Distributed rate-limiting middleware

## Project description

Turnstile is a piece of WSGI middleware that performs true distributed rate-limiting. System administrators can run an API on multiple nodes, then place this middleware in the pipeline prior to the application. Turnstile uses a Redis database to track the rate at which users are hitting the API, and can then apply configured rate limits, even if each request was made against a different API node.

Turnstile is intended for use with PasteDeploy-style configuration files. It is a filter, and should be placed in an appropriate place in the WSGI pipeline such that the limit classes used with Turnstile can access the information necessary to make rate-limiting decisions. (With the turnstile.limits:Limit class provided by Turnstile, no additional information is required, as that class does not differentiate between users of your application.)

The filter section of the PasteDeploy configuration file will also need to contain enough information to allow Turnstile to access the database. Other options may be configured from here as well, such as the preprocess configuration variable. The simplest example of Turnstile configuration would be:

[filter:turnstile]
paste.filter_factory = turnstile.middleware:turnstile_filter
redis.host = <your Redis database host name or IP>

The following are the recognized configuration options:

control.channel

Specifies the channel that the control daemon listens on. (See below for more information about the purpose of the control daemon.) This option defaults to “control”.

control.errors_channel

Specifies the channel that the control daemon (see below) reports errors to. This option defaults to “errors”.

control.errors_key

Specifies the key of a set in the Redis database to which errors will be stored. This option defaults to “errors”.

control.limits_key

The key under which the limits are stored in the database. See the section on tools for more information on how to load and dump the limits stored in the Redis database. This option defaults to “limits”.

control.node_name

The name of the node. If provided, this option allows the specification of a recognizable name for the node. Currently, this node name is only reported when issuing a “ping” command to the control daemon (see below), and may be used to verify that all hosts responded to the ping.

When limits are changed in the database, a command is sent to the control daemon (see below) to cause the limits to be reloaded. As having all nodes hit the Redis database simultaneously may overload the database, this option, if set, allows the reload to be spread out randomly within a configured interval. This option should be set to the size of the desired interval, in seconds. If not set, limits will be reloaded immediately by all nodes.

control.shard_hint

Can be used to set a sharding hint which will be provided to the listening thread of the control daemon (see below). This hint is not used by the default Redis Connection class.

preprocess

Contains a list of functions, specified as “module:function” pairs separated by spaces. During each request, each preprocessor will be called in turn, with the middleware object (from which can be obtained the database handle, as well as the configuration as a two-layer dictionary) and the request environment as arguments. Note that any exceptions thrown by the preprocessors will not be caught, and request processing will be halted; this will likely result in a 500 error being returned to the user.

redis.connection_pool

Identifies the connection pool class to use. If not provided, defaults to redis.ConnectionPool. This may be used to allow client-side sharding of the Redis database.

redis.connection_pool.connection_class

Identifies the connection class to use. If not provided, the appropriate redis.Connection subclass for the configured connection is used (redis.Connection if redis.host is specified, else redis.UnixDomainSocketConnection).

redis.connection_pool.max_connections

Allows specification of the maximum number of connections to the Redis database. Optional.

redis.connection_pool.parser_class

Identifies the parser class to use. Optional. This is an advanced feature of the redis package used by Turnstile.

redis.connection_pool.*

Any other configuration value provided in the redis.connection_pool. hierarchy will be passed as keyword arguments to the configured connection pool class. Note that the values passed will be strings.

redis.db

Identifies the specific sub-database of the Redis database to be used by Turnstile. If not provided, defaults to 0.

redis.host

Identifies the host name or IP address of the Redis database to connect to. Either redis.host or redis.unix_socket_path must be provided.

