Skip to content

Implementing GraphQL with joins to avoid the N+1 problem

Notifications You must be signed in to change notification settings

healx/python-graphjoiner

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

GraphJoiner: Implementing GraphQL with joins

GraphJoiner has been superseded by GraphLayer, which uses the same fundamental ideas with a significantly simpler API.

In the reference GraphQL implementation, resolve functions describe how to fulfil some part of the requested data for each instance of an object. If implemented naively with a SQL backend, this results in the N+1 problem. For instance, given the query:

{
    books(genre: "comedy") {
        title
        author {
            name
        }
    }
}

A naive GraphQL implementation would issue one SQL query to get the list of all books in the comedy genre, and then N queries to get the author of each book (where N is the number of books returned by the first query).

There are various solutions proposed to this problem: GraphJoiner suggests that using joins is a natural fit for many use cases. For this specific case, we only need to run two queries: one to find the list of all books in the comedy genre, and one to get the authors of books in the comedy genre.

Installation

pip install graphjoiner

Example

Let's say we have some models defined by SQLAlchemy. A book has an ID, a title, a genre and an author ID. An author has an ID and a name.

from sqlalchemy import Column, Integer, Unicode, ForeignKey
from sqlalchemy.ext.declarative import declarative_base

Base = declarative_base()

class AuthorRecord(Base):
    __tablename__ = "author"

    id = Column(Integer, primary_key=True)
    name = Column(Unicode, nullable=False)

class BookRecord(Base):
    __tablename__ = "book"

    id = Column(Integer, primary_key=True)
    title = Column(Unicode, nullable=False)
    genre = Column(Unicode, nullable=False)
    author_id = Column(Integer, ForeignKey(AuthorRecord.id))

We then define object types for the root, books and authors:

from graphjoiner.declarative import RootType, single, many, select, String
from graphjoiner.declarative.sqlalchemy import SqlAlchemyObjectType, column_field, sql_join

class Author(SqlAlchemyObjectType):
    __model__ = AuthorRecord

    id = column_field(AuthorRecord.id)
    name = column_field(AuthorRecord.name)

class Book(SqlAlchemyObjectType):
    __model__ = BookRecord

    id = column_field(BookRecord.id)
    title = column_field(BookRecord.title)
    genre = column_field(BookRecord.genre)
    author_id = column_field(BookRecord.author_id)
    author = single(lambda: sql_join(Author))

class Root(RootType):
    books = many(lambda: select(Book))

    @books.arg("genre", String)
    def books_arg_genre(query, genre):
        return query.filter(BookRecord.genre == genre)

We create an execute() function by calling executor() with our Root:

from graphjoiner.declarative import executor

execute = executor(Root)

execute can then be used to execute queries:

query = """
    {
        books(genre: "comedy") {
            title
            author {
                name
            }
        }
    }
"""

class Context(object):
    def __init__(self, session):
        self.session = session

result = execute(root, query, context=Context(session))

Where result.data is:

{
    "books": [
        {
            "title": "Leave It to Psmith",
            "author": {
                "name": "PG Wodehouse"
            }
        },
        {
            "title": "Right Ho, Jeeves",
            "author": {
                "name": "PG Wodehouse"
            }
        },
        {
            "title": "Catch-22",
            "author": {
                "name": "Joseph Heller"
            }
        },
    ]
}

Let's break things down a little, starting with the definition of Author:

class Author(SqlAlchemyObjectType):
    __model__ = AuthorRecord

    id = column_field(AuthorRecord.id)
    name = column_field(AuthorRecord.name)

When defining object types that represent SQLAlchemy models, we can inherit from SqlAlchemyObjectType, with the __model__ attribute set to the appropriate model.

Fields that can be fetched without further joining can be defined using column_field(). GraphJoiner will automatically infer the GraphQL type of the field based on the SQL type of the column.

Next is the definition of Book:

class Book(SqlAlchemyObjectType):
    __model__ = BookRecord

    id = column_field(BookRecord.id)
    title = column_field(BookRecord.title)
    genre = column_field(BookRecord.genre)
    author_id = column_field(BookRecord.author_id)
    author = single(lambda: sql_join(Author))

As before, we inherit from SqlAlchemyObjectType, set __model__ to the appropriate class, and define a number of fields that correspond to columns.

