Easy stream processing utilizing Python and tumbling home windows

On this tutorial, I wish to present you find out how to downsample a stream of sensor knowledge utilizing solely Python (and Redpanda as a message dealer). The aim is to indicate you the way easy stream processing will be, and that you just don’t want a heavy-duty stream processing framework to get began.

Till lately, stream processing was a fancy activity that often required some Java experience. However steadily, the Python stream processing ecosystem has matured and there are a couple of extra choices accessible to Python builders — akin to Faust, Bytewax and Quix. Later, I’ll present a bit extra background on why these libraries have emerged to compete with the present Java-centric choices.

However first let’s get to the duty at hand. We are going to use a Python libary known as Quix Streams as our stream processor. Quix Streams is similar to Faust, nevertheless it has been optimized to be extra concise in its syntax and makes use of a Pandas like API known as StreamingDataframes.

You’ll be able to set up the Quix Streams library with the next command:

pip set up quixstreams

What you’ll construct

You’ll construct a easy software that may calculate the rolling aggregations of temperature readings coming from numerous sensors. The temperature readings will are available in at a comparatively excessive frequency and this software will combination the readings and output them at a decrease time decision (each 10 seconds). You’ll be able to consider this as a type of compression since we don’t wish to work on knowledge at an unnecessarily excessive decision.

You’ll be able to entry the entire code on this GitHub repository.

This software consists of code that generates artificial sensor knowledge, however in a real-world situation this knowledge might come from many sorts of sensors, akin to sensors put in in a fleet of automobiles or a warehouse filled with machines.

Right here’s an illustration of the fundamental structure:

The earlier diagram displays the principle elements of a stream processing pipeline: You have got the sensors that are the knowledge producers, Redpanda because the streaming knowledge platform, and Quix because the stream processor.

Knowledge producers

These are bits of code which are connected to programs that generate knowledge akin to firmware on ECUs (Engine Management Models), monitoring modules for cloud platforms, or net servers that log person exercise. They take that uncooked knowledge and ship it to the streaming knowledge platform in a format that that platform can perceive.

Streaming knowledge platform

That is the place you place your streaming knowledge. It performs kind of the identical function as a database does for static knowledge. However as an alternative of tables, you employ subjects. In any other case, it has related options to a static database. You’ll wish to handle who can eat and produce knowledge, what schemas the info ought to adhere to. Not like a database although, the info is consistently in flux, so it’s not designed to be queried. You’d often use a stream processor to remodel the info and put it elsewhere for knowledge scientists to discover or sink the uncooked knowledge right into a queryable system optimized for streaming knowledge akin to RisingWave or Apache Pinot. Nevertheless, for automated programs which are triggered by patterns in streaming knowledge (akin to suggestion engines), this isn’t a perfect resolution. On this case, you positively wish to use a devoted stream processor.

Stream processors

These are engines that carry out steady operations on the info because it arrives. They could possibly be in comparison with simply common previous microservices that course of knowledge in any software again finish, however there’s one huge distinction. For microservices, knowledge arrives in drips like droplets of rain, and every “drip” is processed discreetly. Even when it “rains” closely, it’s not too arduous for the service to maintain up with the “drops” with out overflowing (consider a filtration system that filters out impurities within the water).

For a stream processor, the info arrives as a steady, large gush of water. A filtration system can be shortly overwhelmed except you alter the design. I.e. break the stream up and route smaller streams to a battery of filtration programs. That’s form of how stream processors work. They’re designed to be horizontally scaled and work in parallel as a battery. And so they by no means cease, they course of the info repeatedly, outputting the filtered knowledge to the streaming knowledge platform, which acts as a form of reservoir for streaming knowledge. To make issues extra sophisticated, stream processors usually have to maintain observe of knowledge that was acquired beforehand, akin to within the windowing instance you’ll check out right here.

Be aware that there are additionally “knowledge customers” and “knowledge sinks” — programs that eat the processed knowledge (akin to entrance finish purposes and cell apps) or retailer it for offline evaluation (knowledge warehouses like Snowflake or AWS Redshift). Since we gained’t be protecting these on this tutorial, I’ll skip over them for now.

