Facilitating the Spread of Knowledge and Innovation in Professional Software Development

Write for InfoQ


Choose your language

InfoQ Homepage Presentations Blazing Fast, Minimal Change - Speed up Your Code by Refactoring to Rust

Blazing Fast, Minimal Change - Speed up Your Code by Refactoring to Rust



Lily Mara takes an existing codebase and rewrites part of it in Rust, focusing on writing a Rust reimplementation, cross-language regression testing, performance benchmarking of the new code.


Lily Mara is an Engineering Manager at OneSignal in San Mateo, CA. She manages the Infrastructure Services team, which is responsible for in-house services used by other OneSignal engineering teams. Lily is the author of Refactoring to Rust, an early-access book by Manning Publications about improving the performance of existing software systems through the gradual addition of Rust code.

About the conference


Mara: My name is Lily Mara. I'm an engineering manager at OneSignal in San Mateo, California. I've been using Rust professionally on side of desk projects since about 2017. In 2019, I started at OneSignal, where it's used as a primary language. I spoke at RustConf 2021, about the importance of not over-optimizing our Rust programs. I've spoken at many Rust meetup groups. I'm the author of the book, "Refactoring to Rust," available at

Rust - Full Rewrites

If you've heard of Rust before, you have probably heard one thing over again, you've probably heard that it's pretty fast. It generally performs on par with something like C or C++. It's way faster than some dynamic languages that we're maybe using like Python or Ruby. If you have an older, monolithic application written in one of these languages, maybe with something like Django or Rails, then it might be tempting if you're performance constrained to say, let's throw this old thing out, and let's start over with Rust. Rust is fast, we want to go fast, let's rewrite it in Rust. This is a very tempting idea. The reality is often as they do, encroach on our perfect vision. Full rewrite projects can often be quite problematic for a number of reasons. Some of which, they can often take a lot longer than we expect. We think something's going to take a month, it ends up taking three years. We think something is going to be really easy, we realize that the problem was so much more complicated than we realized. We can introduce new bugs, because different programming languages have different paradigms. When you're trying to adapt old code into a new system, you can misunderstand the way the old thing worked. Full rewrites often also do not fix underlying architectural problems. This might be things like using the wrong database technology. It might also be things in the code. We think a piece of code looks ugly, because we don't understand everything that's going on, and we try and rewrite it in a simpler way. We realize that, there actually was a reason we did all those things in a very strange way in the original code. Full rewrite projects are often problematic.

Microservices, and FFI Refactoring

What else is available to us? There's also microservices, of course. We can break out our monolith into multiple services. We can put the performance where it needs to be put, and leave the monolithic stuff in the monolith. If you're working at a place that doesn't really have a robust infrastructure for managing a bunch of microservices, and maybe you're working at a place where there really is just one monolith rolled out on a couple of boxes, and you're really not ready for the architectural shift of going to a bunch of microservices. What do you do in this case? We're going to discuss the feasibility aspect in a little bit more detail later. The gist of it is that microservices are always the best option for everybody in every circumstance, of course. I would like to propose an alternative that we can use. For the purposes of having a term for it, I'm going to refer to it as FFI refactoring. FFI refactoring is where we take a little piece of the code, we rewrite it in a faster language, in this case, Rust, and we connect it to the original codebase using CFFI. This is something that's going to be a little bit abstracted for us, and it's going to be made a lot easier by some of the binding libraries that we're going to be using. The underlying technology is CFFI. We're going to refer to it as FFI refactoring. We're also going to be using the terms host language to refer to the original programming language, and guest language to refer to the new programming language.

When is this an acceptable strategy for us to use? If you're working at a place that has a really robust infrastructure for working with lots of microservices, then maybe consider using a microservice. If there's an existing pattern, sitting there ready for you to use, maybe just use that existing pattern. It's generally much easier to follow what's already sitting there than to try and blaze a new trail. If you're in a place where there isn't a robust infrastructure for microservices, if architecturally shifting to microservices would be really difficult for you, or maybe you're running code on an end user device and you don't necessarily want to have a bunch of binaries talking to each other over local loopback networking, in order to run your program, then maybe FFI refactoring is a better option for you.

