DuckDbSharp

DuckDbSharp is a bidirectional interoperability layer between DuckDB and .NET.

Features

• Support for deeply nested structures and lists
• Expose .NET methods/collections as table UDFs, or as scalar functions
• Execute DuckDB queries from .NET
• Generates static types from SQL (incl. field nullness detection)
• Dynamic results are supported as well (as dynamic assemblies/types)
• Performance-oriented with minimal allocations
• Support for both normal and [Flags] enums
• Native AOT support
• Pass .NET collections as SQL parameters (either as array or as table)
• Results are streamed as IEnumerable<>
• Write DuckDB loadable extensions in C# (work in progress)

Notes:

• This is not an ADO.NET (System.Data) provider
  • Rationale: ADO.NET is flat-table oriented (sublists/subfields are not first class citizens, despite these being probably among the best features of DuckDB).
  • Additionally, ADO.NET is very unergonomic to use unless you add an ORM on top of it. Most existing ORMs however don't work well with sublists/subfields. This library deserializes/serializes directly on top of CLR POCO objects, and can generate (and keep up to date) the type definitions for a better IDE experience.

Usage

Calling DuckDB from .NET (auto-generated types)

-- my_query.sql
SELECT
    42 AS column1,
    [1, 2, 3] AS column2,
    [{a: 1, b: 2}] as column3

foreach (var user in db.ExecuteQuery_my_query())
{
    // "user" has an auto-generated type with all the fields and sub-fields of the SQL query above.
}

See detailed instructions below.

Calling DuckDB from .NET (inline sql)

using var db = ThreadSafeTypedDuckDbConnection.CreateInMemory();
foreach (var user in db.Execute<User>("select * from user"))
{
    
}

You can also use value tuples (Execute<(string A, int B)>("select 'a', 42")), but keep in mind that only column order matters, since tuple member names are erased at runtime.

Calling .NET from DuckDB

[DuckDbFunction]
public static IEnumerable<User> GetUsers(string country) { /*...*/ }

SELECT * FROM GetUsers('US')

Getting started

• Add a reference to DuckDbSharp
• Create a queries directory with the .sql queries you want to be able to call from .NET
  • File name must be: ReturnType query_name(paramtype1, paramtype2).sql
  • If ReturnType is not specified, a type will be automatically generated based on the SQL schema of the result.
  • If your query is parameterized, specify the types of the parameters (e.g. string or long). Otherwise, parens are unnecessary.
• Call GenerateCSharpTypes as shown in the example, and run.
• Start using the now generated extension methods of TypedDuckDbConnectionBase (one for each query).
• Remember to commit the generated files as well. This is very important in order to be able to recompile old versions of the repository.

Benchmarks

Time to read ~100,000 rows of Northwind customers. In all 4 cases, the final result is a List<Customer>.

Library	Mean	Error	StdDev	Description
DuckDbSharp (this project)	145.4 ms	2.63 ms	2.46 ms	SELECT * FROM customer
DuckDB.NET + Dapper	177.2 ms	2.88 ms	2.69 ms	SELECT * FROM customer
Protobuf-net	131.3 ms	2.52 ms	2.81 ms	Deserialize from MemoryStream of protos
Newtonsoft JSON	241.7 ms	2.93 ms	2.60 ms	Deserialize from MemoryStream of JSON

Note: while protobuf-net is slightly faster, its use cases is very different (serialization/deserialization only, with no query support)

Advanced features

Customizing (de)serialization

• [DuckDbInclude] and [DuckDbIgnore] always take the precedence over other rules. In their absence, [ProtoMember] from protobuf-net is also taken into account. Otherwise, only public fields and properties are taken into account.
• Enums are serialized as DuckDb enums, you can override this with [DuckDbSerializeAs(typeof(string))] or [DuckDbSerializeAs(typeof(int))] (or whatever their underlying type is).
• [Flags] enums are always serialized as structs of booleans, one for each bit.
• [DuckDbDefaultValueIsNullish] can be applied to structs, and it means that default(SomeStruct) should be represented as NULL in DuckDB.

Reading and writing parquets

• You can use DuckDbUtils.QueryParquet<T>() and DuckDbUtils.WriteParquet<T>() to directly read/write .parquet files (no database required).