We also define an author field that allows a book to be joined to an author. GraphJoiner will automatically inspect BookRecord and AuthorRecord and use the foreign keys to determine how they should be joined together. To override this behaviour, you can pass in an explicit join argument:

author = single(lambda: sql_join(Author, join={Book.author_id: Author.id}))

This explicitly tells GraphJoiner that authors can be joined to books by equality between the fields Book.author_id and Author.id. When defining relationships such as this, we call single() with a lambda to defer evaluation until all of the types and fields have been defined.

Finally, we can create a root object:

class Root(RootType):
    books = many(lambda: select(Book))

    @books.arg("genre", String)
    def books_arg_genre(query, genre):
        return query.filter(BookRecord.genre == genre)

The root has only one field, books, which we define using many(). Using select tells GraphJoiner to select all of the books in the database, rather than trying to perform a join.

Using books.arg() adds an optional argument to the field.

For completeness, we can tweak the definition of Author so we can request the books by an author:

class Author(SqlAlchemyObjectType):
    __model__ = AuthorRecord

    id = column_field(AuthorRecord.id)
    name = column_field(AuthorRecord.name)
    books = many(lambda: sql_join(Book))

API

graphjoiner.declarative

ObjectType

Represents a GraphQL object type. Fields can be declared as attributes. For instance, to create an object type called User with a name and emailAddress field:

from graphqjoiner import NonNull, ObjectType, String

class User(ObjectType):
    name = NonNull(String)
    email_address = NonNull(String)

Field names are inferred from attribute names, converting from snake case to camel case. In the example above, the attribute name email_address is converted to the field name emailAddress.

To create a type that can be joined to, implement __fetch_immediates__ as a static or class method.

  • __fetch_immediates__(selections, query, context): fetch the values for the selected fields that aren't defined as relationships.

    Receives the arguments:

    • selections: an iterable of the selections, where each selection has the attributes:
      • field: the field being selected
      • args: the arguments for the selection
      • selections: the sub-selections of that selection
    • query: the query for the records to select.
    • context: the context as passed into the executor

    Should return a list of tuples, where each tuple contains the value for each selection in the same order.

Implementing __select_all__ allows the object to be used with select(). __select_all__() takes no arguments, and should return a query that represents all instances of the object.

For instance, to implement a base type for static data:

import collections

from graphjoiner.declarative import ObjectType, RootType, select, single, String

class StaticDataObjectType(ObjectType):
    @classmethod
    def __select_all__(cls):
        return cls.__records__

    @classmethod
    def __fetch_immediates__(cls, selections, records, context):
        return [
            tuple(
                getattr(record, selection.field.attr_name)
                for selection in selections
            )
            for record in records
        ]

AuthorRecord = collections.namedtuple("AuthorRecord", ["name"])

class Author(StaticDataObjectType):
    __records__ = [AuthorRecord("PG Wodehouse")]

    name = field(type=String)

class Root(RootType):
    author = single(lambda: select(Author))

Relationships

Use single, single_or_null, first_or_null and many to create fields that are joined to other types. For instance, to select all books from the root type:

from graphjoiner.declarative import many, RootType, select

class Root(RootType):
    ...
    books = many(lambda: select(Book))

Each relationship function accepts a joiner: a value that describes how to join the left type to the right type. The joiner is always wrapped in a lambda to defer evaluation until all types are defined. In this case, the left type is Root, the right type is Book, and the joiner is select(Book). Calling select() with just the right type tells GraphJoiner to select all values, in this case all books.

All joiners accept a filter argument that allow the query to be tweaked. For instance, supposing books are selected using SQLAlchemy queries, and we want the books field to be sorted by title:

from graphjoiner.declarative import many, RootType, select
from graphjoiner.declarative.sqlalchemy import SqlAlchemyObjectType

class Book(SqlAlchemyObjectType):
    __model__ = BookRecord

    ...

    @staticmethod
    def order_by_title(query):
        # query is an instance of sqlalchemy.orm.Query
        return query.order_by(BookRecord.title)

class Root(RootType):
    ...

    books = many(lambda: select(
        Book,
        filter=Book.order_by_title,
    ))

Arguments can be added using the arg() decorator. If the GraphQL selection for that field includes a value for the argument, the query is updated using the decorated function. For instance, to allow books to be filtered by title:

from graphjoiner.declarative import many, RootType, select, String
from graphjoiner.declarative.sqlalchemy import SqlAlchemyObjectType

class Book(SqlAlchemyObjectType):
    __model__ = BookRecord

    ...