There are some unfortunate realities that we're going to need to discuss when we're going to FFI refactoring. Because we're going to move to multiple languages for our program, we are probably going to be complicating our deployments a bit. Because we're going to have to ship not just a bunch of Python and Ruby files to the servers, we're going to have to compile Rust beforehand. We're going to have to ship Rust dynamic library files to our servers. We're going to have to make sure OS versions and compiler versions match up. This will get slightly more complicated. It's possible that we can add bugs, just like with a full rewrite project, all we're doing now is a rewrite on a smaller scale. Because we're reimplementing code, we can of course create bugs. Because it's on a smaller scale, the chance for that is maybe a little bit less. We will definitely have to watch out for translation bugs, because we're moving between multiple programming languages. We don't have to just worry about the quirks of Python, we also have to worry about the quirks of Python and Rust and the quirks of translating Python data structures to Rust data structures.

What makes a good project for an FFI refactor? As I've kept hammering home, it's very similar to the discussion that we've been having for several years now of microservices versus monoliths. Do we want to make our big deploy even bigger or do we want to split it out into a bunch of stuff? Microservices can be great. You can scale independently. You can upgrade independently. You can deploy independently. One thing going down doesn't necessarily take everything else down. There's lots of reasons to use microservices. There's also reasons to consider doing an FFI refactor. As I've said, if you have a monolith, and it will be difficult for you to go to a microservices based approach, maybe do FFI refactoring. Or, it may also be the case that you need to do a very slight performance bump and that could be maybe dwarfed by networking overhead. If you use an FFI refactor, you can keep everything within memory, within a single process, and you can get some serious performance benefits by doing that.

You should also consider what language you're using as your host language. If we take a look at this little compass here, we can see there are some languages that are going to be slower than Rust, where doing an FFI refactor to Rust will probably improve performance. C and C++ generally are like on par with Rust as far as performance goes, maybe slightly faster, so doing an FFI refactor to Rust might actually decrease performance a little bit. You should also consider how the tooling is for your language. Some languages, Ruby, Python, Node.js, have really good tooling for integrating with Rust. Lua also has quite good tooling for integrating with Rust. Languages in this upper right quadrant right here are going to be really good choices for us to use for an FFI refactor.

If you're dealing with something that has poor tooling, or something that's really not going to get much of a performance benefit from refactoring to Rust, then you should maybe consider other options. Go is a pretty interesting choice, because Go is at a similar performance level to Rust. Rust is generally faster because it doesn't have a garbage collector, it doesn't have quite as heavy of a runtime. However, the tooling is not great, because if we wanted to integrate Go with Rust, we have to rely on the CFFI. Go developers can tell you that once you have to invoke the CFFI in Go, it slows down a lot. There hasn't been a whole lot of development work on building out great Go bindings that I'm aware of at least, because people are aware that there's this huge performance penalty that will have to be paid, if you want to do an FFI linking between Go and Rust. Generally speaking, a language like Ruby, Python, Node, Lua is going to be a really good choice, and others not so good.


For the purposes of having a concrete example to talk through in the Rust, of this talk, we're going to imagine that you're a developer, you're working on a Flask HTTP server application that's written in Python. We're just going to take a look at this one handler just so that we have something really small and concise that we can deal with. This handler, it takes in a list of numbers in a JSON request POST body. It computes several statistical properties about those numbers. It computes the range, which is the difference between the maximum and the minimum values. It computes the quartiles, which are the 25th, 50th, and 75th percentiles. It computes the mean, the average of all the numbers. It computes the standard deviation, which is something that I don't exactly know the definition of, but statisticians tell me it's important. Let's see how we can do an FFI refactor, let's see how we can redo this in Rust. Let's go.

There is a free resource at This is the book, "The Rust Programming Language," written by Steve Klabnik, and Carol Nichols, Goulding. It's available for free online. There's a few places where I'm going to be calling out which chapter in "The Rust Programming Language" you should read through if you would like to get some more information on one of these subjects. If you are interested in Rust more generally, I would highly recommend reading through the book because it's a pretty good book.

