Sep 0.2.0-preview.3

This is a prerelease version of Sep.
There is a newer version of this package available.
See the version list below for details.
dotnet add package Sep --version 0.2.0-preview.3
NuGet\Install-Package Sep -Version 0.2.0-preview.3
This command is intended to be used within the Package Manager Console in Visual Studio, as it uses the NuGet module's version of Install-Package.
<PackageReference Include="Sep" Version="0.2.0-preview.3" />
For projects that support PackageReference, copy this XML node into the project file to reference the package.
paket add Sep --version 0.2.0-preview.3
#r "nuget: Sep, 0.2.0-preview.3"
#r directive can be used in F# Interactive and Polyglot Notebooks. Copy this into the interactive tool or source code of the script to reference the package.
// Install Sep as a Cake Addin
#addin nuget:?package=Sep&version=0.2.0-preview.3&prerelease

// Install Sep as a Cake Tool
#tool nuget:?package=Sep&version=0.2.0-preview.3&prerelease

Sep - Possibly the World's Fastest .NET CSV Parser

Build Status codecov NuGet

Modern, minimal, fast, zero allocation, reading and writing of separated values (csv, tsv etc.). Cross-platform, trimmable and AOT/NativeAOT compatible. Featuring an opinionated API design and pragmatic implementation targetted at machine learning use cases.

⭐ Please star this project if you like it. ⭐

  • 🌃 Modern - utilizes features such as Span<T>, Generic Math (ISpanParsable<T>/ISpanFormattable), ref struct, ArrayPool<T> and similar from .NET 7+ and C# 11+ for a modern and highly efficient implementation.
  • 🔎 Minimal - a succinct yet expressive API with few options and no hidden changes to input or output. What you read/write is what you get. This means there is no "automatic" escaping/unescaping of quotes, for example.
  • 🚀 Fast - blazing fast with both architecture specific and cross-platform SIMD vectorized parsing. Uses csFastFloat for fast parsing of floating points. Reads or writes one row at a time efficiently with detailed benchmarks to prove it.
  • 🗑️ Zero allocation - intelligent and efficient memory management allowing for zero allocations after warmup incl. supporting use cases of reading or writing arrays of values (e.g. features) easily without repeated allocations.
  • 🌐 Cross-platform - works on any platform, any architecture supported by .NET. 100% managed and written in beautiful modern C#.
  • ✂️ Trimmable and AOT/NativeAOT compatible - no problematic reflection or dynamic code generation. Hence, fully trimmable and Ahead-of-Time compatible. With a simple console tester program executable possible in just a few MBs. 💾
  • 🗣️ Opinionated and pragmatic - conforms to the essentials of RFC-4180, but takes an opinionated and pragmatic approach towards this especially with regards to quoting and line ends. See section RFC-4180.

Example | Naming and Terminology | API | Limitations and Constraints | Comparison Benchmarks | Example Catalogue | RFC-4180 | FAQ

Example

var text = """
           A;B;C;D;E;F
           Sep;🚀;1;1.2;0.1;0.5
           CSV;✅;2;2.2;0.2;1.5
           """;

using var reader = Sep.Reader().FromText(text);   // Infers separator 'Sep' from header
using var writer = reader.Spec.Writer().ToText(); // Writer defined from reader 'Spec'
                                                  // Use .FromFile(...)/ToFile(...) for files
var idx = reader.Header.IndexOf("B");
var nms = new[] { "E", "F" };

foreach (var readRow in reader)           // Read one row at a time
{
    var a = readRow["A"].Span;            // Column as ReadOnlySpan<char>
    var b = readRow[idx].ToString();      // Column to string (might be pooled)
    var c = readRow["C"].Parse<int>();    // Parse any T : ISpanParsable<T>
    var d = readRow["D"].Parse<float>();  // Parse float/double fast via csFastFloat
    var s = readRow[nms].Parse<double>(); // Parse multiple columns as Span<T>
                                          // - Sep handles array allocation and reuse
    foreach (ref var v in s) { v *= 10; }

    using var writeRow = writer.NewRow(); // Start new row. Row written on Dispose.
    writeRow["A"].Set(a);                 // Set by ReadOnlySpan<char>
    writeRow["B"].Set(b);                 // Set by string
    writeRow["C"].Set($"{c * 2}");        // Set via InterpolatedStringHandler, no allocs
    writeRow["D"].Format(d / 2);          // Format any T : ISpanFormattable
    writeRow[nms].Format(s);              // Format multiple columns directly
    // Columns are added on first access as ordered, header written when first row written
}

var expected = """
               A;B;C;D;E;F
               Sep;🚀;2;0.6;1;5
               CSV;✅;4;1.1;2;15
               
               """;                       // Empty line at end is for line ending,
                                          // which is always written.
Assert.AreEqual(expected, writer.ToString());

// Above example code is for demonstration purposes only.
// Short names and repeated constants are only for demonstration.

Naming and Terminology

Sep uses naming and terminology that is not based on RFC-4180, but is more tailored to usage in machine learning or similar. Additionally, Sep takes a pragmatic approach towards names by using short names and abbreviations where it makes sense and there should be no ambiguity given the context. That is, using Sep for Separator and Col for Column to keep code succinct.

Term Description
Sep Short for separator, also called delimiter. E.g. comma (,) is the separator for the separated values in a csv-file.
Header Optional first row defining names of columns.
Row A row is a collection of col(umn)s, which may span multiple lines. Also called record.
Col Short for column, also called field.
Line Horizontal set of characters until a line ending; \r\n, \r, \n.
Index 0-based that is RowIndex will be 0 for first row (or the header if present).
Number 1-based that is LineNumber will be 1 for the first line (as in notepad). Given a row may span multiple lines a row can have a From line number and an ToExcl line number matching the C# range indexing syntax [LineNumberFrom..LineNumberToExcl].

