The following describes our approach to change feeds on HQ. For related content see Cory’s brown bag on the topic
What they are¶
A change feed is modeled after the CouchDB
It can be thought of as a real-time log of “changes” to our database.
Anything that creates such a log is called a “(change) publisher”.
Other processes can listen to a change feed and then do something with the results. Processes that listen to changes are called “subscribers”. In the HQ codebase “subscribers” are referred to as “pillows” and most of the change feed functionality is provided via the pillowtop module. This document refers to pillows and subscribers interchangeably.
Common use cases for change subscribers:
- ETL (our main use case)
- Saving docs to ElasticSearch
- Custom report tables
- UCR data sources
- Cache invalidation
- Real-time visualizations (e.g. dimagisphere)
We use kafka as our primary back-end to facilitate change feeds.
This allows us to decouple our subscribers from the underlying source of changes so that they can be database-agnostic.
For legacy reasons there are still change feeds that run off of CouchDB’s
_changes feed however these are in the process of being phased out.
Topics are a kafka concept that are used to create logical groups (or “topics”) of data.
In the HQ codebase we use topics primarily as a 1:N mapping to HQ document classes (or
Forms and cases currently have their own topics, while everything else is lumped in to a “meta” topic.
This allows certain pillows to subscribe to the exact category of change/data they are interested in (e.g. a pillow that sends cases to elasticsearch would only subscribe to the “cases” topic).
Published changes are just “stubs” but do not contain the full data that was affected.
Each change should be associated with a “document store” which is an abstraction that represents a way to retrieve the document from its original database.
This allows the subscribers to retrieve the full document while not needing to have the underlying source hard-coded (so that it can be changed).
To add a new document store, you can use one of the existing subclasses of
DocumentStore or roll your own.
Publishing changes is the act of putting them into kafka from somewhere else.
Publishing changes from couch is easy since couch already has a great change feed implementation with the
For any database that you want to publish changes from the steps are very simple.
Just create a
ConstructedPillow with a
CouchChangeFeed feed pointed at the database you wish to publish from and a
KafkaProcessor to publish the changes.
There is a utility function (
get_change_feed_pillow_for_db) which creates this pillow object for you.
Currently SQL-based change feeds are published from the app layer.
Basically, you can just call a function that publishes the change in a
.save() function (or a
See the functions in form_processors.change_publishers and their usages for an example of how that’s done.
It is planned (though unclear on what timeline) to find an option to publish changes directly from SQL to kafka to avoid race conditions and other issues with doing it at the app layer. However, this change can be rolled out independently at any time in the future with (hopefully) zero impact to change subscribers.
From anywhere else¶
There is not yet a need/precedent for publishing changes from anywhere else, but it can always be done at the app layer.
Subscribing to changes¶
It is recommended that all new change subscribers be instances (or subclasses) of
You can use the
KafkaChangeFeed object as the change provider for that pillow, and configure it to subscribe to one or more topics.
Look at usages of the
ConstructedPillow class for examples on how this is done.
Porting a new pillow¶
Porting a new pillow to kafka will typically involve the following steps. Depending on the data being published, some of these may be able to be skipped (e.g. if there is already a publisher for the source data, then that can be skipped).
- Setup a publisher, following the instructions above.
- Setup a subscriber, following the instructions above.
- For non-couch-based data sources, you must setup a
DocumentStoreclass for the pillow, and include it in the published feed.
- For any pillows that require additional bootstrap logic (e.g. setting up UCR data tables or bootstrapping elasticsearch indexes) this must be hooked up manually.