We're going to get started by creating a new Rust project by running, cargo new --lib rstats. This is going to create a couple new files for us. The first one is carg.toml, which is like the package manager's registry file, sort of like package JSON in a Node project. It's going to create a file. This is the entry point for our crate. Let's open up that cargo.toml file, and we're going to add a couple of dependencies. The first one is we're going to add a statistics crate version 0.15 of a crate called statrs. If you noticed, in the Python code, we were actually using the statistics module from Python's standard library. Rust has a much smaller standard library than Python's. It basically only includes OS primitives, things like files, threads, basic timer functionality, and some networking code, as well as some generic data structures. Python has a very large standard library by comparison. We're bringing in this statistics crate so that we have access to some statistical functions.

We're also going to be bringing in version 0.16 of the pyo3 crate. Pyo3 is going to be used to generate the bindings that talk between Rust and Python. We're also going to need to enable the extension module feature. This is required for making an extension module, making something that compiles Rust code into something that Python knows how to deal with. There are other features available for doing different things. You can, for example, write Rust code that runs Python code. Lots of different options available to us. We're also going to add a little bit of metadata further up in the cargo.toml. We're going to set the crate type to be a cdylib. Normally, when we compile Rust, we're actually compiling code that is only useful to the same version of the Rust compiler on the same hardware architecture. Setting the crate type to cdylib will actually cause us to use C calling conventions, and this is necessary so that the Python interpreter knows how to call our functions.

Now, the general architecture of what we're going to do here. We have our Python code. We have our Rust code. The Flask library is going to call into our Python HTTP handler, which is going to deserialize the JSON request body. It's going to send that over across the FFI boundary into Rust, which is going to compute the statistics. Then we're going to send that back across the FFI boundary to Python. That is going to be serialized back into JSON, which is then going to go back out to the HTTP client. We could more easily have the JSON deserialize and serialize steps happen inside of Rust. I didn't want to do that because it goes against the spirit of this talk. The idea is that we can take one piece of functionality, and we can rewrite that one piece of functionality in Rust. In my mind, the JSON serialize, deserializing is some extra piece of work that needs to stay in Python for some reason. We're going to keep that in Python. It's also going to give us the opportunity to see how we can parse structured data back and forth between these two languages. Because if we were doing the JSON parsing and serializing in Rust, then we would actually just be parsing strings back and forth. That's a little bit less interesting.

Now let's go ahead and jump into the code. We're going to open up the file in the source directory, and there's going to be a bunch of starter code in there. We're just going to go ahead and delete all that. We're going to create a new function called compute_stats. It's going to take in a Vec of f64s, that is a growable array of 64-bit floating point numbers that lives on the heap. We're going to call that numbers. What is our return type going to be on this function?

Let's look back to the Python code. The Python code returns a JSON object that has these four properties. It has a range, quartiles, means, and stddev, standard deviation. In Rust, we generally don't parse around anonymous dictionaries that have complex types for the values. Generally speaking, we use structs that have well typed fields. We're going to create a new struct in our Rust code, we're going to call it StatisticsResponse. It's going to have those expected four fields in it. It's going to have three f64 values for the range, the mean, and the standard deviation. It's also going to have a quartiles field that has an array of three f64 values. This is going to match the structure of our Python code. Then we'll set the return type of our compute_stats function to be that StatisticsResponse type. We're now going to need to bring in a couple of types from the statrs library. These are all necessary, and I know there's a lot of them. We're going to bring in data, distribution, max, min, and OrderStatistics. Some of these are types. Some of these are traits, but we need to bring all of them in so that we can compute the statistics that we need to.