Application Programming Interface (API)

Besides being the succinct name of the library, Sep is both the main entry point to using the library and the container for a validated separator. That is, Sep is basically defined as:

public readonly record struct Sep(char Separator);

The separator char is validated upon construction and is guaranteed to be within a limited range and not being a char like " (quote) or similar. This can be seen in src/Sep/Sep.cs. The separator is constrained also for internal optimizations, so you cannot use any char as a separator.

⚠ Note that all types are within the namespace nietras.SeparatedValues and not Sep since it is problematic to have a type and a namespace with the same name.

To get started you can use Sep as the static entry point to building either a reader or writer. That is, for SepReader:

using var reader = Sep.Reader().FromFile("titanic.csv");

where .Reader() is a convenience method corresponding to:

using var reader = Sep.Auto.Reader().FromFile("titanic.csv");

where Sep? Auto => null; is a static property that returns null for a nullable Sep to signify that the separator should be inferred from the first row, which might be a header. If the first row does not contain any of the by default supported separators or there are no rows, the default separator will be used.

⚠ Note Sep uses ; as the default separator, since this is what was used in an internal proprietary library which Sep was built to replace. This is also to avoid issues with comma , being used as a decimal separator in some locales. Without having to resort to quoting.

If you want to specify the separator you can write:

using var reader = Sep.New(',').Reader().FromFile("titanic.csv");

or

var sep = new Sep(',');
using var reader = sep.Reader().FromFile("titanic.csv");

Similarly, for SepWriter:

using var writer = Sep.Writer().ToFile("titanic.csv");

or

using var writer = Sep.New(',').Writer().ToFile("titanic.csv");

where you have to specify a valid separator, since it cannot be inferred. To fascillitate easy flow of the separator and CultureInfo both SepReader and SepWriter expose a Spec property of type SepSpec that simply defines those two. This means you can write:

using var reader = Sep.Reader().FromFile("titanic.csv");
using var writer = reader.Spec.Writer().ToFile("titanic-survivors.csv");

where the writer then will use the separator inferred by the reader, for example.

API Pattern

In general, both reading and writing follow a similar pattern:

Sep/Spec => SepReaderOptions => SepReader => Row => Col(s) => Span/ToString/Parse
Sep/Spec => SepWriterOptions => SepWriter => Row => Col(s) => Set/Format

where each continuation flows fluently from the preceding type. For example, Reader() is an extension method to Sep or SepSpec that returns a SepReaderOptions. Similarly, Writer() is an extension method to Sep or SepSpec that returns a SepWriterOptions.

SepReaderOptions and SepWriterOptions are optionally configurable. That and the APIs for reader and writer is covered in the following sections.

For a complete example, see the example above or the ReadMeTest.cs.

⚠ Note that it is important to understand that Sep Row/Col/Cols are ref structs (please follow the ref struct link and understand how this limits the usage of those). This is due to these types being simple facades or indirections to the underlying reader or writer. That means you cannot use LINQ or create an array of all rows like reader.ToArray() as the reader is not IEnumerable<> either since ref structs cannot be used in interfaces, which is in fact the point. Hence, you need to parse or copy to different types instead. The same applies to Col/Cols which point to internal state that is also reused. This is to avoid repeated allocations for each row and get the best possible performance, while still defining a well structured and straightforward API that guides users to relevant functionality. See Why SepReader Is Not IEnumerable and LINQ Compatible for more.

SepReader API

SepReader API has the following structure (in pseudo-C# code):

using var reader = Sep.Reader(o => o).FromFile/FromText/From...;
var header = reader.Header;
var _ = header.IndexOf/IndicesOf/NamesStartingWith...;
foreach (var row in reader)
{
    var _ = row[colName/colNames].Span/ToString/Parse<T>...;
    var _ = row[colIndex/colIndices].Span/ToString/Parse<T>...;
}

That is, to use SepReader follow the points below:

  1. Optionally define Sep or use default automatically inferred separator.
  2. Specify reader with optional configuration of SepReaderOptions. For example, if a csv-file does not have a header this can be configured via:
    Sep.Reader(o => o with { HasHeader = false })
    
    For all options consult the properties on the options type.
  3. Specify source e.g. file, text (string), TextWriter, etc. via From extension methods.
  4. Optionally access the header. For example, to get all columns starting with GT_ use:
    var colNames = header.NamesStarting("GT_");
    var colIndices = header.IndicesOf(colNames);
    
  5. Enumerate rows. One row at a time.
  6. Access a column by name or index. Or access multiple columns with names and indices. Sep internally handles pooled allocation and reuse of arrays for multiple columns.
  7. Use Span to access the column directly as a ReadOnlySpan<char>. Or use ToString to convert to a string. Or use Parse<T> where T : ISpanParsable<T> to parse the column chars to a specific type.
Why SepReader Is Not IEnumerable and LINQ Compatible

As mentioned earlier Sep only allows enumeration and access to one row at a time and SepReader.Row is just a simple facade or indirection to the underlying reader. This is why it is defined as a ref struct. In fact, the following code:

using var reader = Sep.Reader().FromText(text);
foreach (var row in reader)
{ }

can also be rewritten as:

using var reader = Sep.Reader().FromText(text);
while (reader.MoveNext())
{
    var row = reader.Current;
}

where row is just a facade for exposing row specific functionality. That is, row is still basically the reader underneath. Hence, let's imagine if SepReader did implement IEnumerable<SepReader.Row> and the Row was not a ref struct. Then, you would be able to write something like below:

using var reader = Sep.Reader().FromText(text);
SepReader.Row[] rows = reader.ToArray();

Given Row is just a facade for the reader, this would be equivalent to writing:

using var reader = Sep.Reader().FromText(text);
SepReader[] rows = reader.ToArray();

which hopefully makes it clear why this is not a good thing. The array would effectively be the reader repeated several times. If this would have to be supported one would have to allocate memory for each row always, which would basically be no different than a ReadLine approach as benchmarked in Comparison Benchmarks.

