swift/doc/source/development_middleware.rst

Middleware and Metadata

Using Middleware

Python WSGI Middleware (or just "middleware") can be used to "wrap" the request and response of a Python WSGI application (i.e. a webapp, or REST/HTTP API), like Swift's WSGI servers (proxy-server, account-server, container-server, object-server). Swift uses middleware to add (sometimes optional) behaviors to the Swift WSGI servers.

Middleware can be added to the Swift WSGI servers by modifying their paste configuration file. The majority of Swift middleware is applied to the proxy-server.

Given the following basic configuration:

[DEFAULT]
log_level = DEBUG
user = <your-user-name>

[pipeline:main]
pipeline = proxy-server

[app:proxy-server]
use = egg:swift#proxy

You could add the healthcheck middleware by adding a section for that filter and adding it to the pipeline:

[DEFAULT]
log_level = DEBUG
user = <your-user-name>

[pipeline:main]
pipeline = healthcheck proxy-server

[filter:healthcheck]
use = egg:swift#healthcheck

[app:proxy-server]
use = egg:swift#proxy

Some middleware is required and will be inserted into your pipeline automatically by core swift code (e.g. the proxy-server will insert catch_errors and gatekeeper at the start of the pipeline if they are not already present). You can see which features are available on a given Swift endpoint (including middleware) using the discoverability interface.

Creating Your Own Middleware

The best way to see how to write middleware is to look at examples.

Many optional features in Swift are implemented as common_middleware and provided in swift.common.middleware, but Swift middleware may be packaged and distributed as a separate project. Some examples are listed on the associated_projects page.

A contrived middleware example that modifies request behavior by inspecting custom HTTP headers (e.g. X-Webhook) and uses sysmeta to persist data to backend storage as well as common patterns like a .get_container_info cache/query and .wsgify decorator is presented below:

from swift.common.http import is_success
from swift.common.swob import wsgify
from swift.common.utils import split_path, get_logger
from swift.common.request_helpers import get_sys_meta_prefix
from swift.proxy.controllers.base import get_container_info
from eventlet import Timeout
import six
if six.PY3:
    from eventlet.green.urllib import request as urllib2
else:
    from eventlet.green import urllib2

# x-container-sysmeta-webhook
SYSMETA_WEBHOOK = get_sys_meta_prefix('container') + 'webhook'


class WebhookMiddleware(object):
    def __init__(self, app, conf):
        self.app = app
        self.logger = get_logger(conf, log_route='webhook')

    @wsgify
    def __call__(self, req):
        obj = None
        try:
            (version, account, container, obj) = \
                split_path(req.path_info, 4, 4, True)
        except ValueError:
            # not an object request
            pass
        if 'x-webhook' in req.headers:
            # translate user's request header to sysmeta
            req.headers[SYSMETA_WEBHOOK] = \
                req.headers['x-webhook']
        if 'x-remove-webhook' in req.headers:
            # empty value will tombstone sysmeta
            req.headers[SYSMETA_WEBHOOK] = ''
        # account and object storage will ignore x-container-sysmeta-*
        resp = req.get_response(self.app)
        if obj and is_success(resp.status_int) and req.method == 'PUT':
            container_info = get_container_info(req.environ, self.app)
            # container_info may have our new sysmeta key
            webhook = container_info['sysmeta'].get('webhook')
            if webhook:
                # create a POST request with obj name as body
                webhook_req = urllib2.Request(webhook, data=obj)
                with Timeout(20):
                    try:
                        urllib2.urlopen(webhook_req).read()
                    except (Exception, Timeout):
                        self.logger.exception(
                            'failed POST to webhook %s' % webhook)
                    else:
                        self.logger.info(
                            'successfully called webhook %s' % webhook)
        if 'x-container-sysmeta-webhook' in resp.headers:
            # translate sysmeta from the backend resp to
            # user-visible client resp header
            resp.headers['x-webhook'] = resp.headers[SYSMETA_WEBHOOK]
        return resp


def webhook_factory(global_conf, **local_conf):
    conf = global_conf.copy()
    conf.update(local_conf)

    def webhook_filter(app):
        return WebhookMiddleware(app, conf)
    return webhook_filter

In practice this middleware will call the URL stored on the container as X-Webhook on all successful object uploads.

