15 February 2012

Saga vs Batch Processing (Spring Batch introduction)

When dealing with business process often some state transitions are not immediately executed as result of human interaction but rather being scheduled for future execution. One example could be expiration of Payment Period (see previous article for domain description).

Modelling such process execution flow explicitly in the code is the right thing to do if we want to keep process logic maintainable and self describing. That's why we should apply bpm tools or Saga pattern (if we prefer simple and light-weight solution) to do the job.

Scheduling activity within Saga


The following code shows how to schedule activity within Saga (Axon implementation):

When (current) Payment Period is created, RenewPaymentPeriodCommand is scheduled for execution on time when the period expires (after validation interval passes). This approach is clean, easy to implement and test but...requires (mind) shifting from procedural way of modelling business logic (see transaction script) towards event driven architecture (EDA). For those who are not yet ready to enter EDA and bpm, there is an old-time heavy-weight, bullet-proof way of executing scheduled tasks on certain time: batch processing.

Batch processing


Batch processing is suitable for optimizing execution of high-volume, repetitive tasks in such way that system is under heavy load only within relatively short time window (batch window). (See http://en.wikipedia.org/wiki/Batch_processing)

Our goal will be to execute Payment Period renewal in batch mode. Instead of scheduling each Payment Period renewal explicitly in the business process (saga), batch job (RenewPaymentPeriodsJob) scheduled to run repeatedly (every hour or day depending on requirements) will invoke RenewPaymentPeriodCommand for all expired Payment Periods that it will find in database.
We will only change processing mode but will not touch business logic that will be still encapsulated inside of aggregate roots and event listeners. We will use the same commands dispatching mechanism that is used for online transactions processing avoiding creation of specialized services or using sql statements to implement batch jobs (as one could expect from batch jobs:) ).

Running jobs in batch mode gives more control over system load. It can be controlled how often and when batch jobs should be run. In Saga, execution model is different, operations are executed automatically and there is no control mechanism at runtime. Which way is better..? As always, depends. But it is good to have both alternatives ready to apply when time comes...

Lets see how to use Spring Batch, modern batch framework for JVM, to apply batch processing using commands and queries as building blocks.

Spring Batch processing model


Batch Stereotypes

The diagram above highlights the key concepts that make up the domain language of batch. A Job has one step or combines multiple steps that belong logically together in a flow. Each step has exactly one ItemReader, ItemProcessor, and ItemWriter. A Job needs to be launched (JobLauncher), and meta data about the currently running process needs to be stored (JobRepository).

Spring Batch uses a Chunk Oriented processing style within its most common implementation. Chunk oriented processing refers to reading the data one at a time, and creating 'chunks' that will be written out, within a transaction boundary. Committing a transaction, at each commit interval, commits a 'chunk'.

The data item could be line in a file or record in a database table but Spring Batch integrates modern to-object mapping frameworks so we don't have to dirty our hands by manipulating low-level data.

Chunk oriented processing


Spring Batch application


Our goal is to use Aggregate Roots as processing items. To build batch step we need to implement an ItemReader that will fetch ARs (entities) from database by executing provided query and an ItemProcessor that will build a command based on AR data and dispatch command to the system. Since batch processing is performed transactionally (chunks are automatically committed by Spring Batch with use of provided transaction manager) commands need to to be dispatched synchronously. This logic of batch step may be shared across different jobs as long as the concept of ItemReader responsible of fetching ARs using provided query and ItemProcessor responsible for dispatching command is preserved. What will distinguish different steps from each other is query specification and command to be executed.

So lets define interface that will describe these responsibilities of batch step:

BatchStepSpecification object should be able to provide query specification (executable by ItemReader, more on query specifications in a moment) and build Command (for each AR to be processed) executable by ItemProcessor.

Now we need to implement ItemReader and ItemWriter that will use step specification to do their job.