Jumping down, back into our compute_stats function, we're going to take that vector of numbers and we're going to put it into a data which is a type that we just pulled out of the statrs crate. This is necessary because a lot of the traits that we just pulled in, they can only be called on a data instance and not on a vector directly. You can also see that we marked our data as being mutable. That is because in order to compute some of these statistics, statrs is actually going to shuffle some of the elements in our data structure around. If you're used to coming from a language like Python, or Ruby, or Java, then this might seem a little bit strange to you, because, normally, I think in those languages, if items need to be shuffled around, it's pretty common for the library to actually make a defensive copy of whatever your input buffer is, so that as a user of that library, you're not going to have your data changed around. Generally speaking, Rust takes the exact opposite approach, where if things need to be mutated under the hood, that will be exposed to the users so that if the original order of your data buffer was not strictly required, you don't have to do any defensive copies at all. Your code can be just a teensy bit little faster. If you have a really big set of numbers that you're computing statistics on, you don't need to copy those at all, you could have a multi-gigabyte vector of numbers to compute statistics on and they'll just be shuffled around in memory as required, instead of needing to be copied just to preserve ordering that we don't necessarily care about.

Now we can get to filling in our StatisticsResponse. We'll put an instance of StatisticsResponse at the end of our function, and we'll start filling in the fields. Computing the range is relatively straightforward, very similar to what we did in Python. We'll subtract the Max from the mean. Computing the quartiles is also pretty straightforward. We can use the lower quartile median and upper quartile functions on our data instance. Computing the mean is very straightforward, but it does have one little extra trick on it. Notice that the call here is data.mean.unwrap. What is this unwrap telling us? For that, we're going to need to jump to the definition of the distribution trait. We can see that the mean function does not actually return a value directly, it returns an Option value. This is something that's unique to Rust and some other ML type languages. If you're used to coming from a different language, you're probably used to dealing with null values. Null values are a special value that can generally be assigned to variables of any type. If you want to write code that correctly handles null values, you basically need to pepper checks all over your code. You need to repeat those checks, because any string instance or array instance or HashMap instance, might actually secretly be holding a non-value.

Rust does not have the concept of null, it doesn't have a secret variable that can be assigned to variables of any type. Instead, Rust has a special type called Option. Instead of being a special value that can be assigned to variables of any type, an Option is a wrapper that goes around a variable. If you have an f64, for example, if you have that 64-bit floating point number, that is always guaranteed to be initialized to something, if you have a Vec, that is always guaranteed to be initialized to something. If you have an Option Vec, or an Option f64, then you have to write the code that deals with the possibility that that thing is not initialized. That could look something like this. We can use a match statement. We need to deal with the case that there's nothing there, if we want to deal with the thing that is inside of the Option.

Comparing the two, Option versus null. Option is strongly typed. You can't get away with forgetting to check something. You can also centralize your checks, which is really nice and really powerful. Because like I said, when you're dealing with null values, you don't necessarily know that the input value to a function, or the return value from a function isn't null. Because according to the type system, it's theoretically possible for any function in Java, or Ruby, or Python to return null. We end up repeating null checks all over the place. With Option, because it's strongly typed, you can convert an Option back into a Vec. Then, as long as you write the rest of your code to deal with Vec, you know that it's initialized and you never have to do that check again. It's very convenient, and it leads to great peace of mind knowing that things are initialized.

Let's jump back to our code and see what that one line that we made all that fuss, was about. On this line, we have data.mean, which returns an Option f64. Then we call unwrap on it. Unwrap is a function for dealing with Options. It will look at the Option, and if there's a value present, it returns the value. If there's no value present, it will actually panic the whole thread and make the thread unwind up to a point where there's a panic handler. Generally speaking, in production code, you don't want to be using unwrap, you want to be using proper handling of our Options with a match statement like we had previously. This is quick and dirty, so we're going to use an unwrap. If you'd like some more information on using Options, you can read chapter six of "The Rust Programming Language."

Similarly, when we calculate the standard deviation, this function also returns an Option, so we're also going to need to use unwrap on it. Now we have our StatisticsResponse. It's got all the fields in it, but it's actually not very useful to us yet, because it's a Rust function that returns a Rust data type, and we need a Python function. We need something that we can run from Python and call from our Flask HTTP handler. We don't have it yet. Let's do that. We're going to need to import some more types from, this time, pyo3. Pyo3 is the Rust crate that allows us to write bindings between Python and Rust. We're going to bring in pyo3::prelude::*. A prelude is a convention, but not necessarily a requirement for Rust crates. If there's a lot of types and traits and macros and things that need to be brought in, in order for your crate to be really useful, it can be common for crate authors to include a module called prelude that includes all the most commonly needed things. You can use a glob import like this, as we're doing here.

