Asyncio DataLoader

NOTICE: This is a fork of aiodataloader by Syrus Akbary to maintain a custom release cycle to match with our product roadmap. Of course, all changes made by Lablup Inc. and its members are made available with the same open source terms of the MIT license. If the upstream repository resumes to a faster release cycle with latest contributions, we will follow it.

DataLoader is a generic utility to be used as part of your application's data fetching layer to provide a simplified and consistent API over various remote data sources such as databases or web services via batching and caching.

A port of the "Loader" API originally developed by @schrockn at Facebook in 2010 as a simplifying force to coalesce the sundry key-value store back-end APIs which existed at the time. At Facebook, "Loader" became one of the implementation details of the "Ent" framework, a privacy-aware data entity loading and caching layer within web server product code. This ultimately became the underpinning for Facebook's GraphQL server implementation and type definitions.

DataLoader is a simplified version of this original idea implemented in Python for AsyncIO services. DataLoader is often used when implementing a graphene service, though it is also broadly useful in other situations.

DataLoader is provided so that it may be useful not just to build GraphQL services with AsyncIO but also as a publicly available reference implementation of this concept in the hopes that it can be ported to other languages. If you port DataLoader to another language, please open an issue to include a link from this repository.

Getting Started

First, install DataLoader using pip.

pip install aiodataloader

To get started, create a DataLoader. Each DataLoader instance represents a unique cache. Typically instances are created per request when used within a web-server like Sanic if different users can see different things.

Note: DataLoader assumes a AsyncIO environment with async/await available only in Python 3.5+.

Batching

Batching is not an advanced feature, it's DataLoader's primary feature. Create loaders by providing a batch loading function.

from aiodataloader import DataLoader

class UserLoader(DataLoader):
    async def batch_load_fn(self, keys):
        return await my_batch_get_users(keys)

user_loader = UserLoader()

A batch loading function accepts a Iterable of keys, and returns a Promise which resolves to a List of values^*.

Then load individual values from the loader. DataLoader will coalesce all individual loads which occur within a single frame of execution (a single tick of the event loop) and then call your batch function with all requested keys.

user1_future = user_loader.load(1)
user2_future = user_loader.load(2)

user1 = await user1_future
user2 = await user2_future

user1_invitedby = user_loader.load(user1.invited_by_id)
user2_invitedby = user_loader.load(user2.invited_by_id)

print("User 1 was invited by", await user1_invitedby)
print("User 2 was invited by", await user2_invitedby)

A naive application may have issued four round-trips to a backend for the required information, but with DataLoader this application will make at most two.

DataLoader allows you to decouple unrelated parts of your application without sacrificing the performance of batch data-loading. While the loader presents an API that loads individual values, all concurrent requests will be coalesced and presented to your batch loading function. This allows your application to safely distribute data fetching requirements throughout your application and maintain minimal outgoing data requests.

Batch Function

A batch loading function accepts an List of keys, and returns a Future which resolves to an List of values. There are a few constraints that must be upheld:

• The List of values must be the same length as the List of keys.
• Each index in the List of values must correspond to the same index in the List of keys.

For example, if your batch function was provided the List of keys: [ 2, 9, 6, 1 ], and loading from a back-end service returned the values:

{ 'id': 9, 'name': 'Chicago' }
{ 'id': 1, 'name': 'New York' }
{ 'id': 2, 'name': 'San Francisco' }

Our back-end service returned results in a different order than we requested, likely because it was more efficient for it to do so. Also, it omitted a result for key 6, which we can interpret as no value existing for that key.

To uphold the constraints of the batch function, it must return an List of values the same length as the List of keys, and re-order them to ensure each index aligns with the original keys [ 2, 9, 6, 1 ]:

[
  { 'id': 2, 'name': 'San Francisco' },
  { 'id': 9, 'name': 'Chicago' },
  None,
  { 'id': 1, 'name': 'New York' }
]

Caching

DataLoader provides a memoization cache for all loads which occur in a single request to your application. After .load() is called once with a given key, the resulting value is cached to eliminate redundant loads.

