategy is configured) except that it doesn’t use a queue. Use this only if for some reason you do not want your processing to be distributed across nodes.

queued-asynchronous-processing-strategy

Uses a queue to decouple the flow’s receiver from the rest of the steps in the flow. It works the same way in a scope as in a flow. Mule applies this strategy unless another is specified. Select this if you want to fine-tune this processing strategy by:

  • Changing the number of threads available to the flow.

  • Limiting the number of messages that can be queued.

  • Specifying a queue store to persist data.

queued-thread-per-processor-processing-strategy

Not applicable to most use cases. Writes messages to a queue, then every processor in the scope runs sequentially in a different thread.

thread-per-processor-processing-strategy

Not applicable to most use cases. Every processor in the scope runs sequentially in a different thread.

For more information about configuring the asynchronous scope element, refer to the Async Scope reference.

Creating a Custom Processing Strategy

If neither the synchronous nor queued-asynchronous processing strategies fit your needs, and fine-tuning the asynchronous strategy is not sufficient, you can create a custom processing strategy. You create the custom strategy through the custom-processing-strategy `element and configure it using Spring bean properties. This custom processing strategy must implement the `org.mule.api.processor.ProcessingStrategy interface.

The following code example illustrates a custom processing strategy:

<custom-processing-strategy name="customStrategy" class="org.mule.CustomProcessingStrategy">
  <spring:property name="threads" value="500"/>
</custom-processing-strategy>

Reusing Processing Strategies

You can use a named processing strategy, either a custom processing strategy that you have created or a fine-tuned processing strategy, on as many flows in an application as you like.

  • declare the processing strategy, as in the following:

    <queued-asynchronous-processing-strategy name="allow500Threads" maxThreads="500"/>
  • refer to it in appropriate flows, for instance:

    <flow name="acceptOrders" processingStrategy="allow500Threads">
      <vm:inbound-endpoint path="acceptOrders" exchange-pattern="one-way"/>
      <vm:outbound-endpoint path="commonProcessing" exchange-pattern="one-way"/>
    </flow>
    
    <flow name="processNewEmployee" processingStrategy="allow500Threads">
      <vm:inbound-endpoint path="processNewEmployee" exchange-pattern="one-way"/>
      <vm:outbound-endpoint path="commonProcessing" exchange-pattern="one-way"/>
    </flow>
    
    <flow name="receiveInvoice" processingStrategy="allow500Threads">
      <vm:inbound-endpoint path="receiveInvoice" exchange-pattern="one-way"/>
      <vm:outbound-endpoint path="commonProcessing" exchange-pattern="one-way"/>
    </flow>