JSON columns

On DuckDB:

• DuckDB has a JSON type, VARCHAR underneath, that rejects malformed documents as they are stored
• The generated SQL casts a JSON column to VARCHAR, so store.JSON holds the document as stored, keys in the order they were written
• ->> extracts text and -> extracts a JSON fragment, as on SQLite
• PlaceOrder takes store.JSON, and the generated code hands duckdb-go the document's text, the only form it binds a JSON value in
• The ->> comparison's parameter is typed with sgl.param(fish, type: varchar), as inference leaves it as any

A JSON column holds a whole document in one value: a basket of items, a set of preferences, a payload you do not want to split into tables. The database can still search inside it. On the Go side you want to keep your own struct for the document, and let the generated code carry it in and out of the database without getting in the way.

Seagull gives every JSON column one generated type, JSON, which holds the raw document as bytes. You fill it with json.Marshal and read it with json.Unmarshal into whatever struct you like. Fragments extracted inside a query come back as JSON too, and values extracted as text come back as plain strings.

This example keeps a chip shop's orders. Each has a basket document, and a delivery document only when the order is delivered.

The schema

db/migrations/0001_orders.up.sql

CREATE SEQUENCE order_ids;

CREATE TABLE orders (
    id       BIGINT  PRIMARY KEY DEFAULT nextval('order_ids'),
    customer VARCHAR NOT NULL,
    basket   JSON    NOT NULL,
    delivery JSON
);

The JSON type

sgl generate writes the JSON type into its own file in your package, so there is nothing extra to import:

internal/store/sgl_json.go

// Code generated by sgl. DO NOT EDIT.

package store

import (
	"database/sql/driver"
	"encoding/json"
	"fmt"
)

type JSON []byte

func (j *JSON) Scan(source any) error {
	switch value := source.(type) {
	case nil:
		*j = nil
		return nil
	case []byte:
		clone := make(JSON, len(value))
		copy(clone, value)
		*j = clone
		return nil
	case string:
		*j = JSON(value)
		return nil
	default:
		encoded, err := json.Marshal(value)
		if err != nil {
			return fmt.Errorf("sgl: encode %T to JSON: %w", source, err)
		}
		*j = encoded
		return nil
	}
}

func (j JSON) Value() (driver.Value, error) {
	if j == nil {
		return nil, nil
	}
	return string(j), nil
}

func (j JSON) MarshalJSON() ([]byte, error) {
	if j == nil {
		return []byte("null"), nil
	}
	return j, nil
}

func (j *JSON) UnmarshalJSON(data []byte) error {
	clone := make(JSON, len(data))
	copy(clone, data)
	*j = clone
	return nil
}

It is a []byte with the methods database/sql and encoding/json look for. Scan accepts whatever shape the driver returns for a JSON column, and a nil JSON is SQL NULL. That is why the nullable delivery column is a plain JSON, not a pointer: nil already says "no document".

Reading a document

internal/store

type GetOrderRow struct {
	ID       int64  `json:"id"`
	Customer string `json:"customer"`
	Basket   JSON   `json:"basket"`
	Delivery JSON   `json:"delivery"`
}

type JSON []byte

func (queries *Queries) GetOrder(ctx context.Context, id int64) (GetOrderRow, error) {
	var row GetOrderRow
	err := queries.reader.QueryRowContext(ctx, getOrderSQL, id).Scan(&row.ID, &row.Customer, &row.Basket, &row.Delivery)
	if err != nil {
		return GetOrderRow{}, err
	}
	return row, nil
}

Searching inside documents

OrdersWithFish finds the orders for one kind of fish, and pulls two things out of each document as it goes.

db/queries/orders.sql

-- sgl.query(OrdersWithFish, many)
-- $1 as sgl.param(fish, type: varchar)
SELECT
    id,
    customer,
    basket -> '$.extras' AS extras,
    delivery ->> '$.address' AS address
FROM orders
WHERE basket ->> '$.fish' = $1
ORDER BY id;

->> extracts a value as SQL text, and -> extracts a fragment as JSON, each at a $. path. So Extras is JSON again, ready to decode, and Address is text. It is a *string, because ->> gives NULL when the delivery column or its address is missing.