Before we can make a Python function, we actually need to make a module first, a Python module. A thing that can be imported in Python. In order to do that, we need to write a function that has the same name as our crate, which is rstats. Write a function called rstats, and we're going to add this little annotation above it, PyModule. This is coming from the prelude of pyo3, and it is going to automatically expand at compile time into a bunch of C stuff that the Python interpreter knows how to read and knows how to turn into a module. This is going to require us to add a couple of parameters to this function that are not both going to be used, but they're both required, based on the definition of the PyModule macro. The first one is just called Python. This is the type that comes from pyo3. It represents taking the GIL, the Global Interpreter Lock of the Python interpreter. A lot of times, if you're constructing a Python type, you need access to the Python type. This is to prove to pyo3 that you are holding on to the GIL, because it's easy to misuse the GIL when you're writing Python C code. Since we're not actually using it in this function, we're going to prefix it with an underscore so that the Rust compiler doesn't complain and say, you have an unused parameter on this function. Next up, we are going to add a parameter called m, and this is going to be a reference to a PyModule type. As implied, this is a reference to an empty Python module. Inside the body of this function, we are going to add our new compute_stats function to the Python module. We also need to set a return type for our PyModule function, for our rstats function, and that return type is going to be PyResult(). There's a couple of interesting things going on in here. We're going to jump through them real quick.

The Result type is the way that we handle errors in Rust. Rust does not have an exception system that bubbles values up and lets you catch exceptions with handlers. Instead, much like with Options, we have a result type that has two branches. It has an Ok branch, which also contains a success value inside of it. There's an error branch that also contains an error value inside of it. These are strongly typed. If you want to assume that your function returned a successful result and get the successful result out, you have to deal with the possibility that your function returned an error. That code generally looks like this. Just like with Option, we would use a match statement. I'm going to say if it's ok then pull the value out and do something with it, if there was an error, pull the error out and do something with it. That is Result. If you'd like more information on using the result type for error handling, you can read chapter nine of "The Rust Programming Language." We also had something inside of the result type, we had that open parenthesis, close parenthesis. This is something called the unit type, which is an empty tuple. It's an interesting thing that's somewhat unique to Rust. This is somewhat similar but not exactly similar to a null value. A null value can generally be assigned to values of any type, but the unit type is actually a type in and of itself. You cannot actually assign the unit value to anything other than a variable of the unit type. It represents nothing.

If we jump back to our function, it has a PyResult return type which is actually just a wrapper type, an alias type that comes from the pyo3 crate, and it has its error side always set to being a pyo3 Python error. You just have to fill in the success side. We have our success side set to the unit type, because a result is going to communicate either a success or an error. We really only have side effects in this function. We have the side effect of defining a function in here, putting a function onto our module. There's not like a value that we can return. We're not fetching something from a database that might fail. There's not a great sentinel value that we could return, so we're going to use the unit type instead. The body of this function, we're just going to put that Ok, that success case with the unit value inside of it. This isn't going to define our compute_stats function in a way that Python knows how to deal with but it is going to define a Python module called rstats.

Let's try and use it. We'll jump back over to the Python code, we will add import rstats to the top, and we'll try to run our Python code. We're going to get a giant error because there's no module named rstats that Python knows how to import. If you nested your rstats folder directly under the folder where the Python code is, this is actually going to work but it's not actually going to be importing the module that we care about. It's just going to be importing the directory in a way that Python can default to sometimes. The Python module system is a little confusing. We want to write something that is actually going to be importing our Rust code, not just the directory. Still on the CLI, we're going to install a developer tool that's created by the pyo3 team called maturin. We're going to jump into our rstats folder, the folder with our Rust code, and we're going to run maturin develop. This is going to compile our Rust code and generate some Python bindings for it. Now, if we run flask run, it's going to start up successfully. There's not going to be an error, because that Python module is going to be present. It is going to know how to import rstats.

