This repository implements the same simple backend API in a variety of languages. It's just a personal project of mine to get a feel for the languages, and shouldn't be taken too seriously. So far I've built it in C#, Typescript (Deno), Elixir, Go, Rust, and Scala. Star the repository and/or follow me on Twitter (@losvedir) if you want updates on the project. I hope to eventually get to Swift, Kotlin, ordinary Java, Nim, and Zig. And feel free to open an issue if you want to suggest another language, or a PR if you want to implement one!
All the apps read in the MBTA's GTFS data, which is the standard spec for
transit data - stuff like the routes, stops, and schedules for a system. The
apps look for files in an MBTA_GTFS folder, but could be easily updated to
work with any transit system's data. To get the MBTA data, the following
commands can be run in the repo's root directory:
> curl -o MBTA_GTFS.zip https://cdn.mbta.com/MBTA_GTFS.zip
> unzip -d MBTA_GTFS MBTA_GTFS.zip
The apps are all named some mashup of "Transit" and the programming language name.
For now, the apps only read in the GTFS trips and stop_times data. They parse
the files, which are .txt but CSV, into an in-memory list of structs. I was
interested to see how long this takes, as it's a bunch of IO - there are roughly
75k trips and 2 million stop_times in the MBTA data. In a future iteration, I'd
also like to handle "services", which specify which trips run on which days.
The apps set up a simple webserver that responds to /schedules/:route
requests, and returns a JSON response of all the "schedules" (trips with
included stop_times) for that route. This involves a "join" through trips, and
for some routes serializes a bunch of data. (The most, I think, is for the Red
line at about 7MB of a response.) I represented it this way because even though
so far I've only implemented some of the functionality, I think it makes sense
conceptually to want to look up trips by route_id and/or service_id, and to look
up stop_times by trip_id and/or stop_id. So rather than just storing the data as
a hashmap, I figured it was better to store the data as a big list, and have
various handles into it.
Currently, I'm collecting two things. The first is how long it takes the apps to
load the GTFS stop_times.txt file into an in-memory structure (together with a
hashmap "index" to access it more efficiently). The second is the requests per
second that the webservers can field, as measured using the k6 tool.
This is the time it takes for the app to load the stop_times.txt file, which is
roughly 100MB and 2M records and parse it into a big vector/list/array of a
structured StopTime structs, together with an "index" on trips, which is a map
from the trip ID to a list of indices into the big stop time list.
| Language | Time (ms) |
|---|---|
| C# | 1,390 |
| Deno | 2,905 |
| Elixir | 5,986 |
| Go | 842 |
| Rust | 565 |
| Scala | 931 |
| SQLite | ~ 4,000 |
This is tested using the k6 tool, which I
installed via homebrew. There is a loadTest.js script in the root of the repo,
and I ran the test as follows:
k6 run -u 50 --duration 30s loadTest.js
That sets up 50 "virtual users" concurrently accessing the server, and the test itself has them sequentially issuing requests, of the schedules for roughly a hundred routes, some pretty hefty, some pretty small, in random order.
I'm not trying for a perfectly uniform test environment, but I close most of my usual apps and just run it on my laptop by itself. You shouldn't really test on the same machine, but the requests are decently beefy overall that I figured the load from the test harness wouldn't disrupt the response data too badly. Here I report the requests/sec that k6 spits out, and also an eyeball at the highest RAM and CPU usage I see in ActivityMonitor just out of curiosity.
These use the loadTest.js file which includes about a hundred of the MBTA
routes, many of which have JSON schedule data in the megabytes. Consequently,
the performance here is largely a reflection of how fast JSON can be serialized.
All these were with 50 concurrent virtual users.
| Language | Requests/sec | Max CPU (%) | Max RAM (MB) |
|---|---|---|---|
| C# | 1,534 | 654 | 1,750 |
| Deno | 286 | 280 | 400 |
| Elixir | 388 | 750 | 3,700 |
| Go | 2,715 | 620 | 1,100 |
| Rust | 2,839 | 619 | 603 |
| Scala | 432 | 715 | 3,150 |
These use the loadTestSmallResponses.js runner, and only use about a dozen
routes whose schedule data is in the ~50KB to ~200KB range, so the requests are
a lot higher, and less dominated by JSON encoding. Since the responses are
smaller and more requests can be handled, I also tried it with different number
of concurrent "virtual users".