The value compared with ->> could be any type as far as the SQL says, so inference leaves the parameter as any. The binding line sgl.param(fish, type: varchar) pins it to a string.

internal/store

type OrdersWithFishRow struct {
	ID       int64   `json:"id"`
	Customer string  `json:"customer"`
	Extras   JSON    `json:"extras"`
	Address  *string `json:"address"`
}

type JSON []byte

func (queries *Queries) OrdersWithFish(ctx context.Context, fish string) ([]OrdersWithFishRow, error) {
	rows, err := queries.reader.QueryContext(ctx, ordersWithFishSQL, fish)
	if err != nil {
		return nil, err
	}
	defer rows.Close()
	var results []OrdersWithFishRow
	for rows.Next() {
		var row OrdersWithFishRow
		if err := rows.Scan(&row.ID, &row.Customer, &row.Extras, &row.Address); err != nil {
			return nil, err
		}
		results = append(results, row)
	}
	if err := rows.Err(); err != nil {
		return nil, err
	}
	return results, nil
}

Writing documents

The DuckDB driver cannot bind the generated JSON type yet, so PlaceOrder types its two documents as varchar with binding lines, and the program passes the encoded JSON text. DuckDB's JSON type is VARCHAR underneath, and checks the text as it is stored.

db/queries/orders.sql

-- sgl.query(PlaceOrder, exec)
INSERT INTO orders (customer, basket, delivery) VALUES ($1, $2, $3);

Using it

The program encodes each basket with json.Marshal and places three orders, only one of them for delivery. It reads the first order back and decodes its basket, then searches for the cod orders.

example.go

package main

import (
	"context"
	"database/sql"
	"encoding/json"
	"fmt"
	"io"

	"seagull.sh/seagull/examples/duckdb/006_json/internal/store"
)

// basket is the document an order's basket column holds.
type basket struct {
	// Fish is the fish the order is for.
	Fish string `json:"fish"`

	// Portions is how many portions of fish and chips.
	Portions int `json:"portions"`

	// Extras lists the sides.
	Extras []string `json:"extras"`
}

// delivery is the document an order's delivery column holds, when the order is delivered.
type delivery struct {
	// Address is where the order goes.
	Address string `json:"address"`
}

// run places three chip shop orders and reads them back. A JSON column read by a query
// becomes the generated store.JSON type, which holds the raw document as bytes: decode it
// with json.Unmarshal. A nil store.JSON is SQL NULL.
//
// The generated SQL casts each JSON column it reads to VARCHAR, so store.JSON holds the
// document as DuckDB stored it, the keys in the order they were written.
//
// PlaceOrder takes its documents as store.JSON too, and the generated code hands
// duckdb-go their text; DuckDB's JSON type is VARCHAR underneath and checks the text as
// it is stored.
//
// OrdersWithFish filters with ->>, which extracts a value as text, and selects with ->,
// which extracts a JSON fragment, so Extras is store.JSON again. Address is *string
// because ->> gives NULL when the delivery column or its address is missing.
//
// Takes database (*sql.DB) which is a migrated database.
// Takes output (io.Writer) which receives the decoded orders and the search.
//
// Returns error when a query fails or a document does not encode or decode.
func run(ctx context.Context, database *sql.DB, output io.Writer) error {
	queries := store.New(database)

	orders := []struct {
		customer string
		basket   basket
		delivery *delivery
	}{
		{customer: "Steven", basket: basket{Fish: "cod", Portions: 2, Extras: []string{"mushy peas", "curry sauce"}}},
		{customer: "Gwen", basket: basket{Fish: "haddock", Portions: 1, Extras: []string{}}},
		{customer: "Morwenna", basket: basket{Fish: "cod", Portions: 1, Extras: []string{"pickled egg"}}, delivery: &delivery{Address: "4 Harbour Row"}},
	}
	for _, order := range orders {
		encodedBasket, err := json.Marshal(order.basket)
		if err != nil {
			return fmt.Errorf("encoding %s's basket: %w", order.customer, err)
		}
		parameters := store.PlaceOrderParams{Customer: order.customer, Basket: encodedBasket, Delivery: nil}
		if order.delivery != nil {
			encodedDelivery, err := json.Marshal(order.delivery)
			if err != nil {
				return fmt.Errorf("encoding %s's delivery: %w", order.customer, err)
			}
			parameters.Delivery = encodedDelivery
		}
		if err := queries.PlaceOrder(ctx, parameters); err != nil {
			return fmt.Errorf("placing %s's order: %w", order.customer, err)
		}
	}

