ADR-015: Sep — High-Performance CSV Parsing

Date: 2026-03-01 Authors: Jean-Francois Meyers Scope: granit-dotnet (Granit.DataExchange.Csv)

Context

The Granit.DataExchange.Csv module requires a CSV parser capable of processing files with 100,000+ rows in streaming, without loading the entire file into memory. Use cases include: patient data import, roundtrip reimport, initial reference data loading.

The library must support:

• Streaming: native IAsyncEnumerable for the DataExchange pipeline
• Performance: large files without degradation
• Encodings: UTF-8, UTF-8 BOM, Latin-1, Windows-1252
• RFC 4180: quoted fields, configurable separators
• License: compatible with commercial use without recurring costs
• Target: explicit .NET 10

Decision

Sep (nietras) for CSV parsing in Granit.DataExchange.Csv.

Alternatives considered

Option 1: Sep (selected)

• License: MIT
• Advantage: zero-allocation after warmup, SIMD vectorization (AVX-512, SSE, ARM NEON), explicit net10.0 target, native IAsyncEnumerable (.NET 9+), Span<T> / ISpanParsable<T>, 9-35x faster than CsvHelper, AOT-compatible
• Maturity: active releases in 2025 (0.9.0 to 0.12.2), extensive tests
• Disadvantage: lower-level API (Span-oriented), 0.x version number (conservative versioning by the author, not a sign of instability)

Option 2: CsvHelper

• License: MS-PL / Apache-2.0
• Advantage: de facto standard (508M NuGet downloads), high-level API (GetRecordsAsync<T>(), ClassMap, TypeConverter), native IAsyncEnumerable, excellent documentation, error callbacks (BadDataFound)
• Disadvantage: one string allocation per column (significant at 100K+ rows), no SIMD vectorization, no explicit net10.0 target (via netstandard2.0), 9-35x slower than Sep

Option 3: Sylvan.Data.Csv

• License: MIT
• Advantage: 2-3x faster than CsvHelper, familiar DbDataReader API, auto-detection of delimiter, Lax mode for malformed data
• Disadvantage: no explicit net10.0 target, no SIMD, no IAsyncEnumerable (only ReadAsync()), intermediate performance without decisive advantage over Sep or CsvHelper

Option 4: RecordParser

• License: MIT
• Advantage: near-zero allocation via expression trees, Span<char>
• Disadvantage: last release November 2023 (18+ months), 116K downloads, no .NET 8/9/10 target, no IAsyncEnumerable, minimal documentation, stagnant maintenance

Justification

Criterion	Sep	CsvHelper	Sylvan.Data.Csv	RecordParser
License	MIT	MS-PL/Apache-2.0	MIT	MIT
Performance vs CsvHelper	9-35x	1x (baseline)	2-3x	~2x
Zero-allocation	Yes	No	Low-alloc	Near-zero
SIMD (AVX-512/NEON)	Yes	No	No	No
Target net10.0	Yes	No (netstandard)	No (net6.0)	No
IAsyncEnumerable	Yes (.NET 9+)	Yes	No	No
Span/Memory API	Yes	No	Partial	Yes
NuGet downloads	~1.4M	~508M	~3.1M	~116K
Active maintenance	Yes (2025)	Yes	Yes	No (2023)
AOT/Trimming	Yes	Partial	Partial	Unknown
API ergonomics	Medium	Excellent	Good	Low

The decisive criterion is streaming performance for large files (100K+ rows). Sep is 9-35x faster than CsvHelper thanks to SIMD vectorization and zero allocations. The lower-level API is not a disadvantage as it is encapsulated behind the IFileParser interface — consumers never see the Sep API directly.

Consequences

Positive

• Fastest CSV parsing in the .NET ecosystem (SIMD vectorized)
• Zero-allocation: no GC pressure on large imports
• Explicit net10.0 target: runtime optimizations leveraged
• Native IAsyncEnumerable: natural integration with the DataExchange pipeline
• MIT: no cost, compatible with commercial use
• AOT-compatible: no runtime reflection

Negative

• Span-oriented API more verbose than CsvHelper for internal code
• 0.x version (though stable and actively maintained)
• Smaller community than CsvHelper (1.4M vs 508M downloads)
• No native ClassMap — mapping is done by IDataMapper<T> (by design)
• No built-in error callbacks (BadDataFound) — handled via try/catch in the SepCsvFileParser implementation