Mizzle.Generators 0.1.0-alpha.20

Mizzle

Fluent, type-safe SQL for .NET, inspired by Drizzle ORM.

Declare your schema in C#, build queries against typed columns, and let Mizzle emit SQL for PostgreSQL or SQL Server. It does not use raw SQL strings or LINQ translation. Statically-visible queries skip reflection entirely by baking SQL and a projection mapper at build time; runtime-mapped paths (an untraceable dynamic SelectBuilder, or a write's Returning(...) projection) map into T by reflection instead.

Status: experimental. This is an early alpha. The API will change between releases. PostgreSQL and SQL Server are the only supported databases.

Packages

Installing a dialect package brings in everything you need:

dotnet add package Mizzle.Postgres

Quick start

Declare a table:

using Mizzle.Postgres;

public sealed class Users : PgTable<Users>
{
    public Users() : base("users", "public") { }

    public PgColumn<int> Id { get; } = Identity("id").PrimaryKey();
    public PgColumn<string> Email { get; } = Text("email").NotNull().Unique();
    public PgColumn<string> DisplayName { get; } = Varchar("display_name", 120);
    public PgColumn<bool> IsActive { get; } = Boolean("is_active").NotNull();
    public PgColumn<Guid> PublicId { get; } = Uuid("public_id").NotNull();
    public PgColumn<DateTimeOffset> CreatedAt { get; } = Timestamptz("created_at").NotNull();
    public PgColumn<DateOnly> Birthday { get; } = Date("birthday");
    public PgColumn<long> LoginCount { get; } = BigInt("login_count");
}

The table alias defaults to the table name. Use WithAlias at the query site when one table needs another name for a self-join or repeated lookup.

PostgreSQL tables include factories such as Text, Varchar, Char, Integer, BigInt, Boolean, Uuid, Date, Timestamptz, and Identity. SQL Server tables include NVarChar, NVarCharMax, VarChar, Char, Text, NText, Int, SmallInt, TinyInt, BigInt, Decimal, Numeric, Real, Float, Bit, UniqueIdentifier, Date, DateTime, DateTime2, Timestamp, and Identity.

Register and query:

services.AddMizzlePostgres(connectionString);

var users = new Users();

// If UserRow does not exist, the generator declares it from the select shape.
var found = await db.Select(users.Id, users.Email)
    .From(users)
    .Where(users.Email.Eq("a@b.com"))
    .ToListAsync<UserRow>();

// Existing DTOs are mapped by normalized member name.
var profile = await db.Select(users.Id, users.Email)
    .From(users)
    .Where(users.Id.Eq(42))
    .FirstOrDefaultAsync<MyExistingDto>();

var id = await db.InsertInto(users)
    .Value(users.Email, "new@example.com")
    .Returning(users.Id)
    .SingleAsync<int>();

var inserted = await db.InsertInto(users)
    .Value(users.Email, "new@example.com")
    .Returning(users.Id, users.Email)
    .SingleAsync<UserRow>();

await db.Transaction(async tx =>
{
    await tx.LockAsync("invoice:123");
    // Queries here run on the transaction connection.
    // Nested Transaction calls become savepoints.
});

SQL Server tables and registration look the same, just with SqlTable<T> and AddMizzleSqlServer:

using Mizzle.SqlServer;

public sealed class Users : SqlTable<Users>
{
    public Users() : base("Users", "dbo") { }

    public SqlColumn<int> Id { get; } = Identity("Id").PrimaryKey();
    public SqlColumn<string> Email { get; } = NVarChar("Email", 256).NotNull().Unique();
    public SqlColumn<bool> IsActive { get; } = Bit("IsActive").NotNull();
}

services.AddMizzleSqlServer(connectionString);

var users = new Users();
var found = await db.Select(users.Id, users.Email)
    .From(users)
    .Where(users.Email.Eq("a@b.com"))
    .ToListAsync<UserRow>();

