Rate Limit Plugin

The rate_limit plugin provides basic mechanism for how much traffic a particular service (remap rule) is allowed to do. Currently, the only implementation is a limit on how many active client transactions a service can have. However, it would be easy to refactor this plugin to allow for adding new limiter policies later on.

The limit counters and queues are per remap rule only, i.e. there is (currently) no way to group transaction limits from different remap rules into a single rate limiter.

Note

This is still work in progress, in particularly the configuration and the IP reputation system needs some work. In particular:

  • We need a proper YAML configuration overall, allowing us to configure better per service controls as well as sharing resources between remap rules or SNI.

  • We need reloadable configurations.

  • The IP reputation currently only works with the global plugin settings.

  • There is no support for adding allow listed IPs to the IP reputation.

Remap Plugin

All configuration is done via remap.config, and the following options are available:

--limit

The maximum number of active client transactions.

--queue

When the limit (above) has been reached, all new transactions are placed on a FIFO queue. This option (optional) sets an upper bound on how many queued transactions we will allow. When this threshold is reached, all additional transactions are immediately served with an error message.

The queue is effectively disabled if this is set to 0, which implies that when the transaction limit is reached, we immediately start serving error responses.

The default queue size is UINT_MAX, which is essentially unlimited.

--error

An optional HTTP status error code, to be used together with the --queue option above. The default is 429.

--retry

An optional retry-after value, which if set will cause rejected (e.g. 429) responses to also include a header Retry-After.

--header

This is an optional HTTP header name, which will be added to the client request header IF the transaction was delayed (queued). The value of the header is the delay, in milliseconds. This can be useful to for example log the delays for later analysis.

It is recommended that an @ header is used here, e.g. @RateLimit-Delay, since this header will not leave the ATS server instance.

--maxage

An optional max-age for how long a transaction can sit in the delay queue. The value (default 0) is the age in milliseconds.

--prefix

An optional metric prefix to use instead of the default (plugin.rate_limiter).

--tag

An optional metric tag to use instead of the default. When a tag is not specified the plugin will use the scheme, FQDN, and port when it is non-standard. For example a default plugin tag might be “https.example.com” or “http.example.com:8080” noting that in the latter exampe, the non-standard scheme and port led to “:8080” being appended to the string.

Global Plugin

As a global plugin, the rate limiting currently applies only for TLS enabled connections, based on the SNI from the TLS handshake. As a global plugin we also have the support of an IP reputation system, see below for configurations.

The basic use is as:

rate_limit.so SNI=www1.example.com,www2.example.com --limit=2 --queue=2 --maxage=10000

Note

As a global plugin, it’s highly recommended to also reduce the Keep-Alive inactive timeout for the service(s) controlled by this plugin. This avoids the risk of having idle connections consume too many of the available resources. This is easily done using e.g. the conf_remap plugin, proxy.config.http.keep_alive_no_activity_timeout_in.

The following options are available:

--limit

The maximum number of active client transactions.

--queue

When the limit (above) has been reached, all new connections are placed on a FIFO queue. This option (optional) sets an upper bound on how many queued transactions we will allow. When this threshold is reached, all additional connections are immediately errored out in the TLS handshake.

The queue is effectively disabled if this is set to 0, which implies that when the transaction limit is reached, we immediately start serving error responses.

The default queue size is UINT_MAX, which is essentially unlimited.

--maxage

An optional max-age for how long a transaction can sit in the delay queue. The value (default 0) is the age in milliseconds.

--prefix

An optional metric prefix to use instead of the default (plugin.rate_limiter).

--tag

An optional metric tag to use instead of the default. When a tag is not specified the plugin will use the FQDN of the SNI associated with each rate limiter instance created during plugin initialization.

--iprep_buckets

The number of LRU buckets to use for the IP reputation. A good number here is 10, which is the default, but can be configured. The reason for the different buckets is to account for a pseudo-sorted list of IPs on the frequency seen. Too few buckets will not be enough to keep such sorting, rendering the algorithm useless. To function in our setup, the number of buckets must be less than 100.

--iprep_bucketsize

This is the size of the largest LRU bucket (the entry bucket), 15 is a good value. This is a power of 2, so 15 means the largest LRU can hold 32768 entries. Note that this option must be bigger then the --iprep_buckets setting, for the bucket halfing to function.

The default here is 0, which means the IP reputation filter is not enabled!

--iprep_percentage

This is the minimum percentage of the limit that the pressure must be at, before we start blocking IPs. The default is 0.9 which means 90% of the limit.

--iprep_maxage

This is used for aging out entries out of the LRU, the default is 0 which means no aging happens. Even with no aging, entries will eventually fall out of buckets because of the LRU mechanism that kicks in. The aging is here to make sure a spike in traffic from an IP doesn’t keep the entry for too long in the LRUs.

--iprep_permablock_limit

The minimum number of hits an IP must reach to get moved to the permanent bucket. In this bucket, entries will stay for 2x

--iprep_permablock_pressure

This option specifies from which bucket an IP is allowed to move from into the perma block bucket. A good value here is likely 0 or 1, which is very conservative.

--iprep_permablock_maxage

Similar to --iprep_maxage above, but only applies to the long term (perma-block) bucket. Default is 0, which means no aging to this bucket is applied.

