Network Working Group I. Cooper
Request for Comments: 3143 Equinix, Inc.
Category: Informational J. Dilley
Akamai Technologies, Inc.
June 2001

Known HTTP Proxy/Caching Problems

Status of this Memo

This memo provides information for the Internet community. It does
not specify an Internet standard of any kind. Distribution of this
memo is unlimited.

Copyright Notice

Copyright (C) The Internet Society (2001). All Rights Reserved.

Abstract

This document catalogs a number of known problems with World Wide Web
(WWW) (caching) proxies and cache servers. The goal of the document
is to provide a discussion of the problems and proposed workarounds,
and ultimately to improve conditions by illustrating problems. The
constrUCtion of this document is a joint effort of the Web caching
community.

Table of Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . 2
1.1 Problem Template . . . . . . . . . . . . . . . . . . . . . . 2
2. Known Problems . . . . . . . . . . . . . . . . . . . . . . . 4
2.1 Known Specification Problems . . . . . . . . . . . . . . . . 5
2.1.1 Vary header is underspecified and/or misleading . . . . . . 5
2.1.2 Client Chaining Loses Valuable Length Meta-Data . . . . . . 9
2.2 Known Architectural Problems . . . . . . . . . . . . . . . . 10
2.2.1 Interception proxies break client cache directives . . . . . 10
2.2.2 Interception proxies prevent introduction of new HTTP
methods . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.3 Interception proxies break IP address-based authentication . 12
2.2.4 Caching proxy peer selection in heterogeneous networks . . . 13
2.2.5 ICP Performance . . . . . . . . . . . . . . . . . . . . . . 15
2.2.6 Caching proxy meshes can break HTTP serialization of content 16
2.3 Known Implementation Problems . . . . . . . . . . . . . . . 17
2.3.1 User agent/proxy failover . . . . . . . . . . . . . . . . . 17
2.3.2 Some servers send bad Content-Length headers for files that
contain CR . . . . . . . . . . . . . . . . . . . . . . . 18

3. Security Considerations . . . . . . . . . . . . . . . . . . 18
References . . . . . . . . . . . . . . . . . . . . . . . . . 19
Authors' Addresses . . . . . . . . . . . . . . . . . . . . . 20
A. Archived Known Problems . . . . . . . . . . . . . . . . . . 21
A.1 Architectural . . . . . . . . . . . . . . . . . . . . . . . 21
A.1.1 Cannot specify multiple URIs for replicated resources . . . 21
A.1.2 Replica distance is unknown . . . . . . . . . . . . . . . . 22
A.1.3 Proxy resource location . . . . . . . . . . . . . . . . . . 23
A.2 Implementation . . . . . . . . . . . . . . . . . . . . . . . 23
A.2.1 Use of Cache-Control headers . . . . . . . . . . . . . . . . 23
A.2.2 Lack of HTTP/1.1 compliance for caching proxies . . . . . . 24
A.2.3 ETag support . . . . . . . . . . . . . . . . . . . . . . . . 25
A.2.4 Servers and content should be optimized for caching . . . . 26
A.3 Administration . . . . . . . . . . . . . . . . . . . . . . . 27
A.3.1 Lack of fine-grained, standardized hierarchy controls . . . 27
A.3.2 Proxy/Server exhaustive log format standard for analysis . . 27
A.3.3 Trace log timestamps . . . . . . . . . . . . . . . . . . . . 28
A.3.4 Exchange format for log summaries . . . . . . . . . . . . . 29
Full Copyright Statement . . . . . . . . . . . . . . . . . . 32

1. Introduction

This memo discusses problems with proxies - which act as
application-level intermediaries for Web requests - and more
specifically with caching proxies, which retain copies of previously
requested resources in the hope of improving overall quality of
service by serving the content locally. Commonly used terminology in
this memo can be found in the "Internet Web Replication and Caching
Taxonomy"[2].

No individual or organization has complete knowledge of the known
problems in Web caching, and the editors are grateful to the
contributors to this document.

1.1 Problem Template

A common problem template is used within the following sections. We
gratefully acknowledge RFC2525 [1] which helped define an initial
format for this known problems list. The template format is
summarized in the following table and described in more detail below.

Name: short, descriptive name of the problem (3-5 Words)
Classification: classifies the problem: performance, security, etc
Description: describes the problem succinctly
Significance: magnitude of problem, environments where it exists
Implications: the impact of the problem on systems and networks
See Also: a reference to a related known problem
Indications: states how to detect the presence of this problem

Solution(s): describe the solution(s) to this problem, if any
Workaround: practical workaround for the problem
References: information about the problem or solution
Contact: contact name and email address for this section

Name
A short, descriptive, name (3-5 words) name associated with the
problem.

Classification
Problems are grouped into categories of similar problems for ease
of reading of this memo. Choose the category that best describes
the problem. The suggested categories include three general
categories and several more specific categories.

* Architecture: the fundamental design is incomplete, or
incorrect

* Specification: the spec is ambiguous, incomplete, or incorrect.