This is perhaps also the reason why no other efficient .NET CSV parser (known to author) implements an API pattern like Sep, but instead let the reader define all functionality directly and hence only let's you access the current row and cols on that. This API, however, is in this authors opinion not ideal and can be a bit confusing, which is why Sep is designed like it is. The downside is the above caveat.

If you want to use LINQ or similar you have to first parse or transform the rows into some other type and enumerate it. This is easy to do and instead of counting lines you should focus on how such enumeration can be easily expressed using C# iterators (aka yield return). With local functions this can be done inside a method like:

var text = """
           Key;Value
           A;1.1
           B;2.2
           """;
var expected = new (string Key, double Value)[] {
    ("A", 1.1),
    ("B", 2.2),
};

using var reader = Sep.Reader().FromText(text);
var actual = Enumerate(reader).ToArray();

CollectionAssert.AreEqual(expected, actual);

static IEnumerable<(string Key, double Value)> Enumerate(SepReader reader)
{
    foreach (var row in reader)
    {
        yield return (row["Key"].ToString(), row["Value"].Parse<double>());
    }
}

Now if instead refactoring this to something LINQ-compatible by defining a common Enumerate or similar method it could be:

var text = """
           Key;Value
           A;1.1
           B;2.2
           """;
var expected = new (string Key, double Value)[] {
    ("A", 1.1),
    ("B", 2.2),
};

using var reader = Sep.Reader().FromText(text);
var actual = Enumerate(reader,
    row => (row["Key"].ToString(), row["Value"].Parse<double>()))
    .ToArray();

CollectionAssert.AreEqual(expected, actual);

static IEnumerable<T> Enumerate<T>(SepReader reader, SepReader.RowFunc<T> func)
{
    foreach (var row in reader)
    {
        yield return func(row);
    }
}

Which discounting the Enumerate method (which could naturally be an extension method), does have less boilerplate, but not really more effective lines of code. The issue here is that this tends to favor factoring code in a way that can become very inefficient quickly. Consider if one wanted to only enumerate rows matching a predicate on Key which meant only 1% of rows were to be enumerated e.g.:

var text = """
           Key;Value
           A;1.1
           B;2.2
           """;
var expected = new (string Key, double Value)[] {
    ("B", 2.2),
};

using var reader = Sep.Reader().FromText(text);
var actual = Enumerate(reader,
    row => (row["Key"].ToString(), row["Value"].Parse<double>()))
    .Where(kv => kv.Item1.StartsWith("B", StringComparison.Ordinal))
    .ToArray();

CollectionAssert.AreEqual(expected, actual);

static IEnumerable<T> Enumerate<T>(SepReader reader, SepReader.RowFunc<T> func)
{
    foreach (var row in reader)
    {
        yield return func(row);
    }
}

This means you are still parsing the double (which is magnitudes slower than getting just the key) for all rows. Imagine if this was an array of floating points or similar. Not only would you then be parsing a lot of values you would also be allocated 99x arrays that aren't used after filtering with Where.

Instead, you should focus on how to express the enumeration in a way that is both efficient and easy to read. For example, the above could be rewritten as:

var text = """
           Key;Value
           A;1.1
           B;2.2
           """;
var expected = new (string Key, double Value)[] {
    ("B", 2.2),
};

using var reader = Sep.Reader().FromText(text);
var actual = Enumerate(reader).ToArray();

CollectionAssert.AreEqual(expected, actual);

static IEnumerable<(string Key, double Value)> Enumerate(SepReader reader)
{
    foreach (var row in reader)
    {
        var keyCol = row["Key"];
        if (keyCol.Span.StartsWith("B"))
        {
            yield return (keyCol.ToString(), row["Value"].Parse<double>());
        }
    }
}

This does not take significantly longer to write and is a lot more efficient (also avoids allocating a string for key for each row) and is easier to debug and perhaps even read. All examples above can be seen in ReadMeTest.cs.

SepWriter API

SepWriter API has the following structure (in pseudo-C# code):

using var writer = Sep.Writer(o => o).ToFile/ToText/To...;
foreach (var data in EnumerateData())
{
    using var row = writer.NewRow();
    var _ = row[colName/colNames].Set/Format<T>...;
    var _ = row[colIndex/colIndices].Set/Format<T>...;
}

That is, to use SepWriter follow the points below:

  1. Optionally define Sep or use default automatically inferred separator.
  2. Specify writer with optional configuration of SepWriterOptions. For all options consult the properties on the options type.
  3. Specify destination e.g. file, text (string via StringWriter), TextWriter, etc. via To extension methods.
  4. MISSING: SepWriter currently does not allow you to define the header up front. Instead, header is defined by the order in which column names are accessed/created when defining the row.
  5. Define new rows with NewRow. ⚠ Be sure to dispose any new rows before starting the next! For convenience Sep provides an overload for NewRow that takes a SepReader.Row and copies the columns from that row to the new row:
    using var reader = Sep.Reader().FromText(text);
    using var writer = reader.Spec.Writer().ToText();
    foreach (var readRow in reader)
    {   using var writeRow = writer.NewRow(readRow); }
    
  6. Create a column by selecting by name or index. Or multiple columns via indices and names. Sep internally handles pooled allocation and reuse of arrays for multiple columns.
  7. Use Set to set the column value either as a ReadOnlySpan<char>, string or via an interpolated string. Or use Format<T> where T : IFormattable to format T to the column value.
  8. Row is written when Dispose is called on the row.

    Note this is to allow a row to be defined flexibly with both column removal, moves and renames in the future. This is not yet supported.

Limitations and Constraints

Sep is designed to be minimal and fast. As such, it has some limitations and constraints, since these are not needed for the initial intended usage:

  • Automatic escaping and unescaping quotes is not supported. Use Trim extension method to remove surrounding quotes, for example.
  • Comments # are not directly supported. You can skip a row by:
    foreach (var row in reader)
    {
         // Skip row if starts with #
         if (!row.Span.StartsWith("#"))
         {
              // ...
         }
    }
    
    This does not allow skipping a header row starting with # though.
  • SepWriter is not yet fully featured and one cannot skip writing a header currently.

