I read a lot about how to enable parallel processing and chunking of an individual job, using Master/Slave paradigm. Consider an already implemented Spring Batch solution that was intended to run on a standalone server. With minimal refactoring I would like to enable this to horizontally scale and be more resilient in production operation. Speed and efficiency is not a goal.
http://www.mkyong.com/spring-batch/spring-batch-hello-world-example/
In the following example a Job Repository is used that connects to an initializes a database schema for the Job Repository. Job initiation requests are fed to a message queue, that a single server, with a single Java process is listening on via Spring JMS. When encountering this it executes a new Java process that is the Spring Batch job. If the job has not been started according to the Job Repository it will begin. If the job had failed it will pick up where the job left off. If the job is in process it will ignore.
The single point of failure is the single server and single listening process for job initiation. I would like to increase resiliency by horizontally scaling identical server instances all competing for who can first grab the job initiation message when it first appears in the queue. That server instance will now attempt to run the job.
I was conceiving that all instances of the JobRepository would share the same schema, so they can all query for when the status is currently in process and decide what they will do. I am unsure though if this schema or JobRepository implementation is meant to be utilized by multiple instances.
Is there a risk in pursuing this that this approach could result in deadlocking the database? There are other constraints to where the Partition features of Spring Batch will not work for my application.
I decided to build a prototype to test if the condition that the Spring Batch Job Repository schema and SimpleJobRepository can be used in a load balanced way with multiple Spring Batch Java processes running concurrently. I was afraid that deadlock scenarios might have occurred at the database to where all running job processes get stuck.
My Test
I started with the mkyong Spring Batch HelloWorld example and made some changes to it where it could be packaged into a Jar that can be executed from the command line. I also removed the initialize database step defined in the database.config file and manually established a local MySQL server with the proper schema elements. I added a Job parameter for time to be the current time in millis so that each job instance would be unique.
Next, I wrote a separate Java main class that used Apache Commons Exec framework to create 50 sub processes with no wait between them. Each of these processes have a Thread.sleep for 1 second within their Processor objects as well so that a number of processes will all kick off at the same time and all attempt to access the database at the same time.
Results
After running this test a number of times in a row I see that all 50 Spring batch processes consistently complete successfully and update the same database schema correctly. I don't see any indication that if there were multiple Spring Batch job processes running on multiple servers connecting to the same database that they would interfere with each other on the schema nor do I see any indication that a deadlock could happen at this time.
So it sounds as if load balancing of Spring Batch jobs without the use of advanced Master/Slave and Step Partitioning approaches is a valid use case.
If anybody would like to comment on my test or suggest ways to improve it I would appreciate it.
Here is excerpt from Spring Batch docs on how Spring Batch handles database updates for its repository:
Spring Batch employs an optimistic locking strategy when dealing with updates to the database. This means that each time a record is 'touched' (updated) the value in the version column is incremented by one. When the repository goes back to save the value, if the version number has changed it throws an OptimisticLockingFailureException, indicating there has been an error with concurrent access. This check is necessary, since, even though different batch jobs may be running in different machines, they all use the same database tables.
来源：https://stackoverflow.com/questions/34336880/spring-batch-horizontal-scaling-of-job-repository