Requests per second, by language and concurrent virtual user count (higher is better).
| Language | 1 VU | 10 VU | 50 VU | 100 VU |
|---|---|---|---|---|
| C# | 2,227 | 11,663 | 13,005 | 13,102 |
| Deno | 2,808 | 3,882 | 3,852 | 3,753 |
| Elixir | 623 | 3,306 | 3,809 | 3,852 |
| Go | 2,283 | 10,551 | 11,078 | 11,091 |
| Rust | 3,501 | 20,496 | 22,437 | 22,283 |
| Scala | 705 | 4,204 | 4,289 | 4,332 |
Response times in milliseconds: median / p95 / max, by language and concurrent virtual user count (lower is better):
| Language | 1 VU | 10 VU | 50 VU | 100 VU |
|---|---|---|---|---|
| C# | .3 / 1 / 88 | .6 / 2 / 27 | 2 / 12 / 138 | 6 / 17 / 75 |
| Deno | .3 / .8 / 5 | 2 / 4 / 254 | 13 / 16 / 218 | 26 / 33 / 265 |
| Elixir | 1 / 4 / 11 | 2 / 7 / 22 | 13 / 23 / 58 | 24 / 47 / 132 |
| Go | .3 / 1 / 19 | .6 / 2 / 36 | 3 / 15 / 111 | 6 / 29 / 140 |
| Rust | .2 / .6 / 3 | .4 / 1 / 9 | 2 / 4 / 36 | 4 / 9 / 85 |
| Scala | 1 / 3 / 109 | 2 / 5 / 129 | 3 / 58 / 394 | 10 / 109 / 587 |
This metric I collected from a previous commit, and involved simply counting the number of StopTimes for the Red line. I removed this code in favor of the webserver approach, but am keeping the stats here for posterity.
| Language | Time (ms) |
|---|---|
| C# | 1.0 |
| Deno | 1.4 |
| Elixir | 3.2 |
| Go | 0.4 |
| Rust | 0.7 |
| Scala | 2.5 |
| SQLite | 13 |
Here are some scattered thoughts while I went about writing this.
Where to begin! First, I went into this very confused at just a jargon level of
what all the different pieces of the Microsoft ecosystem are. C# is the language
and it runs on the ".NET CLR". The build/run tool is dotnet so that's kind of
the main term, but I also saw "CLR" thrown around. I ended up working with ".NET
6.0", which is what all the guides and docs called it, and which was cross
platform. I didn't see ".NET Core" anywhere like I was expecting, which I
believe is what used to be the explicitly cross-platform piece? Amusingly, I
spent a fair bit of time trying to look up the standard ".NET web framework"
before eventually finally realizing that that's what ASP.NET is. So that was
useful to connect for me, since I've seen "ASP" a lot but had had no idea how it
fit into the picture.
I wasn't entirely sure I'd even be able to complete this project. I wasn't sure how truly cross platform .NET was, in reality, though development went off without a hitch! I'm going to say, yes, at least for my simple use case of using the standard library plus ASP.NET, it's truly cross platform. I didn't try bringing in any 3rd party libraries, and I imagine there could be some incompatibilities there. In the future I'd like to explore F#, which is a language more inline with my sensibilities, but I wanted to try more "vanilla .NET" first. The developer experience in VSCode was great, the language server worked well, and the code formatter worked (though I despise the convention of opening curly braces on the next line).
As for the language, C# is... all right, I guess. It kind of reminds me of Dart;
it works fine, the tooling is good, it's verbose and very object oriented, but
it doesn't really spark joy. The "billion dollar mistake" is important to me,
and while C# has non-nullability sugar in its typesystem (i.e. with ? after a
number of types), the type system wasn't as rigorous as I was maybe hoping. At
one point I had a bug because I did a stopWatch.Elapsed / 1000 by accident
instead of stopWatch.ElapsedTicks / 1000. The former is a TimeSpan struct
instead of a long like ElapsedTicks, so intuitively it feels like I
shouldn't be able to divide it, though it did a best effort and did something
to it, though I'm not quite sure what.
ASP.NET has a lot of conventions and magic. I don't personally love all that magic but if you're experienced with it, I could see how it would make designing web apps pretty quick.
