Highly Distributed Coordinated Notifications


In Kilo, support for coordinated notifications agents was added. This enabled users to deploy multiple notification agents and ensured related messages within a pipeline were funnelled into the same agent to allow for proper aggregation calculations.

Problem description

In the initial implementation, all the data relating to a pipeline was funneled into a single queue per pipeline. While it ensured all corresponding data is sent to the same place, it removed the ability to scale horizontally as the pipeline queues can only have a single consumer listening to it. This means that while multiple agents/handlers could be used to pull data off main OpenStack queue, once the data reached pipeline processing, it was relegated to a single worker.

Data can be handled in parallel even at the pipeline level. For example, when there are no transformers, datapoints do not need to be handled sequentially. Additionally, when transformers are present, datapoints of different resources have no relevance to each other and can be handled in parallel.

Proposed change

To parallelise and scale out processing, we will create multiple copies of each pipeline. When a datapoint arrives, we will bucketise each datapoint by a hashed grouping key. Each transfromer will have a grouping key assigned to it to note any dependency requirements (ie. transformers that work on resource ids). When setting up pipeline, all the keys of transformers in the pipeline will be combined to ensure that related datapoints will be consistently sent to the same pipeline for processing.

The basic workflow is as follows:

* on notification agent startup, create a listener for main queue
* for each pipeline definition, we create x queues and x listeners where x
  corresponds to the number of notification agents registered to group
* when a datapoint is received, agent builds sample.
* after sample is built, we hash fields defined by transformer requirements
  and mod by number of agents.
* using mod value we push datapoint to corresponding pipeline queue.
* same processing steps here on out (listener grabs data -> pipeline -> pub)

This solution CANNOT handle multiple grouping_keys in pipeline. To properly handle multiple grouping_keys in a pipeline we need to requeue after each transform. The logic would become: main queue -> build sample -> pipe1.transform1 queue -> pipe1.transform2 queue -> etc -> publish.

Studying the existing transformers we have:

* Accumulator - this does not really have any grouping requirements, it just
                batches samples.
* Arithmetic - this is grouped by resource_id
* RateOfChange - this is grouped by name+resource_id. but it can more
                 be more generally grouped by just resource_id
* Aggregator - this is grouped by name+resource_id+<custom> but can also be
               more generally grouped by just resource_id

Based on the above, it seems like resource_id is always a valid general grouping key and the transformers may do more granular groupings themselves. Because of this, it seems safe to assume we don’t need to support multiple grouping keys (for now).


  1. Dumb easy fix is to detect whether a pipeline has transformers. If not, it can be consumed by any number of consumers and thus we can assign multiple listeners to those queues. This doesn’t help distribute load of pipelines with transformers but allows for transformers spanning resources.
  2. We implement a shared memory/storage mechanism so any worker can discover the historical context. This is hard. I feel like i would end up recreating Storm/Spark
  3. Magic

Data model impact


REST API impact


Security impact


Pipeline impact

Nothing from user point of view. Internally, we will have more pipeline queues.

Other end user impact

None. Unless we decide to add option to define number of copies of pipeline queues.

Performance/Scalability Impacts

Positive. It will distribute pipeline processing when running in coordinated mode. Message queues are consistently used with thousands of queues so creating copies of pipeline queues (a relatively small set) should not be an issue.

Other deployer impact


Developer impact




Primary assignee:
Ongoing maintainer:

Work Items

  • add grouping key to each transformer to define grouping of datapoints * will be just resource_id
  • add support to build pipeline hashing from above grouping keys
  • add functionality to create pipelines queues per agent and distribution logic.

Future lifecycle

Support different grouping keys in a pipeline.




Test already exists. Just need to validate that we create appropriate amount of copies.

Documentation Impact

None. Maybe dev docs.