Joins keep the same style. Conditions are typed, and chained Where calls are combined with AND:

var rows = await db.Select(authors.DisplayName, tags.Label)
    .From(authors)
    .LeftJoin(tags).On(authors.FavoriteTagId.Eq(tags.TagId), tags.Kind.Eq("topic"))
    .Where(authors.BlogId.Eq(blogId))
    .OrderBy(authors.DisplayName)
    .ToListAsync<AuthorTagRow>();

Guarded updates for optimistic concurrency:

await db.Update(users)
    .Set(users.Email, "renamed@example.com")
    .Where(users.Id, 42)
    .Expect(1)          // throws ConcurrencyException if affected rows != 1
    .ExecuteAsync();

Paging on any ordered query:

var page = await db.Select(users.Id, users.Email)
    .From(users)
    .OrderBy(users.Email)
    .Page(2, 25)
    .ToPageAsync<UserRow>(includeTotal: true);

Storage Conversions

Some databases store values in a shape you do not want in your application: GUIDs in char(36), dates in char(8), or flags as 'Y' and 'N'. Declare that conversion on the column:

public SqlColumn<Guid> AccountId { get; } =
    Char("account_id", 36)
        .Map(AccountConvert.ToGuid, AccountConvert.FromGuid)
        .PrimaryKey();

Queries use the mapped type:

await db.Select(accounts.AccountId, accounts.Email)
    .From(accounts)
    .Where(accounts.AccountId.Eq(accountId))
    .SingleAsync<AccountRow>();

The write side receives the storage value, so the predicate above sends a string to the database. Generated mappers call the read converter directly. Converters must be static method references, not lambdas, so the generator can see and bake the call. Lambdas report MIZ008.

Projecting into domain types

Typed terminators come in two modes.

If the result type does not exist, Mizzle generates a record from the selected columns:

var rows = await db.Select(users.Id, users.Email)
    .From(users)
    .ToListAsync<UserRow>();

If the result type already exists, Mizzle maps into it by normalized member name. Underscores and casing are ignored, so display_name can match DisplayName.

When a column name and member name do not line up, use As at the select site:

var row = await db.Select(
        books.BookId.As("Id"),
        books.DisplayTitle.As("Title"),
        authors.DisplayName.As("Author"))
    .From(books)
    .InnerJoin(authors).On(books.AuthorId.Eq(authors.AuthorId))
    .SingleAsync<BookSummary>(ct);

As returns a copy of the column. The table's column stays unchanged, and SQL gets a real alias such as AS [Title]. Projection diagnostics use the aliased name, so a typo reports MIZ003 and a type mismatch reports MIZ010 at the call site.

As also accepts nameof(...), which keeps the alias in sync if the target member is renamed:

books.BookId.As(nameof(BookSummary.Id))

Anything else -- a field, a variable, string concatenation -- is not a compile-time constant and falls back to the runtime path.

The delegate overloads are always runtime mapped:

var rows = await db.Select(users.Id, users.Email)
    .From(users)
    .ToListAsync(r => new UserRow(r.GetInt32(0), r.GetString(1)));

The delegate-free typed terminators need the query's projection to be statically visible, even when the SQL itself can't be baked. A chain built by reassigning a local -- adding Where, OrderBy, paging, and the like -- still gets a generated mapper; only building a genuinely different projection (a new or additional Select) breaks that link. An unsupported call reports MIZ014 rather than silently falling back, in Strict mode or not. If you pass around a dynamic SelectBuilder in a way the generator can't trace at all, use the delegate overload.

Reusing Tables

Use WithAlias when one table appears more than once in a query. It works for self-joins and for lookup tables used in several roles:

var primaryTag = new Tags().WithAlias("primary_tag");
var secondaryTag = new Tags().WithAlias("secondary_tag");

var rows = await db.Select(
        posts.Title,
        primaryTag.Label.As("PrimaryTag"),
        secondaryTag.Label.As("SecondaryTag"))
    .From(posts)
    .LeftJoin(primaryTag).On(posts.PrimaryTagId.Eq(primaryTag.TagId))
    .LeftJoin(secondaryTag).On(posts.SecondaryTagId.Eq(secondaryTag.TagId))
    .ToListAsync<PostTagRow>();

WithAlias returns a new table instance. The original table keeps its default alias and can still be shared. If two table instances in one generated query use the same alias, Mizzle reports MIZ011.