Comparison Benchmarks

To investigate the performance of Sep it is compared to:

  • CsvHelper - the most commonly used CSV library with a staggering downloads downloads on NuGet. Fully featured and battle tested.
  • Sylvan - is well-known and has previously been shown to be the fastest CSV libraries for parsing (Sep changes that 😉).
  • ReadLine/WriteLine - basic naive implementations that read line by line and split on separator. While writing columns, separators and line endings directly. Does not handle quotes or similar correctly.

All benchmarks are run from/to memory either with:

  • StringReader or StreamReader + MemoryStream
  • StringWriter or StreamWriter + MemoryStream

This to avoid confounding factors from reading from or writing to disk.

When using StringReader/StringWriter each char counts as 2 bytes, when measuring throughput e.g. MB/s. When using StreamReader/StreamWriter content is UTF-8 encoded and each char typically counts as 1 byte, as content usually limited to 1 byte per char in UTF-8. Note that in .NET for TextReader and TextWriter data is converted to/from char, but for reading such conversion can often be just as fast as Memmove.

By default only StringReader/StringWriter results are shown, if a result is based on StreamReader/StreamWriter it will be called out. Usually, results for StreamReader/StreamWriter are in line with StringReader/StringWriter but with half the throughput due to 1 byte vs 2 bytes. For brevity they are not shown here.

For all benchmark results, Sep has been defined as the Baseline in BenchmarkDotNet. This means Ratio will be 1.00 for Sep. For the others Ratio will then show how many times faster Sep is than that. Or how many times more bytes are allocated in Alloc Ratio.

Disclaimer: Any comparison made is based on a number of preconditions and assumptions. Sep is a new library written from the ground up to use the latest and greatest features in .NET. CsvHelper has a long history and has to take into account backwards compatibility and still supporting older runtimes, so may not be able to easily utilize more recent features. Same goes for Sylvan. Additionally, Sep has a different feature set compared to the two. Performance is a feature, but not the only feature. Keep that in mind when evaluating results.

Runtime and Platforms

The following runtime is used for benchmarking:

  • NET 7.0.5 (7.0.523.17405)

The following platforms are used for benchmarking:

  • AMD 5950X X64 Platform Information
    OS=Windows 10 (10.0.19044.2846/21H2/November2021Update)
    AMD Ryzen 9 5950X, 1 CPU, 32 logical and 16 physical cores
    
  • Neoverse N1 ARM64 Platform Information (cloud instance)
    OS=ubuntu 22.04
    Neoverse N1, ARM, 4 vCPU
    

Reader Comparison Benchmarks

The following reader scenarios are benchmarked:

Details for each can be found in the following. However, for each of these 3 different scopes are benchmarked to better assertain the low-level performance of each library and approach and what parts of the parsing consume the most time:

  • Row - for this scope only the row is enumerated. That is, for Sep all that is done is:
    foreach (var row in reader) { }
    
    this should capture parsing both row and columns but without accessing these. Note that some libraries (like Sylvan) will defer work for columns to when these are accessed.
  • Cols - for this scope all rows and all columns are enumerated. If possible columns are accessed as spans, if not as strings, which then might mean a string has to be allocated. That is, for Sep this is:
    foreach (var row in reader)
    {
        for (var i = 0; i < row.ColCount; i++)
        {
            var span = row[i].Span;
        }
    }
    
  • XYZ - finally the full scope is performed which is specific to each of the scenarios.
NCsvPerf PackageAssets Reader Comparison Benchmarks

NCsvPerf from The fastest CSV parser in .NET is a benchmark which in Joel Verhagen own words was defined with:

My goal was to find the fastest low-level CSV parser. Essentially, all I wanted was a library that gave me a string[] for each line where each field in the line was an element in the array.

What is great about this work is it tests a whole of 35 different libraries and approaches to this. Providing a great overview of those and their performance on this specific scenario. Given Sylvan is the fastest of those it is used as the one to beat here, while CsvHelper is used to compare to the most commonly used library.

The source used for this benchmark PackageAssetsBench.cs is a PackageAssets.csv with NuGet package information in 25 columns with rows like:

75fcf875-017d-4579-bfd9-791d3e6767f0,2020-11-28T01:50:41.2449947+00:00,Akinzekeel.BlazorGrid,0.9.1-preview,2020-11-27T22:42:54.3100000+00:00,AvailableAssets,RuntimeAssemblies,,,net5.0,,,,,,lib/net5.0/BlazorGrid.dll,BlazorGrid.dll,.dll,lib,net5.0,.NETCoreApp,5.0.0.0,,,0.0.0.0
75fcf875-017d-4579-bfd9-791d3e6767f0,2020-11-28T01:50:41.2449947+00:00,Akinzekeel.BlazorGrid,0.9.1-preview,2020-11-27T22:42:54.3100000+00:00,AvailableAssets,CompileLibAssemblies,,,net5.0,,,,,,lib/net5.0/BlazorGrid.dll,BlazorGrid.dll,.dll,lib,net5.0,.NETCoreApp,5.0.0.0,,,0.0.0.0
75fcf875-017d-4579-bfd9-791d3e6767f0,2020-11-28T01:50:41.2449947+00:00,Akinzekeel.BlazorGrid,0.9.1-preview,2020-11-27T22:42:54.3100000+00:00,AvailableAssets,ResourceAssemblies,,,net5.0,,,,,,lib/net5.0/de/BlazorGrid.resources.dll,BlazorGrid.resources.dll,.dll,lib,net5.0,.NETCoreApp,5.0.0.0,,,0.0.0.0
75fcf875-017d-4579-bfd9-791d3e6767f0,2020-11-28T01:50:41.2449947+00:00,Akinzekeel.BlazorGrid,0.9.1-preview,2020-11-27T22:42:54.3100000+00:00,AvailableAssets,MSBuildFiles,,,any,,,,,,build/Microsoft.AspNetCore.StaticWebAssets.props,Microsoft.AspNetCore.StaticWebAssets.props,.props,build,any,Any,0.0.0.0,,,0.0.0.0
75fcf875-017d-4579-bfd9-791d3e6767f0,2020-11-28T01:50:41.2449947+00:00,Akinzekeel.BlazorGrid,0.9.1-preview,2020-11-27T22:42:54.3100000+00:00,AvailableAssets,MSBuildFiles,,,any,,,,,,build/Akinzekeel.BlazorGrid.props,Akinzekeel.BlazorGrid.props,.props,build,any,Any,0.0.0.0,,,0.0.0.0

