Rust and CSV parsing


With a beta of csv 1.0 just released, the time is ripe for a tutorial on how to read and write CSV data in Rust. This tutorial is targeted toward beginning Rust programmers, and is therefore full of examples and spends some time on basic concepts. Experienced Rust programmers may find parts of this useful, but would probably be happier with a quick skim.

For an introduction to Rust, please see the official book. If you haven’t written any Rust code yet but have written code in another language, then this tutorial might be accessible to you without needing to read the book first.

The CSV library is available on Github and has comprehensive API documentation.

Finally, a version of this blog post is included as a tutorial in the API documentation, and is more likely to be updated as time passes.

Target audience: Beginning Rust programmers.

Table of contents

CSV 1.0 beta release

Before diving into the tutorial, I wanted to briefly recount the road to 1.0. The first commit to the rust-csv repository was on March 22, 2014, which is a little more than an entire year before the Rust 1.0 release. For those of you that have been involved with Rust since before 1.0, you will surely remember just how much the language changed. And of course, I changed as well, as I became more familiar with the language and its idioms. However, the CSV library API mostly remained unchanged since its original version. The API made it difficult to increase performance and fix several outstanding bugs, and worst of all, was using the old serialization infrastructure.

The CSV 1.0 release marks a library that is faster, has a better API and brings support for Serde, Rust’s serialization framework.

The new CSV library comes with a csv-core crate, which can parse CSV data without Rust’s standard library and is predominantly responsible for the performance improvements. In particular, the old CSV library used a type of finite state machine that had a lot of overhead. The csv-core crate compiles its parser down to a table based DFA that occupies only a few hundred bytes on the stack. As a result, we get approximately a factor of 2 improvement across the board:

count_game_deserialize_owned_bytes  30,404,805 (85 MB/s)   23,878,089 (108 MB/s)    -6,526,716  -21.47%   x 1.27
count_game_deserialize_owned_str    30,431,169 (85 MB/s)   22,861,276 (113 MB/s)    -7,569,893  -24.88%   x 1.33
count_game_iter_bytes               21,751,711 (119 MB/s)  11,873,257 (218 MB/s)    -9,878,454  -45.41%   x 1.83
count_game_iter_str                 25,609,184 (101 MB/s)  13,769,390 (188 MB/s)   -11,839,794  -46.23%   x 1.86
count_game_read_bytes               12,110,082 (214 MB/s)  6,686,121 (388 MB/s)     -5,423,961  -44.79%   x 1.81
count_game_read_str                 15,497,249 (167 MB/s)  8,269,207 (314 MB/s)     -7,228,042  -46.64%   x 1.87
count_mbta_deserialize_owned_bytes  5,779,138 (125 MB/s)   3,775,874 (191 MB/s)     -2,003,264  -34.66%   x 1.53
count_mbta_deserialize_owned_str    5,777,055 (125 MB/s)   4,353,921 (166 MB/s)     -1,423,134  -24.63%   x 1.33
count_mbta_iter_bytes               3,991,047 (181 MB/s)   1,805,387 (400 MB/s)     -2,185,660  -54.76%   x 2.21
count_mbta_iter_str                 4,726,647 (153 MB/s)   2,354,842 (307 MB/s)     -2,371,805  -50.18%   x 2.01
count_mbta_read_bytes               2,690,641 (268 MB/s)   1,253,111 (577 MB/s)     -1,437,530  -53.43%   x 2.15
count_mbta_read_str                 3,399,631 (212 MB/s)   1,743,035 (415 MB/s)     -1,656,596  -48.73%   x 1.95
count_nfl_deserialize_owned_bytes   10,608,513 (128 MB/s)  5,828,747 (234 MB/s)     -4,779,766  -45.06%   x 1.82
count_nfl_deserialize_owned_str     10,612,366 (128 MB/s)  6,814,770 (200 MB/s)     -3,797,596  -35.78%   x 1.56
count_nfl_iter_bytes                6,798,767 (200 MB/s)   2,564,448 (532 MB/s)     -4,234,319  -62.28%   x 2.65
count_nfl_iter_str                  7,888,662 (172 MB/s)   3,579,865 (381 MB/s)     -4,308,797  -54.62%   x 2.20
count_nfl_read_bytes                4,588,369 (297 MB/s)   1,911,120 (714 MB/s)     -2,677,249  -58.35%   x 2.40
count_nfl_read_str                  5,755,926 (237 MB/s)   2,847,833 (479 MB/s)     -2,908,093  -50.52%   x 2.02
count_pop_deserialize_owned_bytes   11,052,436 (86 MB/s)   8,848,364 (108 MB/s)     -2,204,072  -19.94%   x 1.25
count_pop_deserialize_owned_str     11,054,638 (86 MB/s)   9,184,678 (104 MB/s)     -1,869,960  -16.92%   x 1.20
count_pop_iter_bytes                6,190,345 (154 MB/s)   3,110,704 (307 MB/s)     -3,079,641  -49.75%   x 1.99
count_pop_iter_str                  7,679,804 (124 MB/s)   4,274,842 (223 MB/s)     -3,404,962  -44.34%   x 1.80
count_pop_read_bytes                3,898,119 (245 MB/s)   2,218,535 (430 MB/s)     -1,679,584  -43.09%   x 1.76
count_pop_read_str                  5,195,237 (183 MB/s)   3,209,998 (297 MB/s)     -1,985,239  -38.21%   x 1.62

With all the goodies out of the way, let’s begin!

Setup

In this section, we’ll get you setup with a simple program that reads CSV data and prints a “debug” version of each record. This assumes that you have the Rust toolchain installed, which includes both Rust and Cargo.

We’ll start by creating a new Cargo project:

$ cargo new --bin csvtutor
$ cd csvtutor

Once inside csvtutor, open Cargo.toml in your favorite text editor and add csv = "1.0.0-beta.3" to your [dependencies] section. At this point, your Cargo.toml should look something like this:

[package]
name = "csvtutor"
version = "0.1.0"
authors = ["Your Name"]

[dependencies]
csv = "1.0.0-beta.3"

Next, let’s build your project. Since you added the csv crate as a dependency, Cargo will automatically download it and compile it for you. To build your project, use Cargo:

$ cargo build

This will produce a new binary, csvtutor, in your target/debug directory. It won’t do much at this point, but you can run it:

$ ./target/debug/csvtutor
Hello, world!

Let’s make our program do something useful. Our program will read CSV data on stdin and print debug output for each record on stdout. To write this program, open src/main.rs in your favorite text editor and replace its contents with this:

//tutorial-setup-01.rs
// This makes the csv crate accessible to your program.
extern crate csv;

// Import the standard library's I/O module so we can read from stdin.
use std::io;

// The `main` function is where your program starts executing.
fn main() {
    // Create a CSV parser that reads data from stdin.
    let mut rdr = csv::Reader::from_reader(io::stdin());
    // Loop over each record.
    for result in rdr.records() {
        // An error may occur, so abort the program in an unfriendly way.
        // We will make this more friendly later!
        let record = result.expect("a CSV record");
        // Print a debug version of the record.
        println!("{:?}", record);
    }
}

Don’t worry too much about what this code means; we’ll dissect it in the next section. For now, try rebuilding your project:

$ cargo build

Assuming that succeeds, let’s try running our program. But first, we will need some CSV data to play with! For that, we will use a random selection of 100 US cities, along with their population size and geographical coordinates. (We will use this same CSV data throughout the entire tutorial.) To get the data, download it from github:

$ curl -LO 'https://raw.githubusercontent.com/BurntSushi/rust-csv/master/examples/data/uspop.csv'

And now finally, run your program on uspop.csv:

$ ./target/debug/csvtutor < uspop.csv
StringRecord(["Davidsons Landing", "AK", "", "65.2419444", "-165.2716667"])
StringRecord(["Kenai", "AK", "7610", "60.5544444", "-151.2583333"])
StringRecord(["Oakman", "AL", "", "33.7133333", "-87.3886111"])
# ... and much more

Basic error handling

Since reading CSV data can result in errors, error handling is pervasive throughout the examples in this tutorial. Therefore, we’re going to spend a little bit of time going over basic error handling, and in particular, fix our previous example to show errors in a more friendly way. If you’re already comfortable with things like Result and try!/? in Rust, then you can safely skip this section.

Note that The Rust Programming Language Book contains an introduction to general error handling. For a deeper dive, see my blog post on error handling in Rust. The blog post is especially important if you plan on building Rust libraries.

With that out of the way, error handling in Rust comes in two different forms: unrecoverable errors and recoverable errors.

Unrecoverable errors generally correspond to things like bugs in your program, which might occur when an invariant or contract is broken. At that point, the state of your program is unpredictable, and there’s typically little recourse other than panicking. In Rust, a panic is similar to simply aborting your program, but it will unwind the stack and clean up resources before your program exits.

On the other hand, recoverable errors generally correspond to predictable errors. A non-existent file or invalid CSV data are examples of recoverable errors. In Rust, recoverable errors are handled via Result. A Result represents the state of a computation that has either succeeded or failed. It is defined like so:

enum Result<T, E> {
    Ok(T),
    Err(E),
}

That is, a Result either contains a value of type T when the computation succeeds, or it contains a value of type E when the computation fails.