Trimming Strings

Databases with fixed-width text columns often return padded strings. You can opt generated mappers into trimming string reads:

<PropertyGroup>
  <MizzleTrimStrings>true</MizzleTrimStrings>
</PropertyGroup>

Trimming applies to string storage reads before any Map converter. It does not run on writes. Exclude a column when trailing whitespace is meaningful:

public SqlColumn<string> Signature { get; } = VarChar("signature", 500).Untrimmed();

Common table expressions

With and WithRecursive attach a CTE to any statement -- select, insert, update or delete:

var body = db.Select(o.OrderId, o.Ndc)
    .From(o)
    .Where(o.Status.Eq("abandoned"))
    .Build();

var stale = CteBuilder.Named<StaleOrders>("stale", body);
var s = new StaleOrders();

var rows = await db.Select(s.OrderId, s.Ndc)
    .With(stale)
    .From(s)
    .ToListAsync<StaleOrderRow>();

CteBuilder.Named<T> declares T as a table type from the CTE body's own select list -- one typed column per projected column, matched to the body's real SQL names and types. StaleOrders above needs no hand-written schema; the source generator produces it from stale's shape, so it can't drift out of sync with the query that defines it. From there the CTE behaves like any other table: typed columns, As(...), left-join nullability, and the projection diagnostics.

A CTE whose body is a statically visible chain is baked along with the outer query, so CTE queries stay on the interceptor path instead of falling back to runtime compilation.

If you already have a hand-declared table type for a CTE shape -- one shared across several queries, say -- CteBuilder.Named("name", body) also accepts a plain string and skips generation. Nothing then checks the declared columns against the body's select list, so a mismatch surfaces at the database instead of at build time; prefer the generic overload unless you have a specific reason to hand-declare.

Ranking within a partition

Sql.RowNumber() builds ROW_NUMBER() OVER (PARTITION BY ... ORDER BY ...), for picking one "best" row per group. There is no separate "top 1 per group" query shape -- combine it with a CTE and a second, typed CTE that filters the rank column, same as any other computed column:

var body = db.Select(
        o.Ndc,
        Sql.As(Sql.RowNumber().PartitionBy(o.Ndc).OrderByDesc(o.EffectiveDate), "rn"))
    .From(o)
    .Build();

var ranked = CteBuilder.Named<Ranked>("ranked", body);
var r = new Ranked();

var rows = await db.Select(r.Ndc)
    .With(ranked)
    .From(r)
    .Where(r.rn.Eq(1))
    .ToListAsync<BestNdc>();

Ranked needs no hand-declared schema -- CteBuilder.Named<T> generates it from body's shape, including the computed rn column's real SQL type (bigint/long, matching what ROW_NUMBER() actually returns).

PartitionBy/OrderBy/OrderByDesc take either a plain column or a computed expression (Sql.Case(...), TSql.RTrim(...), ...); mixing both kinds in one call needs an explicit .ToRef() on the column. ROW_NUMBER() is the only window function Mizzle supports today -- no RANK, no frame-based aggregates (SUM() OVER, ...), no QUALIFY.

Returning rows from writes

Insert, update and delete expose the same typed terminators as select -- ToListAsync<T>, FirstAsync<T>, FirstOrDefaultAsync<T>, SingleAsync<T>, SingleOrDefaultAsync<T> -- over their Returning(...) columns, and As(...) works there too:

var updated = await db.Update(o)
    .Set(o.Status, "shipped")
    .Where(o.OrderId.Eq(id))
    .Returning(o.OrderId.As("Id"), o.Status)
    .SingleAsync<ShippedOrder>(ct);

Write projections are mapped at runtime rather than baked, but the projection diagnostics still run at build time, so a returning-into-T mismatch is a compile error rather than a runtime throw.

How queries execute

Every query builds an immutable IR graph. That is just a small object model for the query: selected columns, source table, joins, predicates, ordering, limits, and values. Each builder call returns the next query shape instead of mutating the old one.

