API No Pipeline

https://blueprints.launchpad.net/ceilometer/+spec/api-no-pipeline

The ceilometer API server allows new samples to be posted via the API. These are then processed through the pipeline and then dispatched as required. This means the API server must use the pipeline code and have a pipeline.yaml file, raising the complexity of the code and the testing and install profile of the API. An alternative would be for the API server to publish notifications for the meters it receives which would then be processed by the notification agent (which has a pipeline of its own) and then dispatched.

Problem description

Including the transformation pipeline in the API server increases complexity on at least three dimensions for a single feature in the API (creating samples) which is not commonly used (most samples are produced via polling agents and notifications).

The use of the pipeline creates dependencies in code, testing and installation that could be avoided. In code we must include the pipeline modules and classes. In testing and installation we must have a pipeline.yaml.

In the installation case, if there are multiple API servers, the pipeline.yaml must be kept in sync across the hosts on which those servers are installed. That’s an unnecessary burden.

Proposed change

One way to resolve this complexity is to not use the pipeline in the API server at all. Instead when new samples are posted, format them as notifications and push them onto the notification bus to be retrieved by the notification agent. The pipeline within that agent will then do any expected transformations.

For convenience of testing, an optional flag will be add that can enable posting samples directly to storage to avoid notification-agent process.

Alternatives

An alternative which accomplishes the state goal would be:

  • A more complex solution would be create notifications on the bus that are targeted to a specific transformation service. Such a service would only listen for ceilometer generated measurement notifications, transform them according to a pipeline description and then dispatch them onward to storage or other processing. This model would skip the notification agent entirely.

    Pro: extends the long term plan of decomposing to smaller pieces. Con: adds to the complexity of this change, increasing time to completion and risk of error.

Another alternative which are related but not quite the same include:

  • The issues with testing of the API requiring a pipeline.yaml could be addressed by having a code-based set of pipeline defaults which are used in the absence of a pipeline.yaml file (and could also be used to generated the default pipeline.yaml).

Data model impact

The proposal does not impact the structure of the data, but it would impact the flow of the data. Testing will be required to see what kind of impact this will have on the timeliness of data acquisition.

REST API impact

Samples created via the API will take a slightly different path between the API and their eventual storage.

An optional boolean parameter ‘direct’ will be added to the post-samples API that allow users posting samples directly to storage and avoid notification agent processing(will skip pipeline), this is mainly used for testing.

Security impact

None beyond existing concerns.

Pipeline impact

Pipeline transformations that had been happening in the api server will happen in the notification agent.

Other end user impact

None.

Performance/Scalability Impacts

This will potentially introduce some latency in the time between when a sample is posted to the API and when it lands in the data store.

Other deployer impact

Deployers of the API server will no longer need to deploy a pipeline.yaml alongside the API server.

And deployers need to config the ‘notification_driver’ in ceilometer.conf if need posting samples through notification-agent, such as:

notification_driver = messaging

Developer impact

This change may increase the amount of asynchrony in some tests. For example, the time between posting a sample and being able to retrieve that sample may become more unpredictable. As it is already unpredictable any tests which rely on immediate retrieval are bad anyway, so we should fix that.

Implementation

Assignee(s)

Primary assignee:

liusheng

Other contributors:

chdent

Ongoing maintainer:

liusheng, chdent

Work Items

  • Ensure existence of robust tests of posting meters that are transformed by the pipeline.

  • Replace pipeline-based publishing in the API code with a notifier.

  • Update API tests to deal with things like mocked pipelines that no longer exist.

  • API option to perform direct writing to storage

Future lifecycle

The members of the Telemetry program will do any required ongoing maintenance for this feature.

Dependencies

No new dependencies.

Testing

In addition to existing unit tests of the API, the new declarative http tests can be used to confirm that posted samples continue to have expected transformations.

Documentation Impact

Deployment documentation may need some slight updates to indicate the change in configuration file requirements for the api server.

References

None