RFC2227 - Simple Hit-Metering and Usage-Limiting for HTTP

Network Working Group J. Mogul
Request for Comments: 2227 DECWRL
Category: Standards Track P. Leach
Microsoft
October 1997
Simple Hit-Metering and Usage-Limiting for HTTP
Status of this Memo
This document specifies an Internet standards track protocol for the
Internet community, and requests discussion and suggestions for
improvements. Please refer to the current edition of the "Internet
Official Protocol Standards" (STD 1) for the standardization state
and status of this protocol. Distribution of this memo is unlimited.
Copyright Notice
Copyright (C) The Internet Society (1997). All Rights Reserved.
ABSTRACT
This document proposes a simple extension to HTTP, using a new
"Meter" header, which permits a limited form of demographic
information (colloquially called "hit-counts") to be reported by
caches to origin servers, in a more efficient manner than the
"cache-busting" techniques currently used. It also permits an origin
server to control the number of times a cache uses a cached response,
and outlines a technique that origin servers can use to capture
referral information without "cache-busting."
TABLE OF CONTENTS
1 IntrodUCtion 2
1.1 Goals, non-goals, and limitations 3
1.2 Brief summary of the design 4
1.3 Terminology 5
2 Overview 5
2.1 Discussion 7
3 Design concepts 8
3.1 Implementation of the "metering suBTree" 8
3.2 Format of the Meter header 10
3.3 Negotiation of hit-metering and usage-limiting 10
3.4 Transmission of usage reports 14
3.5 When to send usage reports 15
3.6 Subdivision of usage-limits 16
4 Analysis 17
4.1 Approximation accuracy for counting users 18
4.2 What about "Network Computers"? 19
4.3 Critical-path delay analysis 19
5 Specification 20
5.1 Specification of Meter header and directives 20
5.2 Abbreviations for Meter directives 23
5.3 Counting rules 24
5.3.1 Counting rules for hit-metering 24
5.3.2 Counting rules for usage-limiting 25
5.3.3 Equivalent algorithms are allowed 26
5.4 Counting rules: interaction with Range requests 27
5.5 Implementation by non-caching proxies 27
5.6 Implementation by cooperating caches 28
6 Examples 28
6.1 Example of a complete set of exchanges 28
6.2 Protecting against HTTP/1.0 proxies 30
6.3 More elaborate examples 30
7 Interactions with content negotiation 31
7.1 Treatment of responses carrying a Vary header 31
7.2 Interaction with Transparent Content Negotiation 32
8 A Note on Capturing Referrals 32
9 Alternative proposals 33
10 Security Considerations 34
11 Acknowledgments 35
12 References 35
13 Authors" Addresses 36
14 Full Copyright Statement 37
1 Introduction
For a variety of reasons, content providers want to be able to
collect information on the frequency with which their content is
Accessed. This desire leads to some of the "cache-busting" done by
existing servers. ("Cache-busting" is the use by servers of
techniques intended to prevent caching of responses; it is unknown
exactly how common this is.) This kind of cache-busting is done not
for the purpose of maintaining transparency or security properties,
but simply to collect demographic information. Some cache-busting is
also done to provide different advertising images to appear on the
same page (i.e., each retrieval of the page sees a different ad).
This proposal supports a model similar to that of publishers of
hard-copy publications: such publishers (try to) report to their
advertisers how many people read an issue of a publication at least
once; they don"t (try to) report how many times a reader re-reads an
issue. They do this by counting copies published, and then try to
estimate, for their publication, on average how many people read a
single copy at least once. The key point is that the results aren"t
exact, but are still useful. Another model is that of coding
inquiries in such a way that the advertiser can tell which
publication produced the inquiry.
1.1 Goals, non-goals, and limitations
HTTP/1.1 already allows origin servers to prevent caching of
responses, and evidence exists [9] that at least some of the time,
this is being done for the sole purpose of collecting counts of the
number of accesses of specific pages. Some of this evidence is
inferred from the study of proxy traces; some is based on eXPlicit
statements of the intention of the operators of Web servers.
Information collected this way might or might not be of actual use to
the people who collect it; the fact is that they want to collect it,
or already do so.
The goal of this proposal is to provide an optional performance
optimization for this use of HTTP/1.1.
This specification is:
- Optional: no server or proxy is required to implement it.
- Proxy-centered: there is no involvement on the part of
end-client implementations.
- Solely a performance optimization: it provides no
information or functionality that is not already available
in HTTP/1.1. The intent is to improve performance overall,
and reduce latency for almost all interactions; latency
might be increased for a small fraction of HTTP
interactions.
- Best-efforts: it does not guarantee the accuracy of the
reported information, although it does provide accurate
results in the absence of persistent network failures or
host crashes.
- Neutral with respect to privacy: it reveals to servers no
information about clients that is not already available
through the existing features of HTTP/1.1.
The goals of this specification do not include:
- Solving the entire problem of efficiently obtaining
extensive information about requests made via proxies.
- Improving the protection of user privacy (although our
proposal may reduce the transfer of user-specific
information to servers, it does not prevent it).
- Preventing or encouraging the use of log-exchange
mechanisms.
- Avoiding all forms of "cache-busting", or even all
cache-busting done for gathering counts.
This design has certain potential limitations:
- If it is not deployed widely in both proxies and servers,
it will provide little benefit.
- It may, by partially solving the hit-counting problem,
reduce the pressure to adopt more complete solutions, if
any become available.
- Even if widely deployed, it might not be widely used, and
so might not significantly improve performance.
These potential limitations might not be problems in actual practice.
1.2 Brief summary of the design
This section is included for people not wishing to read the entire
document; it is not a specification for the proposed design, and
over-simplifies many ASPects of the design.
The goal of this design is to eliminate the need for origin servers
to use "cache-busting" techniques, when this is done just for the
purpose of counting the number of users of a resource. (Cache-
busting includes techniques such as setting immediate Expiration
dates, or sending "Cache-control: private" in each response.)
The design adds a new "Meter" header to HTTP; the header is always
protected by the "Connection" header, and so is always hop-by-hop.