* Implementation: the implementation of the spec is incorrect.

* Performance: perceived page response at the client is
excessive; network bandwidth consumption is excessive; demand
on origin or proxy servers exceed reasonable bounds.

* Administration: care and feeding of caches is, or causes, a
problem.

* Security: privacy, integrity, or authentication concerns.

Description
A definition of the problem, succinct but including necessary
background information.

Significance (High, Medium, Low)
May include a brief summary of the environments for which the
problem is significant.

Implications
Why the problem is viewed as a problem. What inappropriate
behavior results from it? This section should substantiate the
magnitude of any problem indicated with High significance.

See Also
Optional. List of other known problems that are related to this
one.

Indications
How to detect the presence of the problem. This may include
references to one or more substantiating documents that
demonstrate the problem. This should include the network
configuration that led to the problem such that it can be
reproduced. Problems that are not reproducible will not appear in
this memo.

Solution(s)
Solutions that permanently fix the problem, if such are known. For
example, what version of the software does not exhibit the
problem? Indicate if the solution is accepted by the community,
one of several solutions pending agreement, or open possibly with
eXPerimental solutions.

Workaround
Practical workaround if no solution is available or usable. The
workaround should have sufficient detail for someone experiencing
the problem to get around it.

References
References to related information in technical publications or on
the web. Where can someone interested in learning more go to find
out more about this problem, its solution, or workarounds?

Contact
Contact name and email address of the person who supplied the
information for this section. The editors are listed as contacts
for anonymous submissions.

2. Known Problems

The remaining sections of this document present the currently
documented known problems. The problems are ordered by
classification and significance. Issues with protocol specification
or architecture are first, followed by implementation issues. Issues
of high significance are first, followed by lower significance.

Some of the problems initially identified in the previous versions of
this document have been moved to Appendix A since they discuss issues
where resolution primarily involves education rather than protocol
work.

A full list of the problems is available in the table of contents.

2.1 Known Specification Problems

2.1.1 Vary header is underspecified and/or misleading

Name
The "Vary" header is underspecified and/or misleading

Classification
Specification

Description
The Vary header in HTTP/1.1 was designed to allow a caching proxy
to safely cache responses even if the server's choice of variants
is not entirely understood. As RFC2616 says:

The Vary header field can be used to express the parameters the
server uses to select a representation that is subject to
server-driven negotiation.

One might expect that this mechanism is useful in general for
extensions that change the response message based on some ASPects
of the request. However, that is not true.

During the design of the HTTP delta encoding specification[9] it
was realized that an HTTP/1.1 proxy that does not understand delta
encoding might cache a delta-encoded response and then later
deliver it to a non-delta-capable client, unless the extension
included some mechanism to prevent this. Initially, it was
thought that Vary would suffice, but the following scenario proves
this wrong.

NOTE: It is likely that other scenarios exhibiting the same basic
problem with "Vary" could be devised, without reference to delta
encoding. This is simply a concrete scenario used to explain the
problem.

A complete description of the IM and A-IM headers may be found in
the "Delta encoding in HTTP" specification. For the purpose of
this problem description, the relevant details are:

1. The concept of an "instance manipulation" is introduced. In
some ways, this is similar to a content-coding, but there are
differences. One example of an instance manipulation name is
"vcdiff".

2. A client signals its willingness to accept one or more
instance-manipulations using the A-IM header.

3. A server indicates which instance-manipulations are used to
encode the body of a response using the IM header.

4. Existing implementations will ignore the A-IM and IM headers,
following the usual HTTP rules for handling unknown headers.

5. Responses encoded with an instance-manipulation are sent using
the (proposed) 226 status code, "IM Used".

6. In response to a conditional request that carries an IM header,
if the request-URI has been modified then a server may transmit
a compact encoding of the modifications using a delta-encoding
instead of a status-200 response. The encoded response cannot
be understood by an implementation that does not support delta
encodings.

This summary omits many details.

Suppose client A sends this request via proxy P:

GET http://example.com/foo.Html HTTP/1.1
Host: example.com
If-None-Match: "abc"
A-IM: vcdiff

and the origin server returns, via P, this response:

HTTP/1.1 226 IM Used
Etag: "def"
Date: Wed, 19 Apr 2000 18:46:13 GMT
IM: vcdiff
Cache-Control: max-age-60
Vary: A-IM, If-None-Match

the body of which is a delta-encoded response (it encodes the
difference between the Etag "abc" instance of foo.html, and the
"def" instance). Assume that P stores this response in its cache,
and that P does not understand the vcdiff encoding.

Later, client B, also ignorant of delta-encoding, sends this
request via P:

GET http://example.com/foo.html HTTP/1.1
Host: example.com

What can P do now? According to the specification for the Vary
header in RFC2616,

The Vary field value indicates the set of request-header fields
that fully determines, while the response is fresh, whether a
cache is permitted to use the response to reply to a subsequent
request without revalidation.

Implicitly, however, the cache would be allowed to use the stored
response in response to client B WITH "revalidation". This is the
potential bug.

