spring-integration/src/reference/docbook/transactions.xml

<?xml version="1.0" encoding="UTF-8"?>
<chapter xmlns="http://docbook.org/ns/docbook" version="5.0" xml:id="transactions"
    xmlns:xlink="http://www.w3.org/1999/xlink">
  <title>Transaction Support</title>

  <section id="transaction-support">
    <title>Understanding Transactions in Message flows</title>
    <para>
      Spring Integration exposes several hooks to address transactional needs of you message flows.
      But to better understand these hooks and how you can benefit from them we must first revisit the 6 mechanisms
      that could be used to initiate Message flows and see how transactional needs of these flows
      could be addressed within each of these mechanisms.
    </para>
    <para>
    Here are the 6 mechanisms to initiate a Message flow and their short summary (details for each are provided throughout this manual):
    <itemizedlist>
          <listitem>
            <para><emphasis>Gateway Proxy</emphasis> - Your basic Messaging Gateway</para>
          </listitem>
          <listitem>
            <para><emphasis>MessageChannel</emphasis> - Direct interactions with MessageChannel methods (e.g., channel.send(message))</para>
          </listitem>
          <listitem>
            <para><emphasis>Message Publisher</emphasis> - the way to initiate message flow as the by-product of method invocations on Spring beans</para>
          </listitem>
          <listitem>
            <para><emphasis>Inbound Channel Adapters/Gateways</emphasis> - the way to initiate message flow based on connecting third-party
            system with Spring Integration messaging system(e.g., [JmsMessage] -> Jms Inbound Adapter[SI Message] -> SI Channel)</para>
          </listitem>
          <listitem>
            <para><emphasis>Scheduler</emphasis> - the way to initiate message flow based on scheduling events distributed
            by a pre-configured Scheduler</para>
          </listitem>
          <listitem>
            <para><emphasis>Poller</emphasis> - similar to the Scheduler and is the way to initiate message flow based on scheduling
            or interval-based events distributed by a pre-configured Poller</para>
          </listitem>
      </itemizedlist>
    </para>
    <para>
    These 6 could be split in 2 general categories:
     <itemizedlist>
          <listitem>
            <para><emphasis>Message flows initiated by a USER process</emphasis> - Example scenarios in this category
            would be invoking a Gateway method or explicitly sending a Message to a MessageChannel. In other words, these message flows depend on a third
            party process (e.g., some code that we wrote) to be initiated.</para>
          </listitem>
          <listitem>
            <para><emphasis>Message flows initiated by a DAEMON process</emphasis> - Example scenarios in this category would be a Poller
            polling a Message queue to initiate a new Message flow with the polled Message or a Scheduler scheduling the
            process by creating a new Message and initiating a message flow at a predefined time.</para>
          </listitem>
      </itemizedlist>
      </para>
      <para>
      Clearly the <emphasis>Gateway Proxy</emphasis>, <emphasis>MessageChannel.send(..)</emphasis> and <emphasis>MessagePublisher</emphasis>
      all belong to the 1st category and <emphasis>Inbound Adapters/Gateways</emphasis>, <emphasis>Scheduler</emphasis> and <emphasis>Poller</emphasis>
      belong to the 2nd.
    </para>
    <para>
    So, how do we address transactional needs in various scenarios within each category and is there a need for Spring Integration
    to provide something explicitly with regard to transactions for a particular scenario? Or, can Spring's Transaction Support be leveraged instead?.
    </para>

    <para>
    The first and most obvious goal is NOT to re-invent something that has already been invented unless you can provide a better solution.
    In our case Spring itself provides first class support for transaction management. So our goal here is not to provide something new but rather
    delegate/use Spring to benefit from the existing support for transactions. In other words as a framework we must expose hooks to the Transaction management functionality
    provided by Spring. But since Spring Integration configuration is based on Spring Configuration it is not always necessary to expose these hooks as they
    are already exposed via Spring natively. Remember every Spring Integration component is a Spring Bean after all.
    </para>
    <para>
    With this goal in mind let's look at the two scenarios. 
    </para>
    <para>
    If you think about it, Message flows that are initiated by the <emphasis>USER process</emphasis> (Category 1) and obviously configured in a Spring Application Context,
    are subject to transactional configuration of such processes and therefore don't need to be explicitly configured by Spring Integration to support transactions.
    The transaction could and should be initiated through standard Transaction support provided by Spring. The Spring Integration message flow will honor
    the transactional semantics of the components naturally because it is Spring configured. For example, a Gateway or ServiceActivator method could
    be annotated with <classname>@Transactional</classname> or <classname>TransactionInterceptor</classname> could be defined in an XML configuration
    with a point-cut expression pointing to specific methods that should be transactional.
  The bottom line is that you have full control over transaction configuration and boundaries in these scenarios.
    </para>

    <para>
    However, things are a bit different when it comes to Message flows initiated by the <emphasis>DAEMON process</emphasis> (Category 2).
    Although configured by the developer these flows do not directly involve a human or some other process to be initiated. These are trigger-based flows
    that are initiated by a trigger process (DAEMON process) based on the configuration of such process. For example, we could have a Scheduler
    initiating a message flow every Friday night of every week. We can also configure a trigger that initiates a Message flow every second, etc.
    So, we obviously need a way to let these trigger-based processes know of our intention to make the resulting Message flows transactional so that
    a Transaction context could be created whenever a new Message flow is initiated. In other words we need to expose some Transaction configuration, but ONLY enough
    to delegate to Transaction support already provided by Spring (as we do in other scenarios).
    </para>