	first, err := queries.GetOrder(ctx, 1)
	if err != nil {
		return fmt.Errorf("getting order 1: %w", err)
	}
	var decoded basket
	if err := json.Unmarshal(first.Basket, &decoded); err != nil {
		return fmt.Errorf("decoding order 1's basket: %w", err)
	}
	fmt.Fprintf(output, "order 1 stored %s\n", first.Basket)
	fmt.Fprintf(output, "decoded: %d x %s with %v, delivery is NULL: %t\n", decoded.Portions, decoded.Fish, decoded.Extras, first.Delivery == nil)

	cod, err := queries.OrdersWithFish(ctx, "cod")
	if err != nil {
		return fmt.Errorf("finding cod orders: %w", err)
	}
	fmt.Fprintln(output, "\ncod orders:")
	for _, order := range cod {
		address := "collecting"
		if order.Address != nil {
			address = "to " + *order.Address
		}
		fmt.Fprintf(output, "  order %d  %-9s extras %-30s %s\n", order.ID, order.Customer, order.Extras, address)
	}
	return nil
}

What it prints

The DuckDB driver hands back a JSON column already decoded, so JSON encodes it again, and the keys come back in alphabetical order rather than the order they were written in. Compare documents by decoding them, not by their bytes.

order 1 stored {"fish":"cod","portions":2,"extras":["mushy peas","curry sauce"]}
decoded: 2 x cod with [mushy peas curry sauce], delivery is NULL: true

cod orders:
  order 1  Steven    extras ["mushy peas","curry sauce"]   collecting
  order 3  Morwenna  extras ["pickled egg"]                to 4 Harbour Row

Across engines

Only these Go types of OrdersWithFish differ between engines; the rest are the same everywhere.

	DuckDB	SQLite	PostgreSQL	MySQL	MariaDB	ClickHouse
id column	int64	int64	int64	int64	int64	uint64
extras column	JSON	JSON	JSON	JSON	JSON	string
address column	*string	*string	·	*string	*string	*string
delivery_address column	·	·	*string	·	·	·

A · marks a parameter or column that version of the example does not have.

CockroachDB and TimescaleDB run the PostgreSQL version unchanged.

Compare the whole example side by side with SQLite, PostgreSQL, MySQL, MariaDB and ClickHouse.

Run it locally

From the directory you cloned the Seagull repository into:

cd seagull/examples/duckdb && go run ./006_json

Its test runs it too, and checks what it prints against output.txt:

cd seagull/examples/duckdb && go test ./006_json

Needs cgo and a C compiler (the DuckDB driver).

The files

Every file of the example, as it is in the Seagull repository.

sgl.yaml

# Copyright 2026 PolitePixels Limited
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# This project stands against fascism, authoritarianism, and all forms of
# oppression. We built this to empower people, not to enable those who would
# strip others of their rights and dignity.

version: 1

databases:
  - name: chippy
    dialect: duckdb
    migrations: db/migrations
    queries: db/queries
    go:
      package: store
      out: internal/store

db/migrations/0001_orders.up.sql

CREATE SEQUENCE order_ids;

CREATE TABLE orders (
    id       BIGINT  PRIMARY KEY DEFAULT nextval('order_ids'),
    customer VARCHAR NOT NULL,
    basket   JSON    NOT NULL,
    delivery JSON
);

db/migrations/0001_orders.down.sql

DROP TABLE orders;
DROP SEQUENCE order_ids;

db/queries/orders.sql

-- sgl.query(PlaceOrder, exec)
INSERT INTO orders (customer, basket, delivery) VALUES ($1, $2, $3);

-- sgl.query(GetOrder, one)
SELECT id, customer, basket, delivery FROM orders WHERE id = $1;