An obvious implementation of the proxy would send this request to
test whether its cache entry is fresh (i.e., to revalidate the
entry):

GET /foo.html HTTP/1.1
Host: example.com
If-None-Match: "def"

That is, the proxy simply forwards the new request, after doing
the usual transformation on the URL and tacking on the "obvious"
If-None-Match header.

If the origin server's Etag for the current instance is still
"def", it would naturally respond:

HTTP/1.1 304 Not Modified
Etag: "def"
Date: Wed, 19 Apr 2000 18:46:14 GMT

thus telling the proxy P that it can use its stored response. But
this cache response actually involves a delta-encoding that would
not be sensible to client B, signaled by a header field that would
be ignored by B, and so the client displays garbage.

The problem here is that the original request (from client A)
generated a response that is not sensible to client B, not merely
one that is not "the appropriate representation" (as the result of
server-driven negotiation).

One might argue that the proxy P shouldn't be storing status-226
responses in the first place. True in theory, perhaps, but
unfortunately RFC2616, section 13.4, says:

A response received with any [status code other than 200, 203,
206, 300, 301 or 410] MUST NOT be returned in a reply to a
subsequent request unless there are cache-control directives or
another header(s) that explicitly allow it. For example, these

include the following: an Expires header (section 14.21); a
"max-age", "s-maxage", "must-revalidate", "proxy-revalidate",
"public" or "private" cache-control directive (section 14.9).

In other words, the specification allows caching of responses with
yet-to-be-defined status codes if the response carries a plausible
Cache-Control directive. So unless we ban servers implementing
this kind of extension from using these Cache-Control directives
at all, the Vary header just won't work.

Significance
Medium

Implications
Certain plausible extensions to the HTTP/1.1 protocol might not
interoperate correctly with older HTTP/1.1 caches, if the
extensions depend on an interpretation of Vary that is not the
same as is used by the cache implementer.

This would have the effect either of causing hard-to-debug cache
transparency failures, or of discouraging the deployment of such
extensions, or of encouraging the implementers of such extensions
to disable caching entirely.

Indications
The problem is visible when hand-simulating plausible message
exchanges, especially when using the proposed delta encoding
extension. It probably has not been visible in practice yet.

Solution(s)

1. Section 13.4 of the HTTP/1.1 specification should probably be
changed to prohibit caching of responses with status codes that
the cache doesn't understand, whether or not they include
Expires headers and the like. (It might require some care to
define what "understands" means, leaving room for future
extensions with new status codes.) The behavior in this case
needs to be defined as equivalent to "Cache-Control: no-store"
rather than "no-cache", since the latter allows revalidation.

Possibly the specification of Vary should require that it be
treated as "Cache-Control: no-store" whenever the status code
is unknown - that should solve the problem in the scenario
given here.

2. Designers of HTTP/1.1 extensions should consider using
mechanisms other than Vary to prevent false caching.

It is not clear whether the Vary mechanism is widely
implemented in caches; if not, this favors solution #1.

Workaround
A cache could treat the presence of a Vary header in a response as
an implicit "Cache-control: no-store", except for "known" status
codes, even though this is not required by RFC2616. This would
avoid any transparency failures. "Known status codes" for basic
HTTP/1.1 caches probably include: 200, 203, 206, 300, 301, 410
(although this list should be re-evaluated in light of the problem
discussed here).

References
See [9] for the specification of the delta encoding extension, as
well as for an example of the use of a Cache-Control extension
instead of "Vary."

Contact
Jeff Mogul <mogul@pa.dec.com>

2.1.2 Client Chaining Loses Valuable Length Meta-Data

Name
Client Chaining Loses Valuable Length Meta-Data

Classification
Performance

Description
HTTP/1.1[3] implementations are prohibited from sending Content-
Length headers with any message whose body has been Transfer-
Encoded. Because 1.0 clients cannot accept chunked Transfer-
Encodings, receiving 1.1 implementations must forward the body to
1.0 clients must do so without the benefit of information that was
discarded earlier in the chain.

Significance
Low

Implications
Lacking either a chunked transfer encoding or Content-Length
indication creates negative performance implications for how the
proxy must forward the message body.

In the case of response bodies, the server may either forward the
response while closing the connection to indicate the end of the
response or must utilize store and forward semantics to buffer the
entire response in order to calculate a Content-Length. The
former option defeats the performance benefits of persistent
connections in HTTP/1.1 (and their Keep-Alive cousin in HTTP/1.0)
as well as creating some ambiguously lengthed responses. The
latter store and forward option may not even be feasible given the
size of the resource and it will always introduce increased
latency.

Request bodies must undertake the store and forward process as 1.0
request bodies must be delimited by Content-Length headers. As
with response bodies this may place unacceptable resource
constraints on the proxy and the request may not be able to be
satisfied.

Indications
The lack of HTTP/1.0 style persistent connections between 1.0
clients and 1.1 proxies, only when Accessing 1.1 servers, is a
strong indication of this problem.