This mechanism allows the construction of a "metering subtree", which
is a connected subtree of proxies, rooted at an origin server. Only
those proxies that explicitly volunteer to join in the metering
subtree for a resource participate in hit-metering, but those proxies
that do volunteer are required to make their best effort to provide
accurate counts. When a hit-metered response is forwarded outside of
the metering subtree, the forwarding proxy adds "Cache-control: s-
maxage=0", so that other proxies (outside the metering subtree) are
forced to forward all requests to a server in the metering subtree.
NOTE: the HTTP/1.1 specification does not currently define a "s-
maxage" Cache-control directive. The HTTP working group has
decided to add such a directive to the next revision of the
HTTP/1.1 specification [7].
The Meter header carries zero or more directives, similar to the way
that the Cache-control header carries directives. Proxies may use
certain Meter directives to volunteer to do hit-metering for a
resource. If a proxy does volunteer, the server may use certain
directives to require that a response be hit-metered. Finally,
proxies use a "count" Meter directive to report the accumulated hit
counts.
The Meter mechanism can also be used by a server to limit the number
of uses that a cache may make of a cached response, before
revalidating it.
The full specification includes complete rules for counting "uses" of
a response (e.g., non-conditional GETs) and "reuses" (conditional
GETs). These rules ensure that the results are entirely consistent
in all cases, except when systems or networks fail.
1.3 Terminology
This document uses terms defined and explained in the HTTP/1.1
specification [4], including "origin server," "resource," "hop-by-
hop," "unconditional GET," and "conditional GET." The reader is
expected to be familiar with the HTTP/1.1 specification and its
terminology.
The key Words "MUST", "MUST NOT", "REQUIRED", "SHOULD", SHOULD NOT",
"RECOMMENDED", "MAY", and "OPTIONAL" in this document are to be
interpreted as described in RFC2119 [1].
2 Overview
The design described in this document introduces several new features
to HTTP:
- Hit-metering: allows an origin server to obtain reasonably
accurate counts of the number of clients using a resource
instance via a proxy cache, or a hierarchy of proxy caches.
- Usage-limiting: allows an origin server to control the
number of times a cached response may be used by a proxy
cache, or a hierarchy of proxy caches, before revalidation
with the origin server.
These new non-mandatory features require minimal new protocol
support, no change in protocol version, relatively little overhead in
message headers. The design adds no additional network round-trips
in any critical path that directly affects user-perceived latency
(see section 4.3 for an analysis).
The primary goal of hit-metering and usage-limiting is to obviate the
need for an origin server to send "Cache-control: s-maxage=0" with
responses for resources whose value is not likely to change
immediately. In other words, in cases where the only reason for
contacting the origin server on every request that might otherwise be
satisfied by a proxy cache entry is to allow the server to collect
demographic information or to control the number of times a cache
entry is used, the extension proposed here will avoid a significant
amount of unnecessary network traffic and latency.
This design introduces one new "Meter" header, which is used both in
HTTP request messages and HTTP response messages. The Meter header
is used to transmit a number of directives and reports. In
particular, all negotiation of the use of hit-metering and usage
limits is done using this header. No other changes to the existing
HTTP/1.1 specification [4] are proposed in this document.
This design also introduces several new concepts:
1. The concepts of a "use" of a cache entry, which is when a
proxy returns its entity-body in response to a conditional
or non-conditional request, and the "reuse" of a cache
entry, which is when a proxy returns a 304 (Not Modified)
response to a conditional request which is satisfied by
that cache entry.
2. The concept of a hit-metered resource, for which proxy
caches make a best-effort attempt to report accurate
counts of uses and/or reuses to the origin server.
3. The concept of a usage-limited resource, for which the
origin server expects proxy caches to limit the number of
uses and/or reuses.
The new Meter directives and reports interact to allow proxy caches
and servers to cooperate in the collection of demographic data. The
goal is a best-efforts approximation of the true number of uses
and/or reuses, not a guaranteed exact count.
The new Meter directives also allow a server to bound the inaccuracy
of a particular hit-count, by bounding the number of uses between
reports. It can also, for example, bound the number of times the
same ad is shown because of caching.
Section 7.1 describes a way to use server-driven content negotiation
(the Vary header) that allows an HTTP origin server to flexibly
separate requests into categories and count requests by category.
Implementation of such a categorized hit counting is likely to be a
very small modification to most implementations of Vary; some
implementations may not require any modification at all.
2.1 Discussion
Mapping this onto the publishing model, a proxy cache would increment
the use-count for a cache entry once for each unconditional GET done
for the entry, and once for each conditional GET that results in
sending a copy of the entry to update a client"s invalid cached copy.
Conditional GETs that result in 304 (Not Modified) are not included
in the use-count, because they do not result in a new user seeing the
page, but instead signify a repeat view by a user that had seen it
before. However, 304 responses are counted in the reuse-count.
HEADs are not counted at all, because their responses do not contain
an entity-body.
The Meter directives apply only to shared proxy caches, not to end-
client (or other single-user) caches. Single user caches should not
use Meter, because their hits will be automatically counted as a
result of the unconditional GET with which they first fetch the page,
from either the origin-server or from a proxy cache. Their
subsequent conditional GETs do not result in a new user seeing the
page.
The mechanism specified here counts GETs; other methods either do not
result in a page for the user to read, aren"t cached, or are
"written-through" and so can be directly counted by the origin
server. (If, in the future, a "cachable POST" came into existence,
whereby the entity-body in the POST request was used to select a
cached response, then such POSTs would have to be treated just like
GETs.) The applicability of hit-metering to any new HTTP methods
that might be defined in the future is currently unspecifiable.
In the case of multiple caches along a path, a proxy cache does the
obvious summation when it receives a use-count or reuse-count in a
request from another cache.
3 Design concepts
In order to allow the introduction of hit-metering and usage-limiting
without requiring a protocol revision, and to ensure a reasonably
close approximation of accurate counts, the negotiation of metering
and usage-limiting is done hop-by-hop, not end-to-end. If one
considers the "tree" of proxies that receive, store, and forward a
specific response, the intent of this design is that within some
(possibly null) "metering subtree", rooted at the origin server, all
proxies are using the hit-metering and/or usage-limiting requested by
the origin server.
Proxies at the leaves of this subtree will insert a "Cache-control:
s-maxage=0" directive, which forces all other proxies (below this
subtree) to check with a leaf of the metering subtree on every
request. However, it does not prevent them from storing and using
the response, if the revalidation succeeds.
