Skip to content

billyrrr/onto

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

onto

Build Status Coverage Status Documentation Status

Demo:

When you change the attendance status of one of the participants in the meeting, all other participants receive an updated version of the list of people attending the meeting.

Untitled_2

Some reasons that you may want to use this framework or architectual practice:

  • You want to build a reactive system and not just a reactive view.
  • You want to build a scalable app that is native to distributed systems.
  • You want a framework with a higher level of abstraction, so you can exchange components such as transportation protocols
  • You want your code to be readable and clear and written mostly in python, while maintaining compatibility to different APIs.
  • You have constantly-shifting requirements, and want to have the flexibility to migrate different layers, for example, switch from REST API to WebSocket to serve a resource.

This framework is at beta testing stage. API is not guaranteed and may change.

Documentations: readthedocs

Quickstart: Quickstart

API Documentations: API Docs

Example of a Project using onto (now named onto): gravitate-backend

Related Technologies

Connectors supported

Implemented:

  • REST API (Flask and Flasgger)
  • GraphQL (Starlette)
  • Firestore
  • Firebase Functions
  • JsonRPC (flask-jsonrpc)
  • Leancloud Engine
  • WebSocket (flask socketio)

To be supported:

  • Flink Table API
  • Kafka

What I am currently trying to build

Front end creates mutations in graphql. "Onto" receives the view model, and triggers action on domain model. A method in domain model is called (which lives in Flink Stateful Functions runtime). Different domain models communicate to persist a change, and save the output view into Kafka. Another set of system statically interprets "view model definition code" as SQL, and submit jobs with Flink SQL to assemble "view model". Eventually, the 1NF view of the data is sent to Kafka, and eventually delivered to front end in forms of GraphQL Subscription.

(Write side has Serializable-level consistency, and read side has eventual consistency)

What it already does

  • Serialization and deserialization
  • GraphQL/Flask server
  • Multiple table join
  • ...

Installation

In your project directory,

pip install onto

See more in Quickstart.

State Management

You can combine information gathered in domain models and serve them in Firestore, so that front end can read all data required from a single document or collection, without client-side queries and excessive server roundtrip time.

There is a medium article that explains a similar architecture called "reSolve" architecture.

See examples/meeting_room/view_models on how to use onto to expose a "view model" in firestore that can be queried directly by front end without aggregation.

Processor Modes

onto is essentially a framework for source-sink operations:

Source(s) -> Processor -> Sink(s)

Take query as an example,

  • Boiler
  • NoSQL
  • Flink
    • staticmethods: converts to UDF
    • classmethods: converts to operators and aggregator's

Declare View Model

from onto.attrs import attrs

class CityView(ViewModel):

    name: str = attrs.nothing
    country: str = attrs.nothing

    @classmethod
    def new(cls, snapshot):
        store = CityStore()
        store.add_snapshot("city", dm_cls=City, snapshot=snapshot)
        store.refresh()
        return cls(store=store)

    @name.getter
    def name(self):
        return self.store.city.city_name

    @country.getter
    def country(self):
        return self.store.city.country

    @property
    def doc_ref(self):
        return CTX.db.document(f"cityView/{self.store.city.doc_id}")

Document View

class MeetingSessionGet(Mediator):

    from onto import source, sink

    source = source.domain_model(Meeting)
    sink = sink.firestore()  # TODO: check variable resolution order

    @source.triggers.on_update
    @source.triggers.on_create
    def materialize_meeting_session(self, obj):
        meeting = obj
        assert isinstance(meeting, Meeting)

        def notify(obj):
            for ref in obj._view_refs:
                self.sink.emit(reference=ref, snapshot=obj.to_snapshot())

        _ = MeetingSession.get(
            doc_id=meeting.doc_id,
            once=False,
            f_notify=notify
        )
        # mediator.notify(obj=obj)

    @classmethod
    def start(cls):
        cls.source.start()

Create Flask View

You can use a RestMediator to create a REST API. OpenAPI3 docs will be automatically generated in <site_url>/apidocs when you run _ = Swagger(app).

app = Flask(__name__)

class MeetingSessionRest(Mediator):

    # from onto import source, sink

    view_model_cls = MeetingSessionC

    rest = RestViewModelSource()

    @rest.route('/<doc_id>', methods=('GET',))
    def materialize_meeting_session(self, doc_id):

        meeting = Meeting.get(doc_id=doc_id)

        def notify(obj):
            d = obj.to_snapshot().to_dict()
            content = jsonify(d)
            self.rest.emit(content)

        _ = MeetingSessionC.get(
            doc_id=meeting.doc_id,
            once=False,
            f_notify=notify
        )

    # @rest.route('/', methods=('GET',))
    # def list_meeting_ids(self):
    #     return [meeting.to_snapshot().to_dict() for meeting in Meeting.all()]

    @classmethod
    def start(cls, app):
        cls.rest.start(app)

swagger = Swagger(app)

app.run(debug=True)

(currently under implementation)

Object Lifecycle

Once

Object created with cls.new -> Object exported with obj.to_view_dict.