Before any SQL is written, a capability pass checks the target dialect and throws UnsupportedFeatureException for anything it cannot do. For example, ILike is valid on PostgreSQL but not SQL Server. Mizzle reports that instead of emitting a different query.

What that means in practice:

• SQL and parameters stay separate until execution.
• PostgreSQL and SQL Server share the same fluent surface where they can.
• Dialect-only behavior fails clearly instead of turning into surprise SQL.
• The same query shape can run dynamically or be picked up by the source generator.
• Statically-visible list and single-row queries can skip runtime SQL emission.

Statically-visible query chains are compiled at build time. A source generator reconstructs the query, bakes the SQL string, generates the projection mapper, and intercepts the call site. Dynamic queries fall back to the runtime pipeline with a shape cache. Typed paging uses the generated mapper with the normal paging executor so includeTotal and cursor behavior stay in one place.

Setting <MizzleQueryMode>Strict</MizzleQueryMode> in your project turns any non-compilable query into a build error.

What it deliberately doesn't do

• No raw SQL. If a construct isn't in the IR, it isn't expressible. This is a guarantee, not a gap. Check that the current surface covers your needs before adopting it.
• No LINQ / IQueryable, no sync APIs, no migrations, no MySQL (yet).

CLI

Mizzle.Cli ships as a .NET tool. It is installed through NuGet and runs as mizzle:

dotnet tool install --global Mizzle.Cli --prerelease

mizzle version
mizzle version --verbose
mizzle type-map --provider postgres
mizzle inspect --connection "Host=localhost;Database=app;Username=postgres;Password=..." --schema public --all
mizzle scaffold --connection "Host=localhost;Database=app;Username=postgres;Password=..." --schema public --tables users,posts --namespace MyApp.Data --output ./Data/Tables
mizzle doctor
mizzle doctor --project ./MyApp.csproj
mizzle doctor --solution ./MyApp.slnx
mizzle diff --connection "Host=localhost;Database=app;Username=postgres;Password=..." --schema public --source ./Data/Tables
mizzle explain --provider postgres --sql-file ./query.sql
mizzle translate-query --provider postgres --sql-file ./query.sql

Database commands infer postgres or sqlserver from common connection string shapes. Pass --provider when the connection string is ambiguous.

doctor is intentionally read-only. If you run it from a directory with one solution, it checks the projects in that solution. If there is no solution, it uses the single project in the current directory. You can also pass --project or --solution explicitly.

Project and solution paths must be inside the current working directory. Inherited project files are read only up to that directory: Directory.Build.props, Directory.Build.targets, and Directory.Packages.props. The command checks dialect references, generator references, nullable settings, MizzleQueryMode, old constructor alias syntax, non-literal column names, lambda column maps, and mismatched Mizzle package versions. Test and benchmark projects are shown when they use Mizzle, but their app configuration checks are skipped.

The CLI stops on unsupported database types or SQL shapes with MZCLI### messages. That is intentional: it should point at what Mizzle needs to learn next, not generate code that quietly guesses.

Current commands:

• version: show the installed Mizzle CLI version. Pass --verbose to include build metadata.
• type-map: show the database types Mizzle knows how to scaffold.
• inspect: list tables, columns, database types, nullability, keys, and unsupported mappings.
• scaffold: generate PgTable<> or SqlTable<> classes from an existing database.
• doctor: check a project or solution for common Mizzle setup problems.
• diff: compare live database columns with existing Mizzle table classes.
• explain: summarize SQL features and likely Mizzle support.
• translate-query: translate a small SQL subset into Mizzle query syntax.

Building

dotnet test Mizzle.slnx

Unit and generator tests run anywhere; the integration tests use Testcontainers and need Docker.

Benchmarks

Benchmarks live in benchmarks/Mizzle.Benchmarks.

They compare Mizzle's runtime and source-generated query paths with raw Npgsql and Dapper. The goal is to show the overhead of Mizzle's typed query API, not to claim a universal winner.

dotnet run -c Release --project benchmarks/Mizzle.Benchmarks

By default the benchmarks start PostgreSQL with Testcontainers. To use an existing database instead, set MIZZLE_BENCH_POSTGRES to a PostgreSQL connection string.

License

MIT