- Dependencies:
- a running MongoDB accessible to minibatch
- Python 3.x
- see extras & optional dependencies below for specific requirements
minibatch provides a straight-forward, Python-native approach to mini-batch streaming and complex-event processing that is easily scalable. Streaming primarily consists of
- a producer, which is some function inserting data into the stream
- a consumer, which is some function retrieving data from the stream
- transform and windowing functions to process the data in small batches and in parallel
minibatch is an integral part of omega|ml, however also works independently. omega|ml is the Python DataOps and MLOps platform for humans.
- native Python producers and consumers
- includes three basic Window strategies: CountWindow, FixedTimeWindow, RelaxedTimeWindow
- extensible Window strategies by subclassing and overriding a few methods
- scalable, persistent streams - parallel inserts, parallel processing of windows
A few hightlights
- creating a stream and appending data is just 2 lines of code
- producer and consumer stream code runs anywhere
- no dependencies other than mongoengine, pymongo
- extensible sources and sinks (already available: Kafka, MQTT, MongoDB collections, omega|ml datasets)
- a fully functional streaming web app can be built in less than 15 lines of code (using Flask)
Why is it called minibatch? Because it focuses on getting things done by using existing technology, and making it easy to use this techonlogy. It may be minimalistic in approach, but maximises results.
Install and setup
$ pip install minibatch $ docker run -d -p 27017:27017 mongo
See extras & optional dependencies below to select specific packages according to your deployment needs, e.g. for MQTT, Kafka, omega|ml
Create a stream producer or attach to a source
import minibatch as mb stream = mb.stream('test') for i in range(100): stream.append({'date': datetime.datetime.utcnow().isoformat()}) sleep(.5)
Currently there is support for Kafka and MQTT sources. However arbitrary other sources can be added.
from minibatch.contrib.kafka import KafkaSource source = KafkaSource('topic', urls=['kafka:port']) stream.attach(source)
Consume the stream
from minibatch import streaming @streaming('test', size=2, keep=True) def myprocess(window): print(window.data) return window => [{'date': '2018-04-30T20:18:22.918060'}, {'date': '2018-04-30T20:18:23.481320'}] [{'date': '2018-04-30T20:18:24.041337'}, {'date': '2018-04-30T20:18:24.593545'} ...
myprocess is called for every N-tuple of items (size=2) appended to the stream by the producer(s). The frequency is determined by the emitter strategy. This can be configured or changed for a custom emitter strategy, as shown in the next step.
Configure the emitter strategy
Note the @streaming decorator. It implements a blocking consumer that delivers batches of data according to some strategy implemented by a WindowEmitter. Currently @streaming provides the following interface:
- size=N - uses the
CountWindow
emitter - interval=SECONDS - uses the
RelaxedTimeWindow
emitter - interval=SECONDS, relaxed=False - uses the
FixedTimeWindow
emitter - emitter=CLASS:WindowEmitter - uses the given subclass of a
WindowEmitter
- workers=N - set the number of workers to process the decorated function, defaults to number of CPUs
- executor=CLASS:Executor - the asynchronous executor to use, defaults to
concurrent.futures.ProcessPoolExecutor
- size=N - uses the
Write a flask app as a streaming source
This is a simple helloworld-style streaming application that is fully functional and distributable.
# app.py def consumer(url): @streaming('test-stream', url=url) def processing(window): ... # whatever processing you need to do if __name__ == '__main__': app = StreamingApp() app.start_streaming(consumer) app.run() # run the app (check status at http://localhost:5000/status) $ python app.py # in an other process, stream data $ python [] import minibatch as mb stream = mb.stream('test-stream') stream.append(dict(data='foobar') Note there is no UI in this example, however the data is processed as it comes in. To add a UI, specify using @app.route, as for any flask app, write the processed data into a sink that the UI can access. For a full example see help(minibatch.contrib.apps.omegaml.StreamingApp)
Currently provided in minibatch.contrib
:
- KafkaSource - attach a stream to a Apache Kafka topic
- MQTTSource - attach to an MQTT broker
- MongoSource - attach to a MongoDB collection
- DatasetSource - attach to a omega|ml dataset
- CeleryEventSource - attach to a Celery app event dispatcher
Stream sources are arbitrary objects that support the stream()
method, as follows.
class SomeSource:
...
def stream(self, stream):
for data in source:
stream.append(data)
The result of a stream can be forwarded to a sink. Currently
provided sinks in minibatch.contrib
are:
- KafkaSink - forward messagess to a Apache Kafka topic
- MQTTSink - forward messages to an MQTT broker
- MongoSink - forward messages to a MongoDB collection
- DatasetSink - write to a omega|ml dataset
Stream sinks are arbitrary objects that support the put()
method, as follows.
class SomeSink:
...
def put(self, message):
sink.send(message)
minibatch provides the following window emitters out of the box:
CountWindow
- emit fixed-sized windows. Waits until at least n messages are- available before emitting a new window
FixedTimeWindow
- emit all messages retrieved within specific, time-fixed windows of- a given interval of n seconds. This guarantees that messages were received in the specific window.