Multi

Object created when a new domain model is created in database -> Object changed when underlying datasource changes -> Object calls self.notify

Typical ViewMediator Use Cases

Data flow direction is described as Source -> Sink. "Read" describes the flow of data where front end would find data in Sink useful. "Write" describes the flow of data where the Sink is the single source of truth.

Rest

Read: Request -> Response
Write: Request -> Document

  1. Front end sends HTTP request to Server
  2. Server queries datastore
  3. Server returns response

Query

Read: Document -> Document
Write: Document -> Document

  1. Datastore triggers update function
  2. Server rebuilds ViewModel that may be changed as a result
  3. Server saves newly built ViewModel to datastore

Query+Task

Read: Document -> Document
Write: Document -> Document

  1. Datastore triggers update function for document d at time t
  2. Server starts a transaction
  3. Server sets write_option to only allow commit if documents are last updated at time t (still under design)
  4. Server builds ViewModel with transaction
  5. Server saves ViewModel with transaction
  6. Server marks document d as processed (remove document or update a field)
  7. Server retries up to MAX_RETRIES from step 2 if precondition failed

WebSocket

Read: Document -> WebSocket Event
Write: WebSocket Event -> Document

  1. Front end subscribes to a ViewModel by sending a WebSocket event to server
  2. Server attaches listener to the result of the query
  3. Every time the result of the query is changed and consistent:
    1. Server rebuilds ViewModel that may be changed as a result
    2. Server publishes newly built ViewModel
  4. Front end ends the session
  5. Document listeners are released

Document

Read: Document -> Document
Write: Document -> Document

Comparisons

Rest Query Query+Task WebSocket Document
Guarantees ≤1 (At-Most-Once) ≥ 1 (At-Least-Once) =11 (Exactly-Once) ≤1 (At-Most-Once) ≥ 1 (At-Least-Once)
Idempotence If Implemented No Yes, with transaction1 If Implemented No
Designed For Stateless Lambda Stateful Container Stateless Lambda Stateless Lambda Stateful Container
Latency Higher Higher Higher Lower Higher
Throughput Higher when Scaled Lower2 Lower Higher when Scaled Lower2
Stateful No If Implemented If Implemented Yes Yes
Reactive No Yes Yes Yes Yes

Comparisons

GraphQL

In GraphQL, the fields are evaluated with each query, but onto evaluates the fields if and only if the underlying data source changes. This leads to faster read for data that has not changed for a while. Also, the data source is expected to be consistent, as the field evaluation are triggered after all changes made in one transaction to firestore is read.

GraphQL, however, lets front-end customize the return. You must define the exact structure you want to return in onto. This nevertheless has its advantage as most documentations of the request and response can be done the same way as REST API.

REST API / Flask

REST API does not cache or store the response. When a view model is evaluated by onto, the response is stored in firestore forever until update or manual removal.

onto controls role-based access with security rules integrated with Firestore. REST API usually controls these access with a JWT token.

Redux

Redux is implemented mostly in front end. onto targets back end and is more scalable, since all data are communicated with Firestore, a infinitely scalable NoSQL datastore.

onto is declarative, and Redux is imperative. The design pattern of REDUX requires you to write functional programming in domain models, but onto favors a different approach: ViewModel reads and calculates data from domain models and exposes the attribute as a property getter. (When writing to DomainModel, the view model changes domain model and exposes the operation as a property setter). Nevertheless, you can still add function callbacks that are triggered after a domain model is updated, but this may introduce concurrency issues and is not perfectly supported due to the design tradeoff in onto.

Architecture Diagram:

Architecture Diagram

Contributing

Pull requests are welcome.

Please make sure to update tests as appropriate.

License

MIT

Footnotes

  1. A message may be received and processed by multiple consumer, but only one consumer can successfully commit change and mark the event as processed. 2

  2. Scalability is limited by the number of listeners you can attach to the datastore. 2