ItemReader


To avoid keeping all entities to be processed in memory (this could be a large set) Spring Batch offers two solutions: Cursor and Paging database ItemReaders. Let's go with the letter one. For loading entities from database in a paging fashion Spring Batch provides several implementations of AbstractPagingItemReader one of them being JpaPagingItemReader for fetching JPA entities. JPAPagingItemReader allows you to define query by providing JPQL statement. But we want our queries be more maintainable and composable. Therefore I recommend to represent JPA queries as Specifications. This is possible with use of Spring Data JPA module. Specification can be translated into JPA Query in the following way:

More about specifications you can read here: Advanced Spring Data JPA – Specifications And QueryDSL

Implementing SpecificationPagingReader is straightforward. Main thing to do is to implement doReadPage() method:

Finally, we can define our ItemReader in Spring context:

The instance of SpecificationPagingReader will be created automatically by Spring whenever batch step is executed. This is the magic of step scope provided by Spring Batch. It allows late (dynamic) binding of properties. In our case stepSpecification is unknown until particular step is executed (more details ahead).

ItemProcessor


ItemProcessor interface defines just one method. The implementation in our case is simple:

And Spring bean definition:


ItemWriter


ItemWriter must be provided as well but it can be empty implementation assuming all changes to ARs are flushed to database as a result of command processing within the service.

Finally we are able to build batch step that will be reused by concrete batch jobs.

Batch Step


Here, we compose three beans (reader, processor, writer) and additionally provide following parameters:
  • skip-limit - the maximum number of items that will be allowed to be skipped (in case the processing of item ends with exception). Once the skip limit is reached, the next exception found will cause the step to fail.
  • commit-interval - how many items will be processed in one transaction (chunk size), value > 1 might increase performance, but also could result in rollback of successfully processed ARs (if exception skip limit is reached)
  • skippable-exception-classes - exceptions that will result in skipping the processed entity instead of step failure

Job registry


Now we can define our jobs. The easiest way is to use AutomaticJobRegistrar class. Registration in this case is performed automatically on aplication startup, based on defined path under which spring context files containing jobs definitions are located. By putting each job bean into separate spring context file we are able to provide its stepSpecification bean that will be created when job's step is executed. If all those files were imported into the same context, the stepSpecification definitions would clash and override one another, but with the automatic registrar this is avoided.

Please see example definition of batch job:

As shown above, creating new batch job is just a matter of creating new spring context file containing job id (globally unique), step id and step specification bean. Nothing more is required. The solution is powerful and simple but has one limitation. We can not create jobs with several steps, each one configured with different BatchScopeSpecification. For example:

I found a solution for this problem, in case you are interested, let me know by dropping in a comment.

Job execution


To execute a Job, we need to create JobParameters object and use JobLauncher to run Job with created parameters. Please refer to description of artifacts related to batch job execution. It is important to understand differences between Job, JobInstance and JobExecution.


Things to remember:
  • JobInstance - the concept of a logical job run (Job + JobParameters)
  • JobExecution - a signle attempt to run JobInstance. JobInstance corresponding to a given execution will not be considered complete unless the execution completes successfully. There can be more than one failed JobExecutions but only one successful execution of given JobInstance.

Batch jobs scheduling


Spring Batch is not a scheduling framework. It is entirely up to the scheduler to determine when a Job should be run. There is no requirement that one JobInstance be kicked off after another, unless there is potential for the two job instances to attempt to access the same data, causing issues with locking at the database level. But attempting to run the same JobInstance while another is already running will result in a JobExecutionAlreadyRunningException being thrown.

Other Spring Batch goodies

  • Restartability - the framework periodically persists the ExecutionContext at commit points. This allows the ItemReader to store its state in case a fatal error occurs during the run, or even if the power goes out. JpaPagingItemReader supports restart by storing item count, therefore requires item ordering to be preserved between runs.
  • Non Sequential Step Execution - conditional flow of steps
  • Partitioning
  • JobOperator interface for common monitoring tasks such as stopping, restarting, or summarizing a Job, as is commonly done by batch operators.


No comments:

Post a comment