Transactions Guide

Providing support for distributed (XA) transactions

Immutant encapsulates the distributed transaction support provided by Narayana within the immutant.transactions namespace. This allows you to define XA (2PC) transactions involving Immutant components either inside the WildFly application server or embedded within your standalone apps.

A distributed transaction is one in which multiple types of resources may participate. The most common example of a transactional resource is a relational database, but in Immutant both caching (backed by Infinispan) and messaging (backed by HornetQ) resources are transactional. Technically speaking, they provide implementations of the XA protocol, and any resource that does may participate in an Immutant transaction.

This allows your application to say, tie the success of a SQL database update or the storage of an entry on a replicated cache to the delivery of a message to a remote queue, i.e. the message is only sent if the database and cache update successfully. If any single component of an XA transaction fails, all of them rollback.

Defining a transaction

If you’re familiar with JTA, we make a TransactionManager available via the manager function, and everything else provided by the library is mostly a syntactic sugary glaze slathered on that instance. For example, the transaction macro will query the manager to see if a transaction is active. If so, it’ll simply invoke its body. Otherwise, it’ll start a new transaction, execute its body, and commit the transaction unless either an exception is caught or set-rollback-only is called, in which case the transaction is rolled back. Either way, it relies on the transactional components within the body to automatically enlist themselves as XA resources. For example,

(def queue (msg/queue "/queue/test"))
(def cache (csh/cache "test" :transactional true))

(transaction
  (msg/publish queue "message")
  (csh/swap-in! cache :count (fnil inc 0)))

Here, we’ve combined a messaging call and a caching call into a single, atomic operation. Either both succeed or neither succeeds. Note that the cache had to be created with its :transactional flag set. The publish function will detect an active transaction and create an XA-capable context with which to send the message. If you pass a context for publish to use, be sure its :xa? flag is set if you intend to publish within a transaction.

Transaction Scope

When transactional components interact, the state of a transaction when a particular function is invoked isn’t always predictable. For example, can a function that requires a transaction assume one has been started prior to its invocation? In JEE container-managed persistence, a developer answers these questions using the @TransactionAttribute annotation.

But annotations are gross, my friend!

So in Immutant, JEE transaction attributes are represented as Clojure macros. In fact, the transaction macro is merely an alias for immutant.transactions.scope/required, which is the implicit attribute used in JEE. There are a total of 6 macros in the immutant.transactions.scope namespace:

  • required- Execute within current transaction, if any, otherwise start a new one, execute, commit or rollback
  • requires-new- Suspend current transaction, if any, start a new one, execute, commit or rollback, and resume the suspended one
  • not-supported - Suspend current transaction, if any, and execute without a transaction
  • supports- Execute the body whether there’s a transaction or not; may lead to unpredictable results
  • mandatory - Toss an exception if there’s no active transaction
  • never - Toss an exception if there is an active transaction

These macros give the developer complete declarative control over the transactional semantics of their application as its functional chunks are combined.

Pseudo-Transactions

While messaging and caching resources in Immutant are XA capable, not all data sources are, and even a JDBC library providing an XA driver is insufficient outside of an application server.

But non-XA resources like Datomic or JDBC connections running outside of a container can still participate in a distributed transaction as long as their local transaction is the last operation in the body defining the XA transaction. This technique, known as a pseudo transaction, relies on the non-XA resource throwing an exception in the event of failure. This exception will cause the operations on the XA resources to rollback. For example,

(immutant.transactions/transaction
  (immutant.messaging/publish queue message)
  (immutant.caching/swap-in! cache :count (fnil inc 0))
  (clojure.java.jdbc/with-db-transaction [t spec]
    (write-sql-things t data)))

Note that we define a local transaction using with-db-transaction assuming write-sql-things invokes multiple SQL statements, so that they all rollback if any throw an exception. If you’re only calling a single SQL statement, e.g. one INSERT, the local transaction isn’t necessary.

JDBC DataSources in WildFly

Within WildFly, it’s possible to use real, honest-to-goodness JDBC XA datasources, thereby obviating the need for pseudo transactions.

Configuring datasources in WildFly is beyond the scope of this guide, but once you have the JNDI name for the XA datasource, you can pass that as the :name option in the “database spec” required by java.jdbc or any of its many DSL wrappers.

Unfortunately, java.jdbc attempts to invoke operations on the connection returned from the datasource bound to that name that are illegal within an XA transaction, specifically: commit, rollback, and setAutoCommit. It’s up to the TransactionManager to commit/rollback a distributed transaction, not its constituents. So Immutant provides a factory function, immutant.transactions.jdbc/factory, that wraps the connection to turn those calls into no-ops. This means the database spec you pass to java.jdbc should look something like this:

(def db-spec {:factory immutant.transactions.jdbc/factory
              :name "java:jboss/datasources/ExampleDS"})

Transaction Configuration

Narayana will look for a configuration file on the classpath called jbossts-properties.xml. There is a default one embedded within its distribution jar that you can use as a basis for your own.

One of the properties, in particular, has an annoying default value. The location of the persistent object store is PutObjectStoreDirHere by default. This can be overridden with the following system property if you’d prefer not to create the XML file:

"-DObjectStoreEnvironmentBean.objectStoreDir=target/ObjectStore"

The “now obsolete but somewhat equivalent and occasionally still relevant” analog to the above is this:

"-Dcom.arjuna.ats.arjuna.objectstore.objectStoreDir=target/ObjectStore"

So you may need to set them both.