No proxy is required to implement hit-metering or usage-limiting.
However, any proxy that transmits the Meter header in a request MUST
implement every unconditional requirement of this specification,
without exception or amendment.
This is a conservative design, which may sometimes fail to take
advantage of hit-metering support in proxies outside the metering
subtree. However, it is likely that without the reliability offered
by a conservative design, managers of origin servers with
requirements for accurate approximations will not take advantage of
any hit-metering proposal.
The hit-metering/usage-limiting mechanism is designed to avoid any
extra network round-trips in the critical path of any client request,
and (as much as possible) to avoid excessively lengthening HTTP
messages.
The Meter header is used to transmit both negotiation information and
numeric information.
A formal specification for the Meter header appears in section 5; the
following discussion uses an informal approach to improve clarity.
3.1 Implementation of the "metering subtree"
The "metering subtree" approach is implemented in a simple,
straightforward way by defining the new "Meter" header as one that
MUST always be protected by a Connection header in every request or
response. I.e., if the Meter header is present in an HTTP message,
that message:
1. MUST contain "Connection: meter", and MUST be handled
according to the HTTP/1.1 specification of the Connection
header.
2. MUST NOT be sent in response to a request from a client
whose version number is less than HTTP/1.1.
3. MUST NOT be accepted from a client whose version number is
less than HTTP/1.1.
The reason for the latter two restrictions is to protect against
proxies that might not properly implement the Connection header.
Otherwise, a subtree that includes an HTTP/1.0 proxy might
erroneously appear to be a metering subtree.
Note: It appears that for the Connection header mechanism to
function correctly, a system receiving an HTTP/1.0 (or lower-
version) message that includes a Connection header must act as if
this header, and all of the headers it protects, ought to have
been removed from the message by an intermediate proxy.
Although RFC2068 does not specifically require this behavior, it
appears to be implied. Otherwise, one could not depend on the
stated property (section 14.10) that the protected options "MUST
NOT be communicated by proxies over further connections." This
should probably be clarified in a subsequent draft of the HTTP/1.1
specification.
This specification does not, in any way, propose a modification of
the specification of the Connection header.
From the point of view of an origin server, the proxies in a metering
subtree work together to obey usage limits and to maintain accurate
usage counts. When an origin server specifies a usage limit, a proxy
in the metering subtree may subdivide this limit among its children
in the subtree as it sees fit. Similarly, when a proxy in the
subtree receives a usage report, it ensures that the hits represented
by this report are summed properly and reported to the origin server.
When a proxy forwards a hit-metered or usage-limited response to a
client (proxy or end-client) not in the metering subtree, it MUST
omit the Meter header, and it MUST add "Cache-control: s-maxage=0" to
the response.
3.2 Format of the Meter header
The Meter header is used to carry zero or more directives. Multiple
Meter headers may occur in an HTTP message, but according to the
rules in section 4.2 of the HTTP/1.1 specification [4], they may be
combined into a single header (and should be so combined, to reduce
overhead).
For example, the following sequence of Meter headers
Meter: max-uses=3
Meter: max-reuses=10
Meter: do-report
may be expressed as
Meter: max-uses=3, max-reuses=10, do-report
3.3 Negotiation of hit-metering and usage-limiting
An origin server that wants to collect hit counts for a resource, by
simply forcing all requests to bypass any proxy caches, would respond
to requests on the resource with "Cache-control: s-maxage=0". (An
origin server wishing to prevent HTTP/1.0 proxies from improperly
caching the response could also send both "Expires: <now>", to
prevent such caching, and "Cache-control: max-age=NNNN", to allow
newer proxies to cache the response).
The purpose of the Meter header is to obviate the need for "Cache-
control: s-maxage=0" within a metering subtree. Thus, any proxy may
negotiate the use of hit-metering and/or usage-limiting with the
next-hop server. If this server is the origin server, or is already
part of a metering subtree (rooted at the origin server), then it may
complete the negotiation, thereby extending the metering subtree to
include the new proxy.
To start the negotiation, a proxy sends its request with one of the
following Meter directives:
will-report-and-limit
indicates that the proxy is willing and able to
return usage reports and will obey any usage-limits.
wont-report indicates that the proxy will obey usage-limits but
will not send usage reports.
wont-limit indicates that the proxy will not obey usage-limits
but will send usage reports.
A proxy willing to neither obey usage-limits nor send usage reports
MUST NOT transmit a Meter header in the request.
By definition, an empty Meter header:
Meter:
is equivalent to "Meter: will-report-and-limit", and so, by the
definition of the Connection header (see section 14.10 of the
HTTP/1.1 specification [4]), a request that contains
Connection: Meter
and no explicit Meter header is equivalent to a request that contains
Connection: Meter
Meter: will-report-and-limit
This makes the default case more efficient.
An origin server that is not interested in metering or usage-limiting
the requested resource simply ignores the Meter header.
If the server wants the proxy to do hit-metering and/or usage-
limiting, its response should include one or more of the following
Meter directives:
For hit-metering:
do-report specifies that the proxy MUST send usage reports to
the server.
dont-report specifies that the proxy SHOULD NOT send usage
reports to the server.
timeout=NNN sets a metering timeout of NNN minutes, from the time
that this response was originated, for the reporting
of a hit-count. If the proxy has a non-zero hit
count for this response when the timeout expires, it
MUST send a report to the server at or before that
time. Implies "do-report".
By definition, an empty Meter header in a response, or any Meter
header that does not contain "dont-report", means "Meter: do-report";
this makes a common case more efficient.
Note: an origin server using the metering timeout mechanism to
bound the collection period over which hit-counts are obtained
should adjust the timeout values in the responses it sends so that
all responses generated within that period reach their metering
timeouts at or before the end of that period.
If the origin server simply sends a constant metering timeout T
with each response for a resource, the reports that it receives
will reflect activity over a period whose duration is between T
and N*T (in the worst case), where N is the maximum depth of the
metering subtree.
For usage-limiting
max-uses=NNN sets an upper limit of NNN "uses" of the response,
not counting its immediate forwarding to the
requesting end-client, for all proxies in the
following subtree taken together.
max-reuses=NNN sets an upper limit of NNN "reuses" of the response
for all proxies in the following subtree taken
together.