On this tutorial, I’ll present you find out how to use a neighborhood set up of Redpanda for managing your streaming knowledge. I’ve chosen Redpanda as a result of it’s very straightforward to run regionally.

You’ll use Docker compose to shortly spin up a cluster, together with the Redpanda console, so be sure to have Docker put in first.

First, you’ll create separate information to provide and course of your streaming knowledge. This makes it simpler to handle the working processes independently. I.e. you’ll be able to cease the producer with out stopping the stream processor too. Right here’s an outline of the 2 information that you just’ll create:

• The stream producer: sensor_stream_producer.py
Generates artificial temperature knowledge and produces (i.e. writes) that knowledge to a “uncooked knowledge” supply matter in Redpanda. Similar to the Faust instance, it produces the info at a decision of roughly 20 readings each 5 seconds, or round 4 readings a second.
• The stream processor: sensor_stream_processor.py
Consumes (reads) the uncooked temperature knowledge from the “supply” matter, performs a tumbling window calculation to lower the decision of the info. It calculates the common of the info acquired in 10-second home windows so that you get a studying for each 10 seconds. It then produces these aggregated readings to the agg-temperatures matter in Redpanda.

As you’ll be able to see the stream processor does many of the heavy lifting and is the core of this tutorial. The stream producer is a stand-in for a correct knowledge ingestion course of. For instance, in a manufacturing situation, you may use one thing like this MQTT connector to get knowledge out of your sensors and produce it to a subject.

• For a tutorial, it’s easier to simulate the info, so let’s get that arrange first.

You’ll begin by creating a brand new file known as sensor_stream_producer.py and outline the principle Quix software. (This instance has been developed on Python 3.10, however totally different variations of Python 3 ought to work as properly, so long as you’ll be able to run pip set up quixstreams.)

Create the file sensor_stream_producer.py and add all of the required dependencies (together with Quix Streams)

from dataclasses import dataclass, asdict # used to outline the info schema
from datetime import datetime # used to handle timestamps
from time import sleep # used to decelerate the info generator
import uuid # used for message id creation
import json # used for serializing knowledge
from quixstreams import Software

Then, outline a Quix software and vacation spot matter to ship the info.

app = Software(broker_address='localhost:19092')
destination_topic = app.matter(title='raw-temp-data', value_serializer="json")

The value_serializer parameter defines the format of the anticipated supply knowledge (to be serialized into bytes). On this case, you’ll be sending JSON.

Let’s use the dataclass module to outline a really primary schema for the temperature knowledge and add a operate to serialize it to JSON.

@dataclass
class Temperature:
ts: datetime
worth: int
def to_json(self):
# Convert the dataclass to a dictionary
knowledge = asdict(self)
# Format the datetime object as a string
knowledge['ts'] = self.ts.isoformat()
# Serialize the dictionary to a JSON string
return json.dumps(knowledge)

Subsequent, add the code that might be accountable for sending the mock temperature sensor knowledge into our Redpanda supply matter.

i = 0
with app.get_producer() as producer:
whereas i < 10000:
sensor_id = random.alternative(["Sensor1", "Sensor2", "Sensor3", "Sensor4", "Sensor5"])
temperature = Temperature(datetime.now(), random.randint(0, 100))
worth = temperature.to_json()
print(f"Producing worth {worth}")
serialized = destination_topic.serialize(
key=sensor_id, worth=worth, headers={"uuid": str(uuid.uuid4())}
)
producer.produce(
matter=destination_topic.title,
headers=serialized.headers,
key=serialized.key,
worth=serialized.worth,
)
i += 1
sleep(random.randint(0, 1000) / 1000)

This generates 1000 data separated by random time intervals between 0 and 1 second. It additionally randomly selects a sensor title from an inventory of 5 choices.

Now, check out the producer by working the next within the command line

python sensor_stream_producer.py

It is best to see knowledge being logged to the console like this:

[data produced]

When you’ve confirmed that it really works, cease the method for now (you’ll run it alongside the stream processing course of later).