-- sgl.query(OrdersWithFish, many)
-- $1 as sgl.param(fish, type: varchar)
SELECT
    id,
    customer,
    basket -> '$.extras' AS extras,
    delivery ->> '$.address' AS address
FROM orders
WHERE basket ->> '$.fish' = $1
ORDER BY id;

example.go

package main

import (
	"context"
	"database/sql"
	"encoding/json"
	"fmt"
	"io"

	"seagull.sh/seagull/examples/duckdb/006_json/internal/store"
)

// basket is the document an order's basket column holds.
type basket struct {
	// Fish is the fish the order is for.
	Fish string `json:"fish"`

	// Portions is how many portions of fish and chips.
	Portions int `json:"portions"`

	// Extras lists the sides.
	Extras []string `json:"extras"`
}

// delivery is the document an order's delivery column holds, when the order is delivered.
type delivery struct {
	// Address is where the order goes.
	Address string `json:"address"`
}

// run places three chip shop orders and reads them back. A JSON column read by a query
// becomes the generated store.JSON type, which holds the raw document as bytes: decode it
// with json.Unmarshal. A nil store.JSON is SQL NULL.
//
// The generated SQL casts each JSON column it reads to VARCHAR, so store.JSON holds the
// document as DuckDB stored it, the keys in the order they were written.
//
// PlaceOrder takes its documents as store.JSON too, and the generated code hands
// duckdb-go their text; DuckDB's JSON type is VARCHAR underneath and checks the text as
// it is stored.
//
// OrdersWithFish filters with ->>, which extracts a value as text, and selects with ->,
// which extracts a JSON fragment, so Extras is store.JSON again. Address is *string
// because ->> gives NULL when the delivery column or its address is missing.
//
// Takes database (*sql.DB) which is a migrated database.
// Takes output (io.Writer) which receives the decoded orders and the search.
//
// Returns error when a query fails or a document does not encode or decode.
func run(ctx context.Context, database *sql.DB, output io.Writer) error {
	queries := store.New(database)

	orders := []struct {
		customer string
		basket   basket
		delivery *delivery
	}{
		{customer: "Steven", basket: basket{Fish: "cod", Portions: 2, Extras: []string{"mushy peas", "curry sauce"}}},
		{customer: "Gwen", basket: basket{Fish: "haddock", Portions: 1, Extras: []string{}}},
		{customer: "Morwenna", basket: basket{Fish: "cod", Portions: 1, Extras: []string{"pickled egg"}}, delivery: &delivery{Address: "4 Harbour Row"}},
	}
	for _, order := range orders {
		encodedBasket, err := json.Marshal(order.basket)
		if err != nil {
			return fmt.Errorf("encoding %s's basket: %w", order.customer, err)
		}
		parameters := store.PlaceOrderParams{Customer: order.customer, Basket: encodedBasket, Delivery: nil}
		if order.delivery != nil {
			encodedDelivery, err := json.Marshal(order.delivery)
			if err != nil {
				return fmt.Errorf("encoding %s's delivery: %w", order.customer, err)
			}
			parameters.Delivery = encodedDelivery
		}
		if err := queries.PlaceOrder(ctx, parameters); err != nil {
			return fmt.Errorf("placing %s's order: %w", order.customer, err)
		}
	}

	first, err := queries.GetOrder(ctx, 1)
	if err != nil {
		return fmt.Errorf("getting order 1: %w", err)
	}
	var decoded basket
	if err := json.Unmarshal(first.Basket, &decoded); err != nil {
		return fmt.Errorf("decoding order 1's basket: %w", err)
	}
	fmt.Fprintf(output, "order 1 stored %s\n", first.Basket)
	fmt.Fprintf(output, "decoded: %d x %s with %v, delivery is NULL: %t\n", decoded.Portions, decoded.Fish, decoded.Extras, first.Delivery == nil)

	cod, err := queries.OrdersWithFish(ctx, "cod")
	if err != nil {
		return fmt.Errorf("finding cod orders: %w", err)
	}
	fmt.Fprintln(output, "\ncod orders:")
	for _, order := range cod {
		address := "collecting"
		if order.Address != nil {
			address = "to " + *order.Address
		}
		fmt.Fprintf(output, "  order %d  %-9s extras %-30s %s\n", order.ID, order.Customer, order.Extras, address)
	}
	return nil
}

main.go

// Command 006_json applies its migrations to an in-memory DuckDB database, then runs the
// example in example.go: three chip shop orders are stored as JSON documents, read back
// and searched with -> and ->>.
//
// Run it with `go run .`, or `go run . -db example.duckdb` to keep the database in a
// file.
package main

import (
	"context"
	"database/sql"
	"embed"
	"flag"
	"fmt"
	"log"
	"os"