When a proxy has exhausted its allocation of "uses" or "reuses" for a
cache entry, it MUST revalidate the cache entry (using a conditional
request) before returning it in a response. (The proxy SHOULD use
this revalidation message to send a usage report, if one was
requested and it is time to send it. See sections 3.4 and 3.5.)
These Meter response-directives apply only to the specific response
that they are attached to.
Note that the limit on "uses" set by the max-uses directive does
not include the use of the response to satisfy the end-client
request that caused the proxy"s request to the server. This
counting rule supports the notion of a cache-initiated prefetch: a
cache may issue a prefetch request, receive a max-uses=0 response,
store that response, and then return that response (without
revalidation) when a client makes an actual request for the
resource. However, each such response may be used at most once in
this way, so the origin server maintains precise control over the
number of actual uses.
A server MUST NOT send a Meter header that would require a proxy to
do something that it has not yet offered to do. A proxy receiving a
Meter response-directive aSKINg the proxy to do something it did not
volunteer to do SHOULD ignore that directive.
A proxy receiving a Meter header in a response MUST either obey it,
or it MUST revalidate the corresponding cache entry on every access.
(I.e., if it chooses not to obey the Meter header in a response, it
MUST act as if the response included "Cache-control: s-maxage=0".)
Note: a proxy that has not sent the Meter header in a request for
the given resource, and which has therefore not volunteered to
honor Meter directives in a response, is not required to honor
them. If, in this situation, the server does send a Meter header
in a response, this is a protocol error. However, based on the
robustness principle, the proxy may choose to interpret the Meter
header as an implicit request to include "Cache-control: s-
maxage=0" when it forwards the response, since this preserves the
apparent intention of the server.
A proxy that receives the Meter header in a request may ignore it
only to the extent that this is consistent with its own duty to the
next-hop server. If the received Meter request header is
inconsistent with that duty, or if no Meter request header is
received and the response from the next-hop server requests any form
of metering or limiting, then the proxy MUST add "Cache-control: s-
maxage=0" to any response it forwards for that request. (A proxy
SHOULD NOT add or change the Expires header or max-age Cache-control
directive.)
For example, if proxy A receives a GET request from proxy B for
URL X with "Connection: Meter", but proxy A"s cached response for
URL does not include any Meter directives, then proxy A may ignore
the metering offer from proxy B.
However, if proxy A has previously told the origin server "Meter:
wont-limit" (implying will-report), and the cached response
contains "Meter: do-report", and proxy B"s request includes
"Meter: wont-report", then proxy B"s offer is inconsistent with
proxy A"s duty to the origin server. Therefore, in this case
proxy A must add "Cache-control: s-maxage=0" when it returns the
cached response to proxy B, and must not include a Meter header in
this response.
If a server does not want to use the Meter mechanism, and will not
want to use it any time soon, it may send this directive:
wont-ask recommends that the proxy SHOULD NOT send any Meter
directives to this server.
The proxy SHOULD remember this fact for up to 24 hours. This avoids
virtually all unnecessary overheads for servers that do not wish to
use or support the Meter header. (This directive also implies
"dont-report".)
3.4 Transmission of usage reports
To transmit a usage report, a proxy sends the following Meter header
in a request on the appropriate resource:
Meter: count=NNN/MMM
The first integer indicates the count of uses of the cache entry
since the last report; the second integer indicates the count of
reuses of the entry (see section 5.3 for rules on counting uses and
reuses). The transmission of a "count" directive in a request with
no other Meter directive is also defined as an implicit transmission
of a "will-report-and-limit" directive, to optimize the common case.
(A proxy not willing to honor usage-limits would send "Meter:
count=NNN/MMM, wont-limit" for its reports.)
Note that when a proxy forwards a client"s request and receives a
response, the response that the proxy sends immediately to the
requesting client is not counted as a "use". I.e., the reported
count is the number of times the cache entry was used, and not the
number of times that the response was used.
A proxy SHOULD NOT transmit "Meter: count=0/0", since this conveys no
useful information.
Usage reports MUST always be transmitted as part of a conditional
request (such as a GET or HEAD), since the information in the
conditional header (e.g., If-Modified-Since or If-None-Match) is
required for the origin server to know which instance of a resource
is being counted. Proxys forwarding usage reports up the metering
subtree MUST NOT change the contents of the conditional header, since
otherwise this would result in incorrect counting.
A usage report MUST NOT be transmitted as part of a forwarded request
that includes multiple entity tags in an If-None-Match or If-Match
header.
Note: a proxy that offers its willingness to do hit-metering
(report usage) must count both uses and reuses. It is not
possible to negotiate the reporting of one but not the other.
3.5 When to send usage reports
A proxy that has offered to send usage reports to its parent in the
metering subtree MUST send a usage report in each of these
situations:
1. When it forwards a conditional GET on the resource
instance on behalf of one of its clients (if the GET is
conditional on at most one entity-tag).
2. When it forwards a conditional HEAD on the resource
instance on behalf of one of its clients.
3. When it must generate a conditional GET to satisfy a
client request because the max-uses limit has been
exceeded.
4. Upon expiration of a metering timeout associated with a
cache entry that has a non-zero hit-count.
5. When it removes the corresponding non-zero hit-count entry
from its storage for any reason including:
- the proxy needs the storage space for another
hit-count entry.
- the proxy is not able to store more than one response
per resource, and a request forwarded on behalf of a
client has resulted in the receipt of a new response
(one with a different entity-tag or last-modified
time).
Note that a cache might continue to store hit-count information
even after having deleted the body of the response, so it is
not necessary to report the hit-count when deleting the body;
it is only necessary to report it if the proxy is about to
"forget" a non-zero value.
(Section 5.3 explains how hit-counts become zero or non-zero.)
If the usage report is being sent because the proxy is about to
remove the hit-count entry from its storage, or because of an expired
metering timeout:
- The proxy MUST send the report as part of a conditional
HEAD request on the resource instance.
- The proxy is not required to retry the HEAD request if it
fails (this is a best-efforts design). To improve
accuracy, however, the proxy SHOULD retry failed HEAD
requests, subject to resource constraints.
- The proxy is not required to serialize any other operation
on the completion of this request.