Solution(s)
An HTTP specification clarification that would allow origin known
identity document Content-Lengths to be carried end to end would
alleviate this issue.

Workaround
None.

Contact
Patrick McManus <mcmanus@AppliedTheory.com>

2.2 Known Architectural Problems

2.2.1 Interception proxies break client cache directives

Name
Interception proxies break client cache directives

Classification
Architecture

Description
HTTP[3] is designed for the user agent to be aware if it is
connected to an origin server or to a proxy. User agents
believing they are transacting with an origin server but which are

really in a connection with an interception proxy may fail to send
critical cache-control information they would have otherwise
included in their request.

Significance
High

Implications
Clients may receive data that is not synchronized with the origin
even when they request an end to end refresh, because of the lack
of inclusion of either a "Cache-control: no-cache" or "must-
revalidate" header. These headers have no impact on origin server
behavior so may not be included by the browser if it believes it
is connected to that resource. Other related data implications
are possible as well. For instance, data security may be
compromised by the lack of inclusion of "private" or "no-store"
clauses of the Cache-control header under similar conditions.

Indications
Easily detected by placing fresh (un-expired) content on a caching
proxy while changing the authoritative copy, then requesting an
end-to-end reload of the data through a proxy in both interception
and explicit modes.

Solution(s)
Eliminate the need for interception proxies and IP spoofing, which
will return correct context awareness to the client.

Workaround
Include relevant Cache-Control directives in every request at the
cost of increased bandwidth and CPU requirements.

Contact
Patrick McManus <mcmanus@AppliedTheory.com>

2.2.2 Interception proxies prevent introduction of new HTTP methods

Name
Interception proxies prevent introduction of new HTTP methods

Classification
Architecture

Description
A proxy that receives a request with a method unknown to it is
required to generate an HTTP 501 Error as a response. HTTP
methods are designed to be extensible so there may be applications
deployed with initial support just for the user agent and origin

server. An interception proxy that hijacks requests which include
new methods destined for servers that have implemented those
methods creates a de-facto firewall where none may be intended.

Significance
Medium within interception proxy environments.

Implications
Renders new compliant applications useless unless modifications
are made to proxy software. Because new methods are not required
to be globally standardized it is impossible to keep up to date in
the general case.

Solution(s)
Eliminate the need for interception proxies. A client receiving a
501 in a traditional HTTP environment may either choose to repeat
the request to the origin server directly, or perhaps be
configured to use a different proxy.

Workaround
Level 5 switches (sometimes called Level 7 or application layer
switches) can be used to keep HTTP traffic with unknown methods
out of the proxy. However, these devices have heavy buffering
responsibilities, still require TCP sequence number spoofing, and
do not interact well with persistent connections.

The HTTP/1.1 specification allows a proxy to switch over to tunnel
mode when it receives a request with a method or HTTP version it
does not understand how to handle.

Contact
Patrick McManus <mcmanus@AppliedTheory.com>
Henrik Nordstrom <hno@hem.passagen.se> (HTTP/1.1 clarification)

2.2.3 Interception proxies break IP address-based authentication

Name
Interception proxies break IP address-based authentication

Classification
Architecture

Description
Some web servers are not open for public access, but restrict
themselves to accept only requests from certain IP address ranges
for security reasons. Interception proxies alter the source
(client) IP addresses to that of the proxy itself, without the

knowledge of the client/user. This breaks such authentication
mechanisms and prohibits otherwise allowed clients access to the
servers.

Significance
Medium

Implications
Creates end user confusion and frustration.

Indications
Users may start to see refused connections to servers after
interception proxies are deployed.

Solution(s)
Use user-based authentication instead of (IP) address-based
authentication.

Workaround
Using IP filters at the intercepting device (L4 switch) and bypass
all requests to such servers concerned.

Contact
Keith K. Chau <keithc@unitechnetworks.com>

2.2.4 Caching proxy peer selection in heterogeneous networks

Name
Caching proxy peer selection in heterogeneous networks

Classification
Architecture

Description
ICP[4] based caching proxy peer selection in networks with large
variance in latency and bandwidth between peers can lead to non-
optimal peer selection. For example take Proxy C with two
siblings, Sib1 and Sib2, and the following network topology
(summarized).

* Cache C's link to Sib1, 2 Mbit/sec with 300 msec latency

* Cache C's link to Sib2, 64 Kbit/sec with 10 msec latency.

ICP[4] does not work well in this context. If a user submits a
request to Proxy C for page P that results in a miss, C will send
an ICP request to Sib1 and Sib2. Assume both siblings have the

requested object P. The ICP_HIT reply will always come from Sib2
before Sib1. However, it is clear that the retrieval of large
objects will be faster from Sib1, rather than Sib2.

The problem is more complex because Sib1 and Sib2 can't have a
100% hit ratio. With a hit rate of 10%, it is more efficient to
use Sib1 with resources larger than 48K. The best choice depends
on at least the hit rate and link characteristics; maybe other
parameters as well.

