Upgrading to Clients

Background

Version 0.66.0 of botocore added support for clients (10/16/2014) added initial support for clients. This provided an alternate interface to making AWS calls that provided a number of benefits over the existing interface. At the time, both interfaces were added so that users could opt in to trying the new clients.

Below is an example of the old interface:

import botocore.session
session = botocore.session.get_session()
s3 = session.get_service('s3')
endpoint = s3.get_endpoint('us-west-2')
list_objects = s3.get_operation('ListObjects')
http, response = list_objects.call(endpoint, Bucket='mybucket')
if http.status_code == 200:
    print("Contents: %s" % response['Contents])
else:
    print("API call failed, status code: %s, error: %s" % (
        http.status_code, http.content))

Here's an example of the newer (preferred) client interface:

import botocore.session
session = botocore.session.get_session()
s3 = session.create_client('s3', 'us-west-2')
response = s3.list_objects(Bucket='mybucket')
print("Contents: %s" % response['Contents'])

While there are many improvements with the new interface, here's a few notable improvements:

  • Less boilerplate. With clients, the Endpoints are abstracted from the user. The Operations are also abstracted. The user only needs to deal with a single class, the client class.
  • Exceptions. With clients, exceptions are raised on failed requests. You do not have to check the error code of the http response object
  • Return values. The http response is abstracted from the user. The client method makes an API call and returns a simple python dict.
  • More granular context. Clients allow you to have a context that's smaller than on a Service/Operation. For example, you can now create a client and add specific customizations and event handlers that apply to only that specific client, instead of any S3 client.

Deprecation Timeline

Botocore has not had a GA (1.0) release yet. In version 1.0, the Service/Operation object will be removed. Leading up to 1.0, we will be deprecating the old interface according to this schedule:

  • (done) Version 0.66.0 introduced client
  • (done) Version 0.96.0 emitted a PendingDeprecationWarning when using the old interface. By default, these warnings are not printed to stderr. A developer will need to opt in to these warnings to see any emitted warnings.
  • (done) Version 0.99.0 emitted a DeprecationWarning when using the old interface. This has the same behavior as PendingDeprecationWarning, in which by default nothing is printed to stderr.
  • Version 0.104.0 emitted an ImminentRemovalWarning when using the old interface. This warning is now printed to stderr by default. A user can still turn off these warnings via the warnings module.
  • The develop branch on github will be updated to completely remove the old interface. At this point the client interface will be the only interface available when using the develop branch on github.
  • A 1.0 alpha version will be released. By default, pip will not pick up any alpha/beta releases.
  • A 1.0 beta version will be released.
  • A GA (1.0) version of botocore will be released.

How To Upgrade

Migrating to clients is straightforward. This section outlines how to upgrade and some common scenarios and how to upgrade.

Keyword Arg Casing

In botocore's old interface, kwargs could either be snake_cased or CamelCased and botocore would map them to the correct version needed by the service which is CamelCase.

Botocore's client interface accepts only CamelCased args. More specifically, botocore uses the same casing used by the respective AWS service. Most AWS services uses CamelCasing. Some AWS services use lowerCamelCase.

Old

list_objects.call(endpoint, bucket='foo', key='bar')

New

client.list_objects(Bucket='foo', Key='bar')

Connection Pooling

In the old interface, the connection pooling was tied to an endpoint object. To reuse existing HTTP connection, you needed to keep a reference to the endpoint objects. With clients, connection pooling is tied to a client. Use a single client to make multiple API calls.

Old

service = session.get_service('s3')
endpoint = service.get_endpoint('us-west-2')
operation = service.get_operation('ListObjects')
head_object = service.get_operation('HeadObject')
parsed = operation.call(endpoint, Bucket='mybucket')[1]
for obj in parsed['Contents']:
    name = obj['Key']
    # Use existing connection be passing in the same endpoint.
    print(head_object.call(endpoint, Bucket='mybucket', Key=name))

New

s3 = session.get_client('s3', 'us-west-2')
for obj in s3.list_objects(Bucket='mybucket')['Contents']:
    name = obj['Key']
    # Using the same client will reuse any existing HTTP
    # connections the client was using.
    print(s3.head_object(Bucket='mybucket', Key=name))

Operation and Method Names

In the old interface, you would retrieve an API operation using the casing defined by the service, which is typically CamelCase. For example, you'd use service.get_operation('ListObjects'), not service.get_operation('list_objects'). With clients, method names, which map 1 - 1 to operation names are snake_cased, as is common in python code.

Old

service = session.get_service('s3')
list_objects = service.get_operation('ListObjects')
head_object = service.get_operation('HeadObject')
get_object = service.get_operation('GetObject')

New

s3 = session.get_client('s3', 'us-west-2')
list_objects = s3.list_objects
head_object = s3.head_object
get_object = s3.get_object

Return Values

In the old interface, the return value for an operation.call invocation is a tuple of the HTTP response object, and the parsed dict that results from parsing the HTTP object. In the client interface, only the parsed response is returned. The HTTP response object is not returned.

Old

ec2 = session.get_service('ec2')
endpoint = ec2.get_endpoint('ec2')
describe_instances = ec2.get_operation('DescribeInstances')
http, parsed = describe_instances.call(endpoint)

New

ec2 = session.get_client('ec2', 'us-west-2')
parsed = ec2.describe_instances()

The main reason for returning the HTTP response was to check if an error occurred via the HTTP response status code. This is now no longer required (discussed below). Exception are automatically raised.

Note that if for some reason you do need to see the response status code, it is available via the ResponseMetadata in the parsed dict that's returned.

ec2 = session.get_client('ec2', 'us-west-2')
parsed = ec2.describe_instances()
print("The status code is:",
      parsed['ResponseMetadata']['HTTPStatusCode']))

Error Handling

With clients, exceptions are now raised on any non 2xx response. A ClientError exception has both a .msg attribute as well as the parsed error response (which is a dictionary). There isn't really equivalent functionality in the old interface, but below is an example of how you can handle an error:

New

from botocore.exceptions import ClientError

ec2 = session.get_client('ec2', 'us-west-2')
try:
    parsed = ec2.describe_instances(InstanceIds=['i-badid'])
except ClientError as e:
    logger.error("Received error: %s", e, exc_info=True)
    # Only worry about a specific service error code
    if e.response['Error']['Code'] == 'InvalidInstanceID.NotFound':
        raise

If you run into any issues migrating from the old interface to the newer client interface, please file an issue on github and let us know. We'd be happy to help.