For Scope = Asset the columns are parsed into a PackageAsset class, which consists of 25 properties of which 22 are strings. Each asset is accumulated into a List<PackageAsset>. Each column is accessed as a string regardless.

This means this benchmark is dominated by turning columns into strings for the decently fast parsers. Hence, the fastest libraries in this test employ string pooling. That is, basically a custom dictionary from ReadOnlySpan<char> to string, which avoids allocating a new string for repeated values. And as can be seen in the csv-file there are a lot of repeated values. Both Sylvan and CsvHelper do this in the benchmark. So does Sep and as with Sep this is an optional configuration that has to be explicitly enable. For Sep this means the reader is created with something like:

using var reader = Sep.Reader(o => o with
{
    HasHeader = false,
    CreateToString = SepToString.PoolPerCol(maximumStringLength: 128),
})
.From(CreateReader());

What is unique for Sep is that it allows defining a pool per column e.g. via SepToString.PoolPerCol(...). This is based on the fact that often each column has its own set of values or strings that may be repeated without any overlap to other columns. This also allows one to define per column specific handling of ToString behavior. Whether to pool or not. Or even to use a statically defined pool.

PackageAssets Benchmark Results

The results below show Sep is now the fastest .NET CSV Parser (for this benchmark on these platforms and machines 😀). While for pure parsing allocating only a fraction of the memory due to extensive use of pooling and the ArrayPool<T>.

This is in many aspects due to Sep having extremely optimized string pooling and optimized hashing of ReadOnlySpan<char>, and thus not really due the the csv-parsing itself, since that is not a big part of the time consumed. At least not for a decently fast csv-parser.

AMD 5950X - PackageAssets Benchmark Results (Sep 0.2.0, Sylvan 1.3.2, CsvHelper 30.0.1)
Method Scope Rows Mean Ratio MB MB/s ns/row Allocated Alloc Ratio
Sep______ Row 1000000 57.28 ms 1.00 583 10191.6 57.3 1.33 KB 1.00
Sylvan___ Row 1000000 70.99 ms 1.24 583 8223.5 71.0 7.33 KB 5.51
ReadLine_ Row 1000000 244.88 ms 4.28 583 2384.0 244.9 1772445.54 KB 1,332,587.55
CsvHelper Row 1000000 1,046.25 ms 18.25 583 558.0 1046.3 20.65 KB 15.53
Sep______ Cols 1000000 74.69 ms 1.00 583 7815.7 74.7 1.98 KB 1.00
Sylvan___ Cols 1000000 127.64 ms 1.71 583 4573.7 127.6 7.84 KB 3.95
ReadLine_ Cols 1000000 255.66 ms 3.42 583 2283.5 255.7 1772445.91 KB 893,201.09
CsvHelper Cols 1000000 1,516.77 ms 20.31 583 384.9 1516.8 446.74 KB 225.13
Sep______ Asset 1000000 720.17 ms 1.00 583 810.6 720.2 266666.69 KB 1.00
Sylvan___ Asset 1000000 900.52 ms 1.25 583 648.3 900.5 266890.6 KB 1.00
ReadLine_ Asset 1000000 1,991.64 ms 2.77 583 293.1 1991.6 2038832.79 KB 7.65
CsvHelper Asset 1000000 1,962.81 ms 2.72 583 297.4 1962.8 266834.63 KB 1.00
Neoverse N1 - PackageAssets Benchmark Results (Sep 0.2.0, Sylvan 1.3.2, CsvHelper 30.0.1)
Method Scope Rows Mean Ratio MB MB/s ns/row Allocated Alloc Ratio
Sep______ Row 1000000 237.7 ms 1.00 581 2448.0 237.7 1.35 KB 1.00
Sylvan___ Row 1000000 691.9 ms 2.91 581 841.0 691.9 6.25 KB 4.63
ReadLine_ Row 1000000 817.9 ms 3.44 581 711.4 817.9 1772445.63 KB 1,313,302.69
CsvHelper Row 1000000 2,176.4 ms 9.16 581 267.4 2176.4 20.74 KB 15.36
Sep______ Cols 1000000 296.4 ms 1.00 581 1963.2 296.4 2.25 KB 1.00
Sylvan___ Cols 1000000 823.7 ms 2.78 581 706.4 823.7 7.01 KB 3.11
ReadLine_ Cols 1000000 861.6 ms 2.92 581 675.4 861.6 1772446.24 KB 787,753.89
CsvHelper Cols 1000000 3,132.2 ms 10.56 581 185.8 3132.2 447.07 KB 198.70
Sep______ Asset 1000000 1,367.3 ms 1.00 581 425.6 1367.3 266667.05 KB 1.00
Sylvan___ Asset 1000000 2,102.3 ms 1.54 581 276.8 2102.3 266893.73 KB 1.00
ReadLine_ Asset 1000000 3,095.6 ms 2.26 581 188.0 3095.6 2038835.03 KB 7.65
CsvHelper Asset 1000000 3,767.4 ms 2.76 581 154.4 3767.4 266847.42 KB 1.00
PackageAssets with Quotes Benchmark Results

NCsvPerf does not examine performance in the face of quotes in the csv. This is relevant since some libraries like Sylvan will revert to a slower (not SIMD vectorized) parsing code path if it encounters quotes. Sep was designed to always use SIMD vectorization no matter what.