    @staticmethod
    def filter_by_title(query, title):
        # query is an instance of sqlalchemy.orm.Query
        return query.filter(BookRecord.title == title)

class Root(RootType):
    ...

    books = many(lambda: select(Book))
    @books.arg("title", String)
    def books_arg_title(query, title):
        return Book.filter_by_title(query, title)

select(target, join_query=None, join_fields=None)

Creates a joiner to the target type. When given no additional arguments, it will select all values of the target type using target.__select_all__(). All left values are joined onto all right values i.e. the join is the cartesian product. Unless the left type is the root type, this probably isn't what you want.

Set join_fields to describe which fields to use to join together the left and right types. Each item in the dictionary should map a field from the left type to a field from the right type. For instance, supposing each author has a unique ID, and each book has an author ID:

from graphjoiner.declarative import field, Int, ObjectType, select, single

class Book(ObjectType):
    ...
    author_id = field(type=Int)
    author = single(lambda: select(
        Author,
        join_fields={Book.author_id: Author.id},
    ))

Set join_query to describe how to join the left query and the right query. This should be a function that accepts a left query and a right query, and returns a right query filtered to the values relevant to the left query. This avoids the cost of fetching all values of the right type only to discard those that don't join onto any left values. For instance, when using the sqlalchemy module, we'd like to fetch the authors for just the requested book, rather than all available authors:

from graphjoiner.declarative import select, single
from graphjoiner.declarative.sqlalchemy import column_field, SqlAlchemyObjectType

class Book(SqlAlchemyObjectType):
    ...
    author_id = column_field(BookRecord.author_id)

    def join_authors(book_query, author_query):
        author_ids = book_query \
            .add_columns(BookRecord.author_id) \
            .subquery()

        return author_query.join(
            author_ids,
            author_ids.c.author_id == AuthorRecord.id,
        )

    author = single(lambda: select(
        Author,
        join_query=join_authors,
        join_fields={Book.author_id: Author.id},
    ))

In this particular case, using sql_join() would remove much of the boilerplate:

from graphjoiner.declarative import single
from graphjoiner.declarative.sqlalchemy import column_field, sql_join, SqlAlchemyObjectType

class Book(SqlAlchemyObjectType):
    ...
    author_id = column_field(BookRecord.author_id)
    author = single(lambda: sql_join(Author, {Book.author_id: Author.id}))

extract(field, sub_field)

Create a new field by extracting sub_field from field. The arguments for the new field are the same as the arguments for field.

For instance, supposing we have a field books on the root type, each book has a title field, and we want to add a bookTitles field to the root type:

from graphjoiner.declarative import extract, many, RootType, select

class Root(RootType):
    books = many(lambda: select(Book))
    book_titles = extract(books, lambda: Book.title)

If we want to just have the bookTitles field without a books field, we can pass the relationship directly into extract():

from graphjoiner.declarative import extract, many, RootType, select

class Root(RootType):
    book_titles = extract(
        many(lambda: select(Book)),
        lambda: Book.title,
    )

extract() is often useful when modelling many-to-many relationships. For instance, suppose a book may have many publishers, and each publisher may publish many books. We define a type that associates books and publishers:

from graphjoiner.declarative import ObjectType, select, single

class BookPublisherAssociation(ObjectType):
    book = single(lambda: select(Book, ...))
    publisher = single(lambda: select(Publisher, ...))

We can then use extract to define a field for all publishers of a book, and a field for books from a publisher:

from graphjoiner.declarative import extract, many, ObjectType, select

class Book(ObjectType):
    ...
    publishers = extract(
        many(lambda: select(BookPublisherAssociation, ...)),
        lambda: BookPublisherAssociation.publisher,
    )

class Publisher(ObjectType):
    ...
    books = extract(
        many(lambda: select(BookPublisherAssociation, ...)),
        lambda: BookPublisherAssociation.book,
    )

Interfaces

To define an interface, subclass InterfaceType and specify fields using field():

from graphjoiner.declarative import InterfaceType, String

class HasName(InterfaceType):
    name = field(type=String)

To set which interfaces an object implements, set the __interfaces__ attribute:

from graphjoiner.declarative import ObjectType

class Author(ObjectType):
    __interfaces__ = lambda: [HasName]
    ...