In addition to relieving pressure on your data storage, caching results per-request also creates fewer objects which may relieve memory pressure on your application:

user_future1 = user_loader.load(1)
user_future2 = user_loader.load(1)

assert user_future1 == user_future2

Caching per-Request

DataLoader caching does not replace Redis, Memcache, or any other shared application-level cache. DataLoader is first and foremost a data loading mechanism, and its cache only serves the purpose of not repeatedly loading the same data in the context of a single request to your Application. To do this, it maintains a simple in-memory memoization cache (more accurately: .load() is a memoized function).

Avoid multiple requests from different users using the DataLoader instance, which could result in cached data incorrectly appearing in each request. Typically, DataLoader instances are created when a Request begins, and are not used once the Request ends.

For example, when using with Sanic:

def create_loaders(auth_token) {
    return {
      'users': user_loader,
    }
}


app = Sanic(__name__)

@app.route("/")
async def test(request):
    auth_token = authenticate_user(request)
    loaders = create_loaders(auth_token)
    return render_page(request, loaders)

Clearing Cache

In certain uncommon cases, clearing the request cache may be necessary.

The most common example when clearing the loader's cache is necessary is after a mutation or update within the same request, when a cached value could be out of date and future loads should not use any possibly cached value.

Here's a simple example using SQL UPDATE to illustrate.

# Request begins...
user_loader = ...

# And a value happens to be loaded (and cached).
user4 = await user_loader.load(4)

# A mutation occurs, invalidating what might be in cache.
await sql_run('UPDATE users WHERE id=4 SET username="zuck"')
user_loader.clear(4)

# Later the value load is loaded again so the mutated data appears.
user4 = await user_loader.load(4)

# Request completes.

Caching Exceptions

If a batch load fails (that is, a batch function throws or returns a rejected Promise), then the requested values will not be cached. However if a batch function returns an Exception instance for an individual value, that Exception will be cached to avoid frequently loading the same Exception.

In some circumstances you may wish to clear the cache for these individual Errors:

try:
    user_loader.load(1)
except Exception as e:
    user_loader.clear(1)
    raise

Disabling Cache

In certain uncommon cases, a DataLoader which does not cache may be desirable. Calling DataLoader(batch_fn, cache=false) will ensure that every call to .load() will produce a new Future, and requested keys will not be saved in memory.

However, when the memoization cache is disabled, your batch function will receive an array of keys which may contain duplicates! Each key will be associated with each call to .load(). Your batch loader should provide a value for each instance of the requested key.

For example:

class MyLoader(DataLoader):
    cache = False
    async def batch_load_fn(self, keys):
        print(keys)
        return keys

my_loader = MyLoader()

my_loader.load('A')
my_loader.load('B')
my_loader.load('A')

# > [ 'A', 'B', 'A' ]

More complex cache behavior can be achieved by calling .clear() or .clear_all() rather than disabling the cache completely. For example, this DataLoader will provide unique keys to a batch function due to the memoization cache being enabled, but will immediately clear its cache when the batch function is called so later requests will load new values.

class MyLoader(DataLoader):
    cache = False
    async def batch_load_fn(self, keys):
        self.clear_all()
        return keys

API

class DataLoader

DataLoader creates a public API for loading data from a particular data back-end with unique keys such as the id column of a SQL table or document name in a MongoDB database, given a batch loading function.

Each DataLoader instance contains a unique memoized cache. Use caution when used in long-lived applications or those which serve many users with different access permissions and consider creating a new instance per web request.

new DataLoader(batch_load_fn, **options)

Create a new DataLoader given a batch loading function and options.

• batch_load_fn: An async function (coroutine) which accepts an List of keys and returns a Future which resolves to an List of values.

• options:

  • batch: Default True. Set to False to disable batching, instead immediately invoking batch_load_fn with a single load key.

  • max_batch_size: Default Infinity. Limits the number of items that get passed in to the batch_load_fn.

  • cache: Default True. Set to False to disable memoization caching, instead creating a new Promise and new key in the batch_load_fn for every load of the same key.