	_ "github.com/duckdb/duckdb-go/v2"

	"seagull.sh/seagull/sdk/migrate"
	"seagull.sh/seagull/sdk/migrate/migrate_backend_duckdb"
)

// migrations holds the migration files, compiled into the binary.
//
//go:embed db/migrations
var migrations embed.FS

func main() {
	path := flag.String("db", "", "DuckDB database file; an in-memory database when empty")
	flag.Parse()
	if err := execute(context.Background(), *path); err != nil {
		log.Fatal(err)
	}
}

// execute opens and migrates the DuckDB database at path, or an in-memory one when path
// is empty, and runs the example against it.
//
// Takes path (string) which is the database file, or empty.
//
// Returns error when the database cannot be prepared or the example fails.
func execute(ctx context.Context, path string) error {
	database, err := sql.Open("duckdb", path)
	if err != nil {
		return err
	}
	defer func() { _ = database.Close() }()

	migrator, err := migrate.NewMigrator(database, migrate_backend_duckdb.New(), migrations, "db/migrations")
	if err != nil {
		return err
	}
	if _, err := migrator.Up(ctx); err != nil {
		return fmt.Errorf("migrating: %w", err)
	}
	return run(ctx, database, os.Stdout)
}

internal/store/driver_values.go

// Code generated by sgl. DO NOT EDIT.

package store

import "database/sql/driver"

type sglDriverValue struct {
	value any
}

func (wrapped sglDriverValue) Value() (driver.Value, error) {
	return driver.DefaultParameterConverter.ConvertValue(wrapped.value)
}

internal/store/models.go

// Code generated by sgl. DO NOT EDIT.

package store

type Orders struct {
	ID       int64  `json:"id"`
	Customer string `json:"customer"`
	Basket   JSON   `json:"basket"`
	Delivery JSON   `json:"delivery"`
}

internal/store/orders.sql.go

// Code generated by sgl. DO NOT EDIT.

package store

import "context"

const getOrderSQL = `SELECT id, customer, CAST(basket AS VARCHAR) AS "basket", CAST(delivery AS VARCHAR) AS "delivery" FROM orders WHERE id = $1;`

type GetOrderRow struct {
	ID       int64  `json:"id"`
	Customer string `json:"customer"`
	Basket   JSON   `json:"basket"`
	Delivery JSON   `json:"delivery"`
}

func (queries *Queries) GetOrder(ctx context.Context, id int64) (GetOrderRow, error) {
	var row GetOrderRow
	err := queries.reader.QueryRowContext(ctx, getOrderSQL, id).Scan(&row.ID, &row.Customer, &row.Basket, &row.Delivery)
	if err != nil {
		return GetOrderRow{}, err
	}
	return row, nil
}

const ordersWithFishSQL = `SELECT
    id,
    customer,
    CAST(basket -> '$.extras' AS VARCHAR) AS extras,
    delivery ->> '$.address' AS address
FROM orders
WHERE basket ->> '$.fish' = $1
ORDER BY id;`

type OrdersWithFishRow struct {
	ID       int64   `json:"id"`
	Customer string  `json:"customer"`
	Extras   JSON    `json:"extras"`
	Address  *string `json:"address"`
}

func (queries *Queries) OrdersWithFish(ctx context.Context, fish string) ([]OrdersWithFishRow, error) {
	rows, err := queries.reader.QueryContext(ctx, ordersWithFishSQL, fish)
	if err != nil {
		return nil, err
	}
	defer rows.Close()
	var results []OrdersWithFishRow
	for rows.Next() {
		var row OrdersWithFishRow
		if err := rows.Scan(&row.ID, &row.Customer, &row.Extras, &row.Address); err != nil {
			return nil, err
		}
		results = append(results, row)
	}
	if err := rows.Err(); err != nil {
		return nil, err
	}
	return results, nil
}