Note: proxy implementors are strongly encouraged to batch several
HEAD-based reports to the same server, when possible, over a
single persistent connection, to reduce network overhead as much
as possible. This may involve a non-naive algorithm for
scheduling the deletion of hit-count entries.
If the usage count is sent because of an arriving request that also
carries a "count" directive, the proxy MUST combine its own (possibly
zero) use and reuse counts with the arriving counts, and then attempt
to forward the request.
However, the proxy is not required to forward an arriving request
with a "count" directive, provided that:
- it can reply to the request using a cached response, in
compliance with other requirements of the HTTP
specification.
- such a response does not exceed a max-uses limit.
- it is not required to forward the request because of an
expired metering timeout.
If an arriving request carries a "count" directive, and the proxy no
longer has a cache entry for the resource, the proxy MUST forward the
"count" directive. (This is, in any case, what a proxy without a
suitable cache entry would normally do for any valid request it
receives.)
3.6 Subdivision of usage-limits
When an origin server specifies a usage limit, a proxy in the
metering subtree may subdivide this limit among its children in the
subtree as it sees fit.
For example, consider the situation with two proxies P1 and P2, each
of which uses proxy P3 as a way to reach origin server S. Imagine
that S sends P3 a response with
Meter: max-uses=10
The proxies use that response to satisfy the current requesting end-
client. The max-uses directive in this example allows the
combination of P1, P2, and P3 together to satisfy 10 additional end-
client uses (unconditional GETs) for the resource.
This specification does not constrain how P3 divides up that
allocation among itself and the other proxies. For example, P3 could
retain all of max-use allocation for itself. In that case, it would
forward the response to P1 and/or P2 with
Meter: max-uses=0
P3 might also divide the allocation equally among P1 and P2,
retaining none for itself (which may be the right choice if P3 has
few or no other clients). In this case, it could send
Meter: max-uses=5
to the proxy (P1 or P2) that made the initial request, and then
record in some internal data structure that it "owes" the other proxy
the rest of the allocation.
Note that this freedom to choose the max-uses value applies to the
origin server, as well. There is no requirement that an origin
server send the same max-uses value to all caches. For example, it
might make sense to send "max-uses=2" the first time one hears from a
cache, and then double the value (up to some maximum limit) each time
one gets a "use-count" from that cache. The idea is that the faster
a cache is using up its max-use quota, the more likely it will be to
report a use-count value before removing the cache entry. Also, high
and frequent use-counts imply a corresponding high efficiency benefit
from allowing caching.
Again, the details of such heuristics would be outside the scope of
this specification.
4 Analysis
This section includes informal analyses of several aspects of hit-
metering:
1. the accuracy of results when applied to counting users
(section 4.1).
2. the problem of counting users whose browsers do not
include caches, such as Network Computers (section 4.2).
3. delays imposed on "critical paths" for HTTP operations
(section 4.3).
4.1 Approximation accuracy for counting users
For many (but not all) service operators, the single most important
aspect of the request stream is the number of distinct users who have
retrieved a particular entity within a given period (e.g., during a
given day). The hit-metering mechanism is designed to provide an
origin server with an approximation of the number of users that
reference a given resource. The intent of the design is that the
precision of this approximation is consistent with the goals of
simplicity and optional implementation.
Almost all Web users use client software that maintains local caches,
and the state of the art of local-caching technology is quite
effective. (Section 4.2 discusses the case where end-client caches
are small or non-existent.) Therefore, assuming an effective and
persistent end-client cache, each individual user who retrieves an
entity does exactly one GET request that results in a 200 or 203
response, or a 206 response that includes the first byte of the
entity. If a proxy cache maintains and reports an accurate use-count
of such retrievals, then its reported use-count will closely
approximate the number of distinct users who have retrieved the
entity.
There are some circumstances under which this approximation can break
down. For example, if an entity stays in a proxy cache for much
longer than it persists in the typical client cache, and users often
re-reference the entity, then this scheme will tend to over-count the
number of users. Or, if the cache-management policy implemented in
typical client caches is biased against retaining certain kinds of
frequently re-referenced entities (such as very large images), the
use-counts reported will tend to overestimate the user-counts for
such entities.
Browser log analysis has shown that when a user revisits a resource,
this is almost always done very soon after the previous visit, almost
always with fewer than eight intervening references [11]. Although
this result might not apply universally, it implies that almost all
reuses will hit in the end-client cache, and will not be seen as
unconditional GETs by a proxy cache.
The existing (HTTP/1.0) "cache-busting" mechanisms for counting
distinct users will certainly overestimate the number of users behind
a proxy, since it provides no reliable way to distinguish between a
user"s initial request and subsequent repeat requests that might have
been conditional GETs, had not cache-busting been employed. The
"Cache-control: s-maxage=0" feature of HTTP/1.1 does allow the
separation of use-counts and reuse-counts, provided that no HTTP/1.0
proxy caches intervene.
Note that if there is doubt about the validity of the results of
hit-metering a given set of resources, the server can employ cache-
busting techniques for short periods, to establish a baseline for
validating the hit-metering results. Various approaches to this
problem are discussed in a paper by James Pitkow [9].
4.2 What about "Network Computers"?
The analysis in section 4.1 assumed that "almost all Web users" have
client caches. If the Network Computers (NC) model becomes popular,
however, then this assumption may be faulty: most proposed NCs have
no disk storage, and relatively little RAM. Many Personal Digital
Assistants (PDAs), which sometimes have network access, have similar
constraints. Such client systems may do little or no caching of HTTP
responses. This means that a single user might well generate many
unconditional GETs that yield the same response from a proxy cache.
First note that the hit-metering design in this document, even with
such clients, provides an approximation no worse than available with
unmodified HTTP/1.1: the counts that a proxy would return to an
origin server would represent exactly the number of requests that the
proxy would forward to the server, if the server simply specifies
"Cache-control: s-maxage=0".
However, it may be possible to improve the accuracy of these hit-
counts by use of some heuristics at the proxy. For example, the
proxy might note the IP address of the client, and count only one GET
per client address per response. This is not perfect: for example,
it fails to distinguish between NCs and certain other kinds of hosts.
The proxy might also use the heuristic that only those clients that
never send a conditional GET should be treated this way, although we
are not at all certain that NCs will never send conditional GETs.