    <para>
    Spring Integration provides transactional support for Pollers. Pollers are a special type of component because
    we can call receive() within that poller task against a resource that is itself transactional thus including <emphasis>receive()</emphasis>
    call in the the boundaries of the Transaction allowing it to be rolled back in case of a task failure. If we were to add the same support
    for channels, the added transactions would affect all downstream components starting with that <emphasis>send()</emphasis> call. That is
    providing a rather wide scope for transaction demarcation without any strong reason especially when Spring already provides several ways to
    address the transactional needs of any component downstream. However the <emphasis>receive()</emphasis> method being included in a transaction
    boundary is the "strong reason" for pollers.
    </para>

     <section id="transaction-poller">
      <title>Poller Transaction Support</title>
        <para>
        Any time you configure a Poller you can provide transactional configuration via the <emphasis>transactional</emphasis> sub-element and its attributes:
        <programlisting language="xml"><![CDATA[<int:poller max-messages-per-poll="1" fixed-rate="1000">
    <transactional transaction-manager="txManager" 
                   isolation="DEFAULT"
                   propagation="REQUIRED" 
                   read-only="true" 
                   timeout="1000"/>
</poller>]]></programlisting>
  As you can see this configuration looks very similar to native Spring transaction configuration. You must still provide a reference to a Transaction manager and specify
  transaction attributes or rely on defaults (e.g., if the 'transaction-manager' attribute is not specified, it will default to the bean with the name 'transactionManager').
  Internally the process would be wrapped in Spring's native Transaction where <classname>TransactionInterceptor</classname> is responsible for handling transactions.
  For more information on how to configure a Transaction Manager, the types of Transaction Managers (e.g., JTA, Datasource etc.) and other details related to
  transaction configuration please refer to Spring's Reference manual (Chapter 10 - Transaction Management).
        </para>
        <para>
        With the above configuration all Message flows initiated by this poller will be transactional. For more information and details on a
        Poller's transactional configuration please refer to section - <emphasis>21.1.1. Polling and Transactions</emphasis>.
        </para>

        <para>
        Along with transactions, several more cross cutting concerns might need to be addressed when running a Poller. To help with that,
        the Poller element accepts an <emphasis>&lt;advice-chain&gt; </emphasis> sub-element which allows you to define a custom chain of Advice
        instances to be applied on the Poller. (see section 4.4 for more details)
         In Spring Integration 2.0, the Poller went through the a refactoring effort and is now using a proxy mechanism to address transactional
         concerns as well as other cross cutting concerns. One of the significant changes evolving from this effort is that we
         made <emphasis>&lt;transactional&gt;</emphasis> and <emphasis>&lt;advice-chain&gt;</emphasis>  elements mutually exclusive.
    The rationale behind this is that if you need more than one advice, and one of them is Transaction advice, then you can simply
    include it in the <emphasis>&lt;advice-chain&gt;</emphasis> with the same convenience as before but with much more control
    since you now have an option to position any advice in the desired order.
 <programlisting language="xml"><![CDATA[<int:poller max-messages-per-poll="1" fixed-rate="10000">
  <advice-chain>
    <ref bean="txAdvice"/>
    <ref bean="someAotherAdviceBean" />
    <beans:bean class="foo.bar.SampleAdvice"/>
  </advice-chain>
</poller>

<tx:advice id="txAdvice" transaction-manager="txManager">
  <tx:attributes>
    <tx:method name="get*" read-only="true"/>
    <tx:method name="*"/>
  </tx:attributes>
</tx:advice>
]]></programlisting>

As yo can see from the example above, we have provided a very basic XML-based configuration of Spring Transaction advice  - "txAdvice" and
included it within the <emphasis>&lt;advice-chain&gt;</emphasis> defined by the Poller.
If you only need to address transactional concerns of the Poller, then you can still use the <emphasis>&lt;transactional&gt;</emphasis> element
as a convinience.
        </para>
    </section>
  </section>

  <section id="transaction-boundaries">
      <title>Transaction Boundaries</title>
      <para>
      Another important factor is the boundaries of Transactions within a Message flow.
      When a transaction is started, the transaction context is bound to the current thread. So regardless of how many endpoints and channels you have in your
      Message flow your transaction context will be preserved as long as you are ensuring that the flow continues on the same thread.
      As soon as you break it by introducing a <emphasis>Pollable Channel</emphasis> or <emphasis>Executor Channel</emphasis> or initiate a new thread manually
      in some service, the Transactional boundary will be broken as well. Essentially the Transaction will END right there, and if
      a successful handoff has transpired between the threads, the flow would be considered a success and a COMMIT signal would be sent
      even though the flow will continue and might still result in an Exception somewhere downstream. If such a flow were synchronous, that Exception could
      be thrown back to the initiator of the Message flow who is also the initiator of the transactional context and the transaction would result in a ROLLBACK.
      The middle ground is to use transactional channels at any point where a thread boundary is being broken. For example, you can use a Queue-backed Channel
      that delegates to a transactional MessageStore strategy, or you could use a JMS-backed channel.
      </para>
    </section>
</chapter>