Field sets

Field sets can be used to define multiple fields using a single attribute. For instance, this definition without field sets:

from graphjoiner.declarative import field, Int, ObjectType, String

class Book(ObjectType):
    title = field(type=String)
    author_id = field(type=Int)

is roughly equivalent to this definition using field sets:

from graphjoiner.declarative import field, field_set, ObjectType, String

class Book(ObjectType):
    fields = field_set(
        title=field(type=String),
        author_id=field(type=String),
    )

Field sets are useful when a set of fields needs to be generated dynamically.

Input object types

Define input types by inheriting from InputObjectType, and defining fields using field(). For instance:

from graphjoiner.declarative import InputObjectType, String

class BookSelectionInput(InputObjectType):
    title = field(type=String, default=None)

The fields on input object values are available as attributes. For instance:

from graphjoiner.declarative import many, RootType, select, String
from graphjoiner.declarative.sqlalchemy import SqlAlchemyObjectType

class Book(SqlAlchemyObjectType):
    __model__ = BookRecord

    ...

    @staticmethod
    def filter_by_title(query, title):
        # query is an instance of sqlalchemy.orm.Query
        return query.filter(BookRecord.title == title)

class Root(RootType):
    ...

    books = many(lambda: select(Book))
    @books.arg("selection", BookSelectionInput)
    def books_arg_title(query, title):
        if selection.title is not None:
            query = Book.filter_by_title(query, title)

        return query

The default value for each field can be set by passing the default argument to each field. To allow the absence of a value to be distinguished from an explicit null value, the default value for a field is undefined if the default argument is not set. For instance, to allow books to be filtered by title, including null titles:

from graphjoiner.declarative import field, InputObjectType, many, RootType, select, String, undefined
from graphjoiner.declarative.sqlalchemy import SqlAlchemyObjectType

class BookSelectionInput(InputObjectType):
    title = field(type=String)

class Book(SqlAlchemyObjectType):
    __model__ = BookRecord

    ...

    @staticmethod
    def filter_by_title(query, title):
        # query is an instance of sqlalchemy.orm.Query
        return query.filter(BookRecord.title == title)

class Root(RootType):
    ...

    books = many(lambda: select(Book))
    @books.arg("selection", BookSelectionInput)
    def books_arg_title(query, title):
        if selection.title is not undefined:
            query = Book.filter_by_title(query, title)

        return query

Core Example

The declarative API of GraphJoiner is built on top of a core API. The core API exposes the fundamentals of how GraphJoiner works, giving greater flexibility at the cost of being rather verbose to use directly. The below shows how the original example could be written using the core API. In general, using the declarative API should be preferred, either by using the built-in tools or adding your own.

Let's say we have some models defined by SQLAlchemy. A book has an ID, a title, a genre and an author ID. An author has an ID and a name.

from sqlalchemy import Column, Integer, Unicode, ForeignKey
from sqlalchemy.ext.declarative import declarative_base

Base = declarative_base()

class Author(Base):
    __tablename__ = "author"

    id = Column(Integer, primary_key=True)
    name = Column(Unicode, nullable=False)

class Book(Base):
    __tablename__ = "book"

    id = Column(Integer, primary_key=True)
    title = Column(Unicode, nullable=False)
    genre = Column(Unicode, nullable=False)
    author_id = Column(Integer, ForeignKey(Author.id))

We then define object types for the root, books and authors:

from graphql import GraphQLInt, GraphQLString, GraphQLArgument
from graphjoiner import JoinType, RootJoinType, single, many, field
from sqlalchemy.orm import Query

def create_root():
    def fields():
        return {
            "books": many(
                book_join_type,
                books_query,
                args={"genre": GraphQLArgument(type=GraphQLString)}
            )
        }

    def books_query(args, _):
        query = Query([]).select_from(Book)

        if "genre" in args:
            query = query.filter(Book.genre == args["genre"])

        return query

    return RootJoinType(name="Root", fields=fields)

root = create_root()

def fetch_immediates_from_database(selections, query, context):
    query = query.with_entities(*(
        selection.field.column_name
        for selection in selections
    ))

    return query.with_session(context.session).all()

def create_book_join_type():
    def fields():
        return {
            "id": field(column_name="id", type=GraphQLInt),
            "title": field(column_name="title", type=GraphQLString),
            "genre": field(column_name="genre", type=GraphQLString),
            "authorId": field(column_name="author_id", type=GraphQLInt),
            "author": single(author_join_type, author_query, join={"authorId": "id"}),
        }

    def author_query(args, book_query):
        books = book_query.with_entities(Book.author_id).distinct().subquery()
        return Query([]) \
            .select_from(Author) \
            .join(books, books.c.author_id == Author.id)

    return JoinType(
        name="Book",
        fields=fields,
        fetch_immediates=fetch_immediates_from_database,
    )

book_join_type = create_book_join_type()

def create_author_join_type():
    def fields():
        return {
            "id": field(column_name="id", type=GraphQLInt),
            "name": field(column_name="name", type=GraphQLString),
        }

    return JoinType(
        name="Author",
        fields=fields,
        fetch_immediates=fetch_immediates_from_database,
    )
author_join_type = create_author_join_type()

