I have master and nodes infrastructure. Master executes remote jobs on nodes. Each job returns success/failure message back to master. So the flow of execution is the following:

  1. master receives command
  2. master creates remote job and sends message to node to execute remote job
  3. master returns 'accepted'
  4. remote job is executing on node
  5. remote job sends message to master notifying about the result

What happens if master receives new command on same resource sometime during 4, i.e. while remote job is executing?

I have 2 options:

  1. Refuse the new command. However, i must ensure that remote job always returns and if not, some timeout will fail the command - basically, to prevent resource to be forever locked by some stalling task.

  2. Accept new command and queue it. After the first task on the resource finishes, we executes the next one. However, this solution is more complex - as chain of commands may fail in the middle; need to keep track of future tasks etc.

My major concern is high availability of the system, to be more robust then super-powerful. Obviously, I am leaning towards solution #1.

Is there anything I am missing here? Is some other approach better?


More concrete example: 1 Master, 1 Node. Master may receives message: create, update and delete. Each message is executed on a Node as remote job. These jobs create, update and delete some resources on the Nodes. So I want to prevent e.g. to execute delete/update while resource is being created and similar situations.

The order of commands execution, obviously, must be the same as received.

  • It's very unclear what you want this system to do. Are the "nodes" interchangeable workers or is the master directing to specific nodes because certain resources are only accessible from those nodes? Is there any requirement the jobs execute in the order received? Any mutual exclusion requirements of some sort? It might be easier to say what you are trying to accomplish overall with this system than play 20 questions. Commented Feb 28, 2016 at 17:41
  • @DerekElkins Ive added business rule in the edit. It's really simple as that.
    – igor
    Commented Feb 28, 2016 at 17:49

1 Answer 1


For even your 1 node scenario, if you want the above to work correctly in the face of machine/network failure then what you are asking for is to solve consensus. This is impossible to achieve with the technical meaning of "availability". This is the CAP theorem. (Beware the misleading wording describing "availability" on that page. A better wording is: "any valid request will receive a successful response no matter what". In particular, your system is not available if it returns an error message for every request or, in fact, any valid request. An invalid request, like reading a key that doesn't exist, can obviously have an error response.) By "high availability", though, you probably mean something like "will work even if N servers fail". This is achievable and is what algorithms like Paxos, Viewstamped Replication, and Raft do.

That said, you really don't want to implement these algorithms yourself for production use. It's very easy to make very subtle errors when implementing them. Luckily, others have already built production-ready implementations. In particular, I recommend using Apache ZooKeeper, or possibly Apache Kafka (which was built on top of ZooKeeper). If you decide to use ZooKeeper "directly", then you will probably want to look at Apache Curator.

If your operations are idempotent (and it sounds like they may well be) and the order jobs are executed on different nodes doesn't matter, then having the master be a Kafka cluster and having the nodes subscribe to independent topics will handle this nicely with pretty good efficiency. Simply commit the offset after handling each request. If order matters between nodes, a single topic that all nodes subscribe to to get requests can be used. Then all the nodes send their responses to a separate Kafka topic. Finally, all nodes consume the response topic as well and only process the next request when a response to the previous request is consumed. Duplicate responses can easily be ignored and compacted.

To get exactly-once semantics you need to be able to atomically commit the log offset with any persistent state of the node. If your node's state is a deterministic function of the history of requests it's received, then exactly-once semantics is trivial: all the necessary information is in the Kafka log, you just replay it. This scenario is extremely desirable, fits very naturally with Kafka, and simplifies things dramatically. In this scenario, replicating the node is trivial; you can just run multiple copies, you can spin them up whenever and they'll recreate their state, no coordination between them is needed.

If your nodes' state is not a deterministic function of the history of requests it receives, then you are in a potentially hairy place of essentially implementing a replication log. The good news is that ZooKeeper and Kafka can handle a lot of the trickier aspects. You can use ZooKeeper to elect a leader and Kafka to store and distribute the replication log. The leader will consume the request log, perform any non-deterministic actions such as reading the local clock or making web requests to other servers, calculate a delta between the current state and the tentative new state, then publish that delta, which can be deterministically applied, to a Kafka topic representing the replication log. The nodes (both the followers and the leader) then consume the replication log applying the deltas to their local state. The leader publishes its response to the appropriate topic when it applies the delta it read from the replication log. It's possible multiple deltas get published for the same request if there is a leadership change during the processing. That's not a problem since the replication log is strictly ordered, and any deltas after the first can be ignored. (In the scenario of the previous paragraph, the request log essentially was the replication log which is why so much was simplified.)

The final concern is that you'll probably want to do some form of checkpointing so you don't need to keep the full log around and replay it from the beginning. Another nice thing about this log-centered design is that this can be done as a background process if desired.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.