Since the solution to this problem appears to require heuristics
based on the actual behavior of NCs (or perhaps a new HTTP protocol
feature that allows unambiguous detection of cacheless clients), it
appears to be premature to specify a solution.
4.3 Critical-path delay analysis
In systems (such as the Web) where latency is at issue, there is
usually a tree of steps which depend on one another, in such a way
that the final result cannot be accomplished until all of its
predecessors have been. Since the tree structure admits some
parallelism, it is not necessary to add up the timings for each step
to discover the latency for the entire process. But any single path
through this dependency tree cannot be parallelized, and the longest
such path is the one whose length (in units of seconds) determines
the overall latency. This is the "critical path", because no matter
how much shorter one makes any other path, that cannot change the
overall latency for the final result.
If one views the final result, for a Web request, as rendering a page
at a browser, or otherwise acting on the result of a request, clearly
some network round trips (e.g., exchanging TCP SYN packets if the
connection doesn"t already exist) are on the critical path. This
hit-metering design does add some round-trips for reporting non-zero
counts when a cache entry is removed, but, by design, these are off
any critical path: they may be done in parallel with any other
operation, and require only "best efforts", so a proxy does not have
to serialize other operations with their success or failure.
Clearly, anything that changes network utilization (either increasing
or decreasing it) can indirectly affect user-perceived latency. Our
expectation is that hit-metering, on average, will reduce loading and
so even its indirect effects should not add network round-trips in
any critical path. But there might be a few specific instances where
the added non-critical-path operations (specifically, usage reports
upon cache-entry removal) delay an operation on a critical path.
This is an unavoidable problem in datagram networks.
5 Specification
5.1 Specification of Meter header and directives
The Meter general-header field is used to:
- Negotiate the use of hit-metering and usage-limiting among
origin servers and proxy caches.
- Report use counts and reuse counts.
Implementation of the Meter header is optional for both proxies and
origin servers. However, any proxy that transmits the Meter header
in a request MUST implement every requirement of this specification,
without exception or amendment.
The Meter header MUST always be protected by a Connection header. A
proxy that does not implement the Meter header MUST NOT pass it
through to another system (see section 5.5 for how a non-caching
proxy may comply with this specification). If a Meter header is
received in a message whose version is less than HTTP/1.1, it MUST be
ignored (because it has clearly flowed through a proxy that does not
implement Meter).
A proxy that has received a response with a version less than
HTTP/1.1, and therefore from a server (or another proxy) that does
not implement the Meter header, SHOULD NOT send Meter request
directives to that server, because these would simply waste
bandwidth. This recommendation does not apply if the proxy is
currently hit-metering or usage-limiting any responses from that
server. If the proxy receives a HTTP/1.1 or higher response from
such a server, it should cease its suppression of the Meter
directives.
All proxies sending the Meter header MUST adhere to the "metering
subtree" design described in section 3.
Meter = "Meter" ":" 0#meter-directive
meter-directive = meter-request-directive
meter-response-directive
meter-report-directive
meter-request-directive =
"will-report-and-limit"
"wont-report"
"wont-limit"
meter-report-directive =
"count" "=" 1*DIGIT "/" 1*DIGIT
meter-response-directive =
"max-uses" "=" 1*DIGIT
"max-reuses" "=" 1*DIGIT
"do-report"
"dont-report"
"timeout" "=" 1*DIGIT
"wont-ask"
A meter-request-directive or meter-report-directive may only appear
in an HTTP request message. A meter-response-directive may only
appear in an HTTP response directive.
An empty Meter header in a request means "Meter: will-report-and-
limit". An empty Meter header in a response, or any other response
including one or more Meter headers without the "dont-report" or
"wont-ask" directive, implies "Meter: do-report".
The meaning of the meter-request-directives are as follows:
will-report-and-limit
indicates that the proxy is willing and able to
return usage reports and will obey any usage-limits.
wont-report indicates that the proxy will obey usage-limits but
will not send usage reports.
wont-limit indicates that the proxy will not obey usage-limits
but will send usage reports.
A proxy willing neither to obey usage-limits nor to send usage
reports MUST NOT transmit a Meter header in the request.
The meaning of the meter-report-directives are as follows:
count "=" 1*DIGIT "/" 1*DIGIT
Both digit strings encode decimal integers. The
first integer indicates the count of uses of the
cache entry since the last report; the second integer
indicates the count of reuses of the entry.
Section 5.3 specifies the counting rules.
The meaning of the meter-response-directives are as follows:
max-uses "=" 1*DIGIT
sets an upper limit on the number of "uses" of the
response, not counting its immediate forwarding to
the requesting end-client, for all proxies in the
following subtree taken together.
max-reuses "=" 1*DIGIT
sets an upper limit on the number of "reuses" of the
response for all proxies in the following subtree
taken together.
do-report specifies that the proxy MUST send usage reports to
the server.
dont-report specifies that the proxy SHOULD NOT send usage
reports to the server.
timeout "=" 1*DIGIT
sets a metering timeout of the specified number of
minutes (not seconds) after the origination of this
response (as indicated by its "Date" header). If the
proxy has a non-zero hit count for this response when
the timeout expires, it MUST send a report to the
server at or before that time. Timeouts should be
implemented with an accuracy of plus or minus one
minute. Implies "do-report".
wont-ask specifies that the proxy SHOULD NOT send any Meter
headers to the server. The proxy should forget this
advice after a period of no more than 24 hours.
Section 5.3 specifies the counting rules, and in particular specifies
a somewhat non-obvious interpretation of the max-uses value.
5.2 Abbreviations for Meter directives
To allow for the most efficient possible encoding of Meter headers,
we define abbreviated forms of all Meter directives. These are
exactly semantically equivalent to their non-abbreviated
counterparts. All systems implementing the Meter header MUST
implement both the abbreviated and non-abbreviated forms.
Implementations SHOULD use the abbreviated forms in normal use.