We can execute the query by calling execute:

from graphjoiner import execute

query = """
    {
        books(genre: "comedy") {
            title
            author {
                name
            }
        }
    }
"""

class Context(object):
    def __init__(self, session):
        self.session = session

execute(root, query, context=Context(session))

Which produces:

{
    "books": [
        {
            "title": "Leave It to Psmith",
            "author": {
                "name": "PG Wodehouse"
            }
        },
        {
            "title": "Right Ho, Jeeves",
            "author": {
                "name": "PG Wodehouse"
            }
        },
        {
            "title": "Catch-22",
            "author": {
                "name": "Joseph Heller"
            }
        },
    ]
}

Let's break things down a little, starting with the definition of the root object:

def create_root():
    def fields():
        return {
            "books": many(
                book_join_type,
                books_query,
                args={"genre": GraphQLArgument(type=GraphQLString)}
            )
        }

    def books_query(args, _):
        query = Query([]).select_from(Book)

        if "genre" in args:
            query = query.filter(Book.genre == args["genre"])

        return query

    return RootJoinType(name="Root", fields=fields)

root = create_root()

For each object type, we need to define its fields. The root has only one field, books, a one-to-many relationship, which we define using many(). The first argument, book_join_type, is the type we're defining a relationship to. The second argument to describes how to create a query representing all of those related books: in this case all books, potentially filtered by a genre argument.

This means we need to define book_join_type:

def create_book_join_type():
    def fields():
        return {
            "id": field(column_name="id", type=GraphQLInt),
            "title": field(column_name="title", type=GraphQLString),
            "genre": field(column_name="genre", type=GraphQLString),
            "authorId": field(column_name="author_id", type=GraphQLInt),
            "author": single(author_join_type, author_query, join={"authorId": "id"}),
        }

    def author_query(args, book_query):
        books = book_query.with_entities(Book.author_id).distinct().subquery()
        return Query([]) \
            .select_from(Author) \
            .join(books, books.c.author_id == Author.id)

    return JoinType(
        name="Book",
        fields=fields,
        fetch_immediates=fetch_immediates_from_database,
    )

book_join_type = create_book_join_type()

The author field is defined as a one-to-one mapping from book to author. As before, we define a function that generates a query for the requested authors. We also provide a join argument to single() so that GraphJoiner knows how to join together the results of the author query and the book query: in this case, the authorId field on books corresponds to the id field on authors. (If we leave out the join argument, then GraphJoiner will perform a cross join i.e. a cartesian product. Since there's always exactly one root instance, this is fine for relationships defined on the root.)

The remaining fields define a mapping from the GraphQL field to the database column. This mapping is handled by fetch_immediates_from_database. The value of selections in fetch_immediates() is the selections of fields that aren't defined as relationships (using single or many) that were either explicitly requested in the original GraphQL query, or are required as part of the join.

def fetch_immediates_from_database(selections, query, context):
    query = query.with_entities(*(
        fields[selection.field_name].column_name
        for selection in selections
    ))

    return query.with_session(context.session).all()

For completeness, we can tweak the definition of author_join_type so we can request the books by an author:

def create_author_join_type():
    def fields():
        return {
            "id": field(column_name="id", type=GraphQLInt),
            "name": field(column_name="name", type=GraphQLString),
            "author": many(book_join_type, book_query, join={"id": "authorId"}),
        }

    def book_query(args, author_query):
        authors = author_query.with_entities(Author.id).distinct().subquery()
        return Query([]) \
            .select_from(Book) \
            .join(authors, authors.c.id == Book.author_id)

    return JoinType(
        name="Author",
        fields=fields,
        fetch_immediates=fetch_immediates_from_database,
    )

author_join_type = create_author_join_type()

About

Implementing GraphQL with joins to avoid the N+1 problem

Resources

Stars

Watchers

Forks

Packages

No packages published

Languages

  • Python 99.6%
  • Makefile 0.4%