The stream processor performs three major duties: 1) eat the uncooked temperature readings from the supply matter, 2) repeatedly combination the info, and three) produce the aggregated outcomes to a sink matter.

Let’s add the code for every of those duties. In your IDE, create a brand new file known as sensor_stream_processor.py.

First, add the dependencies as earlier than:

import os
import random
import json
from datetime import datetime, timedelta
from dataclasses import dataclass
import logging
from quixstreams import Software
logging.basicConfig(degree=logging.INFO)
logger = logging.getLogger(__name__)

Let’s additionally set some variables that our stream processing software wants:

TOPIC = "raw-temperature" # defines the enter matter
SINK = "agg-temperature"  # defines the output matter
WINDOW = 10  # defines the size of the time window in seconds
WINDOW_EXPIRES = 1 # defines, in seconds, how late knowledge can arrive earlier than it's excluded from the window

We’ll go into extra element on what the window variables imply a bit later, however for now, let’s crack on with defining the principle Quix software.

app = Software(
broker_address='localhost:19092',
consumer_group="quix-stream-processor",
auto_offset_reset="earliest",
)

Be aware that there are a couple of extra software variables this time round, specifically consumer_group and auto_offset_reset. To be taught extra concerning the interaction between these settings, try the article “Understanding Kafka’s auto offset reset configuration: Use instances and pitfalls“

Subsequent, outline the enter and output subjects on both facet of the core stream processing operate and add a operate to place the incoming knowledge right into a DataFrame.

input_topic = app.matter(TOPIC, value_deserializer="json")
output_topic = app.matter(SINK, value_serializer="json")
sdf = app.dataframe(input_topic)
sdf = sdf.replace(lambda worth: logger.data(f"Enter worth acquired: {worth}"))

We’ve additionally added a logging line to ensure the incoming knowledge is undamaged.

Subsequent, let’s add a customized timestamp extractor to make use of the timestamp from the message payload as an alternative of Kafka timestamp. On your aggregations, this principally signifies that you wish to use the time that the studying was generated somewhat than the time that it was acquired by Redpanda. Or in even easier phrases “Use the sensor’s definition of time somewhat than Redpanda’s”.

def custom_ts_extractor(worth):
# Extract the sensor's timestamp and convert to a datetime object
dt_obj = datetime.strptime(worth["ts"], "%Y-%m-%dTpercentH:%M:%S.%f") # 
# Convert to milliseconds because the Unix epoch for efficent procesing with Quix
milliseconds = int(dt_obj.timestamp() * 1000)
worth["timestamp"] = milliseconds
logger.data(f"Worth of latest timestamp is: {worth['timestamp']}")
return worth["timestamp"]
# Override the beforehand outlined input_topic variable in order that it makes use of the customized timestamp extractor 
input_topic = app.matter(TOPIC, timestamp_extractor=custom_ts_extractor, value_deserializer="json") 

Why are we doing this? Properly, we might get right into a philosophical rabbit gap about which form of time to make use of for processing, however that’s a topic for an additional article. With the customized timestamp, I simply needed for instance that there are numerous methods to interpret time in stream processing, and also you don’t essentially have to make use of the time of knowledge arrival.

Subsequent, initialize the state for the aggregation when a brand new window begins. It should prime the aggregation when the primary file arrives within the window.

def initializer(worth: dict) -> dict:
value_dict = json.hundreds(worth)
return {
'depend': 1,
'min': value_dict['value'],
'max': value_dict['value'],
'imply': value_dict['value'],
}

This units the preliminary values for the window. Within the case of min, max, and imply, they’re all equivalent since you’re simply taking the primary sensor studying as the place to begin.

Now, let’s add the aggregation logic within the type of a “reducer” operate.

def reducer(aggregated: dict, worth: dict) -> dict:
aggcount = aggregated['count'] + 1
value_dict = json.hundreds(worth)
return {
'depend': aggcount,
'min': min(aggregated['min'], value_dict['value']),
'max': max(aggregated['max'], value_dict['value']),
'imply': (aggregated['mean'] * aggregated['count'] + value_dict['value']) / (aggregated['count'] + 1)
}

This operate is barely obligatory if you’re performing a number of aggregations on a window. In our case, we’re creating depend, min, max, and imply values for every window, so we have to outline these prematurely.