The abbreviated forms of Meter directive are shown below, with the
corresponding non-abbreviated literals in the comments:
Abb-Meter = "Meter" ":" 0#abb-meter-directive
abb-meter-directive = abb-meter-request-directive
abb-meter-response-directive
abb-meter-report-directive
abb-meter-request-directive =
"w" ; "will-report-and-limit"
"x" ; "wont-report"
"y" ; "wont-limit"
abb-meter-report-directive =
"c" "=" 1*DIGIT "/" 1*DIGIT ; "count"
abb-meter-response-directive =
"u" "=" 1*DIGIT ; "max-uses"
"r" "=" 1*DIGIT ; "max-reuses"
"d" ; "do-report"
"e" ; "dont-report"
"t" "=" 1*DIGIT ; "timeout"
"n" ; "wont-ask"
Note: although the Abb-Meter BNF rule is defined separately from
the Meter rule, one may freely mix abbreviated and non-abbreviated
Meter directives in the same header.
5.3 Counting rules
Note: please remember that hit-counts and usage-counts are
associated with individual responses, not with resources. A cache
entry that, over its lifetime, holds more than one response is
also not a "response", in this particular sense.
Let R be a cached response, and V be the value of the Request-URI and
selecting request-headers (if any, see section 14.43 of the HTTP/1.1
specification [4]) that would select R if contained in a request. We
define a "use" of R as occurring when the proxy returns its stored
copy of R in a response with any of the following status codes: a 200
(OK) status; a 203 (Non-Authoritative Information) status; or a 206
(Partial Content) status when the response contains byte #0 of the
entity (see section 5.4 for a discussion of Range requests).
Note: when a proxy forwards a client"s request and receives a
response, the response that the proxy sends immediately to the
requesting client is not counted as a "use". I.e., the reported
count is the number of times the cache entry was used, and not the
number of times that the response was used.
We define a "reuse" of R as as occurring when the proxy responds to a
request selecting R with a 304 (Not Modified) status, unless that
request is a Range request that does not specify byte #0 of the
entity.
5.3.1 Counting rules for hit-metering
A proxy participating in hit-metering for a cache response R
maintains two counters, CU and CR, associated with R. When a proxy
first stores R in its cache, it sets both CU and CR to 0 (zero).
When a subsequent client request results in a "use" of R, the proxy
increments CU. When a subsequent client request results in a "reuse"
of R, the proxy increments CR. When a subsequent client request
selecting R (i.e., including V) includes a "count" Meter directive,
the proxy increments CU and CR using the corresponding values in the
directive.
When the proxy sends a request selecting R (i.e., including V) to the
inbound server, it includes a "count" Meter directive with the
current CU and CR as the parameter values. If this request was
caused by the proxy"s receipt of a request from a client, upon
receipt of the server"s response, the proxy sets CU and CR to the
number of uses and reuses, respectively, that may have occurred while
the request was in progress. (These numbers are likely, but not
certain, to be zero.) If the proxy"s request was a final HEAD-based
report, it need no longer maintain the CU and CR values, but it may
also set them to the number of intervening uses and reuses and retain
them.
5.3.2 Counting rules for usage-limiting
A proxy participating in usage-limiting for a response R maintains
either or both of two counters TU and TR, as appropriate, for that
resource. TU and TR are incremented in just the same way as CU and
CR, respectively. However, TU is zeroed only upon receipt of a
"max-uses" Meter directive for that response (including the initial
receipt). Similarly, TR is zeroed only upon receipt of a "max-
reuses" Meter directive for that response.
A proxy participating in usage-limiting for a response R also stores
values MU and/or MR associated with R. When it receives a response
including only a max-uses value, it sets MU to that value and MR to
infinity. When it receives a response including only a max-reuses
value, it sets MR to that value and MU to infinity. When it receives
a response including both max-reuses and max-reuses values, it sets
MU and MR to those values, respectively. When it receives a
subsequent response including neither max-reuses nor max-reuses
values, it sets both MU and MR to infinity.
If a proxy participating in usage-limiting for a response R receives
a request that would cause a "use" of R, and TU >= MU, it MUST
forward the request to the server. If it receives a request that
would cause a "reuse" of R, and TR >= MR, it MUST forward the request
to the server. If (in either case) the proxy has already forwarded a
previous request to the server and is waiting for the response, it
should delay further handling of the new request until the response
arrives (or times out); it SHOULD NOT have two revalidation requests
pending at once that select the same response, unless these are Range
requests selecting different subranges.
There is a special case of this rule for the "max-uses" directive: if
the proxy receives a response with "max-uses=0" and does not forward
it to a requesting client, the proxy should set a flag PF associated
with R. If R is true, then when a request arrives while if TU >= MU,
if the PF flag is set, then the request need not be forwarded to the
server (provided that this is not required by other caching rules).
However, the PF flag MUST be cleared on any use of the response.
Note: the "PF" flag is so named because this feature is useful
only for caches that could issue a "prefetch" request before an
actual client request for the response. A proxy not implementing
prefetching need not implement the PF flag.
5.3.3 Equivalent algorithms are allowed
Any other algorithm that exhibits the same external behavior (i.e.,
generates exactly the same requests from the proxy to the server) as
the one in this section is explicitly allowed.
Note: in most cases, TU will be equal to CU, and TR will be
equal to CR. The only two cases where they could differ are:
1. The proxy issues a non-conditional request for the
resource using V, while TU and/or TR are non-zero, and
the server"s response includes a new "max-uses" and/or
"max-reuses" directive (thus zeroing TU and/or TR, but
not CU and CR).
2. The proxy issues a conditional request reporting the
hit-counts (and thus zeroing CU and CR, but not TU or
TR), but the server"s response does not include a new
"max-uses" and/or "max-reuses" directive.
To solve the first case, the proxy has several implementation
options
- Always store TU and TR separately from CU and CR.
- Create "shadow" copies of TU and TR when this situation
arises (analogous to "copy on write").
- Generate a HEAD-based usage report when the
non-conditional request is sent (or when the
"max-uses=0" is received), causing CU and CR to be
zeroed (analogous in some ways to a "memory barrier"
instruction).
In the second case, the server implicitly has removed the
usage-limit(s) on the response (by setting MU and/or MR to
infinity), and so the fact that, say, TU is different from CU
is not significant.
Note: It may also be possible to eliminate the PF flag by
sending extra HEAD-based usage-report requests, but we
recommend against this; it is better to allocate an extra bit
per entry than to transmit extra requests.