If the Redis database has been configured to use a password, this option allows that password to be specified.

redis.port

Identifies the port the Redis database is listening on. If not provided, defaults to 6379.

redis.socket_timeout

If provided, specifies an integer socket timeout for the Redis database connection.

redis.unix_socket_path

Names the UNIX socket on the local host for the local Redis database to connect to. Either redis.host or redis.unix_socket_path must be provided.

status

Contains the status code to return if rate limiting is tripped. This defaults to “413 Request Entity Too Large”. Note that this value must start with the 3-digit HTTP code, followed by a space and the text corresponding to that status code. Also note that, regardless of the status code, Turnstile will include the Retry-After header in the response. (The value of the Retry-After header will be the integer number of seconds until the request can be retried.)

turnstile

If set, identifies an alternate class to use for the Turnstile middleware. This can be used in conjunction with subclassing turnstile.middleware:TurnstileMiddleware, which may be done to override how over-limit conditions are formatted.

Other configuration values are available to the preprocessors and the turnstile.limits:Limit subclasses, but extreme care should be taken that such configurations remain in sync across the entire cluster.

## The Control Daemon

Turnstile stores the limits configuration in the Redis database, in addition to the ephemeral information used to check and enforce the rate limits. This makes it possible to change the limits dynamically from a single, central location. In order to facilitate such changes, each Turnstile instance uses an eventlet thread to run a “control daemon.” The control daemon uses the publish/subscribe support provided by Redis to listen for commands, of which two are currently recognized: ping and reload.

### The Ping Command

The “ping” command is the simplest of the control daemon commands. In its simplest form, the message “ping:<channel>” is written to the control channel, which will cause all running Turnstile instances to return the message “pong” to the specified channel. If the control.node_name configuration option has been set, this node name will be included in the response, as “pong:<node name>”. Finally, additional data (such as a timestamp) can be included in the “ping” command, as in the message “ping:<channel>:<timestamp>”; this data will be appended to the response, i.e., “pong:<node name>:<timestamp>”. This could be used to verify that all nodes are responding and not too heavily loaded.

(Note that if control.node_name is not specified, the response to a “ping” command containing additional data such as a timestamp will be “pong::<timestamp>”.)

The “reload” command is the real reason for the existence of the control daemon. This command causes the current set of limits to be loaded from the database and installed in the middleware. (This operation is done in a thread-safe manner.)

The simplest form of the reload command is simply, “reload”. If the control.reload_spread configuration option was set, the reload will be scheduled for some time within the configured time interval; otherwise, it will be performed immediately.

The next simplest form of the reload command is “reload:immediate”. This causes an immediate reload of the limits, overriding any configured time spread.

Note that the setup_limits tool automatically initiates a reload once the limits are updated in the database. See the section on tools for more information.

## Turnstile Tools

The limits are stored in the Redis database using a sorted set, and they are encoded using Msgpack. (Although the Msgpack format is not human-readable, it is very space and time efficient, which is why it was chosen for this application.) This makes manual management of them more difficult, and so Turnstile ships with two tools to make management of the rate limiting configuration easier.

### The dump_limits Tool

The dump_limits tool may be used to dump the current limits in the database into an XML representation. This tool requires the name of an INI-style configuration file; see the section on configuring the tools below for more information.

A usage summary for dump_limits:

usage: dump_limits [-h] [--debug] config limits_file

Dump the current limits from the Redis database.

positional arguments:
config       Name of the configuration file, for connecting to the Redis
database.
limits_file  Name of the XML file that the limits will be dumped to.

optional arguments:
-h, --help   show this help message and exit
--debug, -d  Run the tool in debug mode.

### The setup_limits Tool

The setup_limits tool may be used to read an XML file (such as that produced by dump_limits) and load the rate limiting configuration into the Redis database. This tools also requires the name of an INI-style configuration file; see the section on configuring the tools below for more information.

A usage summary for setup_limits:

usage: setup_limits [-h] [--debug] [--dryrun] [--noreload]
config limits_file

Set up or update limits in the Redis database.

positional arguments:
config                Name of the configuration file, for connecting to the
Redis database.
limits_file           Name of the XML file describing the limits to
configure.

optional arguments:
-h, --help            show this help message and exit
--debug, -d           Run the tool in debug mode.
--dryrun, --dry_run, --dry-run, -n
database.
Cause all nodes to immediately reload the limits
configuration.
Cause all nodes to reload the limits configuration
over the specified number of seconds.