The relationship between unrecoverable errors and recoverable errors is important. In particular, it is strongly discouraged to treat recoverable errors as if they were unrecoverable. For example, panicking when a file could not be found, or if some CSV data is invalid, is considered bad practice. Instead, predictable errors should be handled using Rust’s Result type.

With our new found knowledge, let’s re-examine our previous example and dissect its error handling.

//tutorial-error-01.rs
extern crate csv;

use std::io;

fn main() {
    let mut rdr = csv::Reader::from_reader(io::stdin());
    for result in rdr.records() {
        let record = result.expect("a CSV record");
        println!("{:?}", record);
    }
}

There are two places where an error can occur in this program. The first is if there was a problem reading a record from stdin. The second is if there is a problem writing to stdout. In general, we will ignore the latter problem in this tutorial, although robust command line applications should probably try to handle it (e.g., when a broken pipe occurs). The former however is worth looking into in more detail. For example, if a user of this program provides invalid CSV data, then the program will panic:

$ cat invalid
header1,header2
foo,bar
quux,baz,foobar
$ ./target/debug/csvtutor < invalid
StringRecord { position: Some(Position { byte: 16, line: 2, record: 1 }), fields: ["foo", "bar"] }
thread 'main' panicked at 'called `Result::unwrap()` on an `Err` value: UnequalLengths { pos: Some(Position { byte: 24, line: 3, record: 2 }), expected_len: 2, len: 3 }', /checkout/src/libcore/result.rs:859
note: Run with `RUST_BACKTRACE=1` for a backtrace.

What happened here? First and foremost, we should talk about why the CSV data is invalid. The CSV data consists of three records: a header and two data records. The header and first data record have two fields, but the second data record has three fields. By default, the csv crate will treat inconsistent record lengths as an error. (This behavior can be toggled using the ReaderBuilder::flexible config knob.) This explains why the first data record is printed in this example, since it has the same number of fields as the header record. That is, we don’t actually hit an error until we parse the second data record.

(Note that the CSV reader automatically interprets the first record as a header. This can be toggled with the ReaderBuilder::has_headers config knob.)

So what actually causes the panic to happen in our program? That would be the first line in our loop:

for result in rdr.records() {
    let record = result.expect("a CSV record"); // this panics
    println!("{:?}", record);
}

The key thing to understand here is that rdr.records() returns an iterator that yields Result values. That is, instead of yielding records, it yields a Result that contains either a record or an error. The expect method, which is defined on Result, unwraps the success value inside the Result. Since the Result might contain an error instead, expect will panic when it does contain an error.

It might help to look at the implementation of expect:

use std::fmt;

// This says, "for all types T and E, where E can be turned into a human
// readable debug message, define the `expect` method."
impl<T, E: fmt::Debug> Result<T, E> {
    fn expect(self, msg: &str) -> T {
        match self {
            Ok(t) => t,
            Err(e) => panic!("{}: {:?}", msg, e),
        }
    }
}

Since this causes a panic if the CSV data is invalid, and invalid CSV data is a perfectly predictable error, we’ve turned what should be a recoverable error into an unrecoverable error. We did this because it is expedient to use unrecoverable errors. Since this is bad practice, we will endeavor to avoid unrecoverable errors throughout the rest of the tutorial.

Switch to recoverable errors

We’ll convert our unrecoverable error to a recoverable error in 3 steps. First, let’s get rid of the panic and print an error message manually:

//tutorial-error-02.rs
extern crate csv;

use std::io;
use std::process;

fn main() {
    let mut rdr = csv::Reader::from_reader(io::stdin());
    for result in rdr.records() {
        // Examine our Result.
        // If there was no problem, print the record.
        // Otherwise, print the error message and quit the program.
        match result {
            Ok(record) => println!("{:?}", record),
            Err(err) => {
                println!("error reading CSV from <stdin>: {}", err);
                process::exit(1);
            }
        }
    }
}

If we run our program again, we’ll still see an error message, but it is no longer a panic message:

$ cat invalid
header1,header2
foo,bar
quux,baz,foobar
$ ./target/debug/csvtutor < invalid
StringRecord { position: Some(Position { byte: 16, line: 2, record: 1 }), fields: ["foo", "bar"] }
error reading CSV from <stdin>: CSV error: record 2 (line: 3, byte: 24): found record with 3 fields, but the previous record has 2 fields

The second step for moving to recoverable errors is to put our CSV record loop into a separate function. This function then has the option of returning an error, which our main function can then inspect and decide what to do with.

//tutorial-error-03.rs
extern crate csv;

use std::error::Error;
use std::io;
use std::process;

fn main() {
    if let Err(err) = run() {
        println!("{}", err);
        process::exit(1);
    }
}

fn run() -> Result<(), Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());
    for result in rdr.records() {
        // Examine our Result.
        // If there was no problem, print the record.
        // Otherwise, convert our error to a Box<Error> and return it.
        match result {
            Err(err) => return Err(From::from(err)),
            Ok(record) => {
              println!("{:?}", record);
            }
        }
    }
    Ok(())
}

Our new function, run, has a return type of Result<(), Box<Error>>. In simple terms, this says that run either returns nothing when successful, or if an error occurred, it returns a Box<Error>, which stands for “any kind of error.” A Box<Error> is hard to inspect if we cared about the specific error that occurred. But for our purposes, all we need to do is gracefully print an error message and exit the program.

The third and final step is to replace our explicit match expression with a special Rust language feature: the question mark.

//tutorial-error-04.rs
extern crate csv;

use std::error::Error;
use std::io;
use std::process;

fn main() {
    if let Err(err) = run() {
        println!("{}", err);
        process::exit(1);
    }
}

fn run() -> Result<(), Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());
    for result in rdr.records() {
        // This is effectively the same code as our `match` in the
        // previous example. In other words, `?` is syntactic sugar.
        let record = result?;
        println!("{:?}", record);
    }
    Ok(())
}

This last step shows how we can use the ? to automatically forward errors to our caller without having to do explicit case analysis with match ourselves. We will use the ? heavily throughout this tutorial, and it’s important to note that it can only be used in functions that return Result.

We’ll end this section with a word of caution: using Box<Error> as our error type is the minimally acceptable thing we can do here. Namely, while it allows our program to gracefully handle errors, it makes it hard for callers to inspect the specific error condition that occurred. However, since this is a tutorial on writing command line programs that do CSV parsing, we will consider ourselves satisfied. If you’d like to know more, or are interested in writing a library that handles CSV data, then you should check out my blog post on error handling.

With all that said, if all you’re doing is writing a one-off program to do CSV transformations, then using methods like expect and panicking when an error occurs is a perfectly reasonable thing to do. Nevertheless, this tutorial will endeavor to show idiomatic code.

Reading CSV

Now that we’ve got you setup and covered basic error handling, it’s time to do what we came here to do: handle CSV data. We’ve already seen how to read CSV data from stdin, but this section will cover how to read CSV data from files and how to configure our CSV reader to data formatted with different delimiters and quoting strategies.

First up, let’s adapt the example we’ve been working with to accept a file path argument instead of stdin.

//tutorial-read-01.rs
extern crate csv;

use std::env;
use std::error::Error;
use std::ffi::OsString;
use std::fs::File;
use std::process;

fn run() -> Result<(), Box<Error>> {
    let file_path = get_first_arg()?;
    let file = File::open(file_path)?;
    let mut rdr = csv::Reader::from_reader(file);
    for result in rdr.records() {
        let record = result?;
        println!("{:?}", record);
    }
    Ok(())
}

/// Returns the first positional argument sent to this process. If there are no
/// positional arguments, then this returns an error.
fn get_first_arg() -> Result<OsString, Box<Error>> {
    match env::args_os().nth(1) {
        None => Err(From::from("expected 1 argument, but got none")),
        Some(file_path) => Ok(file_path),
    }
}

fn main() {
    if let Err(err) = run() {
        println!("{}", err);
        process::exit(1);
    }
}

If you replace the contents of your src/main.rs file with the above code, then you should be able to rebuild your project and try it out:

$ cargo build
$ ./target/debug/csvtutor uspop.csv
StringRecord(["Davidsons Landing", "AK", "", "65.2419444", "-165.2716667"])
StringRecord(["Kenai", "AK", "7610", "60.5544444", "-151.2583333"])
StringRecord(["Oakman", "AL", "", "33.7133333", "-87.3886111"])
# ... and much more

This example contains two new pieces of code:

  1. Code for querying the positional arguments of your program. We put this code into its own funcation called get_first_arg. Our program expects a file path in the first position (which is indexed at 1; the argument at index 0 is the executable name), so if one doesn’t exist, then get_first_arg returns an error.
  2. Code for opening a file. In run, we open a file using File::open. If there was a problem opening the file, we forward the error to the caller of run (which is main in this program). Note that we do not wrap the File in a buffer. The CSV reader does buffering internally, so there’s no need for the caller to do it.

Now is a good time to introduce an alternate CSV reader constructor, which makes it slightly more convenient to open CSV data from a file. That is, instead of:

let file_path = get_first_arg()?;
let file = File::open(file_path)?;
let mut rdr = csv::Reader::from_reader(file);

you can use:

let file_path = get_first_arg()?;
let mut rdr = csv::Reader::from_path(file_path)?;

csv::Reader::from_path will open the file for you and return an error if the file could not be opened.