Let's jump back over to our Rust code and see what we can do. Let's bring in our compute_stats function. On top of the compute_stats function, we're going to add this pyfunction annotation. That is going to add some extra code at compile time. Once again, that is going to transform the input types and the output types into something that Python knows how to deal with. If we actually tried to compile this Rust code right now, it's actually going to give us a huge compiler error. It's going to say it doesn't actually know how to turn a StatisticsResponse into something that Python knows how to deal with. We got to fix that. We can do that by adding some more of these little decorators onto our Rust code. We're going to add the pyclass attribute macro on top of our StatisticsResponse struct. We're going to add the pyo3(get) attribute macro on top of all the fields of our StatisticsResponse. This is necessary so that we can access all these individual fields. Otherwise, they would just be hidden from the Python side.

Next, we're going to jump down into our module definition function, and we are going to put in this somewhat complicated line of code. I know it's a lot to look at, but it is well documented and all those steps are necessary. We're going to call the add_function function on our module. We're going to parse that the results of the wrap_pyfunction macro on our compute_stats function, and that also needs access to the module. Then these question marks that are here at the end are error handling. Those are actually going to be doing an early return if those expressions fail, if they evaluate to error responses. We're almost there. We're so close. We have reimplemented the functionality. We have generated the Python bindings. We have exposed those bindings to Python. We have generated a Python class that we can use in order to get access to our fields. Let's recompile our Rust code. We'll run cargo build from the command line. We don't need to run maturin develop again, because of the symlinks that were created. We can just recompile normally, and this is going to regenerate everything that's required.

Now, we can do the Python refactoring. Over in the Python code, we can change up our handler a little bit. We can call rstats.compute_stats, that's the function that we wrote and we exposed. We'll parse it our numbers, just the normal numbers that come straight out of the requests JSON body. Then we are going to parse all of the fields from our response, from the StatisticsResponse into Flask's jsonify function. We actually do have to work all of the fields here individually, unfortunately. Pyo3 does not automatically generate JSON deserializable Py classes, we could do it with a little bit of extra work. We're just going to work all the fields manually. We can now boot up our flask application, and we can try running it. We can try running our HTTP handler. Let's use curl. We'll hit that stats endpoint, and we do get some numbers back. It's all working, everything's flowing great.

Let's compare the results from our Python handler, the original Python handler that was 100% Python, as well as our refactored handler. There are actually some differences in here. These values are not the same. The quartiles fields are different between Python and Rust. In my research, I learned that there are some differences in statistical libraries and how they compute quartiles of large data series. What do we do? I'm not an engineer anymore, but let me think back on my time as a staff engineer, and give you a great answer to that question. It depends. That's right. You actually have to use your brain. You have to think about the needs of your system. You have to figure out exactly what you need to do. There's a number of strategies that we can take to fix this problem depending on needs.


What can we do? There's two broad things we can do. We can maintain the existing behavior exactly, or we can figure out if there's a way we can deal with it. Maybe this change is acceptable for your system, for some reason. I don't know why it might be, but maybe it is. Maybe you can update your client so that they can deal with this change. Maybe it is possible for you to deal with it. If you want to maintain behavior, you need your code to return exactly the same stuff. What can we do? There's a couple strategies we can explore. We could try using a different library. Maybe there's something other than statrs that has the same return values for quartiles as the Python code. What if that's not an option? Maybe we could reimplement Python statistics library in Rust. Maybe it's not quite as fast as statrs, but even just rewriting the exact same code can often be much faster, because Python is going to have a lot more copying, a lot more GC overhead than something like Rust will. There's another option too. Because we're taking an incremental approach here, we could actually leave the quartile calculation within Python completely. We don't need to do everything in Rust. It's just something that we can do. Based on what your needs are, based on your specific situation, you need to explore one of these options. I left this error in here on purpose, so that we could discuss this. It's very important.