### Configuring the Tools

Both dump_limits and setup_limits require an INI-style configuration file, which specifies how to connect to the Redis database. This file should contain the section “[connection]” and should be populated with the same “redis.*” options as the PasteDeploy configuration file, minus the “redis.” prefix. For example:

[connection]
host = <your Redis database host name or IP>

Each “redis.*” option recognized by the Turnstile middleware is understood by the tools.

### Rate Limit XML

The XML file used for expressing rate limit configuration is relatively straightforward, or at least as straightforward as XML can be. The top-level element is “<limits>”; this should contain a sequence of “<limit>” elements, each containing a number of “<attr>” elements. The specific attributes available for any given limit class depend on the exact class, but that information is documented in the attrs attribute of the limit class. (This information is suitable for introspection.)

The “<limit>” element has one XML attribute which must be specified: the “class” attribute, which must be set to a “module:class” string identifying the desired limit class. The “<attr>” element also has a single XML attribute which must be set: “name”, which identifies the name of the Limit attribute. The contents of the “<attr>” element identify the value for the named attribute.

Some limit attributes are lists; for these attributes, the “<attr>” element must contain one or more “<value>” elements, whose contents identify a single item in the attribute list. Other limit attributes are dictionaries; for these attributes, again the “<attr>” element must contain one or more “<value>” elements, but now those “<value>” elements must have the XML attribute “key” set to the dictionary key corresponding to that value.

As an example, consider the following limits configuration:

<?xml version='1.0' encoding='UTF-8'?>
<limits>
<limit class="turnstile.limits:Limit">
<attr name="requirements">
<value key="pageid">[0-9]+</value>
</attr>
<attr name="unit">second</attr>
<attr name="uri">/page/{pageid}</attr>
<attr name="value">10</attr>
<attr name="verbs">
<value>GET</value>
</attr>
</limit>
</limits>

In this example, GET access to “/page/{pageid}” is rate-limited to 10 per second. The requirements attribute may be used to specify regular expressions to tune the matching of URI components; in this case, the “{pageid}” value must be composed of 1 or more digits. The limit class used is the basic turnstile.limits:Limit limit class.

## Custom Limit Classes

All limit classes must descend from turnstile.limits:Limit. This admittedly un-Pythonic requirement has a number of advantages, including the specific machinery which allows limits to be stored into the Redis database. Most limit classes only need to worry about the attrs class attribute and the filter() method, although the route() and format() methods may also be hooked. For more information about these methods, see the documentation provided for their default implementations in turnstile.limits:Limit.

## Accessing the Turnstile Configuration

The Turnstile configuration is available to preprocessors and to the Limit classes. For preprocessors, it is available directly from the middleware object (the first passed parameter) via the ‘config’ attribute. (The database handle is also available via the ‘db’ attribute, should access to the database be required.) For the filter() method of the Limit classes, the configuration is available in the request environment under the “turnstile.config” key.

The Turnstile configuration is represented as a two-level dictionary. Configuration keys that do not contain a ‘.’ are available in the dictionary accessible via the key ‘None’ in the base configuration. For example, to obtain the configured status value, assuming the Turnstile configuration is available in the “config” variable, the correct code would be:

status = config[None]['status']

For those configuration keys which do contain a ‘.’, the part of the name to the left of the first ‘.’ becomes the first key, and the remainder of the name is the second key. For example, to access the value of the “redis.connection_pool.connection_class” variable, the correct code would be:

connection_class = config['redis']['connection_pool.connection_class']

All values in the configuration are stored as strings. Configuration values do not need to be pre-declared in any way; Turnstile ignores (but maintains) configuration values that it does not use, making these values available for use by preprocessors and Limit classes.

## Project details

Uploaded source