Reading headers

If you had a chance to look at the data inside uspop.csv, you would notice that there is a header record that looks like this:

City,State,Population,Latitude,Longitude

Now, if you look back at the output of the commands you’ve run so far, you’ll notice that the header record is never printed. Why is that? By default, the CSV reader will interpret the first record in CSV data as a header, which is typically distinct from the actual data in the records that follow. Therefore, the header record is always skipped whenever you try to read or iterate over the records in CSV data.

The CSV reader does not try to be smart about the header record and does not employ any heuristics for automatically detecting whether the first record is a header or not. Instead, if you don’t want to treat the first record as a header, you’ll need to tell the CSV reader that there are no headers.

To configure a CSV reader to do this, we’ll need to use a ReaderBuilder to build a CSV reader with our desired configuration. Here’s an example that does just that. (Note that we’ve moved back to reading from stdin, since it produces terser examples.)

//tutorial-read-headers-01.rs
fn run() -> Result<(), Box<Error>> {
    let mut rdr = csv::ReaderBuilder::new()
        .has_headers(false)
        .from_reader(io::stdin());
    for result in rdr.records() {
        let record = result?;
        println!("{:?}", record);
    }
    Ok(())
}

If you compile and run this program with our uspop.csv data, then you’ll see that the header record is now printed:

$ cargo build
$ ./target/debug/csvtutor < uspop.csv
StringRecord(["City", "State", "Population", "Latitude", "Longitude"])
StringRecord(["Davidsons Landing", "AK", "", "65.2419444", "-165.2716667"])
StringRecord(["Kenai", "AK", "7610", "60.5544444", "-151.2583333"])
StringRecord(["Oakman", "AL", "", "33.7133333", "-87.3886111"])

If you ever need to access the header record directly, then you can use the Reader::header method like so:

//tutorial-read-headers-02.rs
fn run() -> Result<(), Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());
    {
        // We nest this call in its own scope because of lifetimes.
        let headers = rdr.headers()?;
        println!("{:?}", headers);
    }
    for result in rdr.records() {
        let record = result?;
        println!("{:?}", record);
    }
    // We can ask for the headers at any time. There's no need to nest this
    // call in its own scope because we never try to borrow the reader again.
    let headers = rdr.headers()?;
    println!("{:?}", headers);
    Ok(())
}

One interesting thing to note in this example is that we put the call to rdr.headers() in its own scope. We do this because rdr.headers() returns a borrow of the reader’s internal header state. The nested scope in this code allows the borrow to end before we try to iterate over the records. If we didn’t nest the call to rdr.headers() in its own scope, then the code wouldn’t compile because we cannot borrow the reader’s headers at the same time that we try to borrow the reader to iterate over its records.

Another way of solving this problem is to clone the header record:

let headers = rdr.headers()?.clone();

This converts it from a borrow of the CSV reader to a new owned value. This makes the code a bit easier to read, but at the cost of copying the header record into a new allocation.

Delimiters, quotes and variable length records

In this section we’ll temporarily depart from our uspop.csv data set and show how to read some CSV data that is a little less clean. This CSV data uses ; as a delimiter, escapes quotes with \" (instead of "") and has records of varying length. Here’s the data, which contains a list of WWE wrestlers and the year they started, if it’s known:

$ cat strange.csv
"\"Hacksaw\" Jim Duggan";1987
"Bret \"Hit Man\" Hart";1984
# We're not sure when Rafael started, so omit the year.
Rafael Halperin
"\"Big Cat\" Ernie Ladd";1964
"\"Macho Man\" Randy Savage";1985
"Jake \"The Snake\" Roberts";1986

To read this CSV data, we’ll want to do the following:

  1. Disable headers, since this data has none.
  2. Change the delimiter from , to ;.
  3. Change the quote strategy from doubled (e.g., "") to escaped (e.g., \").
  4. Permit flexible length records, since some omit the year.
  5. Ignore lines beginning with a #.

All of this (and more!) can be configured with a ReaderBuilder, as seen in the following example:

//tutorial-read-delimiter-01.rs
fn run() -> Result<(), Box<Error>> {
    let mut rdr = csv::ReaderBuilder::new()
        .has_headers(false)
        .delimiter(b';')
        .double_quote(false)
        .escape(Some(b'\\'))
        .flexible(true)
        .comment(Some(b'#'))
        .from_reader(io::stdin());
    for result in rdr.records() {
        let record = result?;
        println!("{:?}", record);
    }
    Ok(())
}

Now re-compile your project and try running the program on strange.csv:

$ cargo build
$ ./target/debug/csvtutor < strange.csv
StringRecord(["\"Hacksaw\" Jim Duggan", "1987"])
StringRecord(["Bret \"Hit Man\" Hart", "1984"])
StringRecord(["Rafael Halperin"])
StringRecord(["\"Big Cat\" Ernie Ladd", "1964"])
StringRecord(["\"Macho Man\" Randy Savage", "1985"])
StringRecord(["Jake \"The Snake\" Roberts", "1986"])

You should feel encouraged to play around with the settings. Some interesting things you might try:

  1. If you remove the escape setting, notice that no CSV errors are reported. Instead, records are still parsed. This is a feature of the CSV parser. Even though it gets the data slightly wrong, it still provides a parse that you might be able to work with. This is a useful property given the messiness of real world CSV data.
  2. If you remove the delimiter setting, parsing still succeeds, although every record has exactly one field.
  3. If you remove the flexible setting, the reader will print the first two records (since they both have the same number of fields), but will return a parse error on the third record, since it has only one field.

This covers most of the things you might want to configure on your CSV reader, although there are a few other knobs. For example, you can change the record terminator from a new line to any other character. (By default, the terminator is CRLF, which treats each of \r\n, \r and \n as single record terminators.) For more details, see the documentation and examples for each of the methods on ReaderBuilder.

Reading with Serde

One of the most convenient features of this crate is its support for Serde. Serde is a framework for automatically serializing and deserializing data into Rust types. In simpler terms, that means instead of iterating over records as an array of string fields, we can iterate over records of a specific type of our choosing.

For example, let’s take a look at some data from our uspop.csv file:

City,State,Population,Latitude,Longitude
Davidsons Landing,AK,,65.2419444,-165.2716667
Kenai,AK,7610,60.5544444,-151.2583333

While some of these fields make sense as strings (City, State), other fields look more like numbers. For example, Population looks like it contains integers while Latitude and Longitude appear to contain decimals. If we wanted to convert these fields to their “proper” types, then we need to do a lot of manual work. This next example shows how.

//tutorial-read-serde-01.rs
fn run() -> Result<(), Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());
    for result in rdr.records() {
        let record = result?;

        let city = &record[0];
        let state = &record[1];
        // Some records are missing population counts, so if we can't
        // parse a number, treat the population count as missing instead
        // of returning an error.
        let pop: Option<u64> = record[2].parse().ok();
        // Lucky us! Latitudes and longitudes are available for every record.
        // Therefore, if one couldn't be parsed, return an error.
        let latitude: f64 = record[3].parse()?;
        let longitude: f64 = record[4].parse()?;

        println!(
            "city: {:?}, state: {:?}, \
             pop: {:?}, latitude: {:?}, longitude: {:?}",
            city, state, pop, latitude, longitude);
    }
    Ok(())
}

The problem here is that we need to parse each individual field manually, which can be labor intensive and repetitive. Serde, however, makes this process automatic. For example, we can ask to deserialize every record into a tuple type: (String, String, Option<u64>, f64, f64).

//tutorial-read-serde-02.rs
// This introduces a type alias so that we can conveniently reference our
// record type.
type Record = (String, String, Option<u64>, f64, f64);

fn run() -> Result<(), Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());
    // Instead of creating an iterator with the `records` method, we create
    // an iterator with the `deserialize` method.
    for result in rdr.deserialize() {
        // We must tell Serde what type we want to deserialize into.
        let record: Record = result?;
        println!("{:?}", record);
    }
    Ok(())
}

Running this code should show similar output as previous examples:

$ cargo build
$ ./target/debug/csvtutor < uspop.csv
("Davidsons Landing", "AK", None, 65.2419444, -165.2716667)
("Kenai", "AK", Some(7610), 60.5544444, -151.2583333)
("Oakman", "AL", None, 33.7133333, -87.3886111)
# ... and much more

One of the downsides of using Serde this way is that the type you use must match the order of fields as they appear in each record. This can be a pain if your CSV data has a header record, since you might tend to think about each field as a value of a particular named field rather than as a numbered field. One way we might achieve this is to deserialize our record into a map type like HashMap or BTreeMap. The next example shows how, and in particular, notice that the only thing that changed from the last example is the definition of the Record type alias and a new use statement that imports HashMap from the standard library:

//tutorial-read-serde-03.rs
use std::collections::HashMap;

// This introduces a type alias so that we can conveniently reference our
// record type.
type Record = HashMap<String, String>;

fn run() -> Result<(), Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());
    for result in rdr.deserialize() {
        let record: Record = result?;
        println!("{:?}", record);
    }
    Ok(())
}

Running this program shows similar results as before, but each record is printed as a map:

$ cargo build
$ ./target/debug/csvtutor < uspop.csv
{"City": "Davidsons Landing", "Latitude": "65.2419444", "State": "AK", "Population": "", "Longitude": "-165.2716667"}
{"City": "Kenai", "Population": "7610", "State": "AK", "Longitude": "-151.2583333", "Latitude": "60.5544444"}
{"State": "AL", "City": "Oakman", "Longitude": "-87.3886111", "Population": "", "Latitude": "33.7133333"}

This method works especially well if you need to read CSV data with header records, but whose exact structure isn’t known until your program runs. However, in our case, we know the structure of the data in uspop.csv. In particular, with the HashMap approach, we’ve lost the specific types we had for each field in the previous example when we deserialized each record into a (String, String, Option<u64>, f64, f64). Is there a way to identify fields by their corresponding header name and assign each field its own unique type? The answer is yes, but we’ll need to bring in a new crate called serde_derive first. You can do that by adding this to the [dependencies] section of your Cargo.toml file:

serde = "1"
serde_derive = "1"

With these crates added to our project, we can now define our own custom struct that represents our record. We then ask Serde to automatically write the glue code required to populate our struct from a CSV record. The next example shows how. Don’t miss the new extern crate lines!