Since there are two extra chars to handle per column, it does have a significant impact on performance, no matter what though. This is expected when looking at the numbers. For each row of 25 columns, there are 24 separators (here ,) and one set of line endings (here \r\n). That's 26 characters. Adding quotes around each of the 25 columns will add 50 characters or almost triple the total to 76.

AMD 5950X - PackageAssets with Quotes Benchmark Results (Sep 0.2.0, Sylvan 1.3.2, CsvHelper 30.0.1)
Method Scope Rows Mean Ratio MB MB/s ns/row Allocated Alloc Ratio
Sep______ Row 1000000 142.9 ms 1.00 667 4672.2 142.9 1.33 KB 1.00
Sylvan___ Row 1000000 418.2 ms 2.93 667 1596.6 418.2 7.33 KB 5.51
ReadLine_ Row 1000000 304.0 ms 2.12 667 2196.5 304.0 2175928.72 KB 1,635,940.53
CsvHelper Row 1000000 1,340.8 ms 9.38 667 498.0 1340.8 20.65 KB 15.53
Sep______ Cols 1000000 159.1 ms 1.00 667 4196.7 159.1 1.98 KB 1.00
Sylvan___ Cols 1000000 487.4 ms 3.06 667 1370.0 487.4 7.84 KB 3.95
ReadLine_ Cols 1000000 300.8 ms 1.89 667 2219.6 300.8 2175929.09 KB 1,096,531.19
CsvHelper Cols 1000000 1,928.4 ms 12.11 667 346.2 1928.4 446.74 KB 225.13
Sep______ Asset 1000000 850.8 ms 1.00 667 784.8 850.8 266720.57 KB 1.00
Sylvan___ Asset 1000000 1,260.3 ms 1.48 667 529.8 1260.3 266890.98 KB 1.00
ReadLine_ Asset 1000000 2,613.8 ms 3.08 667 255.5 2613.8 2442315.96 KB 9.16
CsvHelper Asset 1000000 2,206.5 ms 2.59 667 302.6 2206.5 266839.8 KB 1.00
Neoverse N1 - PackageAssets with Quotes Benchmark Results (Sep 0.2.0, Sylvan 1.3.2, CsvHelper 30.0.1)
Method Scope Rows Mean Ratio MB MB/s ns/row Allocated Alloc Ratio
Sep______ Row 1000000 495.1 ms 1.00 665 1344.7 495.1 1.35 KB 1.00
Sylvan___ Row 1000000 912.3 ms 1.84 665 729.8 912.3 6.25 KB 4.63
ReadLine_ Row 1000000 1,001.4 ms 2.02 665 664.8 1001.4 2175928.8 KB 1,612,265.62
CsvHelper Row 1000000 2,417.3 ms 4.88 665 275.4 2417.3 20.74 KB 15.36
Sep______ Cols 1000000 555.1 ms 1.00 665 1199.4 555.1 2.25 KB 1.00
Sylvan___ Cols 1000000 1,059.2 ms 1.91 665 628.6 1059.2 7.01 KB 3.11
ReadLine_ Cols 1000000 1,024.0 ms 1.85 665 650.2 1024.0 2175929.41 KB 967,079.74
CsvHelper Cols 1000000 3,564.7 ms 6.43 665 186.8 3564.7 447.07 KB 198.70
Sep______ Asset 1000000 1,719.5 ms 1.00 665 387.2 1719.5 266718.38 KB 1.00
Sylvan___ Asset 1000000 2,392.8 ms 1.39 665 278.2 2392.8 266894.75 KB 1.00
ReadLine_ Asset 1000000 4,013.3 ms 2.34 665 165.9 4013.3 2442319.52 KB 9.16
CsvHelper Asset 1000000 4,188.7 ms 2.43 665 159.0 4188.7 266850.03 KB 1.00
Floats Reader Comparison Benchmarks

The FloatsReaderBench.cs benchmark demonstrates what Sep is built for. Namely parsing 32-bit floating points or features as in machine learning. Here a simple CSV-file is randomly generated with N ground truth values, N predicted result values and some typical extra columns leading that, but which aren't used as such in the benchmark. N = 20 here. For example:

Set;FileName;DataSplit;GT_Feature0;GT_Feature1;GT_Feature2;GT_Feature3;GT_Feature4;GT_Feature5;GT_Feature6;GT_Feature7;GT_Feature8;GT_Feature9;GT_Feature10;GT_Feature11;GT_Feature12;GT_Feature13;GT_Feature14;GT_Feature15;GT_Feature16;GT_Feature17;GT_Feature18;GT_Feature19;RE_Feature0;RE_Feature1;RE_Feature2;RE_Feature3;RE_Feature4;RE_Feature5;RE_Feature6;RE_Feature7;RE_Feature8;RE_Feature9;RE_Feature10;RE_Feature11;RE_Feature12;RE_Feature13;RE_Feature14;RE_Feature15;RE_Feature16;RE_Feature17;RE_Feature18;RE_Feature19
SetCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCCC;wwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwwww.png;Train;0.52276427;0.16843422;0.26259267;0.7244084;0.51292276;0.17365117;0.76125056;0.23458846;0.2573214;0.50560355;0.3202332;0.3809696;0.26024464;0.5174511;0.035318818;0.8141374;0.57719684;0.3974705;0.15219308;0.09011261;0.70515215;0.81618196;0.5399706;0.044147138;0.7111546;0.14776127;0.90621275;0.6925897;0.5164137;0.18637845;0.041509967;0.30819967;0.5831603;0.8210651;0.003954861;0.535722;0.8051845;0.7483589;0.3845737;0.14911908
SetAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA;mmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmmm.png;Test;0.6264564;0.11517637;0.24996082;0.77242833;0.2896067;0.6481459;0.14364648;0.044498358;0.6045593;0.51591337;0.050794687;0.42036617;0.7065823;0.6284636;0.21844554;0.013253775;0.36516154;0.2674384;0.06866083;0.71817476;0.07094294;0.46409357;0.012033525;0.7978093;0.43917948;0.5134962;0.4995968;0.008952909;0.82883793;0.012896823;0.0030740085;0.063773096;0.6541431;0.034539033;0.9135142;0.92897075;0.46119377;0.37533295;0.61660606;0.044443816
SetBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB;lllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllllll.png;Validation;0.7922863;0.5323656;0.400699;0.29737252;0.9072584;0.58673894;0.73510516;0.019412167;0.88168067;0.9576787;0.33283427;0.7107;0.1623628;0.10314285;0.4521515;0.33324885;0.7761104;0.14854911;0.13469358;0.21566042;0.59166247;0.5128394;0.98702157;0.766223;0.67204326;0.7149494;0.2894748;0.55206;0.9898286;0.65083236;0.02421702;0.34540752;0.92906284;0.027142895;0.21974725;0.26544374;0.03848049;0.2161237;0.59233844;0.42221397
SetAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA;ssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssssss.png;Train;0.10609442;0.32130885;0.32383907;0.7511514;0.8258279;0.00904226;0.0420841;0.84049565;0.8958947;0.23807365;0.92621964;0.8452882;0.2794469;0.545344;0.63447595;0.62532926;0.19230893;0.29726416;0.18304513;0.029583583;0.23084833;0.93346167;0.98742676;0.78163713;0.13521992;0.8833956;0.18670778;0.29476836;0.5599867;0.5562107;0.7124796;0.121927656;0.5981778;0.39144602;0.88092715;0.4449142;0.34820423;0.96379805;0.46364686;0.54301775

For Scope=Floats the benchmark will parse the features as two spans of floats; one for ground truth values and one for predicted result values. Then calculates the mean squared error (MSE) of those as an example. For Sep this code is succinct and still incredibly efficient:

using var reader = Sep.Reader().From(Reader.CreateReader());

var groundTruthColNames = reader.Header.NamesStartingWith("GT_");
var resultColNames = groundTruthColNames.Select(n =>
    n.Replace("GT_", "RE_", StringComparison.Ordinal))
    .ToArray();

var sum = 0.0;
var count = 0;
foreach (var row in reader)
{
    var gts = row[groundTruthColNames].Parse<float>();
    var res = row[resultColNames].Parse<float>();

    sum += MeanSquaredError(gts, res);
    ++count;
}
return sum / count;

Note how one can access and parse multiple columns easily while there are no repeated allocations for the parsed floating points. Sep internally handles a pool of arrays for handling multiple columns and returns spans for them.

The benchmark is based on an assumption of accessing columns by name per row. Ideally, one would look up the indices of the columns by name before enumerating rows, but this is a repeated nuisance to have to handle and Sep was built to avoid this. Hence, the comparison is based on looking up by name for each, even if this ends up adding a bit more code in the benchmark for other approaches.

As can be seen below, the actual low level parsing of the separated values is a tiny part of the total runtime for Sep for which the run time is dominated by parsing the floating points. Since Sep uses csFastFloat for an integrated fast floating point parser, it is >2x faster than Sylvan for example. If using Sylvan one may consider using csFastFloat if that is an option.

CsvHelper suffers from the fact that one can only access the column as a string so this has to be allocated for each column (ReadLine by definition always allocates a string per column). Still CsvHelper is significantly slower than the naive ReadLine approach. With Sep being >3.8x faster than CsvHelper.

It is a testament to how good the .NET and the .NET GC is that the ReadLine is pretty good compared to CsvHelper regardless of allocating a lot of strings.

AMD 5950X - Floats Benchmark Results (Sep 0.2.0, Sylvan 1.3.2, CsvHelper 30.0.1)
Method Scope Rows Mean Ratio MB MB/s ns/row Allocated Alloc Ratio
Sep______ Row 100000 11.70 ms 1.00 109 9319.2 117.0 1.66 KB 1.00
Sylvan___ Row 100000 13.49 ms 1.15 109 8082.5 134.9 10.64 KB 6.40
ReadLine_ Row 100000 51.94 ms 4.44 109 2098.9 519.4 359865.39 KB 216,511.25
CsvHelper Row 100000 158.64 ms 13.57 109 687.2 1586.4 20.61 KB 12.40
Sep______ Cols 100000 13.42 ms 1.00 109 8125.2 134.2 1.66 KB 1.00
Sylvan___ Cols 100000 21.08 ms 1.57 109 5171.7 210.8 10.64 KB 6.40
ReadLine_ Cols 100000 53.67 ms 4.00 109 2031.1 536.7 359865.39 KB 216,511.25
CsvHelper Cols 100000 173.19 ms 12.86 109 629.4 1731.9 113699.75 KB 68,406.90
Sep______ Floats 100000 136.57 ms 1.00 109 798.2 1365.7 8.87 KB 1.00
Sylvan___ Floats 100000 287.05 ms 2.10 109 379.8 2870.5 18.71 KB 2.11
ReadLine_ Floats 100000 324.53 ms 2.37 109 335.9 3245.3 359871.81 KB 40,553.40
CsvHelper Floats 100000 528.88 ms 3.87 109 206.1 5288.8 87694.14 KB 9,882.12
Neoverse N1 - Floats Benchmark Results (Sep 0.2.0, Sylvan 1.3.2, CsvHelper 30.0.1)
Method Scope Rows Mean Ratio MB MB/s ns/row Allocated Alloc Ratio
Sep______ Row 100000 45.31 ms 1.00 108 2401.8 453.1 1.63 KB 1.00
Sylvan___ Row 100000 145.27 ms 3.20 108 749.1 1452.7 11.49 KB 7.04
ReadLine_ Row 100000 161.71 ms 3.57 108 673.0 1617.1 359865.42 KB 220,527.94
CsvHelper Row 100000 335.98 ms 7.42 108 323.9 3359.8 20.65 KB 12.65
Sep______ Cols 100000 52.50 ms 1.00 108 2072.8 525.0 1.63 KB 1.00
Sylvan___ Cols 100000 165.46 ms 3.15 108 657.7 1654.6 10.28 KB 6.30
ReadLine_ Cols 100000 166.79 ms 3.17 108 652.4 1667.9 359865.42 KB 220,527.94
CsvHelper Cols 100000 358.83 ms 6.83 108 303.3 3588.3 113699.78 KB 69,675.99
Sep______ Floats 100000 258.54 ms 1.00 108 420.9 2585.4 8.84 KB 1.00
Sylvan___ Floats 100000 829.69 ms 3.21 108 131.2 8296.9 18.35 KB 2.07
ReadLine_ Floats 100000 860.02 ms 3.33 108 126.5 8600.2 359871.84 KB 40,687.73
CsvHelper Floats 100000 1,185.77 ms 4.59 108 91.8 11857.7 87694.45 KB 9,914.88