Significance
Medium

Implications
By using the first peer to respond, peer selection algorithms are
not optimizing retrieval latency to end users. Furthermore they
are causing more work for the high-latency peer since it must
respond to such requests but will never be chosen to serve content
if the lower latency peer has a copy.

Indications
Inherent in design of ICP v1, ICP v2, and any cache mesh protocol
that selects peers based upon first response.

This problem is not exhibited by cache digest or other protocols
which (attempt to) maintain knowledge of peer contents and only
hit peers that are believed to have a copy of the requested page.

Solution(s)
This problem is architectural with the peer selection protocols.

Workaround
Cache mesh design when using such a protocol should be done in
such a way that there is not a high latency variance among peers.
In the example presented in the above description the high latency
high bandwidth peer could be used as a parent, but should not be
used as a sibling.

Contact
Ivan Lovric <ivan.lovric@cnet.francetelecom.fr>
John Dilley <jad@akamai.com>

2.2.5 ICP Performance

Name
ICP performance

Classification
Architecture(ICP), Performance

Description
ICP[4] exhibits O(n^2) scaling properties, where n is the number
of participating peer proxies. This can lead ICP traffic to
dominate HTTP traffic within a network.

Significance
Medium

Implications
If a proxy has many ICP peers the bandwidth demand of ICP can be
excessive. System managers must carefully regulate ICP peering.
ICP also leads proxies to become homogeneous in what they serve;
if your proxy does not have a document it is unlikely your peers
will have it either. Therefore, ICP traffic requests are largely
unable to locate a local copy of an object (see [6]).

Indications
Inherent in design of ICP v1, ICP v2.

Solution(s)
This problem is architectural - protocol redesign or replacement
is required to solve it if ICP is to continue to be used.

Workaround
Implementation workarounds exist, for example to turn off use of
ICP, to carefully regulate peering, or to use another mechanism if
available, such as cache digests. A cache digest protocol shares
a summary of cache contents using a Bloom Filter technique. This
allows a cache to estimate whether a peer has a document. Filters
are updated regularly but are not always up-to-date so cannot help
when a spike in popularity occurs. They also increase traffic but
not as much as ICP.

Proxy clustering protocols organize proxies into a mesh provide
another alternative solution. There is ongoing research on this
topic.

Contact
John Dilley <jad@akamai.com>

2.2.6 Caching proxy meshes can break HTTP serialization of content

Name
Caching proxy meshes can break HTTP serialization of content

Classification
Architecture (HTTP protocol)

Description
A caching proxy mesh where a request may travel different paths,
depending on the state of the mesh and associated caches, can
break HTTP content serialization, possibly causing the end user to
receive older content than seen on an earlier request, where the
request traversed another path in the mesh.

Significance
Medium

Implications
Can cause end user confusion. May in some situations (sibling
cache hit, object has changed state from cacheable to uncacheable)
be close to impossible to get the caches properly updated with the
new content.

Indications
Older content is unexpectedly returned from a caching proxy mesh
after some time.

Solutions(s)
Work with caching proxy vendors and researchers to find a suitable
protocol for maintaining proxy relations and object state in a
mesh.

Workaround
When designing a hierarchy/mesh, make sure that for each end-
user/URL combination there is only one single path in the mesh
during normal operation.

Contact
Henrik Nordstrom <hno@hem.passagen.se>

2.3 Known Implementation Problems

2.3.1 User agent/proxy failover

Name
User agent/proxy failover

Classification
Implementation

Description
Failover between proxies at the user agent (using a proxy.pac[8]
file) is erratic and no standard behavior is defined.
Additionally, behavior is hard-coded into the browser, so that
proxy administrators cannot use failover at the user agent
effectively.

Significance
Medium

Implications
Architects are forced to implement failover at the proxy itself,
when it may be more appropriate and economical to do it within the
user agent.

Indications
If a browser detects that its primary proxy is down, it will wait
n minutes before trying the next one it is configured to use. It
will then wait y minutes before aSKINg the user if they'd like to
try the original proxy again. This is very confusing for end
users.

Solution(s)
Work with browser vendors to establish standard extensions to
javascript proxy.pac libraries that will allow configuration of
these timeouts.

Workaround
User education; redundancy at the proxy level.

Contact
Mark Nottingham <mnot@mnot.net>

2.3.2 Some servers send bad Content-Length headers for files that
contain CR

Name
Some servers send bad Content-Length headers for files that
contain CR

Classification
Implementation

Description
Certain web servers send a Content-length value that is larger
than number of bytes in the HTTP message body. This happens when
the server strips off CR characters from text files with lines
terminated with CRLF as the file is written to the client. The
server probably uses the stat() system call to get the file size
for the Content-Length header. Servers that exhibit this behavior
include the GN Web server (version 2.14 at least).

Significance
Low. Surveys indicate only a small number of sites run faulty
servers.

Implications
In this case, an HTTP client (e.g., user agent or proxy) may
believe it received a partial response. HTTP/1.1 [3] advises that
caches MAY store partial responses.