  • cache_key_fn: A function to produce a cache key for a given load key. Defaults to key => key. Useful to provide when Python objects are keys and two similarly shaped objects should be considered equivalent.

  • cache_map: An instance of dict (or an object with a similar API) to be used as the underlying cache for this loader. Default {}.

load(key)

Loads a key, returning a Future for the value represented by that key.

• key: An key value to load.

load_many(keys)

Loads multiple keys, promising an array of values:

a, b = await my_loader.load_many([ 'a', 'b' ]);

This is equivalent to the more verbose:

from asyncio import gather
a, b = await gather(
    my_loader.load('a'),
    my_loader.load('b')
)

• keys: A list of key values to load.

clear(key)

Clears the value at key from the cache, if it exists. Returns itself for method chaining.

• key: An key value to clear.

clear_all()

Clears the entire cache. To be used when some event results in unknown invalidations across this particular DataLoader. Returns itself for method chaining.

prime(key, value)

Primes the cache with the provided key and value. If the key already exists, no change is made. (To forcefully prime the cache, clear the key first with loader.clear(key).prime(key, value).) Returns itself for method chaining.

Using with GraphQL

DataLoader pairs nicely well with GraphQL. GraphQL fields are designed to be stand-alone functions. Without a caching or batching mechanism, it's easy for a naive GraphQL server to issue new database requests each time a field is resolved.

Consider the following GraphQL request:

{
  me {
    name
    bestFriend {
      name
    }
    friends(first: 5) {
      name
      bestFriend {
        name
      }
    }
  }
}

Naively, if me, bestFriend and friends each need to request the backend, there could be at most 13 database requests!

When using DataLoader, we could define the User type using the SQLite example with clearer code and at most 4 database requests, and possibly fewer if there are cache hits.

class User(graphene.ObjectType):
    name = graphene.String()
    best_friend = graphene.Field(lambda: User)
    friends = graphene.List(lambda: User)

    def resolve_best_friend(self, args, context, info):
        return user_loader.load(self.best_friend_id)

    def resolve_friends(self, args, context, info):
        return user_loader.load_many(self.friend_ids)

Common Patterns

Creating a new DataLoader per request.

In many applications, a web server using DataLoader serves requests to many different users with different access permissions. It may be dangerous to use one cache across many users, and is encouraged to create a new DataLoader per request:

def create_loaders(auth_token):
  return {
    'users': DataLoader(lambda ids: gen_users(auth_token, ids)),
    'cdn_urls': DataLoader(lambda raw_urls: gen_cdn_urls(auth_token, raw_urls)),
    'stories': DataLoader(lambda keys: gen_stories(auth_token, keys)),
  }
}

# When handling an incoming web request:
loaders = create_loaders(request.query.auth_token)

# Then, within application logic:
user = await loaders.users.load(4)
pic = await loaders.cdn_urls.load(user.raw_pic_url)

Creating an object where each key is a DataLoader is one common pattern which provides a single value to pass around to code which needs to perform data loading, such as part of the root_value in a [graphql][] request.

Loading by alternative keys.

Occasionally, some kind of value can be accessed in multiple ways. For example, perhaps a "User" type can be loaded not only by an "id" but also by a "username" value. If the same user is loaded by both keys, then it may be useful to fill both caches when a user is loaded from either source:

async def user_by_id_batch_fn(ids):
    users = await gen_users_by_id(ids)
    for user in users:
        username_loader.prime(user.username, user)
    return users

user_by_id_loader = DataLoader(user_by_id_batch_fn)

async def username_batch_fn(names):
    users = await gen_usernames(names)
    for user in users:
        user_by_id_loader.prime(user.id, user)
    return users

username_loader = DataLoader(username_batch_fn)

Custom Caches

DataLoader can optionaly be provided a custom dict instance to use as its memoization cache. More specifically, any object that implements the methods get(), set(), delete() and clear() can be provided. This allows for custom dicts which implement various cache algorithms to be provided. By default, DataLoader uses the standard dict which simply grows until the DataLoader is released. The default is appropriate when requests to your application are short-lived.

Video Source Code Walkthrough

DataLoader Source Code Walkthrough (YouTube):