5.4 Counting rules: interaction with Range requests
HTTP/1.1 allows a client to request sub-ranges of a resource. A
client might end up issuing several requests with the net effect of
receiving one copy of the resource. For uniformity of the results
seen by origin servers, proxies need to observe a rule for counting
these references, although it is not clear that one rule generates
accurate results in every case.
The rule established in this specification is that proxies count as a
"use" or "reuse" only those Range requests that result in the return
of byte #0 of the resource. The rationale for this rule is that in
almost every case, an end-client will retrieve the beginning of any
resource that it references at all, and that it will seldom retrieve
any portion more than once. Therefore, this rule appears to meet the
goal of a "best-efforts" approximation.
5.5 Implementation by non-caching proxies
A non-caching proxy may participate in the metering subtree; this is
strongly recommended.
A non-caching proxy (HTTP/1.1 or higher) that participates in the
metering subtree SHOULD forward Meter headers on both requests and
responses, with the appropriate Connection headers.
If a non-caching proxy forwards Meter headers, it MUST comply with
these restrictions:
1. If the proxy forwards Meter headers in responses, such a
response MUST NOT be returned to any request except the
one that elicited it.
2. Once a non-caching proxy starts forwarding Meter headers,
it should not arbitrarily stop forwarding them (or else
reports may be lost).
A proxy that caches some responses and not others, for whatever
reason, may choose to implement the Meter header as a caching proxy
for the responses that it caches, and as a non-caching proxy for the
responses that it does not cache, as long as its external behavior
with respect to any particularly response is fully consistent with
this specification.
5.6 Implementation by cooperating caches
Several HTTP cache implementations, most notably the Harvest/Squid
cache [2], create cooperative arrangements between several caches.
If such caches use a protocol other than HTTP to communicate between
themselves, such as the Internet Cache Protocol (ICP) [12], and if
they implement the Meter header, then they MUST act to ensure that
their cooperation does not violate the intention of this
specification.
In particular, if one member of a group of cooperating caches agrees
with a server to hit-meter a particular response, and then passes
this response via a non-HTTP protocol to a second cache in the group,
the caches MUST ensure that the server which requested the metering
receives reports that appropriately account for any uses or resues
made by the second cache. Similarly, if the first cache agreed to
usage-limit the response, the total number of uses by the group of
caches MUST be limited to the agreed-upon number.
6 Examples
6.1 Example of a complete set of exchanges
This example shows how the protocol is intended to be used most of
the time: for hit-metering without usage-limiting. Entity bodies are
omitted.
A client sends request to a proxy:
GET http://foo.com/bar.Html HTTP/1.1
The proxy forwards request to the origin server:
GET /bar.html HTTP/1.1
Host: foo.com
Connection: Meter
thus offering (implicitly) "will-report-and-limit".
The server responds to the proxy:
HTTP/1.1 200 OK
Date: Fri, 06 Dec 1996 18:44:29 GMT
Cache-control: max-age=3600
Connection: meter
Etag: "abcde"
thus (implicitly) requiring "do-report" (but not requiring
usage-limiting).
The proxy responds to the client:
HTTP/1.1 200 OK
Date: Fri, 06 Dec 1996 18:44:29 GMT
Etag: "abcde"
Cache-control: max-age=3600, proxy-mustcheck
Age: 1
Since the proxy does not know if its client is an end-system, or a
proxy that doesn"t do metering, it adds the "proxy-mustcheck"
directive.
Another client soon asks for the resource:
GET http://foo.com/bar.html HTTP/1.1
and the proxy sends the same response as it sent to the other client,
except (perhaps) for the Age value.
After an hour has passed, a third client asks for the response:
GET http://foo.com/bar.html HTTP/1.1
But now the response"s max-age has been exceeded, so the proxy
revalidates the response with the origin server:
GET /bar.html HTTP/1.1
If-None-Match: "abcde"
Host: foo.com
Connection: Meter
Meter: count=1/0
thus simultaneously fulfilling its duties to validate the response
and to report the one "use" that wasn"t forwarded.
The origin server responds:
HTTP/1.1 304 Not Modified
Date: Fri, 06 Dec 1996 19:44:29 GMT
Cache-control: max-age=3600
Etag: "abcde"
so the proxy can use the original response to reply to the new
client; the proxy also zeros the use-count it associates with that
response.
Another client soon asks for the resource:
GET http://foo.com/bar.html HTTP/1.1
and the proxy sends the appropriate response.
After another few hours, the proxy decides to remove the cache entry.
When it does so, it sends to the origin server:
HEAD /bar.html HTTP/1.1
If-None-Match: "abcde"
Host: foo.com
Connection: Meter
Meter: count=1/0
reporting that one more use of the response was satisfied from the
cache.
6.2 Protecting against HTTP/1.0 proxies
An origin server that does not want HTTP/1.0 caches to store the
response at all, and is willing to have HTTP/1.0 end-system clients
generate excess GETs (which will be forwarded by HTTP/1.0 proxies)
could send this for its reply:
HTTP/1.1 200 OK
Cache-control: max-age=3600
Connection: meter
Etag: "abcde"
Expires: Sun, 06 Nov 1994 08:49:37 GMT
HTTP/1.0 caches will see the ancient Expires header, but HTTP/1.1
caches will see the max-age directive and will ignore Expires.
Note: although most major HTTP/1.0 proxy implementations observe
the Expires header, it is possible that some are in use that do
not. Use of the Expires header to prevent caching by HTTP/1.0
proxies might not be entirely reliable.
6.3 More elaborate examples
Here is a request from a proxy that is willing to hit-meter but is
not willing to usage-limit:
GET /bar.html HTTP/1.1
Host: foo.com
Connection: Meter
Meter: wont-limit
Here is a response from an origin server that does not want hit
counting, but does want "uses" limited to 3, and "reuses" limited to
6:
HTTP/1.1 200 OK
Cache-control: max-age=3600
Connection: meter
Etag: "abcde"
Expires: Sun, 06 Nov 1994 08:49:37 GMT
Meter: max-uses=3, max-reuses=6, dont-report
Here is the same example with abbreviated Meter directive names:
HTTP/1.1 200 OK
Cache-control: max-age=3600
Connection: meter
Etag: "abcde"
Expires: Sun, 06 Nov 1994 08:49:37 GMT
Meter:u=3,r=6,e
7 Interactions with content negotiation
This section describes two aspects of the interaction between hit-
metering