//tutorial-read-serde-04.rs
extern crate csv;
extern crate serde;
// This lets us write `#[derive(Deserialize)]`.
#[macro_use]
extern crate serde_derive;

use std::error::Error;
use std::io;
use std::process;

// We don't need to derive `Debug` (which doesn't require Serde), but it's a
// good habit to do it for all your types.
//
// Notice that the field names in this struct are NOT in the same order as
// the fields in the CSV data!
#[derive(Debug, Deserialize)]
#[serde(rename_all = "PascalCase")]
struct Record {
    latitude: f64,
    longitude: f64,
    population: Option<u64>,
    city: String,
    state: String,
}

fn run() -> Result<(), Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());
    for result in rdr.deserialize() {
        let record: Record = result?;
        println!("{:?}", record);
        // Try this if you don't like each record smushed on one line:
        // println!("{:#?}", record);
    }
    Ok(())
}

fn main() {
    if let Err(err) = run() {
        println!("{}", err);
        process::exit(1);
    }
}

Compile and run this program to see similar output as before:

$ cargo build
$ ./target/debug/csvtutor < uspop.csv
Record { latitude: 65.2419444, longitude: -165.2716667, population: None, city: "Davidsons Landing", state: "AK" }
Record { latitude: 60.5544444, longitude: -151.2583333, population: Some(7610), city: "Kenai", state: "AK" }
Record { latitude: 33.7133333, longitude: -87.3886111, population: None, city: "Oakman", state: "AL" }

Once again, we didn’t need to change our run function at all: we’re still iterating over records using the deserialize iterator that we started with in the beginning of this section. The only thing that changed in this example was the definition of the Record type and a couple new extern crate statements. Our Record type is now a custom struct that we defined instead of a type alias, and as a result, Serde doesn’t know how to deserialize it by default. However, a special compiler plugin called serde_derive is available, which will read your struct definition at compile time and generate code that will deserialize a CSV record into a Record value. To see what happens if you leave out the automatic derive, change #[derive(Debug, Deserialize)] to #[derive(Debug)].

One other thing worth mentioning in this example is the use of #[serde(rename_all = "PascalCase")]. This directive helps Serde map your struct’s field names to the header names in the CSV data. If you recall, our header record is:

City,State,Population,Latitude,Longitude

Notice that each name is capitalized, but the fields in our struct are not. The #[serde(rename_all = "PascalCase")] directive fixes that by interpreting each field in PascalCase, where the first letter of the field is capitalized. If we didn’t tell Serde about the name remapping, then the program will quit with an error:

$ ./target/debug/csvtutor < uspop.csv
CSV deserialize error: record 1 (line: 2, byte: 41): missing field `latitude`

We could have fixed this through other means. For example, we could have used capital letters in our field names:

#[derive(Debug, Deserialize)]
struct Record {
    Latitude: f64,
    Longitude: f64,
    Population: Option<u64>,
    City: String,
    State: String,
}

However, this violates Rust naming style. (In fact, the Rust compiler will even warn you that the names do not follow convention!)

Another way to fix this is to ask Serde to rename each field individually. This is useful when there is no consistent name mapping from fields to header names:

#[derive(Debug, Deserialize)]
struct Record {
    #[serde(rename = "Latitude")]
    latitude: f64,
    #[serde(rename = "Longitude")]
    longitude: f64,
    #[serde(rename = "Population")]
    population: Option<u64>,
    #[serde(rename = "City")]
    city: String,
    #[serde(rename = "State")]
    state: String,
}

To read more about renaming fields and about other Serde directives, please consult the Serde documentation on attributes.

Handling invalid data with Serde

In this section we will see a brief example of how to deal with data that isn’t clean. To do this exercise, we’ll work with a slightly tweaked version of the US population data we’ve been using throughout this tutorial. This version of the data is slightly messier than what we’ve been using. You can get it like so:

$ curl -LO 'https://raw.githubusercontent.com/BurntSushi/rust-csv/master/examples/data/uspop-null.csv'

Let’s start by running our program from the previous section:

//tutorial-read-serde-invalid-01.rs
#[derive(Debug, Deserialize)]
#[serde(rename_all = "PascalCase")]
struct Record {
    latitude: f64,
    longitude: f64,
    population: Option<u64>,
    city: String,
    state: String,
}

fn run() -> Result<(), Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());
    for result in rdr.deserialize() {
        let record: Record = result?;
        println!("{:?}", record);
    }
    Ok(())
}

Compile and run it on our messier data:

$ cargo build
$ ./target/debug/csvtutor < uspop-null.csv
Record { latitude: 65.2419444, longitude: -165.2716667, population: None, city: "Davidsons Landing", state: "AK" }
Record { latitude: 60.5544444, longitude: -151.2583333, population: Some(7610), city: "Kenai", state: "AK" }
Record { latitude: 33.7133333, longitude: -87.3886111, population: None, city: "Oakman", state: "AL" }
# ... more records
CSV deserialize error: record 42 (line: 43, byte: 1710): field 2: invalid digit found in string

Oops! What happened? The program printed several records, but stopped when it tripped over a deserialization problem. The error message says that it found an invalid digit in the field at index 2 (which is the Population field) on line 43. What does line 43 look like?

$ head -n 43 uspop-null.csv | tail -n1
Flint Springs,KY,NULL,37.3433333,-86.7136111

Ah! The third field (index 2) is supposed to either be empty or contain a population count. However, in this data, it seems that NULL sometimes appears as a value, presumably to indicate that there is no count available.

The problem with our current program is that it fails to read this record because it doesn’t know how to deserialize a NULL string into an Option<u64>. That is, a Option<u64> either corresponds to an empty field or an integer.

To fix this, we tell Serde to convert any deserialization errors on this field to a None value, as shown in this next example:

//tutorial-read-serde-invalid-02.rs
#[derive(Debug, Deserialize)]
#[serde(rename_all = "PascalCase")]
struct Record {
    latitude: f64,
    longitude: f64,
    #[serde(deserialize_with = "csv::invalid_option")]
    population: Option<u64>,
    city: String,
    state: String,
}

fn run() -> Result<(), Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());
    for result in rdr.deserialize() {
        let record: Record = result?;
        println!("{:?}", record);
    }
    Ok(())
}

If you compile and run this example, then it should run to completion just like the other examples:

$ cargo build
$ ./target/debug/csvtutor < uspop-null.csv
Record { latitude: 65.2419444, longitude: -165.2716667, population: None, city: "Davidsons Landing", state: "AK" }
Record { latitude: 60.5544444, longitude: -151.2583333, population: Some(7610), city: "Kenai", state: "AK" }
Record { latitude: 33.7133333, longitude: -87.3886111, population: None, city: "Oakman", state: "AL" }
# ... and more

The only change in this example was adding this attribute to the population field in our Record type:

#[serde(deserialize_with = "csv::invalid_option")]

The invalid_option function is a generic helper function that does one very simple thing: when applied to Option fields, it will convert any deserialization error into a None value. This is useful when you need to work with messy CSV data.

Writing CSV

In this section we’ll show a few examples that write CSV data. Writing CSV data tends to be a bit more straight-forward than reading CSV data, since you get to control the output format.

Let’s start with the most basic example: writing a few CSV records to stdout.