If this example was at <swift-repo>/swift/common/middleware/webhook.py -you could add it to your proxy by creating a new filter section and adding it to the pipeline:

[DEFAULT]
log_level = DEBUG
user = <your-user-name>

[pipeline:main]
pipeline = healthcheck webhook proxy-server

[filter:webhook]
paste.filter_factory = swift.common.middleware.webhook:webhook_factory

[filter:healthcheck]
use = egg:swift#healthcheck

[app:proxy-server]
use = egg:swift#proxy

Most python packages expose middleware as entrypoints. See PasteDeploy documentation for more information about the syntax of the use option. All middleware included with Swift is installed to support the egg:swift syntax.

Middleware may advertize its availability and capabilities via Swift's discoverability support by using .register_swift_info:

from swift.common.registry import register_swift_info
def webhook_factory(global_conf, **local_conf):
    register_swift_info('webhook')
    def webhook_filter(app):
        return WebhookMiddleware(app)
    return webhook_filter

If a middleware handles sensitive information in headers or query parameters that may need redaction when logging, use the .register_sensitive_header and .register_sensitive_param functions. This should be done in the filter factory:

from swift.common.registry import register_sensitive_header
def webhook_factory(global_conf, **local_conf):
    register_sensitive_header('webhook-api-key')
    def webhook_filter(app):
        return WebhookMiddleware(app)
    return webhook_filter

Middlewares can override the status integer that is logged by proxy_logging middleware by setting swift.proxy_logging_status in the request WSGI environment. The value should be an integer. The value will replace the default status integer in the log message, unless the proxy_logging middleware detects a client disconnect or exception while handling the request, in which case swift.proxy_logging_status is overridden by a 499 or 500 respectively.

Swift Metadata

Generally speaking metadata is information about a resource that is associated with the resource but is not the data contained in the resource itself - which is set and retrieved via HTTP headers. (e.g. the "Content-Type" of a Swift object that is returned in HTTP response headers)

All user resources in Swift (i.e. account, container, objects) can have user metadata associated with them. Middleware may also persist custom metadata to accounts and containers safely using System Metadata. Some core Swift features which predate sysmeta have added exceptions for custom non-user metadata headers (e.g. acls, large-objects)

User Metadata

User metadata takes the form of X-<type>-Meta-<key>: <value>, where <type> depends on the resources type (i.e. Account, Container, Object) and <key> and <value> are set by the client.

User metadata should generally be reserved for use by the client or client applications. A perfect example use-case for user metadata is python-swiftclient's X-Object-Meta-Mtime which it stores on object it uploads to implement its --changed option which will only upload files that have changed since the last upload.

New middleware should avoid storing metadata within the User Metadata namespace to avoid potential conflict with existing user metadata when introducing new metadata keys. An example of legacy middleware that borrows the user metadata namespace is tempurl. An example of middleware which uses custom non-user metadata to avoid the user metadata namespace is slo-doc.

User metadata that is stored by a PUT or POST request to a container or account resource persists until it is explicitly removed by a subsequent PUT or POST request that includes a header X-<type>-Meta-<key> with no value or a header X-Remove-<type>-Meta-<key>: <ignored-value>. In the latter case the <ignored-value> is not stored. All user metadata stored with an account or container resource is deleted when the account or container is deleted.

User metadata that is stored with an object resource has a different semantic; object user metadata persists until any subsequent PUT or POST request is made to the same object, at which point all user metadata stored with that object is deleted en-masse and replaced with any user metadata included with the PUT or POST request. As a result, it is not possible to update a subset of the user metadata items stored with an object while leaving some items unchanged.