RelaxedTimeWindow
- every interval of n seconds emit all messages retrieved since- the last window was created. This does not guarantee that messages were received in a given window.
Custom emitter strategies are implemented as a subclass to WindowEmitter
. The main methods
to implement are
window_ready
- returns the tuple(ready, data)
, where ready is True if there is data- to emit
query
- returns the data for the new window. This function retrieves thedata
part- of the return value of
window_ready
See the API reference for more details.
class SortedWindow(WindowEmitter):
"""
sort all data by value and output only multiples of 2 in batches of interval size
"""
def window_ready(self):
qs = Buffer.objects.no_cache().filter(processed=False)
data = []
for obj in sorted(qs, key=lambda obj : obj.data['value']):
if obj.data['value'] % 2 == 0:
data.append(obj)
if len(data) >= self.interval:
break
self._data = data
return len(self._data) == self.interval, ()
def query(self, *args):
return self._data
Concepts
Instead of directly connection producers and consumers, a producer sends messages to a stream. Think of a stream as an endless buffer, or a pipeline, that takes input from many producers on one end, and outputs messages to a consumer on the other end. This transfer of messages happens asynchronously, that is the producer can send messages to the stream independent of whether the consumer is ready to receive, and the consumer can take messages from the stream independent of whether the producer is ready to send.
Unlike usual asynchronous messaging, however, we want the consumer to receive messages in small batches to optimize throughput. That is, we want the pipeline to emit messages only subject to some criteria of grouping messages, where each group is called a mini-batch. The function that determines whether the batching criteria is met (e.g. time elapsed, number of messages in the pipeline) is called emitter strategy, and the output it produces is called window.
Thus in order to connect producers and consumers we need the following parts to our streaming system:
- a
Stream
, keeping metadata for the stream such as its name and when it was created, last read etc. - a
Buffer
acting as the buffer where messages sent by producers are stored until the emitting - a
WindowEmitter
implementing the emitter strategy - a
Window
representing the output produced by the emitter strategy
Note
The producer accepts input from some external system, say an MQTT end point. The producer's responsibility is to enter the data into the streaming buffer. The consumer uses an emitter strategy to produce a Window of data that is then forwarded to the user's processing code.
Implementation
minibatch uses MongoDB to implement Streams, Buffers and Windows. Specifically, the following collections are used:
- stream - represents instances of Stream, each document is a stream with a unique name
- buffer - a virtually endless buffer for all streams in the system, each document contains one message of a stream
- window- each document represents the data as emitted by the particular emitter strategy
By default messages go through the following states
- upon append by a producer: message is inserted into buffer, with flag processed = False
- upon being seen by an emitter: message is marked as processed = True
- upon being emitted: message is copied to window, marked processed = False (in Window)
- upon emit success (no exceptions raised by the emit function): message is deleted from buffer and marked processed = True in window
Notes:
- emitters typically act on a collection of messages, that is steps 2 - 4 are applied to more than one message at a time
- to avoid deleting messages from the buffer, pass @streaming(..., keep=True)
- custom emitters can modify the behavior of both creating windows and handling the buffer by overriding the process(), emit() and commit() methods for each of the above steps 2/3/4, respectively.
minibatch provides the following pip install extras, which come with some additional dependencies. Extras are installed by running
$ pip install minibatch[<extra>|all]
Available extras are:
apps
- adds StreamingApp for easy development & deployment of producers & consumerskafka
- to work with Kafka as a source or a sinkmqtt
- to work with an MQTT broker as a source or a sinkmongodb
- to work with MongoDB as a source or a sinkomegaml
- to work with omega|ml datasets as a source or a sinkall
- all of the abovedev
- all of the above plus a few development packages
Here are a couple of ideas to extend minibatch. Contributions are welcome.
- more examples, following typical streaming examples like word count, filtering
- more emitter strategies, e.g. for sliding windows
- performance testing, benchmarking
- distributed processing of windows via distributed framework such as celery, ray, dask
- extend emitters by typical stream operations e.g. to support operations like count, filter, map, groupby, merge, join
- add other storage backends (e.g. Redis, or some Python-native in-memory db that provides network access and an easy to use ORM layer, like mongoengine does for MongoDB)
We welcome any contributions - examples, issues, bug reports, documentation, code. Please see CONTRIBUTING.md for details.
Apache 2.0 licensed with "No Sell, Consulting Yes" clause. See LICENSE and LICENSE-NOSELLCLAUSE files.