//tutorial-write-01.rs
extern crate csv;

use std::error::Error;
use std::io;
use std::process;

fn run() -> Result<(), Box<Error>> {
    let mut wtr = csv::Writer::from_writer(io::stdout());
    // Since we're writing records manually, we must explicitly write our
    // header record. A header record is written the same way that other
    // records are written.
    wtr.write_record(&["City", "State", "Population", "Latitude", "Longitude"])?;
    wtr.write_record(&["Davidsons Landing", "AK", "", "65.2419444", "-165.2716667"])?;
    wtr.write_record(&["Kenai", "AK", "7610", "60.5544444", "-151.2583333"])?;
    wtr.write_record(&["Oakman", "AL", "", "33.7133333", "-87.3886111"])?;

    // A CSV writer maintains an internal buffer, so it's important
    // to flush the buffer when you're done.
    wtr.flush()?;
    Ok(())
}

fn main() {
    if let Err(err) = run() {
        println!("{}", err);
        process::exit(1);
    }
}

Compiling and running this example results in CSV data being printed:

$ cargo build
$ ./target/debug/csvtutor
City,State,Population,Latitude,Longitude
Davidsons Landing,AK,,65.2419444,-165.2716667
Kenai,AK,7610,60.5544444,-151.2583333
Oakman,AL,,33.7133333,-87.3886111

Before moving on, it’s worth taking a closer look at the write_record method. In this example, it looks rather simple, but if you’re new to Rust then its type signature might look a little daunting:

pub fn write_record<I, T>(&mut self, record: I) -> csv::Result<()>
    where I: IntoIterator<Item=T>, T: AsRef<[u8]>
{
    // implementation elided
}

To understand the type signature, we can break it down piece by piece.

  1. The method takes two parameters: self and record.
  2. self is a special parameter that corresponds to the Writer itself.
  3. record is the CSV record we’d like to write. Its type is I, which is a generic type.
  4. In the method’s where clause, the I type is contrained by the IntoIterator<Item=T> bound. What that means is that I must satisfy the IntoIterator trait. If you look at the documentation of the IntoIterator trait, then we can see that it describes types that can build iterators. In this case, we want an iterator that yields another generic type T, where T is the type of each field we want to write.
  5. T also appears in the method’s where clause, but its constraint is the AsRef<[u8]> bound. The AsRef trait is a way to describe zero cost conversions between types in Rust. In this case, the [u8] in AsRef<[u8]> means that we want to be able to borrow a slice of bytes from T. The CSV writer will take these bytes and write them as a single field. The AsRef<[u8]> bound is useful because types like String, &str, Vec<u8> and &[u8] all satisfy it.
  6. Finally, the method returns a csv::Result<()>, which is short-hand for Result<(), csv::Error>. That means write_record either returns nothing on success or returns a csv::Error on failure.

Now, let’s apply our new found understanding of the type signature of write_record. If you recall, in our previous example, we used it like so:

wtr.write_record(&["field 1", "field 2", "etc"])?;

So how do the types match up? Well, the type of each of our fields in this code is &'static str (which is the type of a string literal in Rust). Since we put them in a slice literal, the type of our parameter is &'static [&'static str], or more succinctly written as &[&str] without the lifetime annotations. Since slices satisfy the IntoIterator bound and strings satisfy the AsRef<[u8]> bound, this ends up being a legal call.

Here are a few more examples of ways you can call write_record:

// A slice of byte strings.
wtr.write_record(&[b"a", b"b", b"c"]);
// A vector.
wtr.write_record(vec!["a", "b", "c"]);
// A string record.
wtr.write_record(&csv::StringRecord::from(vec!["a", "b", "c"]));
// A byte record.
wtr.write_record(&csv::ByteRecord::from(vec!["a", "b", "c"]));

Finally, the example above can be easily adapted to write to a file instead of stdout:

//tutorial-write-02.rs
extern crate csv;

use std::env;
use std::error::Error;
use std::ffi::OsString;
use std::process;

fn run() -> Result<(), Box<Error>> {
    let file_path = get_first_arg()?;
    let mut wtr = csv::Writer::from_path(file_path)?;

    wtr.write_record(&["City", "State", "Population", "Latitude", "Longitude"])?;
    wtr.write_record(&["Davidsons Landing", "AK", "", "65.2419444", "-165.2716667"])?;
    wtr.write_record(&["Kenai", "AK", "7610", "60.5544444", "-151.2583333"])?;
    wtr.write_record(&["Oakman", "AL", "", "33.7133333", "-87.3886111"])?;

    wtr.flush()?;
    Ok(())
}

/// Returns the first positional argument sent to this process. If there are no
/// positional arguments, then this returns an error.
fn get_first_arg() -> Result<OsString, Box<Error>> {
    match env::args_os().nth(1) {
        None => Err(From::from("expected 1 argument, but got none")),
        Some(file_path) => Ok(file_path),
    }
}

fn main() {
    if let Err(err) = run() {
        println!("{}", err);
        process::exit(1);
    }
}

Writing tab separated values

In the previous section, we saw how to write some simple CSV data to stdout that looked like this:

City,State,Population,Latitude,Longitude
Davidsons Landing,AK,,65.2419444,-165.2716667
Kenai,AK,7610,60.5544444,-151.2583333
Oakman,AL,,33.7133333,-87.3886111

You might wonder to yourself: what’s the point of using a CSV writer if the data is so simple? Well, the benefit of a CSV writer is that it can handle all types of data without sacrificing the integrity of your data. That is, it knows when to quote fields that contain special CSV characters (like commas or new lines) or escape literal quotes that appear in your data. The CSV writer can also be easily configured to use different delimiters or quoting strategies.

In this section, we’ll take a look a look at how to tweak some of the settings on a CSV writer. In particular, we’ll write TSV (“tab separated values”) instead of CSV, and we’ll ask the CSV writer to quote all non-numeric fields. Here’s an example:

//tutorial-write-delimiter-01.rs
fn run() -> Result<(), Box<Error>> {
    let mut wtr = csv::WriterBuilder::new()
        .delimiter(b'\t')
        .quote_style(csv::QuoteStyle::NonNumeric)
        .from_writer(io::stdout());

    wtr.write_record(&["City", "State", "Population", "Latitude", "Longitude"])?;
    wtr.write_record(&["Davidsons Landing", "AK", "", "65.2419444", "-165.2716667"])?;
    wtr.write_record(&["Kenai", "AK", "7610", "60.5544444", "-151.2583333"])?;
    wtr.write_record(&["Oakman", "AL", "", "33.7133333", "-87.3886111"])?;

    wtr.flush()?;
    Ok(())
}

Compiling and running this example gives:

$ cargo build
$ ./target/debug/csvtutor
"City"  "State" "Population"    "Latitude"      "Longitude"
"Davidsons Landing"     "AK"    ""      65.2419444      -165.2716667
"Kenai" "AK"    7610    60.5544444      -151.2583333
"Oakman"        "AL"    ""      33.7133333      -87.3886111

In this example, we used a new type QuoteStyle. The QuoteStyle type represents the different quoting strategies available to you. The default is to add quotes to fields only when necessary. This probably works for most use cases, but you can also ask for quotes to always be put around fields, to never be put around fields or to always be put around non-numeric fields.

Writing with Serde

Just like the CSV reader supports automatic deserialization into Rust types with Serde, the CSV writer supports automatic serialization from Rust types into CSV records using Serde. In this section, we’ll learn how to use it.

As with reading, let’s start by seeing how we can serialize a Rust tuple.

//tutorial-write-serde-01.rs
fn run() -> Result<(), Box<Error>> {
    let mut wtr = csv::Writer::from_writer(io::stdout());

    // We still need to write headers manually.
    wtr.write_record(&["City", "State", "Population", "Latitude", "Longitude"])?;

    // But now we can write records by providing a normal Rust value.
    //
    // Note that the odd `None::<u64>` syntax is required because `None` on
    // its own doesn't have a concrete type, but Serde needs a concrete type
    // in order to serialize it. That is, `None` has type `Option<T>` but
    // `None::<u64>` has type `Option<u64>`.
    wtr.serialize(("Davidsons Landing", "AK", None::<u64>, 65.2419444, -165.2716667))?;
    wtr.serialize(("Kenai", "AK", Some(7610), 60.5544444, -151.2583333))?;
    wtr.serialize(("Oakman", "AL", None::<u64>, 33.7133333, -87.3886111))?;

    wtr.flush()?;
    Ok(())
}

Compiling and running this program gives the expected output:

$ cargo build
$ ./target/debug/csvtutor
City,State,Population,Latitude,Longitude
Davidsons Landing,AK,,65.2419444,-165.2716667
Kenai,AK,7610,60.5544444,-151.2583333
Oakman,AL,,33.7133333,-87.3886111

The key thing to note in the above example is the use of serialize instead of write_record to write our data. In particular, write_record is used when writing a simple record that contains string-like data only. On the other hand, serialize is used when your data consists of more complex values like numbers, floats or optional values. Of course, you could always convert the complex values to strings and then use write_record, but Serde can do it for you automatically.

As with reading, we can also serialize custom structs as CSV records. As a bonus, the fields in a struct will automatically be written as a header record!

