Zookeeper

We use zookeeper to track locks between the autonomous components. To make zookeeper easier to use, we use the netflix curator api http://curator.incubator.apache.org/

Zookeeper maintains a tree structure internally. In each leaf in the tree should correspond to some real world thing. A leaf can be locked, and the system is based on the idea that you will lock the leaf before accessing the thing it represents.

We have two "kinds" of leafs. 

The specific lock paths are as follows

SBOI lock = "/SBOI/" + runnable.getComponentName();

batch lock = "/"+runnable.getComponentName() + "/B" + batch.getBatchID() + "-RT" + batch.getRoundTripNumber();

Autonomous component locking procedure 1

Each of the autonomous components will exist as cron jobs. Periodically, often every minute or thereabouts, the cron job will activate. It will then follow the procedure below.

1. Lock "SBOI+component name"
   1. query SBOI for batches
   2. For each returned batch
   3. Attempt to lock the "batch+component name"
   4. if successful, break loop
2. unlock "SBOI+component name"
3. Do work on locked batch (if any)
4. store results in DOMS
5. unlock "batch+component name"

Autonomous component locking procedure 2

Each of the autonomous components will exist as cron jobs. Periodically, often every minute or thereabouts, the cron job will activate. It will then follow the procedure below.

1. Lock "SBOI+component name"
2. query SBOI for batches
3. For each returned batch
   1. Attempt to lock the "batch+component name"
   2. Do work on locked batch (if any). Possibly spawn sub process
   3. store results in DOMS
   4. unlock "batch+component name"
4. unlock "SBOI+component name"

Which one will we use?

Method 1 is simpler to implement (and already done). Method 1 will only maintain a very shortlived lock on SBOI. To run method 1 on three batches, just start it three times.

Method 2 is conceptually simpler to understand. It allows for developer controlled degree of parallelism, but will also force us to code this ourselves.  This is the method described on the Autonomous Components page. Well actually the description is this:

If the robot finds that a batch is ready to work on, it starts working on this batch. The robot will not poll for more work while working.

Some of the robots will be able to multitask, and others will not. Multitasking is done in a slightly different way for robots than for humans. The multitasking robot will build a new little non-multitasking robot child, give him the batch to work on, and then go back to sleep. Next time it polls for work, it will have to remember not the start work on batches that already have been assigned to any of the robots children. This also means that there must only ever be one instance of each type of multitasking robot running.

When a robot have finished work on a batch, it must record this, so the assembly line can move forward. It records the event Batch Object, stored in the Digital Object Management System, DOMS, system, in the form of an event somewhat like "I, <ROBOT>, did this <THIS WORK> on batch <ID> with <THIS RESULT>". The SBOI will then periodically (often) query DOMS for updates to the Batch objects. When the SBOI discover a batch object update, it updates the index, so that robots further along the assembly line can work on the batch.

An autonomous component here needs the logic for multitasking. If we assume there is only one component A running on one server, this is straight forward (and re-usable). This assumption still allows a component to send internal jobs to different servers (e.g.. Hadoop jobs)...