Writer Comparison Benchmarks

Writer benchmarks are still pending, but Sep is unlikely to be the fastest here since it is explicitly designed to make writing more convenient and flexible. Still efficient, but not necessarily fastest. That is, Sep does not require writing header up front and hence having to keep header column order and row values column order the same. This means Sep does not write columns directly upon definition but defers this until a new row has been fully defined and then is ended.

Example Catalogue

The following examples are available in ReadMeTest.cs.

Example - Copy Rows

var text = """
           A;B;C;D;E;F
           Sep;🚀;1;1.2;0.1;0.5
           CSV;✅;2;2.2;0.2;1.5
           
           """; // Empty line at end is for line ending

using var reader = Sep.Reader().FromText(text);
using var writer = reader.Spec.Writer().ToText();
foreach (var readRow in reader)
{
    using var writeRow = writer.NewRow(readRow);
}

Assert.AreEqual(text, writer.ToString());

RFC-4180

While the RFC-4180 requires \r\n (CR,LF) as line ending, the well-known line endings (\r\n, \n and \r) are supported similar to .NET. Environment.NewLine is used when writing. Quoting is supported by simply matching pairs of quotes, no matter what. With no automatic escaping. Hence, you are responsible and in control of this at this time.

Note that some libraries will claim conformance but the RFC is, perhaps naturally, quite strict e.g. only comma is supported as separator/delimiter. Sep defaults to using ; as separator if writing, while auto-detecting supported separators when reading. This is decidedly non-conforming.

The RFC defines the following condensed ABNF grammar:

file = [header CRLF] record *(CRLF record) [CRLF]
header = name *(COMMA name)
record = field *(COMMA field)
name = field
field = (escaped / non-escaped)
escaped = DQUOTE *(TEXTDATA / COMMA / CR / LF / 2DQUOTE) DQUOTE
non-escaped = *TEXTDATA
COMMA = %x2C
CR = %x0D ;as per section 6.1 of RFC 2234 [2]
DQUOTE =  %x22 ;as per section 6.1 of RFC 2234 [2]
LF = %x0A ;as per section 6.1 of RFC 2234 [2]
CRLF = CR LF ;as per section 6.1 of RFC 2234 [2]
TEXTDATA =  %x20-21 / %x23-2B / %x2D-7E

Note how TEXTDATA is restricted too, yet many will allow any character incl. emojis or similar (which Sep supports), but is not in conformance with the RFC.

Quotes inside an escaped field e.g. "fie""ld" are only allowed to be double quotes. Sep currently allows any pairs of quotes and quoting doesn't need to be at start of or end of field (col or column in Sep terminology).

All in all Sep takes a pretty pragmatic approach here as the primary use case is not exchanging data on the internet, but for use in machine learning pipelines or similar.

Frequently Asked Questions (FAQ)

Ask questions on GitHub and this section will be expanded. 😃

SepReader FAQ

SepWriter FAQ

Product Compatible and additional computed target framework versions.
.NET net7.0 is compatible.  net7.0-android was computed.  net7.0-ios was computed.  net7.0-maccatalyst was computed.  net7.0-macos was computed.  net7.0-tvos was computed.  net7.0-windows was computed.  net8.0 was computed.  net8.0-android was computed.  net8.0-browser was computed.  net8.0-ios was computed.  net8.0-maccatalyst was computed.  net8.0-macos was computed.  net8.0-tvos was computed.  net8.0-windows was computed. 
Compatible target framework(s)
Additional computed target framework(s)
Learn more about Target Frameworks and .NET Standard.

NuGet packages

This package is not used by any NuGet packages.

GitHub repositories (2)

Showing the top 2 popular GitHub repositories that depend on Sep:

Repository Stars
DataDog/dd-trace-dotnet
.NET Library for Datadog APM
JasonBock/Rocks
A mocking library based on the Compiler APIs (Roslyn + Mocks)
Version Downloads Last updated
0.4.0 2,829 1/1/2024
0.4.0-preview.1 89 12/23/2023
0.3.0 452,100 11/18/2023
0.2.7 2,325 10/12/2023
0.2.6 443 9/27/2023
0.2.5 236 9/14/2023
0.2.4 219 9/8/2023
0.2.3 246 9/5/2023
0.2.2 475,287 8/10/2023
0.2.1 135 8/10/2023
0.2.0 912 8/7/2023
0.2.0-preview.3 96 7/29/2023
0.1.0 570 5/30/2023
0.1.0-rc.1 84 5/26/2023
0.1.0-preview.8 71 5/26/2023
0.1.0-preview.7 90 5/8/2023
0.1.0-preview.6 87 4/24/2023
0.1.0-preview.5 93 3/19/2023
0.1.0-preview.4 100 12/31/2022
0.1.0-preview.3 100 12/4/2022
0.1.0-preview.2 126 3/21/2022
0.1.0-preview.1 130 1/28/2022