To write custom structs as CSV records, we’ll need to make use of the serde_derive crate again. As in the previous section on reading with Serde, we’ll need to add a couple crates to our [dependencies] section in our Cargo.toml (if they aren’t already there):

serde = "1"
serde_derive = "1"

And we’ll also need to add a couple extra extern crate statements to our code, as shown in the example:

//tutorial-write-serde-02.rs
extern crate csv;
extern crate serde;
#[macro_use]
extern crate serde_derive;

use std::error::Error;
use std::io;
use std::process;

// Note that structs can derive both Serialize and Deserialize!
#[derive(Debug, Serialize)]
#[serde(rename_all = "PascalCase")]
struct Record<'a> {
    city: &'a str,
    state: &'a str,
    population: Option<u64>,
    latitude: f64,
    longitude: f64,
}

fn run() -> Result<(), Box<Error>> {
    let mut wtr = csv::Writer::from_writer(io::stdout());

    wtr.serialize(Record {
        city: "Davidsons Landing",
        state: "AK",
        population: None,
        latitude: 65.2419444,
        longitude: -165.2716667,
    })?;
    wtr.serialize(Record {
        city: "Kenai",
        state: "AK",
        population: Some(7610),
        latitude: 60.5544444,
        longitude: -151.2583333,
    })?;
    wtr.serialize(Record {
        city: "Oakman",
        state: "AL",
        population: None,
        latitude: 33.7133333,
        longitude: -87.3886111,
    })?;

    wtr.flush()?;
    Ok(())
}

fn main() {
    if let Err(err) = run() {
        println!("{}", err);
        process::exit(1);
    }
}

Compiling and running this example has the same output as last time, even though we didn’t explicitly write a header record:

$ cargo build
$ ./target/debug/csvtutor
City,State,Population,Latitude,Longitude
Davidsons Landing,AK,,65.2419444,-165.2716667
Kenai,AK,7610,60.5544444,-151.2583333
Oakman,AL,,33.7133333,-87.3886111

In this case, the serialize method noticed that we were writing a struct with field names. When this happens, serialize will automatically write a header record (only if no other records have been written) that consists of the fields in the struct in the order in which they are defined. Note that this behavior can be disabled with the WriterBuilder::has_headers method.

It’s also worth pointing out the use of a lifetime parameter in our Record struct:

struct Record<'a> {
    city: &'a str,
    state: &'a str,
    population: Option<u64>,
    latitude: f64,
    longitude: f64,
}

The 'a lifetime parameter corresponds to the lifetime of the city and state string slices. This says that the Record struct contains borrowed data. We could have written our struct without borrowing any data, and therefore, without any lifetime parameters:

struct Record {
    city: String,
    state: String,
    population: Option<u64>,
    latitude: f64,
    longitude: f64,
}

However, since we had to replace our borrowed &str types with owned String types, we’re now forced to allocate a new String value for both of city and state for every record that we write. There’s no intrinsic problem with doing that, but it might be a bit wasteful.

For more examples and more details on the rules for serialization, please see the Writer::serialize method.

Pipelining

In this section, we’re going to cover a few examples that demonstrate programs that take CSV data as input, and produce possibly transformed or filtered CSV data as output. This shows how to write a complete program that efficiently reads and writes CSV data. Rust is well positioned to perform this task, since you’ll get great performance with the convenience of a high level CSV library.

The first example of CSV pipelining we’ll look at is a simple filter. It takes as input some CSV data on stdin and a single string query as its only positional argument, and it will produce as output CSV data that only contains rows with a field that matches the query.

//tutorial-pipeline-search-01.rs
extern crate csv;

use std::env;
use std::error::Error;
use std::io;
use std::process;

fn run() -> Result<(), Box<Error>> {
    // Get the query from the positional arguments.
    // If one doesn't exist, return an error.
    let query = match env::args().nth(1) {
        None => return Err(From::from("expected 1 argument, but got none")),
        Some(query) => query,
    };

    // Build CSV readers and writers to stdin and stdout, respectively.
    let mut rdr = csv::Reader::from_reader(io::stdin());
    let mut wtr = csv::Writer::from_writer(io::stdout());

    // Before reading our data records, we should write the header record.
    wtr.write_record(rdr.headers()?)?;

    // Iterate over all the records in `rdr`, and write only records containing
    // `query` to `wtr`.
    for result in rdr.records() {
        let record = result?;
        if record.iter().any(|field| field == &query) {
            wtr.write_record(&record)?;
        }
    }

    // CSV writers use an internal buffer, so we should always flush when done.
    wtr.flush()?;
    Ok(())
}

fn main() {
    if let Err(err) = run() {
        println!("{}", err);
        process::exit(1);
    }
}

If we compile and run this program with a query of MA on uspop.csv, we’ll see that only one record matches:

$ cargo build
$ ./csvtutor MA < uspop.csv
City,State,Population,Latitude,Longitude
Reading,MA,23441,42.5255556,-71.0958333

This example doesn’t actually introduce anything new. It merely combines what you’ve already learned about CSV readers and writers from previous sections.

Let’s add a twist to this example. In the real world, you’re often faced with messy CSV data that might not be encoded correctly. One example you might come across is CSV data encoded in Latin-1. Unfortunately, for the examples we’ve seen so far, our CSV reader assumes that all of the data is UTF-8. Since all of the data we’ve worked on has been ASCII—which is a subset of both Latin-1 and UTF-8—we haven’t had any problems. But let’s introduce a slightly tweaked version of our uspop.csv file that contains an encoding of a Latin-1 character that is invalid UTF-8. You can get the data like so:

$ curl -LO 'https://raw.githubusercontent.com/BurntSushi/rust-csv/master/examples/data/uspop-latin1.csv'

Even though I’ve already given away the problem, let’s see what happen when we try to run our previous example on this new data:

$ ./csvtutor MA < uspop-latin1.csv
City,State,Population,Latitude,Longitude
CSV parse error: record 3 (line 4, field: 0, byte: 125): invalid utf-8: invalid UTF-8 in field 0 near byte index 0

The error message tells us exactly what’s wrong. Let’s take a look at line 4 to see what we’re dealing with:

$ head -n4 uspop-latin1.csv | tail -n1
Õakman,AL,,33.7133333,-87.3886111

In this case, the very first character is the Latin-1 Õ, which is encoded as the byte 0xD5, which is in turn invalid UTF-8. So what do we do now that our CSV parser has choked on our data? You have two choices. The first is to go in and fix up your CSV data so that it’s valid UTF-8. This is probably a good idea anyway, and tools like iconv can help with the task of transcoding. But if you can’t or don’t want to do that, then you can instead read CSV data in a way that is mostly encoding agnostic (so long as ASCII is still a valid subset). The trick is to use byte records instead of string records.

Thus far, we haven’t actually talked much about the type of a record in this library, but now is a good time to introduce them. There are two of them, StringRecord and ByteRecord. Each of them represent a single record in CSV data, where a record is a sequence of an arbitrary number of fields. The only difference between StringRecord and ByteRecord is that StringRecord is guaranteed to be valid UTF-8, where as ByteRecord contains arbitrary bytes. To be clear, the in-memory representation of both types is identical.

Armed with that knowledge, we can now begin to understand why we saw an error when we ran the last example on data that wasn’t UTF-8. Namely, when we call records, we get back an iterator of StringRecord. Since StringRecord is guaranteed to be valid UTF-8, trying to build a StringRecord with invalid UTF-8 will result in the error that we see.

All we need to do to make our example work is to switch from a StringRecord to a ByteRecord. This means using byte_records to create our iterator instead of records, and similarly using byte_headers instead of headers if we think our header data might contain invalid UTF-8 as well. Here’s the change:

//tutorial-pipeline-search-02.rs
fn run() -> Result<(), Box<Error>> {
    let query = match env::args().nth(1) {
        None => return Err(From::from("expected 1 argument, but got none")),
        Some(query) => query,
    };

    let mut rdr = csv::Reader::from_reader(io::stdin());
    let mut wtr = csv::Writer::from_writer(io::stdout());

    wtr.write_record(rdr.byte_headers()?)?;

    for result in rdr.byte_records() {
        let record = result?;
        // `query` is a `String` while `field` is now a `&[u8]`, so we'll
        // need to convert `query` to `&[u8]` before doing a comparison.
        if record.iter().any(|field| field == query.as_bytes()) {
            wtr.write_record(&record)?;
        }
    }

    wtr.flush()?;
    Ok(())
}

Compiling and running this now yields the same results as our first example, but this time it works on data that isn’t valid UTF-8.

$ cargo build
$ ./csvtutor MA < uspop-latin1.csv
City,State,Population,Latitude,Longitude
Reading,MA,23441,42.5255556,-71.0958333

Filter by population count

In this section, we will show another example program that both reads and writes CSV data, but instead of dealing with arbitrary records, we will use Serde to deserialize and serialize records with specific types.

For this program, we’d like to be able to filter records in our population data by population count. Specifically, we’d like to see which records meet a certain population threshold. In addition to using a simple inequality, we must also account for records that have a missing population count. This is where types like Option<T> come in handy, because the compiler will force us to consider the case when the population count is missing.

