This work is supported by Continuum Analytics, and the Data Driven Discovery Initiative from the Moore Foundation.
This blogpost is about experimental software. The project may change or be abandoned without warning. You should not depend on anything within this blogpost.
This week I built a small streaming library for Python. This was originally an exercise to help me understand streaming systems like Storm, Flink, Spark-Streaming, and Beam, but the end result of this experiment is not entirely useless, so I thought I’d share it. This blogpost will talk about my experience building such a system and what I valued when using it. Hopefully it elevates interest in streaming systems among the Python community.
Background with Iterators
If these iterators are infinite, for example if they are coming from some infinite data feed like a hardware sensor or stock market signal then most of these pieces still work except for the final aggregation, which we replace with an accumulating aggregation.
This is usually a fine way to handle infinite data streams. However this approach becomes awkward if you don’t want to block on calling next(seq) and have your program hang until new data comes in. This approach also becomes awkward when you want to branch off your sequence to multiple outputs and consume from multiple inputs. Additionally there are operations like rate limiting, time windowing, etc. that occur frequently but are tricky to implement if you are not comfortable using threads and queues. These complications often push people to a computation model that goes by the name streaming.
To introduce streaming systems in this blogpost I’ll use my new tiny library, currently called streams (better name to come in the future). However if you decide to use streaming systems in your workplace then you should probably use some other more mature library instead. Common recommendations include the following:
- ReactiveX (RxPy)
- Storm (Streamparse)
- Spark Streaming
We make a stream, which is an infinite sequence of data into which we can emit values and from which we can subscribe to make new streams.
From here we replicate our example above. This follows the standard map/filter/reduce chaining API.
Note that we haven’t pushed any data into this stream yet, nor have we said what should happen when data leaves. So that we can look at results, lets make a list and push data into it when data leaves the stream.
And now lets push some data in at the source and see it arrive at the sink:
We’ve accomplished the same result as our infinite iterator, except that rather than pulling data with next we push data through with source.emit. And we’ve done all of this at only a 10x slowdown over normal Python iteators 🙂 (this library takes a few microseconds per element rather than CPython’s normal 100ns overhead).
This will get more interesting in the next few sections.
This approach becomes more interesting if we add multiple inputs and outputs.
Or we can combine streams together
So you may have multiple different input sources updating at different rates and you may have multiple outputs, perhaps some going to a diagnostics dashboard, others going to long-term storage, others going to a database, etc.. A streaming library makes it relatively easy to set up infrastructure and pipe everything to the right locations.
Time and Back Pressure
When dealing with systems that produce and consume data continuously you often want to control the flow so that the rates of production are not greater than the rates of consumption. For example if you can only write data to a database at 10MB/s or if you can only make 5000 web requests an hour then you want to make sure that the other parts of the pipeline don’t feed you too much data, too quickly, which would eventually lead to a buildup in one place.
To deal with this, as our operations push data forward they also accept Tornado Futures as a receipt.
Under normal operation you don’t need to think about Tornado futures at all (many Python users aren’t familiar with asynchronous programming) but it’s nice to know that the library will keep track of balancing out flow. The code below uses @gen.coroutine and yield common for Tornado coroutines. This is similar to the async/await syntax in Python 3. Again, you can safely ignore it if you’re not familiar with asynchronous programming.
There are also a number of operations to help you buffer flow in the right spots, control rate limiting, etc..
I’ve written enough little utilities like timed_window and buffer to discover both that in a full system you would want more of these, and that they are easy to write. Here is the definition of timed_window
If you are comfortable with Tornado coroutines or asyncio then my hope is that this should feel natural.
Recursion and Feedback
By connecting the sink of one stream to the emit function of another we can create feedback loops. Here is stream that produces the Fibonnacci sequence. To stop it from overwhelming our local process we added in a rate limiting step:
Note: due to the time rate-limiting functionality this example relied on an event loop running somewhere in another thread. This is the case for example in a Jupyter notebook, or if you have a Dask Client running.