Now that we have our functionality written, let's talk about how we could test it. I know everybody loves writing tests. Everybody loves having super long test suites, but tests are super important, especially when we're going between multiple languages. We're going to write some subunit tests in our Rust code so that we can do automatic validation. We're going to create a new module at the bottom of our Rust code called tests. The name of the module isn't strictly important, it's just convention. We're going to add an attribute macro on top of the module. This is going to do conditional compilation for us. The test code is not going to be included in any production builds. It's only going to be compiled when we write test code. We are going to import the compute_stats function from the root of the crate into our tests module. We're going to write a new function called test_9_numbers. Adding the test attribute macro on top of our function is going to give us the ability to have our function picked up by Rust's automated test harness, and it will run the function and give us an alert if the function panics, which is going to happen if any of our assertions fail.

Let's pop in some known numbers. Let's calculate the statistics for that set of known numbers. We're going to add in some assertions. These are easy to calculate because it's only 9 numbers. We can run our tests by using cargo test. This is going to compile our code for us. It's going to tell us that we had one test function and it ran successfully. One unit test is obviously not enough to deal with a whole big refactor like this. We're going between multiple languages and we really need to be careful with our testing. You should be leveraging existing tests. You should be leveraging the tests that already exist in Python. You should be updating those so that they're capable of testing, not just the Python code, but the Rust code as well. Because compute_stats is just a normal Python function, you can call it from either place. You can rely on dependency injection as well, so that you can test more code paths with both the Python code and the Rust code. Something else you can do is actually do randomized testing between the old code and the new code. Generate a random input, feed it into the old code, see what it gets you back. then feed it into the new code and compare those two results. They should match up.


Let's also talk about performance. We did this whole thing with the goal of making our code faster? Did we do it? Let's see. We can use Python's timeit module to do some microbenchmarking. For the purposes of this microbenchmark, stats_py is a function that has the original code of our Python HTTP handler in it. We're going to feed in those 9 numbers, and we're going to run this 10,000 times. We're going to do a very similar thing with our Rust code. We're going to take those 9 numbers, and we are going to compute those stats 10,000 times. Let's see what we got. What happened. We can see that the Rust code ran a little more than 100 times faster than the Python code. This seems really promising. This is very cool. This is a benchmark that's running through Python, so we're not just getting the faster code because it's all in Rust, but there is a certain amount of overhead that comes from it being in Python too. This is somewhat fair. There's a little bit of trickery going on in here, because we ran the test 10,000 times. This is actually the total time. Once we add in the average time, it starts to get a little bit less impressive. Because you divide those numbers by 10,000, and you realize, the Python code was pretty quick on its own already. This is not an extremely slow problem. I was feeling a little bit uncreative when I came up with this problem. If you started with something that was taking 500 milliseconds, a second, 5 seconds in Python code you might expect to see significantly more impressive results from refactoring to Rust. We did make the thing 100 times faster. That did actually happen. However, it was already operating at a relatively quick speed in Python, to begin with.

We also have to consider macrobenchmarking. It's really tempting to just want to do a microbenchmark, like just put the teeniest bit of code on the bench and test that. That's going to give you the best-looking results, at least. If we do a macrobenchmark that compares the HTTP performance of the old code and the new code, we can see that it's about a 15% performance difference between Python and Rust. That's because a lot of the time that is spent in our HTTP endpoint is going into the Flask library itself, and its HTTP handling, and the JSON serializing, deserializing. Once again, we should be picking something that is pretty CPU bound, where we're spending a lot of our time in one place, and we should be pulling that into Rust. Microbenchmarks and macrobenchmarks are both super important. They both have their uses. It is important to do macrobenchmarks, and to make sure you know how your system is actually performing under load.


We looked at FFI refactoring and what it is. We looked at how to think about the feasibility of an FFI refactoring project. We learned how we can use the pyo3 library to do an FFI refactoring project. We learned a little bit about how we can test systems using Rust's testing framework. We looked at how we could do some benchmarking strategies for an FFI refactoring project.


See more presentations with transcripts


Recorded at:

Mar 24, 2023