Since we’re using Serde in this example, don’t forget to add the Serde dependencies to your Cargo.toml in your [dependencies] section if they aren’t already there:

serde = "1"
serde_derive = "1"

Now here’s the code:

//tutorial-pipeline-pop-01.rs
extern crate csv;
extern crate serde;
#[macro_use]
extern crate serde_derive;

use std::env;
use std::error::Error;
use std::io;
use std::process;

// Unlike previous examples, we derive both Deserialize and Serialize. This
// means we'll be able to automatically deserialize and serialize this type.
#[derive(Debug, Deserialize, Serialize)]
#[serde(rename_all = "PascalCase")]
struct Record {
    city: String,
    state: String,
    population: Option<u64>,
    latitude: f64,
    longitude: f64,
}

fn run() -> Result<(), Box<Error>> {
    // Get the query from the positional arguments.
    // If one doesn't exist or isn't an integer, return an error.
    let minimum_pop: u64 = match env::args().nth(1) {
        None => return Err(From::from("expected 1 argument, but got none")),
        Some(arg) => arg.parse()?,
    };

    // Build CSV readers and writers to stdin and stdout, respectively.
    // Note that we don't need to write headers explicitly. Since we're
    // serializing a custom struct, that's done for us automatically.
    let mut rdr = csv::Reader::from_reader(io::stdin());
    let mut wtr = csv::Writer::from_writer(io::stdout());

    // Iterate over all the records in `rdr`, and write only records containing
    // a population that is greater than or equal to `minimum_pop`.
    for result in rdr.deserialize() {
        // Remember that when deserializing, we must use a type hint to
        // indicate which type we want to deserialize our record into.
        let record: Record = result?;

        // `map_or` is a combinator on `Option`. It take two parameters:
        // a value to use when the `Option` is `None` (i.e., the record has
        // no population count) and a closure that returns another value of
        // the same type when the `Option` is `Some`. In this case, we test it
        // against our minimum population count that we got from the command
        // line.
        if record.population.map_or(false, |pop| pop >= minimum_pop) {
            wtr.serialize(record)?;
        }
    }

    // CSV writers use an internal buffer, so we should always flush when done.
    wtr.flush()?;
    Ok(())
}

fn main() {
    if let Err(err) = run() {
        println!("{}", err);
        process::exit(1);
    }
}

If we compile and run our program with a minimum threshold of 100000, we should see three matching records. Notice that the headers were added even though we never explicitly wrote them!

$ cargo build
$ ./target/debug/csvtutor 100000 < uspop.csv
City,State,Population,Latitude,Longitude
Fontana,CA,169160,34.0922222,-117.4341667
Bridgeport,CT,139090,41.1669444,-73.2052778
Indianapolis,IN,773283,39.7683333,-86.1580556

Performance

In this section, we’ll go over how to squeeze the most juice out of our CSV reader. As it happens, most of the APIs we’ve seen so far were designed with high level convenience in mind, and that often comes with some costs. For the most part, those costs revolve around unnecessary allocations. Therefore, most of the section will show how to do CSV parsing with as little allocation as possible.

There are two critical preliminaries we must cover.

Firstly, when you care about performance, you should compile your code with cargo build --release instead of cargo build. The --release flag instructs the compiler to spend more time optimizing your code. When compiling with the --release flag, you’ll find your compiled program at target/release/csvtutor instead of target/debug/csvtutor. Throughout this tutorial, we’ve used cargo build because our dataset was small and we weren’t focused on speed. The downside of cargo build --release is that it will take longer than cargo build.

Secondly, the dataset we’ve used throughout this tutorial only has 100 records. We’d have to try really hard to cause our program to run slowly on 100 records, even when we compile without the --release flag. Therefore, in order to actually witness a performance difference, we need a bigger dataset. To get such a dataset, we’ll use the original source of uspop.csv. Warning: the download is 41MB compressed and decompresses to 145MB.

$ curl -LO http://burntsushi.net/stuff/worldcitiespop.csv.gz
$ gunzip worldcitiespop.csv.gz
$ wc worldcitiespop.csv
  3173959   5681543 151492068 worldcitiespop.csv
$ md5sum worldcitiespop.csv
6198bd180b6d6586626ecbf044c1cca5  worldcitiespop.csv

Finally, it’s worth pointing out that this section is not attempting to present a rigorous set of benchmarks. We will stay away from rigorous analysis and instead rely a bit more on wall clock times and intuition.

Amortizing allocations

In order to measure performance, we must be careful about what it is we’re measuring. We must also be careful to not change the thing we’re measuring as we make improvements to the code. For this reason, we will focus on measuring how long it takes to count the number of records corresponding to city population counts in Massachusetts. This represents a very small amount of work that requires us to visit every record, and therefore represents a decent way to measure how long it takes to do CSV parsing.

Before diving into our first optimization, let’s start with a baseline by adapting a previous example to count the number of records in worldcitiespop.csv:

//tutorial-perf-alloc-01.rs
extern crate csv;

use std::error::Error;
use std::io;
use std::process;

fn run() -> Result<u64, Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());

    let mut count = 0;
    for result in rdr.records() {
        let record = result?;
        if &record[0] == "us" && &record[3] == "MA" {
            count += 1;
        }
    }
    Ok(count)
}

fn main() {
    match run() {
        Ok(count) => {
            println!("{}", count);
        }
        Err(err) => {
            println!("{}", err);
            process::exit(1);
        }
    }
}

Now let’s compile and run it and see what kind of timing we get. Don’t forget to compile with the --release flag. (For grins, try compiling without the --release flag and see how long it takes to run the program!)

$ cargo build --release
$ time ./target/release/csvtutor < worldcitiespop.csv
2176

real    0m0.645s
user    0m0.627s
sys     0m0.017s

All right, so what’s the first thing we can do to make this faster? This section promised to speed things up by amortizing allocation, but we can do something even simpler first: iterate over ByteRecords instead of StringRecords. If you recall from a previous section, a StringRecord is guaranteed to be valid UTF-8, and therefore must validate that its contents is actually UTF-8. (If validation fails, then the CSV reader will return an error.) If we remove that validation from our program, then we can realize a nice speed boost as shown in the next example:

//tutorial-perf-alloc-02.rs
fn run() -> Result<u64, Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());

    let mut count = 0;
    for result in rdr.byte_records() {
        let record = result?;
        if &record[0] == b"us" && &record[3] == b"MA" {
            count += 1;
        }
    }
    Ok(count)
}

And now compile and run:

$ cargo build --release
$ time ./target/release/csvtutor < worldcitiespop.csv
2176

real    0m0.429s
user    0m0.403s
sys     0m0.023s

Our program is now approximately 30% faster, all because we removed UTF-8 validation. But was it actually okay to remove UTF-8 validation? What have we lost? In this case, it is perfectly acceptable to drop UTF-8 validation and use ByteRecord instead because all we’re doing with the data in the record is comparing two of its fields to raw bytes:

if &record[0] == b"us" && &record[3] == b"MA" {
    count += 1;
}

In particular, it doesn’t matter whether record is valid UTF-8 or not, since we’re checking for equality on the raw bytes themselves.

UTF-8 validation via StringRecord is useful because it provides access to fields as &str types, where as ByteRecord provides fields as &[u8] types. &str is the type of a borrowed string in Rust, which provides convenient access to string APIs like substring search. Strings are also frequently used in other areas, so they tend to be a useful thing to have. Therefore, sticking with StringRecord is a good default, but if you need the extra speed and can deal with arbitrary bytes, then switching to ByteRecord might be a good idea.

Moving on, let’s try to get another speed boost by amortizing allocation. Amortizing allocation is the technique that creates an allocation once (or very rarely), and then attempts to reuse it instead of creating additional allocations. In the case of the previous examples, we used iterators created by the records and byte_records methods on a CSV reader. These iterators allocate a new record for every item that it yields, which in turn corresponds to a new allocation. It does this because iterators cannot yield items that borrow from the iterator itself, and because creating new allocations tends to be a lot more convenient.

If we’re willing to forgo use of iterators, then we can amortize allocations by creating a single ByteRecord and asking the CSV reader to read into it. We do this by using the Reader::read_byte_record method.

//tutorial-perf-alloc-03.rs
fn run() -> Result<u64, Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());
    let mut record = csv::ByteRecord::new();

    let mut count = 0;
    while rdr.read_byte_record(&mut record)? {
        if &record[0] == b"us" && &record[3] == b"MA" {
            count += 1;
        }
    }
    Ok(count)
}

Compile and run:

$ cargo build --release
$ time ./target/release/csvtutor < worldcitiespop.csv
2176

real    0m0.308s
user    0m0.283s
sys     0m0.023s

Woohoo! This represents another 30% boost over the previous example, which is a 50% boost over the first example.

Let’s dissect this code by taking a look at the type signature of the read_byte_record method:

fn read_byte_record(&mut self, record: &mut ByteRecord) -> csv::Result<bool>;

This method takes as input a CSV reader (the self parameter) and a mutable borrow of a ByteRecord, and returns a csv::Result<bool>. (The csv::Result<bool> is equivalent to Result<bool, csv::Error>.) The return value is true if and only if a record was read. When it’s false, that means the reader has exhausted its input. This method works by copying the contents of the next record into the provided ByteRecord. Since the same ByteRecord is used to read every record, it will already have space allocated for data. When read_byte_record runs, it will overwrite the contents that were there with the new record, which means that it can reuse the space that was allocated. Thus, we have amortized allocation.

An exercise you might consider doing is to use a StringRecord instead of a ByteRecord, and therefore Reader::read_record instead of read_byte_record. This will give you easy access to Rust strings at the cost of UTF-8 validation but without the cost of allocating a new StringRecord for every record.

Serde and zero allocation

In this section, we are going to briefly examine how we use Serde and what we can do to speed it up. The key optimization we’ll want to make is to—you guessed it—amortize allocation.

