An advanced general purpose task distribution system based on Hazelcast with a familiar ExecutorService API.
Hazeltask was originally designed to provide large multi-tenent SaaS applications with an easy to use general purpose work distribution platform focusing on fair customer task load balancing. It has evolved from this mission to be much more powerful and configurable but still really easy to use!
- Fault tolerant with task recovery
- Automatic node task rebalancing
- Configurable task prioritization
- ExecutorService API
- Why should I use this instead of Hazelcast's build in distributed ExecutorService
- Put simply, if you are not using member-targeted DistributedTasks or MultiTasks, you should absolutely use this library. At the very least, it will give you failover recovery. No MemberLeftExceptions. And auto-balancing.
- What can Hazeltask NOT do, that Hazelcast's ExecutorService CAN
- The only thing Hazeltask doesn't do is handle instances of DistributedTask and MultiTask. This is because Hazeltask does the task distribution, and it only submits a task for execution on a single member. Its simply not part of the design.
This project encompasses an advanced distributed work library for Hazelcast.
It is modeled after the ExecutorService API but adds a lot of missing features
the built in Hazelcast executor service doesn't provide such as:
- Failover capabilities when nodes go down. No lost work!
- Work distribution loadbalancing with customizable prioritizers (RoundRobin, Enum, and Fair provided... or create your own)
- Push model is light on Hazelcast with few cluster operations and no cluster locking
- No more MemberLeftExceptions when waiting on Futures! Work will be redistributed to another member and executed returning the result to your Future
- Google Guava ListenableFuture support for easy callbacks and task chaining!
Let's say you have a multi-tenent system where users can submit hundreds of pages of scanned documents that you need to OCR. You want to guarantee that you will OCR all of the pages sent, and furthermore guarantee that one user cannot starve out another. For example, if user A submits 20,000 pages and then user B submits 1 page. You don't want user B to wait a really long time before their 1 page is worked on. You can use this library, group tasks by userId and use the RoundRobin router to solve this.
The easiest way to implement this is by grouping tasks into priorities based on an enum like HIGH, MEDIUM, LOW. Hazeltask comes with built-in support for prioritizing tasks based on an enum.
Oh yes you can. Its easy to combine the examples above and provide per-customer HIGH, MEDIUM, and LOW prioritization. You can implement any kind of prioritization you can think of. And... priorities can CHANGE. This allows you to implement features like: "Bump customer 7 to the highest priority for the next hour".
- Migrate to Hazelcast 3.0 to take advantage of SPI and custom datastructure implementation
- Allow task retries if a task fails to execute due to an exception in user code. Retry after _ time option.
- Refactored for Hazelcast 3.1 support
- Depend on codahale metrics 3.0.1