Indications
Count the number of bytes in the message body and compare to the
Content-length value. If they differ the server exhibits this
problem.

Solutions
Upgrade or replace the buggy server.

Workaround
Some browsers and proxies use one TCP connection per object and
ignore the Content-Length. The document end of file is identified
by the close of the TCP socket.

Contact
Duane Wessels <wessels@measurement-factory.com>

3. Security Considerations

This memo does not raise security considerations in itself. See the
individual submissions for details of security concerns and issues.

References

[1] Paxson, V., Allman, M., Dawson, S., Fenner, W., Griner, J.,
Heavens, I., Lahey, K., Semke, J. and B. Volz, "Known TCP
Implementation Problems", RFC2525, March 1999.

[2] Cooper, I., Melve, I. and G. Tomlinson, "Internet Web
Replication and Caching Taxonomy", RFC3040, January 2001.

[3] Fielding, R., Gettys, J., Mogul, J., Frystyk, H., Masinter, L.,
Leach, P. and T. Berners-Lee, "Hypertext Transfer Protocol --
HTTP/1.1", RFC2616, June 1999.

[4] Wessels, D. and K. Claffy, "Internet Cache Protocol (ICP),
Version 2", RFC2186, September 1997.

[5] Davison, B., "Web Traffic Logs: An Imperfect Resource for
Evaluation", in Proceedings of the Ninth Annual Conference of
the Internet Society (INET'99), July 1999.

[6] Melve, I., "Relation Analysis, Cache Meshes", in Proceedings of
the 3rd International WWW Caching Workshop, June 1998,
<http://wwwcache.ja.net/events/workshop/29/magicnumber.html>.

[7] Krishnamurthy, B. and M. Arlett, "PRO-COW: Protocol Compliance
on the Web", AT&T Labs Technical Report #990803-05-TM, August
1999, <http://www.research.att.com/~bala/papers/procow-1.ps.gz>.

[8] Netscape, Inc., "Navigator Proxy Auto-Config File Format", March
1996,
http://home.netscape.com/eng/mozilla/2.0/relnotes/demo/proxy-
live.html

[9] Mogul, J., Krishnamurthy, B., Douglis, F., Feldmann, A., Goland,
Y., van Hoff, A. and D. Hellerstein, "HTTP Delta in HTTP", Work
in Progress.

Authors' Addresses

Ian Cooper
Equinix, Inc.
2450 Bayshore Parkway
Mountain View, CA 94043
USA

Phone: +1 650 316 6065
EMail: icooper@equinix.com

John Dilley
Akamai Technologies, Inc.
1400 Fashion Island Blvd
Suite 703
San Mateo, CA 94404
USA

Phone: +1 650 627 5244
EMail: jad@akamai.com

Appendix A. Archived Known Problems

The following sub-sections are an archive of problems identified in
the initial production of this memo. These are typically problems
requiring further work/research, or user education. They are
included here for reference purposes only.

A.1 Architectural

A.1.1 Cannot specify multiple URIs for replicated resources

Name
Cannot specify multiple URIs for replicated resources

Classification
Architecture

Description
There is no way to specify that multiple URIs may be used for a
single resource, one for each replica of the resource. Similarly,
there is no way to say that some set of proxies (each identified
by a URI) may be used to resolve a URI.

Significance
Medium

Implications
Forces users to understand the replication model and mechanism.
Makes it difficult to create a replication framework without
protocol support for replication and naming.

Indications
Inherent in HTTP/1.0, HTTP/1.1.

Solution(s)
Architectural - protocol design is necessary.

Workaround
Replication mechanisms force users to locate a replica or mirror
site for replicated content.

Contact
Daniel LaLiberte <liberte@w3.org>

A.1.2 Replica distance is unknown

Name
Replica distance is unknown

Classification
Architecture

Description
There is no recommended way to find out which of several servers
or proxies is closer either to the requesting client or to another
machine, either geographically or in the network topology.

Significance
Medium

Implications
Clients must guess which replica is closer to them when requesting
a copy of a document that may be served from multiple locations.
Users must know the set of servers that can serve a particular
object. This in general is hard to determine and maintain. Users
must understand network topology in order to choose the closest
copy. Note that the closest copy is not always the one that will
result in quickest service. A nearby but heavily loaded server
may be slower than a more distant but lightly loaded server.

Indications
Inherent in HTTP/1.0, HTTP/1.1.

Solution(s)
Architectural - protocol work is necessary. This is a specific
instance of a general problem in widely distributed systems. A
general solution is unlikely, however a specific solution in the
web context is possible.

Workaround
Servers can (many do) provide location hints in a replica
selection web page. Users choose one based upon their location.
Users can learn which replica server gives them best performance.
Note that the closest replica geographically is not necessarily
the closest in terms of network topology. Expecting users to
understand network topology is unreasonable.

Contact
Daniel LaLiberte <liberte@w3.org>

A.1.3 Proxy resource location

Name
Proxy resource location

Classification
Architecture

Description
There is no way for a client or server (including another proxy)
to inform a proxy of an alternate address (perhaps including the
proxy to use to reach that address) to use to fetch a resource.
If the client does not trust where the redirected resource came
from, it may need to validate it or validate where it came from.