As with the previous section, let’s start with a simple baseline based off an example using Serde in a previous section:

//tutorial-perf-serde-01.rs
extern crate csv;
extern crate serde;
#[macro_use]
extern crate serde_derive;

use std::error::Error;
use std::io;
use std::process;

#[derive(Debug, Deserialize)]
#[serde(rename_all = "PascalCase")]
struct Record {
    country: String,
    city: String,
    accent_city: String,
    region: String,
    population: Option<u64>,
    latitude: f64,
    longitude: f64,
}

fn run() -> Result<u64, Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());

    let mut count = 0;
    for result in rdr.deserialize() {
        let record: Record = result?;
        if record.country == "us" && record.region == "MA" {
            count += 1;
        }
    }
    Ok(count)
}

fn main() {
    match run() {
        Ok(count) => {
            println!("{}", count);
        }
        Err(err) => {
            println!("{}", err);
            process::exit(1);
        }
    }
}

Now compile and run this program:

$ cargo build --release
$ ./target/release/csvtutor < worldcitiespop.csv
2176

real    0m1.381s
user    0m1.367s
sys     0m0.013s

The first thing you might notice is that this is quite a bit slower than our programs in the previous section. This is because deserializing each record has a certain amount of overhead to it. In particular, some of the fields need to be parsed as integers or floating point numbers, which isn’t free. However, there is hope yet, because we can speed up this program!

Our first attempt to speed up the program will be to amortize allocation. Doing this with Serde is a bit trickier than before, because we need to change our Record type and use the manual deserialization API. Let’s see what that looks like:

//tutorial-perf-serde-02.rs
#[derive(Debug, Deserialize)]
#[serde(rename_all = "PascalCase")]
struct Record<'a> {
    country: &'a str,
    city: &'a str,
    accent_city: &'a str,
    region: &'a str,
    population: Option<u64>,
    latitude: f64,
    longitude: f64,
}

fn run() -> Result<u64, Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());
    let mut raw_record = csv::StringRecord::new();
    let headers = rdr.headers()?.clone();

    let mut count = 0;
    while rdr.read_record(&mut raw_record)? {
        let record: Record = raw_record.deserialize(Some(&headers))?;
        if record.country == "us" && record.region == "MA" {
            count += 1;
        }
    }
    Ok(count)
}

Compile and run:

$ cargo build --release
$ ./target/release/csvtutor < worldcitiespop.csv
2176

real    0m1.055s
user    0m1.040s
sys     0m0.013s

This corresponds to an approximately 24% increase in performance. To achieve this, we had to make two important changes.

The first was to make our Record type contain &str fields instead of String fields. If you recall from a previous section, &str is a borrowed string where a String is an owned string. A borrowed string points to a already existing allocation where as a String always implies a new allocation. In this case, our &str is borrowing from the CSV record itself.

The second change we had to make was to stop using the Reader::deserialize iterator, and instead deserialize our record into a StringRecord explicitly and then use the StringRecord::deserialize method to deserialize a single record.

The second change is a bit tricky, because in order for it to work, our Record type needs to borrow from the data inside the StringRecord. That means that our Record value cannot outlive the StringRecord that it was created from. Since we overwrite the same StringRecord on each iteration (in order to amortize allocation), that means our Record value must evaporate before the next iteration of the loop. Indeed, the compiler will enforce this!

There is one more optimization we can make: remove UTF-8 validation. In general, this means using &[u8] instead of &str and ByteRecord instead of StringRecord:

//tutorial-perf-serde-03.rs
#[derive(Debug, Deserialize)]
#[serde(rename_all = "PascalCase")]
struct Record<'a> {
    country: &'a [u8],
    city: &'a [u8],
    accent_city: &'a [u8],
    region: &'a [u8],
    population: Option<u64>,
    latitude: f64,
    longitude: f64,
}

fn run() -> Result<u64, Box<Error>> {
    let mut rdr = csv::Reader::from_reader(io::stdin());
    let mut raw_record = csv::ByteRecord::new();
    let headers = rdr.byte_headers()?.clone();

    let mut count = 0;
    while rdr.read_byte_record(&mut raw_record)? {
        let record: Record = raw_record.deserialize(Some(&headers))?;
        if record.country == b"us" && record.region == b"MA" {
            count += 1;
        }
    }
    Ok(count)
}

Compile and run:

$ cargo build --release
$ ./target/release/csvtutor < worldcitiespop.csv
2176

real    0m0.873s
user    0m0.850s
sys     0m0.023s

This corresponds to a 17% increase over the previous example and a 37% increase over the first example.

In sum, Serde parsing is still quite fast, but will generally not be the fastest way to parse CSV since it necessarily needs to do more work.

CSV parsing without the standard library

In this section, we will explore a niche use case: parsing CSV without the standard library. While the csv crate itself requires the standard library, the underlying parser is actually part of the csv-core crate, which does not depend on the standard library. The downside of not depending on the standard library is that CSV parsing becomes a lot more inconvenient.

The csv-core crate is structured similarly to the csv crate. There is a Reader and a Writer, as well as corresponding builders ReaderBuilder and WriterBuilder. The csv-core crate has no record types or iterators. Instead, CSV data can either be read one field at a time or one record at a time. In this section, we’ll focus on reading a field at a time since it is simpler, but it is generally faster to read a record at a time since it does more work per function call.

In keeping with this section on performance, let’s write a program using only csv-core that counts the number of records in the state of Massachusetts.

(Note that we unfortunately use the standard library in this example even though csv-core doesn’t technically require it. We do this for convenient access to I/O, which would be harder without the standard library.)

//tutorial-perf-core-01.rs
extern crate csv_core;

use std::io::{self, Read};
use std::process;

use csv_core::{Reader, ReadFieldResult};

fn run(mut data: &[u8]) -> Option<u64> {
    let mut rdr = Reader::new();

    // Count the number of records in Massachusetts.
    let mut count = 0;
    // Indicates the current field index. Reset to 0 at start of each record.
    let mut fieldidx = 0;
    // True when the current record is in the United States.
    let mut inus = false;
    // Buffer for field data. Must be big enough to hold the largest field.
    let mut field = [0; 1024];
    loop {
        // Attempt to incrementally read the next CSV field.
        let (result, nread, nwrite) = rdr.read_field(data, &mut field);
        // nread is the number of bytes read from our input. We should never
        // pass those bytes to read_field again.
        data = &data[nread..];
        // nwrite is the number of bytes written to the output buffer `field`.
        // The contents of the buffer after this point is unspecified.
        let field = &field[..nwrite];

        match result {
            // We don't need to handle this case because we read all of the
            // data up front. If we were reading data incrementally, then this
            // would be a signal to read more.
            ReadFieldResult::InputEmpty => {}
            // If we get this case, then we found a field that contains more
            // than 1024 bytes. We keep this example simple and just fail.
            ReadFieldResult::OutputFull => {
                return None;
            }
            // This case happens when we've successfully read a field. If the
            // field is the last field in a record, then `record_end` is true.
            ReadFieldResult::Field { record_end } => {
                if fieldidx == 0 && field == b"us" {
                    inus = true;
                } else if inus && fieldidx == 3 && field == b"MA" {
                    count += 1;
                }
                if record_end {
                    fieldidx = 0;
                    inus = false;
                } else {
                    fieldidx += 1;
                }
            }
            // This case happens when the CSV reader has successfully exhausted
            // all input.
            ReadFieldResult::End => {
                break;
            }
        }
    }
    Some(count)
}

fn main() {
    // Read the entire contents of stdin up front.
    let mut data = vec![];
    if let Err(err) = io::stdin().read_to_end(&mut data) {
        println!("{}", err);
        process::exit(1);
    }
    match run(&data) {
        None => {
            println!("error: could not count records, buffer too small");
            process::exit(1);
        }
        Some(count) => {
            println!("{}", count);
        }
    }
}

And compile and run it:

$ cargo build --release
$ time ./target/release/csvtutor < worldcitiespop.csv
2176

real    0m0.572s
user    0m0.513s
sys     0m0.057s

This isn’t as fast as some of our previous examples where we used the csv crate to read into a StringRecord or a ByteRecord. This is mostly because this example reads a field at a time, which incurs more overhead than reading a record at a time. To fix this, you would want to use the Reader::read_record method instead, which is defined on csv_core::Reader.

The other thing to notice here is that the example is considerably longer than the other examples. This is because we need to do more book keeping to keep track of which field we’re reading and how much data we’ve already fed to the reader. There are basically two reasons to use the csv_core crate:

  1. If you’re in an environment where the standard library is not usable.
  2. If you wanted to build your own csv-like library, you could build it on top of csv-core.

Closing thoughts

Congratulations on making it to the end! It seems incredible that one could write so many words on something as basic as CSV parsing. I wanted this guide to be accessible not only to Rust beginners, but to inexperienced programmers as well. My hope is that the large number of examples will help push you in the right direction.

With that said, here are a few more things you might want to look at:

  • The API documentation for the csv crate documents all facets of the library, and is itself littered with even more examples.
  • The csv-index crate provides data structures that can index CSV data that are amenable to writing to disk. (This library is still a work in progress.)
  • The xsv command line tool is a high performance CSV swiss army knife. It can slice, select, search, sort, join, concatenate, index, format and compute statistics on arbitrary CSV data. Give it a try!