Things that this doesn’t do
If you are familiar with streaming systems then you may say the following:
Lets not get ahead of ourselves; there’s way more to a good streaming system than what is presented here. You need to handle parallelism, fault tolerance, out-of-order elements, event/processing times, etc..
… and you would be entirely correct. What is presented here is not in any way a competitor to existing systems like Flink for production-level data engineering problems. There is a lot of logic that hasn’t been built here (and its good to remember that this project was built at night over a week).
Although some of those things, and in particular the distributed computing bits, we may get for free.
So, during the day I work on Dask, a Python library for parallel and distributed computing. The core task schedulers within Dask are more than capable of running these kinds of real-time computations. They handle far more complex real-time systems every day including few-millisecond latencies, node failures, asynchronous computation, etc.. People use these features today inside companies, but they tend to roll their own system rather than use a high-level API (indeed, they chose Dask because their system was complex enough or private enough that rolling their own was a necessity). Dask lacks any kind of high-level streaming API today.
Fortunately, the system we described above can be modified fairly easily to use a Dask Client to submit functions rather than run them locally.
Other things that this doesn’t do, but could with modest effort
There are a variety of ways that we could improve this with modest cost:
- Streams of sequences: We can be more efficient if we pass not individual elements through a Stream, but rather lists of elements. This will let us lose the microseconds of overhead that we have now per element and let us operate at pure Python (100ns) speeds.
- Streams of NumPy arrays / Pandas dataframes: Rather than pass individual records we might pass bits of Pandas dataframes through the stream. So for example rather than filtering elements we would filter out rows of the dataframe. Rather than compute at Python speeds we can compute at C speeds. We’ve built a lot of this logic before for dask.dataframe. Doing this again is straightforward but somewhat time consuming.
- Annotate elements: we want to pass through event time, processing time, and presumably other metadata
- Convenient Data IO utilities: We would need some convenient way to move data in and out of Kafka and other common continuous data streams.
None of these things are hard. Many of them are afternoon or weekend projects if anyone wants to pitch in.
Reasons I like this project
This was originally built strictly for educational purposes. I (and hopefully you) now know a bit more about streaming systems, so I’m calling it a success. It wasn’t designed to compete with existing streaming systems, but still there are some aspects of it that I like quite a bit and want to highlight.
- Lightweight setup: You can import it and go without setting up any infrastructure. It can run (in a limited way) on a Dask cluster or on an event loop, but it’s also fully operational in your local Python thread. There is no magic in the common case. Everything up until time-handling runs with tools that you learn in an introductory programming class.
- Small and maintainable: The codebase is currently a few hundred lines. It is also, I claim, easy for other people to understand. Here is the code for filter:
- Composable with Dask: Handling distributed computing is tricky to do well. Fortunately this project can offload much of that worry to Dask. The dividing line between the two systems is pretty clear and, I think, could lead to a decently powerful and maintainable system if we spend time here.
- Low performance overhead: Because this project is so simple it has overheads in the few-microseconds range when in a single process.
- Pythonic: All other streaming systems were originally designed for Java/Scala engineers. While they have APIs that are clearly well thought through they are sometimes not ideal for Python users or common Python applications.
This project needs both users and developers.
I find it fun and satisfying to work on and so encourage others to play around. The codebase is short and, I think, easily digestible in an hour or two.
This project was built without a real use case (see the project’s examples directory for a basic Daskified web crawler). It could use patient users with real-world use cases to test-drive things and hopefully provide PRs adding necessary features.
I genuinely don’t know if this project is worth pursuing. This blogpost is a test to see if people have sufficient interest to use and contribute to such a library or if the best solution is to carry on with any of the fine solutions that already exist.
Matthew is a graduate student in Computer Science at the University of Chicago. I'm particularly interested in the interface of computational techniques and scientific applications. My background is in Physics, Astronomy, and Engineering. Specialties: Scientific data analysis Numerical linear algebra Statistics and Optimization Complex networks Simple forms of distributed computing