Significance
Medium

Implications
Proxies have no systematic way to locate resources within other
proxies or origin servers. This makes it more difficult to share
information among proxies. Information sharing would improve
global efficiency.

Indications
Inherent in HTTP/1.0, HTTP/1.1.

Solution(s)
Architectural - protocol design is necessary.

Workaround
Certain proxies share location hints in the form of summary
digests of their contents (e.g., Squid). Certain proxy protocols
enable a proxy query another for its contents (e.g., ICP). (See
however "ICP Performance" issue (Section 2.2.5).)

Contact
Daniel LaLiberte <liberte@w3.org>

A.2 Implementation

A.2.1 Use of Cache-Control headers

Name
Use of Cache-Control headers

Classification
Implementation

Description
Many (if not most) implementations incorrectly interpret Cache-
Control response headers.

Significance
High

Implications
Cache-Control headers will be spurned by end users if there are
conflicting or non-standard implementations.

Indications
-

Solution(s)
Work with vendors and others to assure proper application

Workaround
None.

Contact
Mark Nottingham <mnot@mnot.net>

A.2.2 Lack of HTTP/1.1 compliance for caching proxies

Name
Lack of HTTP/1.1 compliance for caching proxies

Classification
Implementation

Description
Although performance benchmarking of caches is starting to be
explored, protocol compliance is just as important.

Significance
High

Implications
Caching proxy vendors implement their interpretation of the
specification; because the specification is very large, sometimes
vague and ambiguous, this can lead to inconsistent behavior
between caching proxies.

Caching proxies need to comply to the specification (or the
specification needs to change).

Indications
There is no currently known compliance test being used.

There is work underway to quantify how closely servers comply with
the current specification. A joint technical report between AT&T
and HP Labs [7] describes the compliance testing. This report
examines how well each of a set of top traffic-producing sites
support certain HTTP/1.1 features.

The Measurement Factory (formerly IRCache) is working to develop
protocol compliance testing software. Running such a conformance
test suite against caching proxy products would measure compliance
and ultimately would help assure they comply to the specification.

Solution(s)
Testing should commence and be reported in an open industry forum.
Proxy implementations should conform to the specification.

Workaround
There is no workaround for non-compliance.

Contact
Mark Nottingham <mnot@mnot.net>
Duane Wessels <wessels@measurement-factory.com>

A.2.3 ETag support

Name
ETag support

Classification
Implementation

Description
Available caching proxies appear not to support ETag (strong)
validation.

Significance
Medium

Implications
Last-Modified/If-Modified-Since validation is inappropriate for
many requirements, both because of its weakness and its use of
dates. Lack of a usable, strong coherency protocol leads
developers and end users not to trust caches.

Indications
-

Solution(s)
Work with vendors to implement ETags; work for better validation
protocols.

Workaround
Use Last-Modified/If-Modified-Since validation.

Contact
Mark Nottingham <mnot@mnot.net>

A.2.4 Servers and content should be optimized for caching

Name
Servers and content should be optimized for caching

Classification
Implementation (Performance)

Description
Many web servers and much web content could be implemented to be
more conducive to caching, reducing bandwidth demand and page load
delay.

Significance
Medium

Implications
By making poor use of caches, origin servers encourage longer load
times, greater load on caching proxies, and increased network
demand.

Indications
The problem is most apparent for pages that have low or zero
expires time, yet do not change.

Solution(s)
-

Workaround
Servers could start using unique object identifiers for write-only
content: if an object changes it gets a new name, otherwise it is
considered to be immutable and therefore have an infinite expire
age. Certain hosting providers do this already.

Contact
Peter Danzig

A.3 Administration

A.3.1 Lack of fine-grained, standardized hierarchy controls

Name
Lack of fine-grained, standardized hierarchy controls

Classification
Administration

Description
There is no standard for instructing a proxy as to how it should
resolve the parent to fetch a given object from. Implementations
therefore vary greatly, and it can be difficult to make them
interoperate correctly in a complex environment.

Significance
Medium

Implications
Complications in deployment of caches in a complex network
(especially corporate networks)

Indications
Inability of some proxies to be configured to direct traffic based
on domain name, reverse lookup IP address, raw IP address, in
normal operation and in failover mode. Inability in some proxies
to set a preferred parent / backup parent configuration.

Solution(s)
-

Workaround
Work with vendors to establish an acceptable configuration within
the limits of their product; standardize on one product.

Contact
Mark Nottingham <mnot@mnot.net>

A.3.2 Proxy/Server exhaustive log format standard for analysis

Name
Proxy/Server exhaustive log format standard for analysis

Classification
Administration

Description
Most proxy or origin server logs used for characterization or
evaluation do not provide sufficient detail to determine
cacheability of responses.

Significance
Low (for operationality; high significance for research efforts)

Implications
Characterizations and simulations are based on non-representative
workloads.