Subsequent up, the juicy half — including the tumbling window performance:

### Outline the window parameters akin to sort and size
sdf = (
# Outline a tumbling window of 10 seconds
sdf.tumbling_window(timedelta(seconds=WINDOW), grace_ms=timedelta(seconds=WINDOW_EXPIRES))
# Create a "cut back" aggregation with "reducer" and "initializer" features
.cut back(reducer=reducer, initializer=initializer)
# Emit outcomes just for closed 10 second home windows
.closing()
)
### Apply the window to the Streaming DataFrame and outline the info factors to incorporate within the output
sdf = sdf.apply(
lambda worth: {
"time": worth["end"], # Use the window finish time because the timestamp for message despatched to the 'agg-temperature' matter
"temperature": worth["value"], # Ship a dictionary of {depend, min, max, imply} values for the temperature parameter
}
)

This defines the Streaming DataFrame as a set of aggregations primarily based on a tumbling window — a set of aggregations carried out on 10-second non-overlapping segments of time.

Tip: When you want a refresher on the various kinds of windowed calculations, try this text: “A information to windowing in stream processing”.

Lastly, produce the outcomes to the downstream output matter:

sdf = sdf.to_topic(output_topic)
sdf = sdf.replace(lambda worth: logger.data(f"Produced worth: {worth}"))
if __name__ == "__main__":
logger.data("Beginning software")
app.run(sdf)

Be aware: You may surprise why the producer code seems very totally different to the producer code used to ship the artificial temperature knowledge (the half that makes use of with app.get_producer() as producer()). It’s because Quix makes use of a unique producer operate for transformation duties (i.e. a activity that sits between enter and output subjects).

As you may discover when following alongside, we iteratively change the Streaming DataFrame (the sdf variable) till it’s the closing kind that we wish to ship downstream. Thus, the sdf.to_topic operate merely streams the ultimate state of the Streaming DataFrame again to the output matter, row-by-row.

The producer operate alternatively, is used to ingest knowledge from an exterior supply akin to a CSV file, an MQTT dealer, or in our case, a generator operate.

Lastly, you get to run our streaming purposes and see if all of the transferring components work in concord.

First, in a terminal window, begin the producer once more:

python sensor_stream_producer.py

Then, in a second terminal window, begin the stream processor:

python sensor_stream_processor.py

Take note of the log output in every window, to ensure all the pieces is working easily.

You can too verify the Redpanda console to guarantee that the aggregated knowledge is being streamed to the sink matter accurately (you’ll high-quality the subject browser at: http://localhost:8080/subjects).

What you’ve tried out right here is only one strategy to do stream processing. Naturally, there are heavy obligation instruments such Apache Flink and Apache Spark Streaming that are have additionally been coated extensively on-line. However — these are predominantly Java-based instruments. Positive, you need to use their Python wrappers, however when issues go fallacious, you’ll nonetheless be debugging Java errors somewhat than Python errors. And Java abilities aren’t precisely ubiquitous amongst knowledge people who’re more and more working alongside software program engineers to tune stream processing algorithms.

On this tutorial, we ran a easy aggregation as our stream processing algorithm, however in actuality, these algorithms usually make use of machine studying fashions to remodel that knowledge — and the software program ecosystem for machine studying is closely dominated by Python.

An oft missed reality is that Python is the lingua franca for knowledge specialists, ML engineers, and software program engineers to work collectively. It’s even higher than SQL as a result of you need to use it to do non-data-related issues like make API calls and set off webhooks. That’s one of many the reason why libraries like Faust, Bytewax and Quix developed — to bridge the so-called impedance hole between these totally different disciplines.

Hopefully, I’ve managed to indicate you that Python is a viable language for stream processing, and that the Python ecosystem for stream processing is maturing at a gentle charge and may maintain its personal in opposition to the older Java-based ecosystem.