Cache Range Requests Plugin
Description
Most origin servers support HTTP/1.1 range requests (rfc 7233). ATS internally handles range request caching in one of 2 ways:
Don’t cache range requests.
Only server range requests from a wholly cached object.
This plugin allows you to remap individual range requests so that they are stored as individual objects in the ATS cache when subsequent range requests are likely to use the same range. This spreads range requests over multiple stripes thereby reducing I/O wait and system load averages.
cache_range_requests reads the range request header byte range
value and then creates a new cache key URL using the original request
url with the range value appended to it. The range header is removed
where appropriate from the requests and the origin server response code
is changed from a 206 to a 200 to insure that the object is written to
cache using the new cache key url. The response code sent to the client
will be changed back to a 206 and all requests to the origin server will
contain the range header so that the correct response is received.
The cache_range_requests plugin by itself has no logic to efficiently manage overlapping ranges. It is best to use this plugin in conjunction with a smart client that only requests predetermined non overlapping cache ranges (request blocking) or as a helper for the slice plugin.
Only requests which contain the Range: <units>= GET header
will be served by the cache_range_requests plugin.
If/when ATS implements partial object caching this plugin will become deprecated.
NOTE Given a multi range request the cache_range_requests only processes the first range and ignores the rest.
How to run the plugin
The plugin can run as a global plugin (a single global instance configured
using plugin.config) or as per-remap plugin (a separate instance
configured per remap rule in remap.config).
Global instance
$ cat plugin.config
cache_range_request.so
Per-remap instance
$cat remap.config
map http://www.example.com http://www.origin.com \
@plugin=cache_range_requests.so
If both global and per-remap instance are used the per-remap configuration would take precedence (per-remap configuration would be applied and the global configuration ignored).
Plugin options
Parent Selection as Cache Key
- --ps-cachekey
- -p
Without this option parent selection is based solely on the hash of a URL Path a URL is requested from the same upstream parent cache listed in parent.config
With this option parent selection is based on the full cache key URL
which includes information about the partial content range. In this mode,
all requests (include partial content) will use consistent hashing method
for parent selection.
X-Crr-Ims header support
- --consider-ims
- -c
- --ims-header=[header name] (default: X-Crr-Ims)
- -i
To support slice plugin self healing an option to force revalidation after cache lookup complete was added. This option is triggered by a special header:
X-Crr-Ims: Tue, 19 Nov 2019 13:26:45 GMT
When this header is provided and a cache hit fresh is encountered the
Date header of the object in cache is compared to this header date
value. If the cache date is less than this IMS date then the object
is marked as STALE and an appropriate If-Modified-Since or If-Match
request along with this X-Crr-Ims header is passed up to the parent.
In order for this to properly work in a CDN each cache in the chain SHOULD also contain a remap rule with the cache_range_requests plugin with this option set.
When used with the slice plugin its –crr-ims-header option must have the same value (or not be defined) in order to work.
Presence of the –ims-header automatically sets the –consider-ims option.
Don’t modify the Cache Key
- --no-modify-cachekey
- -n
With each transaction TSCacheUrlSet may only be called once. When using the cache_range_requests plugin in conjunction with the cachekey plugin the option –include-headers=Range should be added as a cachekey parameter with this option. Configuring this incorrectly WILL result in cache poisoning.
map http://ats/ http://parent/ \
@plugin=cachekey.so @pparam=--include-headers=Range \
@plugin=cache_range_requests.so @pparam=--no-modify-cachekey
Without this `cache_range_requests` plugin option
IF the TSCacheUrlSet call in cache_range_requests fails, an error is generated in the logs and the cache_range_requests plugin will disable transaction caching in order to avoid cache poisoning.
Verify Cacheability
- --verify-cacheability
- -v
This option causes the plugin to verify whether the requested object is cacheable.
By default, an object’s cacheability is not verified after the plugin changes the response code of the upstream response from 206 to 200 to force the object into cache. When this option is enabled, cacheability is considered, and if the object is not cacheable, the status code is reset back to 206, which leads to the object not being cached.
This option is useful when used with other plugins, such as Cache Promote.
Cache Complete Responses
- --cache-complete-responses
- -r
This option causes the plugin to cache complete responses (200 OK). By default, only 206 Partial Content responses are cached by this plugin; without this flag, any 200 OK observed will be marked as not cacheable.
This option is intended to cover the case when an origin responds with a 200 OK when the requested range exceeds the size of the object. For example, if an object is 500 bytes, and the requested range is for bytes 0-5000, some origins will respond with a 206 and a Content-Range header, while others may respond with a 200 OK and no Content-Range header. The same origin that responds with a 200 OK when the requested range exceeds the object size will serve 206s when the range is smaller than or within the bytes of the object.
NOTE: This option should be used carefully with full knowledge of how cache keys are set for a given remap rule that relies on this behavior and origin response mechanics. For example, when this option is the sole argument to cache_range_requests.so and no other plugins are in use, the behavior could be abused, especially if the origin always responds with 200 OKs. This is because the plugin will automatically include the requested Range in the cache key. This means that arbitrary ranges can be used to pollute the cache with different combinations of ranges, which will lead to many copies of the same complete object stored under different cache keys.
For this reason, if the plugin is instructed to cache complete responses, Range request headers coming into the remap should ideally be normalized. Normalization can be accomplished by using the slice plugin without the –ref-relative argument which is disabled by default. The cache key plugin can also be used to tightly control the construction of the cache key itself.
The preferred means of using this plugin option is with the following plugins: - slice to normalize the requested ranges, without the –ref-relative option - cachekey to control the cache key, including the Range header normalized by slice - cache range requests with –no-modify-cachekey and –cache-complete-responses
Configuration examples
Global plugin
cache_range_requests.so --ps-cachekey --consider-ims --no-modify-cachekey
or
cache_range_requests.so -p -c -n
Remap plugin
map http://ats http://parent @plugin=cache_range_requests.so @pparam=--ps-cachekey @pparam=--consider-ims @pparam=--no-modify-cachekey
or
map http://ats http://parent @plugin=cache_range_requests.so @pparam=-p @pparam=-c @pparam=-n