But, wow, I was incredibly surprised and impressed with the performance! It was comparable to my unoptimized Rust (i.e.: treating Rust like a high level language with lots of clones)!
All in all, I was pleasantly surprised and pretty impressed with dotnet and C#.
Deno is pretty neat. I really want it to succeed. I really like TypeScript, and Deno almost gives me what I want: pretending TypeScript is a full-fledged language, with a standard library, that I can build non-frontend apps with. Let's just sweep all that JS-heritage and V8 stuff under the rug...
I wish the standard library weren't at URLs like all the other packages. It
would be great if the deno tool you downloaded also included the standard
library, and you could just reference it without any network stuff. The
documentation is also pretty cryptic (and I think autogenerated?).
The package management stuff I haven't quite wrapped my head around. Obviously, you shouldn't be downloading stuff willy-nilly, but I think with some combination of the conventional deps.ts, import_map, specifying a lock file, the vendor command, --no-remote, etc, I feel like I have all the pieces to kind of build up a reasonable approach, but I don't quite understand it all just yet.
Personally, the --allow-read, --allow-net, etc stuff feels a little gimmicky
to me. I don't think other languages really have that, and I'm not sure what the
threat model is here. I control the backend code, and if I'm worried about my
code doing unexpected things like that I have larger issues. I just run with
-A all the time.
The performance was great when looking at a single virtual user, but sort of topped out there. I don't know if it just can't handle async and multiple cores very well, or if I was doing something wrong.
Elixir is my primary language, so I threw this one in to compare its performance to see what I could be missing. I like Elixir the language and all its nice OTP goodies, but it's known to be a little slow, so I was wondering how much performance I'm leaving on the table.
Normally, my first thought for some state in Elixir would be an Agent or custom GenServer, but that would funnel all requests to the one data source, which would respond sequentially, and I thought under load that could be a bottleneck. So I opted to put the data in ETS, with read concurrency enabled.
ETS stores data as a set (in this case) of Erlang tuples, and wanting to follow the conventions of the other apps, I decided to add an extra "primary key" integer to each tuple, for the purpose of the "indexes". The other languages allow you to simply index into the underlying list, but that's not really possible with the way ETS stores data.
This approach works fine for GTFS static data which is loaded on app start-up, but I'm not entirely sure yet how I will handle when I need to update data, if I extend the apps to poll the real-time vehicle positions and predictions data. In that scenario, I've had issues before with how to handle locking and atomic updates to ETS data. Most likely it would be something like create a whole new ETS table in the background and then swap it out for this one after it's ready.
I used Phoenix here since that's pretty much the standard in Elixir, though it might be a bit heavier weight than the other apps. But my understanding is that it's mostly just plugs that get compiled in, so it's pretty lightweight in how much it actually affects performance vs the minimal possible thing I could do.
The final performance results were unfortunately low, an order of magnitude worse than Rust, and a little behind Deno.
I was super happy to get the work done so far using just the standard library. And the performance was solid! In the JSON-heavy benchmark it actually is the fastest of all the languages, though in the lighter-response benchmark it's more where I expected: fast, but not quite at rust levels.
That said, contrary to my expectations, I found the documentation not great. While the language reference and tour was pretty good and useful (I kept referring to the tour), the library documentation on pkg.go.dev was fairly... bad.
It took me longer than I'd like to admit to figure out how to get a dang
io.Reader, which is what the CSV parsing package takes. I had hoped
searching their docs for io.Reader
would yield a package or function that at a glance would (1) read from the
filesystem and (2) implement the io.Reader interface, but the top result was
simply the definition of the interface, and the rest of the results were random
GitHub repos. And clicking through to the io.Reader definition didn't provide
links to anything that implements it. Eventually I gave up and went the other
direction, trying to figure out how to open and read files. I finally found
os.Open() (though it was my third try after poking around in io and
io.fs). I saw it returns a File, which then sent me on a bit of a goose
change on how to turn it into an io.Reader before realizing that although
it's not mentioned in the docs, the type does implement Read and so it is
already an io.Reader! It was all sort of magical to me, and kind of odd. Now I
realize that in theory I could have searched pkg.go.dev for "Read" to find types
that implement it, and hence satisfy io.Reader and would get me to File and
os.Open() but of course that doesn't work because the search function seems to
be hot garbage.