Metrics

Metric names are generated either using defaults or user-supplied values. In either case, the format of the metric names is as follows:

prefix.type.tag.metric

A user can specify their own prefixes and tags, but not types or metrics.

prefix

The default prefix for all metrics is plugin.rate_limiter.

type

There are two types of metrics: sni and remap. Each type corresponds with the type of configuration used to generate the metric. The global configuration is for rate limiting requests during TLS negotiation, hence, the type of sni. Similarly remap connotes a remap configuration.

tag

By default the metric tag is derived from a description that is set conditionally. When configured in global mode, the SNI argument allows a comma separated list of FQDNs that require rate limiting. Each FQDN is associated with an instance of the rate limiter, and the description of each limiter is set to the FQDN.

When configured on a remap, the plugin will generate a description based on the configuration. When the scheme and port number are standard, the port is omitted from the generated description, however, when the scheme and port combination are non-standard, the port is appended. For example, a standard scheme and port would lead to a description of http.example.com or https.example.com but if a non-standard port was used, a description might be https.example.com:8443 or http.example.com:8080. This approach allows each limiter to increment metrics for the correct remaps.

metric

There are four metrics that may be incremented, depending on which action the plugin takes:

Metric

Definition

queued

Request queued due to being at the limit but under the queue limit.

rejected

Request rejected due to being over the defined limits.

expired

Queued connection is too old to be resumed and is rejected.

resumed

Queued connection is resumed.

IP Reputation

The goal of the IP reputation system is to simply try to identify IPs which are more likely to be abusive than others. It’s not a perfect system, and it relies heavily on the notion of pressure. The Sieve LRUs are always filled, so you have to make sure that you only start using them when the system thinks it’s under pressure.

The Sieve LRU is a chained set of (configurable) LRUs, each with smaller and smaller capacity. This essentially adds a notion of partially sorted elements; All IPs in LRU <n> generally are more active than the IPs in LRU <n+1>. LRU is specially marked for longer term blocking, only the most abusive elements would end up here.

../../_images/SieveLRU.png

Examples

This example shows a simple rate limiting of 128 concurrently active client transactions, with a maximum queue size of 256. The default of HTTP status code 429 is used when queue is full:

map http://cdn.example.com/ http://some-server.example.com \
  @plugin=rate_limit.so @pparam=--limit=128 @pparam=--queue=256

This example would put a hard transaction (in) limit to 256, with no backoff queue, and add a header with the transaction delay if it was queued:

map http://cdn.example.com/ http://some-server.example.com \
  @plugin=rate_limit.so @pparam=--limit=256 @pparam=--queue=0 \
  @pparam=--header=@RateLimit-Delay

This final example will limit the active transaction, queue size, and also add a Retry-After header once the queue is full and we return a 429 error:

map http://cdn.example.com/ http://some-server.example.com \
  @plugin=rate_limit.so @pparam=--limit=256 @pparam=--queue=1024 \
  @pparam=--retry=3600 @pparam=--header=@RateLimit-Delay

In this case, the response would look like this when the queue is full:

HTTP/1.1 429 Too Many Requests
Date: Fri, 26 Mar 2021 22:42:38 GMT
Connection: keep-alive
Server: ATS/10.0.0
Cache-Control: no-store
Content-Type: text/html
Content-Language: en
Retry-After: 3600
Content-Length: 207

Metric Examples

The following examples show the metric names that result from various settings using a hypothetical domain of example.com with both global and remap configurations. Note that in this example the remap configuration contains both TLS and non-TLS remap rules.

Defaults:

proxy.rate_limiter.sni.example.com.queued
proxy.rate_limiter.sni.example.com.rejected
proxy.rate_limiter.sni.example.com.expired
proxy.rate_limiter.sni.example.com.resumed

proxy.rate_limiter.remap.https.example.com.queued
proxy.rate_limiter.remap.https.example.com.rejected
proxy.rate_limiter.remap.https.example.com.expired
proxy.rate_limiter.remap.https.example.com.resumed

proxy.rate_limiter.remap.http.example.com.queued
proxy.rate_limiter.remap.http.example.com.rejected
proxy.rate_limiter.remap.http.example.com.expired
proxy.rate_limiter.remap.http.example.com.resumed

Defaults with non-standard scheme+port combinations in the remap rules:

proxy.rate_limiter.sni.example.com.queued
proxy.rate_limiter.sni.example.com.rejected
proxy.rate_limiter.sni.example.com.expired
proxy.rate_limiter.sni.example.com.resumed

proxy.rate_limiter.remap.https.example.com:8443.queued
proxy.rate_limiter.remap.https.example.com:8443.rejected
proxy.rate_limiter.remap.https.example.com:8443.expired
proxy.rate_limiter.remap.https.example.com:8443.resumed

proxy.rate_limiter.remap.http.example.com:8080.queued
proxy.rate_limiter.remap.http.example.com:8080.rejected
proxy.rate_limiter.remap.http.example.com:8080.expired
proxy.rate_limiter.remap.http.example.com:8080.resumed
With:

  • --prefix=limiter on the global configuration

  • --tag=tls.example.com on the global configuration

  • @pparam=--prefix=limiter on the remap configurations

  • @pparam=--tag=secure.example.com on the TLS-enabled remap configuration

  • @pparam=--tag=insecure.example.com on the non-TLS-enabled remap configuration

limiter.sni.tls.example.com.queued
limiter.sni.tls.example.com.rejected
limiter.sni.tls.example.com.expired
limiter.sni.tls.example.com.resumed

limiter.remap.secure.example.com.queued
limiter.remap.secure.example.com.rejected
limiter.remap.secure.example.com.expired
limiter.remap.secure.example.com.resumed

limiter.remap.insecure.example.com.queued
limiter.remap.insecure.example.com.rejected
limiter.remap.insecure.example.com.expired
limiter.remap.insecure.example.com.resumed