System Metadata (Sysmeta)

System metadata takes the form of X-<type>-Sysmeta-<key>: <value>, where <type> depends on the resources type (i.e. Account, Container, Object) and <key> and <value> are set by trusted code running in a Swift WSGI Server.

All headers on client requests in the form of X-<type>-Sysmeta-<key> will be dropped from the request before being processed by any middleware. All headers on responses from back-end systems in the form of X-<type>-Sysmeta-<key> will be removed after all middlewares have processed the response but before the response is sent to the client. See gatekeeper middleware for more information.

System metadata provides a means to store potentially private custom metadata with associated Swift resources in a safe and secure fashion without actually having to plumb custom metadata through the core swift servers. The incoming filtering ensures that the namespace can not be modified directly by client requests, and the outgoing filter ensures that removing middleware that uses a specific system metadata key renders it benign. New middleware should take advantage of system metadata.

System metadata may be set on accounts and containers by including headers with a PUT or POST request. Where a header name matches the name of an existing item of system metadata, the value of the existing item will be updated. Otherwise existing items are preserved. A system metadata header with an empty value will cause any existing item with the same name to be deleted.

System metadata may be set on objects using only PUT requests. All items of existing system metadata will be deleted and replaced en-masse by any system metadata headers included with the PUT request. System metadata is neither updated nor deleted by a POST request: updating individual items of system metadata with a POST request is not yet supported in the same way that updating individual items of user metadata is not supported. In cases where middleware needs to store its own metadata with a POST request, it may use Object Transient Sysmeta.

Object Metadata

Objects have other metadata in addition to the user metadata and system metadata described above.

Immutable Metadata

Objects have several items of immutable metadata. Like system metadata, these may only be set using PUT requests. However, they do not follow the general X-Object-Sysmeta-<key> naming scheme and they are not automatically removed from client responses.

Object immutable metadata includes:

X-Timestamp
Content-Length
Etag

X-Timestamp and Content-Length metadata MUST be included in PUT requests to object servers. Etag metadata is generated by object servers when they handle a PUT request, but checked against any Etag header sent with the PUT request.

Object immutable metadata, along with Content-Type, is the only object metadata that is stored by container servers and returned in object listings.

Content-Type

Object Content-Type metadata is treated differently from immutable metadata, system metadata and user metadata.

Content-Type MUST be included in PUT requests to object servers. Unlike immutable metadata or system metadata, Content-Type is mutable and may be included in POST requests to object servers. However, unlike object user metadata, existing Content-Type metadata persists if a POST request does not include new Content-Type metadata. This is because an object must have Content-Type metadata, which is also stored by container servers and returned in object listings.

Content-Type is the only item of object metadata that is both mutable and yet also persists when not specified in a POST request.

Object Transient-Sysmeta

If middleware needs to store object metadata with a POST request it may do so using headers of the form X-Object-Transient-Sysmeta-<key>: <value>.

All headers on client requests in the form of X-Object-Transient-Sysmeta-<key> will be dropped from the request before being processed by any middleware. All headers on responses from back-end systems in the form of X-Object-Transient-Sysmeta-<key> will be removed after all middlewares have processed the response but before the response is sent to the client. See gatekeeper middleware for more information.

Transient-sysmeta updates on an object have the same semantic as user metadata updates on an object (see usermeta) i.e. whenever any PUT or POST request is made to an object, all existing items of transient-sysmeta are deleted en-masse and replaced with any transient-sysmeta included with the PUT or POST request. Transient-sysmeta set by a middleware is therefore prone to deletion by a subsequent client-generated POST request unless the middleware is careful to include its transient-sysmeta with every POST. Likewise, user metadata set by a client is prone to deletion by a subsequent middleware-generated POST request, and for that reason middleware should avoid generating POST requests that are independent of any client request.

Transient-sysmeta deliberately uses a different header prefix to user metadata so that middlewares can avoid potential conflict with user metadata keys.

Transient-sysmeta deliberately uses a different header prefix to system metadata to emphasize the fact that the data is only persisted until a subsequent POST.