All that said, actually programming in Go was pretty nice. VSCode support was solid and the build/run cycle was fast! The final result ended up being pretty quick, too. It doesn't have the type richness I appreciate, but I didn't mind it overall.
I started out looking for a "web framework" since that was my expectation of how this works, but it seemed like there was a reasonable consensus that using simply the standard library was a good place to start. That was nice, and helped me avoid the analysis paralysis and reviewing benchmarks and HN and reddit, etc, to decide which framework to use.
This one shocked me in a good way! I was expecting a lot more low level fiddlyness, and was prepared to simply allocate and clone and do all the tricks I've read about to not worry about eking out the most performance possible. After all, I'm comparing against higher level interpreted or GC languages, and am interested in Rust more for its type system than needing to program at a system level.
I've had some experience playing with Rust in the past, so it wasn't brand new
to me, but it has been some time so I was expecting to be a lot more, uh...
rusty. All that said, with my initial approach, I just did a lot of String
cloning and got performance comparable to the best of the other languages
(dotnet or Go, depending on the benchmark). But then after a bit of help from
reddit, I removed some unnecessary String allocations, using &str and dreaded
(to me) lifetime markers, so that the response structs just referenced the
strings allocated in the actual data, and the performance jumped dramatically,
to be far and away the most performant language.
Also, I don't know how much of this is because Rust is special or because BurntSushi is a national treasure and his CSV library is impeccably constructed and documented.
I also was impressed and amused that I got compiler warnings that my Struct had unnecessary fields (I haven't used the Trips' service_ids or the StopTimes' arrival and departure times yet), which wasn't raised for any of the other languages.
For the web server piece, I spent some time trying to decide which framework to
use. When I last looked at Rust, rocket was all the rage, but it seems to have
fallen off the radar almost completely these days! That was mildly concerning.
It seems like actix has taken over as nearly the "default", except there's a
new-ish one called axum that's quite popular. Being a part of the official
tokio project, and guessing that tokio has staying power, I went with axum.
It was a little tricky to get working... I felt like I was playing type tetris a
bit to get my app to compile, and was trying to mindlessly copy documentation
without fully understanding it. I've never quite understood the #[...] syntax,
and so annotating my main function with #[tokio::main] is still black magic
to me. I also got tripped up for a while before realizing that I needed to put
futures as a dependency in my Cargo.toml. That wasn't in the axum docs but I
found it in their examples, though it was quite a wild guess that that was the
thing that allowed the example to compile when mine wasn't. I drew on some
latent knowledge I had buried deep down in there that futures was what the
async ecosystem was built on, and it was a crate rather than part of the
language, but I had thought it was just a temporary thing for experimentation by
the rust folks back in the day.
I also ran into some issues trying to get my shared state to work. My handler
was failing to typecheck and the compiler error was not helpful. The axum docs
actually mention this is a problem and that there's an axum-macros crate that
can help, though. Some of this was my lack of understanding exactly how Arc
works and how to safely have shared state across async requests. In the end, I
appreciate that the flexibility is there; right now I just have an Arc so that
all my handlers can read the data I prepare up front, but I could see how I
could wrap it in an RwLock, for example, to also allow safe updates in the
future. In general, I'm not sure how I feel about Axum's magical
handler/extractor setup, as I still don't really know how it works.
Of all the languages I played with here, Scala is the only one I disliked.
Part of the reason was timing: it seems an ecosystem in flux at the moment. It
didn't work with my JDK 19 out of the box, so I had to downgrade to JDK 17 for
it. I sort of blindly followed the scala-lang.org site and went through the
getting started guide, for Scala 3, and then building the first half of my app
in Scala 3, before realizing that Play (the only Scala webframework I'd heard
of) and Scalatra (the other web framework mentioned in the Getting Started
"ecosystem" section of the guide) don't work on Scala 3 yet. I briefly tried
updating my code to Scala 2 but I wasn't super sure what the differences were,
and I didn't really want to learn Scala 2 if everything is moving to Scala 3
(eventually) anyway. Beyond that I got mixed messages in whether to use sbt
and mill as a build tool.
Beyond that, there seems to be a schism in the community between people who love super sophisticated types (think Haskell style Applicative Functors or whatever) and people who want a nicer Java (these days those people might be moving to Kotlin).
