Why?
The code is still a mess even though it has been converted from an ant to a maven project.
- The Harvester-core module is just a trashcan location for anything remotely harvester related.
Unit tests are separated into their own modules making code coverage close to impossible.
Impractical to use 1-2 weeks of running test cases which should be covered much better by unit testing. (every time anything major is released)
- Registering/deregistering JMS listeners just seems wrong and it is surprising it has worked so far. Nothing in the JMS specifiction suggests that this is even a good idea.
Overall solutions
- Isolate storage implementation completely. It is going to happen sooner or later anyway.
- Remove JMS as much as possible. (Except from legacy code like the bitarchive implementation)
- Split the different parts of NAS into separate modules with little or no dependencies.
Reasoning
- It is easier to build/test/deploy. With improved and localized unit testing you could most likely...
- release much quicker without having to use 1-2 weeks testing.
- release only the module(s) you changed.
- People who don not want to use NAS could instead just use the parts that fit their needs.
- For example people could use the harvest controllers and nothing else.
Harvest Control Manager Component
- Remove the JMS dependency from the controller.
- Instead use a REST interface or some other means of exposing a simple extendable API.
- Remove the notion of channels from the controller.
- The management of organizing controllers into groups is left to the user of these APIs.
- Make the code independent of the rest of NAS so it can be used not only by NAS.
- Controllers should be deployed independently of the rest of NAS.
- Use a plugin architecture for core functionality. (Use classloaders)
- configure harvester
- build progress reports
- build metadata files when the job is complete
- upload data to persistant storage
- A controller is built for a specific harvester; H1, H3, API
- Extendable using custom commands that the plugins add to the controller. (Thinking beyond H3...)
- The API should include all required functions to control the harvest manager
Submit job.
Upload configuration files.
Upload additional files; indexes etc.
Start job.
Get progress/report.
Stop job.
Initiate metadata generation.
Initiate upload.
- uploading new versions of the plugins.
- uploading new versions of the harvester package. (h3 bundle)
- Offer base client implemention. (Used by a job manager/monitor)
The API should make it possible to redo certain operations which occasionally fail and require manual intervention.
Only when the worst happens should i be required for a person to fiddle with the server.
The existing code in the harvest control manager that must be migrated into one or more plugins.
- move files to h1/h3 folder.
- build metadata reports.
- build progress reports.
- upload data to the bitarchive
The plugins can form the base for other people adapting the code to suit their own needs.
The only state the manager should know about is it's harvest.
Everything else is up to the caller. This way you can use "channel" or not. And more importantly you can manage "channel" dynamically without having to reconfigure and redeploying harvest controll managers.