See Also
W3C Web Characterization Activity, since they are also concerned
with collecting high quality logs and building characterizations
from them.

Indications
-

Solution(s)
To properly clean and to accurately determine cacheability of
responses, a complete log is required (including all request
headers as well as all response headers such as "User-agent" [for
removal of spiders] and "Expires", "max-age", "Set-cookie", "no-
cache", etc.)

Workaround
-

References
See "Web Traffic Logs: An Imperfect Resource for Evaluation"[5]
for some discussion of this.

Contact
Brian D. Davison <davison@acm.org>
Terence Kelly <tpkelly@eecs.umich.edu>

A.3.3 Trace log timestamps

Name
Trace log timestamps

Classification
Administration

Description
Some proxies/servers log requests without sufficient timing
detail. Millisecond resolution is often too small to preserve
request ordering and either the servers should record request
reception time in addition to completion time, or elapsed time
plus either one.

Significance
Low (for operationality; medium significance for research efforts)

Implications
Characterization and simulation fidelity is improved with accurate
timing and ordering information. Since logs are generally written
in order of request completion, these logs cannot be re-played
without knowing request generation times and reordering
accordingly.

See Also
-

Indications
Timestamps can be identical for multiple entries (when only
millisecond resolution is used). Request orderings can be jumbled
when clients open additional connections for embedded objects
while still receiving the container object.

Solution(s)
Since request completion time is common (e.g., Squid), recommend
continuing to use it (with microsecond resolution if possible)
plus recording elapsed time since request reception.

Workaround
-

References
See "Web Traffic Logs: An Imperfect Resource for Evaluation"[5]
for some discussion of this.

Contact
Brian D. Davison <davison@acm.org>

A.3.4 Exchange format for log summaries

Name
Exchange format for log summaries

Classification
Administration/Analysis?

Description
Although we have (more or less) a standard log file format for
proxies (plain vanilla Common Logfile and Squid), there isn't a
commonly accepted format for summaries of those log files.
Summaries could be generated by the cache itself, or by post-
processing existing log file formats such as Squid's.

Significance
High, since it means that each log file summarizing/analysis tool
is essentially reinventing the wheel (un-necessary repetition of
code), and the cost of processing a large number of large log
files through a variety of analysis tools is (again for no good
reason) excessive.

Implications
In order to perform a meaningful analysis (e.g., to measure
performance in relation to loading/configuration over time) the
access logs from multiple busy caches, it's often necessary to run
first one tool then another, each against the entire log file (or
a significantly large subset of the log). With log files running
into hundreds of MB even after compression (for a cache dealing
with millions of transactions per day) this is a non-trivial task.

See Also
IP packet/header sniffing - it may be that individual transactions
are at a level of granularity which simply isn't sensible to be
attempting on extremely busy caches. There may also be legal
implications in some countries, e.g., if this analysis identifies
individuals.

Indications
Disks/memory full(!) Stats (using multiple programs) take too long
to run. Stats crunching must be distributed out to multiple
machines because of its high computational cost.

Solution(s)
Have the proxy produce a standardized summary of its activity
either automatically or via an external (e.g., third party) tool,
in a commonly agreed format. The format could be something like
XML or the Extended Common Logfile, but the format and contents
are subjects for discussion. Ideally this approach would permit
individual cache server products to supply subsets of the possible
summary info, since it may not be feasible for all servers to
provide all of the information which people would like to see.

Workaround
Devise a private summary format for your own personal use - but
this complicates or even precludes the exchange of summary info
with other interested parties.

References
See the web pages for the commonly used cache stats analysis
programs, e.g., Calamaris, squidtimes, squidclients, etc.

Contact
Martin Hamilton <martin@wwwcache.ja.net>

Full Copyright Statement

Copyright (C) The Internet Society (2001). All Rights Reserved.

This document and translations of it may be copied and furnished to
others, and derivative works that comment on or otherwise explain it
or assist in its implementation may be prepared, copied, published
and distributed, in whole or in part, without restriction of any
kind, provided that the above copyright notice and this paragraph are
included on all such copies and derivative works. However, this
document itself may not be modified in any way, such as by removing
the copyright notice or references to the Internet Society or other
Internet organizations, except as needed for the purpose of
developing Internet standards in which case the procedures for
copyrights defined in the Internet Standards process must be
followed, or as required to translate it into languages other than
English.

The limited permissions granted above are perpetual and will not be
revoked by the Internet Society or its successors or assigns.

This document and the information contained herein is provided on an
"AS IS" basis and THE INTERNET SOCIETY AND THE INTERNET ENGINEERING
TASK FORCE DISCLAIMS ALL WARRANTIES, EXPRESS OR IMPLIED, INCLUDING
BUT NOT LIMITED TO ANY WARRANTY THAT THE USE OF THE INFORMATION
HEREIN WILL NOT INFRINGE ANY RIGHTS OR ANY IMPLIED WARRANTIES OF
MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.

Acknowledgement

Funding for the RFCEditor function is currently provided by the
Internet Society.