In trying to find a Scala 3 compatible web framework, I saw a lot of people
saying http4s is the new standard, so I tried that one first. But after
generating the skeleton for the app, and trying to add my own routes I gave up.
The "router" part is unwieldy and complicated, though I think I was able to
cargo-cult a route of my own. Here's the given example on how to match against
/hello/:name:
def helloWorldRoutes[F[_]: Sync](H: HelloWorld[F]): HttpRoutes[F] =
val dsl = new Http4sDsl[F]{}
import dsl._
HttpRoutes.of[F] {
case GET -> Root / "hello" / name =>
for {
greeting <- H.hello(HelloWorld.Name(name))
resp <- Ok(greeting)
} yield resp
}But then when it comes to implementing the code that actually returns data there, I got totally flummoxed. Here's the corresponding "HelloWorld" code for the above route:
import cats.Applicative
import cats.implicits._
import io.circe.{Encoder, Json}
import org.http4s.EntityEncoder
import org.http4s.circe._
trait HelloWorld[F[_]]:
def hello(n: HelloWorld.Name): F[HelloWorld.Greeting]
object HelloWorld:
def apply[F[_]](using ev: HelloWorld[F]): HelloWorld[F] = ev
final case class Name(name: String) extends AnyVal
/**
* More generally you will want to decouple your edge representations from
* your internal data structures, however this shows how you can
* create encoders for your data.
**/
final case class Greeting(greeting: String) extends AnyVal
object Greeting:
given Encoder[Greeting] = new Encoder[Greeting]:
final def apply(a: Greeting): Json = Json.obj(
("message", Json.fromString(a.greeting)),
)
given [F[_]]: EntityEncoder[F, Greeting] =
jsonEncoderOf[F, Greeting]
def impl[F[_]: Applicative]: HelloWorld[F] = new HelloWorld[F]:
def hello(n: HelloWorld.Name): F[HelloWorld.Greeting] =
Greeting("Hello, " + n.name).pure[F]That's a lot of both syntax and semantics to grok. trait is an interface,
object is a singleton class, case class is kind of a data record. I don't
know what given or using are. I recognize pure as related to Applicative
but that's a whole complicated library/type concern distinct from
Scala-the-language. I don't really know why the def impl does a
new HelloWorld with a nested def hello.
In the end, I moved on to trying the framework Cask, which
aims to bring simplicity, flexibility and ease-of-use to Scala webservers, avoiding cryptic DSLs or complicated asynchrony
And in the end, I got something that worked! So thanks author of Cask. Cask also
used mill rather than sbt. The latter seems more "official", or at least is
the tool recommended on scala-lang.org, but oh BOY is it slow! mill was nicer
for me to work with.
The performance was not super great, and it used the most memory by far. I don't know if this is because Cask is not performance-focused, but then I couldn't get anything else to work... I liked Scala 3 well enough before dealing with the ecosystem, but I think in the future I'm going to avoid Scala until it finishes its 2 to 3 transition, and only if the non-typenerds win.
Not really an apples-to-apples comparison but I was curious about the order of magnitude performance characteristics of SQLite here.
For importing stop_times I counted (yes, so take that time with a grain of
salt) while running:
sqlite> .mode csv
sqlite> .import MBTA_GTFS/stop_times.txt stop_times
And for scanning the data for the number of Red line schedules, I did:
sqlite> create index stop_times_by_trip on stop_times(trip_id);
sqlite> create index trips_by_route on trips(route_id);
sqlite> .timer on
sqlite> select count(*) from stop_times where trip_id in (select trip_id from trips where route_id = "Red");
Can't beat the convenience! It's an order of magnitude slower than the apps which keep everything in memory, but of course the tradeoff then is it uses much less memory! And while a given read is slow(-ish), I understand that a lot of it is waiting on the filesystem, and that concurrent reads should allow plenty of throughput.
Notes from in-progress work on Swift.
- Had to download many-GB Xcode, which had a host of issues (had to login)
- did
swift initand thenswift runand it crashed. Found a discussion online wheresudo xcode-select --resetwas recommended and that got the HelloWorld to run. - No code formatter?
- Couldn't figure out how to read the relative MBTA_GTFS folder from my project
in xcode, but running
swift runfrom the directory worked. (Though I had toswift package initwhich xcode didn't need.)