const placeOrderSQL = `INSERT INTO orders (customer, basket, delivery) VALUES ($1, $2, $3);`

type PlaceOrderParams struct {
	Customer string
	Basket   JSON
	Delivery JSON
}

func (queries *Queries) PlaceOrder(ctx context.Context, params PlaceOrderParams) error {
	_, err := queries.writer.ExecContext(ctx, placeOrderSQL, params.Customer, sglDriverValue{value: params.Basket}, sglDriverValue{value: params.Delivery})
	return err
}

internal/store/querier.go

// Code generated by sgl. DO NOT EDIT.

package store

import (
	"context"
	"database/sql"
)

type DBTX interface {
	ExecContext(ctx context.Context, query string, args ...any) (sql.Result, error)
	QueryContext(ctx context.Context, query string, args ...any) (*sql.Rows, error)
	QueryRowContext(ctx context.Context, query string, args ...any) *sql.Row
}

type dbtxWrapper interface {
	WrapDBTX(inner any) any
}

type Queries struct {
	reader  DBTX
	writer  DBTX
	wrapper dbtxWrapper
}

func New(db DBTX) *Queries {
	wrapper, _ := db.(dbtxWrapper)
	return &Queries{reader: db, writer: db, wrapper: wrapper}
}

func NewWithReplica(writer DBTX, reader DBTX) *Queries {
	wrapper, _ := writer.(dbtxWrapper)
	return &Queries{reader: reader, writer: writer, wrapper: wrapper}
}

func (queries *Queries) WithTx(transaction *sql.Tx) *Queries {
	if queries == nil {
		return &Queries{reader: transaction, writer: transaction, wrapper: nil}
	}
	db := queries.wrapDBTX(transaction)
	return &Queries{reader: db, writer: db, wrapper: queries.wrapper}
}

func (queries *Queries) wrapDBTX(db DBTX) DBTX {
	if queries.wrapper == nil {
		return db
	}
	wrapped, ok := queries.wrapper.WrapDBTX(db).(DBTX)
	if !ok {
		return db
	}
	return wrapped
}

func (queries *Queries) RunInTx(ctx context.Context, db *sql.DB, fn func(*Queries) error) error {
	transaction, err := db.BeginTx(ctx, nil)
	if err != nil {
		return err
	}
	defer transaction.Rollback()
	if err = fn(queries.WithTx(transaction)); err != nil {
		return err
	}
	return transaction.Commit()
}

internal/store/sgl_json.go

// Code generated by sgl. DO NOT EDIT.

package store

import (
	"database/sql/driver"
	"encoding/json"
	"fmt"
)

type JSON []byte

func (j *JSON) Scan(source any) error {
	switch value := source.(type) {
	case nil:
		*j = nil
		return nil
	case []byte:
		clone := make(JSON, len(value))
		copy(clone, value)
		*j = clone
		return nil
	case string:
		*j = JSON(value)
		return nil
	default:
		encoded, err := json.Marshal(value)
		if err != nil {
			return fmt.Errorf("sgl: encode %T to JSON: %w", source, err)
		}
		*j = encoded
		return nil
	}
}

func (j JSON) Value() (driver.Value, error) {
	if j == nil {
		return nil, nil
	}
	return string(j), nil
}

func (j JSON) MarshalJSON() ([]byte, error) {
	if j == nil {
		return []byte("null"), nil
	}
	return j, nil
}

func (j *JSON) UnmarshalJSON(data []byte) error {
	clone := make(JSON, len(data))
	copy(clone, data)
	*j = clone
	return nil
}

output.txt

order 1 stored {"fish":"cod","portions":2,"extras":["mushy peas","curry sauce"]}
decoded: 2 x cod with [mushy peas curry sauce], delivery is NULL: true

cod orders:
  order 1  Steven    extras ["mushy peas","curry sauce"]   collecting
  order 3  Morwenna  extras ["pickled egg"]                to 4 Harbour Row