Jeffrey C. Mogul
David M. Kristol
Western Research Lab
180 Park Avenue
Compaq Computer Corp.
600 Mountain Ave.
Florham Park, NJ 07932
250 University Avenue
Murray Hill, NJ 07974 USA
Palo Alto, CA 94301
The HTTP/1.1 protocol is the result of four years of discussion and debate among a broad group of Web researchers and developers. It improves upon its phenomenally successful predecessor, HTTP/1.0, in numerous ways. We discuss the differences between HTTP/1.0 and HTTP/1.1, as well as some of the rationale behind these changes.
HTTP/1.0 evolved from the original ``0.9'' version of HTTP (which is still in rare use). The process leading to HTTP/1.0 involved significant debate and experimentation, but never produced a formal specification. The HTTP Working Group (HTTP-WG) of the Internet Engineering Task Force (IETF) produced a document (RFC1945) [BLFF96] that described the ``common usage'' of HTTP/1.0, but did not attempt to create a formal standard out of the many variant implementations. Instead, over a period of roughly four years, the HTTP-WG developed an improved protocol, known as HTTP/1.1. The HTTP/1.1 specification [FGM+98] is soon to become an IETF Draft Standard. Recent versions of some popular agents (MSIE, Apache) claim HTTP/1.1 compliance in their requests or responses, and many implementations have been tested for interoperable compliance with the specification [Mas98,NG98].
The HTTP/1.1 specification states the various requirements for clients, proxies, and servers. However, additional context and rationales for the changed or new features can help developers understand the motivation behind the changes, and provide them with a richer understanding of the protocol. Additionally, these rationales can give implementors a broader feel for the pros and cons of individual features.
In this paper we describe the major changes between the HTTP/1.0 and HTTP/1.1 protocols. The HTTP/1.1 specification is almost three times as long as RFC1945, reflecting an increase in complexity, clarity, and specificity. Even so, numerous rules are implied by the HTTP/1.1 specification, rather than being explicitly stated. While some attempts have been made to document the differences between HTTP/1.0 and HTTP/1.1 ([Mar97,Yap97], Section 19.6.1 of [FGM+98]), we know of no published analysis that covers major differences and the rationale behind them, and that reflects the most recent (and probably near-final) revision of the HTTP/1.1 specification. Because the HTTP-WG, a large and international group of researchers and developers, conducted most of its discussions via its mailing list, the archive of that list [CLF98] documents the history of the HTTP/1.1 effort. But that archive contains over 8500 messages, rendering it opaque to all but the most determined protocol historian.
We structure our discussion by (somewhat arbitrarily) dividing the protocol changes into nine major areas:
The HTTP/1.1 effort assumed, from the outset, that compatibility with the installed base of HTTP implementations (including many that did not conform with [BLFF96]) was mandatory. It seemed unlikely that most software vendors or Web site operators would deploy systems that failed to interoperate with the millions of existing clients, servers, and proxies.
Because the HTTP/1.1 effort took over four years, and generated numerous interim draft documents, many implementors deployed systems using the ``HTTP/1.1'' protocol version before the final version of the specification was finished. This created another compatibility problem: the final version had to be substantially compatible with these pseudo-HTTP/1.1 versions, even if the interim drafts turned out to have errors in them.
These absolute requirements for compatibility with poorly specified prior versions led to a number of idiosyncrasies and non-uniformities in the final design. It is not possible to understand the rationale for all of the HTTP/1.1 features without recognizing this point.
The compatibility issue also underlined the need to include, in HTTP/1.1, as much support as possible for future extensibility. That is, if a future version of HTTP were to be designed, it should not be hamstrung by any additional compatibility problems.
Note that HTTP has always specified that if an implementation receives a header that it does not understand, it must ignore the header. This rule allows a multitude of extensions without any change to the protocol version, although it does not by itself support all possible extensions.
In spite of the confusion over the meaning of the ``HTTP/1.1'' protocol version token (does it imply compatibility with one of the interim drafts, or with the final standard?), in many cases the version number in an HTTP message can be used to deduce the capabilities of the sender. A companion document to the HTTP specification [MFGN97] clearly specified the ground rules for the use and interpretation of HTTP version numbers.
The version number in an HTTP message refers to the hop-by-hop sender of the message, not the end-to-end sender. Thus the message's version number is directly useful in determining hop-by-hop message-level capabilities, but not very useful in determining end-to-end capabilities. For example, if an HTTP/1.1 origin server receives a message forwarded by an HTTP/1.1 proxy, it cannot tell from that message whether the ultimate client uses HTTP/1.0 or HTTP/1.1.
For this reason, as well as to support debugging, HTTP/1.1
Via header that describes the path followed
by a forwarded message. The path information includes the
HTTP version numbers of all senders along the path
and is recorded by each successive recipient.
(Only the last of multiple consecutive HTTP/1.0 senders will
be listed, because HTTP/1.0 proxies will not add information
HTTP/1.1 introduces the
a way for a client to learn about
the capabilities of a server without actually requesting a resource.
For example, a proxy can verify that the server complies with a
specific version of the protocol. Unfortunately, the precise
semantics of the
OPTIONS method were the subject of an intense and
unresolved debate, and we believe that the mechanism is not
yet fully specified.
In order to ease the deployment of incompatible future protocols,
HTTP/1.1 includes the new
Upgrade request-header. By sending the
Upgrade header, a client can inform a server of the set of protocols
it supports as an alternate means of communication. The server may
choose to switch protocols, but this is not mandatory.
Web developers recognized early on that the caching of responses was both possible and highly desirable. Caching is effective because a few resources are requested often by many users, or repeatedly by a given user. Caches are employed in most Web browsers and in many proxy servers; occasionally they are also employed in conjunction with certain origin servers. Web caching products, such as Cisco's cache engine [Cis] and Inktomi's Traffic Server [Ink] (to name two), are now a major business.
Many researchers have studied the effectiveness of HTTP caching [KLM97,DFKM97,ASA+95,IC97]. Caching improves user-perceived latency by eliminating the network communication with the origin server. Caching also reduces bandwidth consumption, by avoiding the transmission of unnecessary network packets. Reduced bandwidth consumption also indirectly reduces latency for uncached interactions, by reducing network congestion. Finally, caching can reduce the load on origin servers (and on intermediate proxies), further improving latency for uncached interactions.
One risk with caching is that the caching mechanism might not be ``semantically transparent'': that is, it might return a response different from what would be returned by direct communication with the origin server. While some applications can tolerate non-transparent responses, many Web applications (electronic commerce, for example) cannot.
HTTP/1.0 provided a simple caching mechanism. An origin
server may mark a response, using the
with a time until which a cache could return the response without
violating semantic transparency.
Further, a cache may check the current validity of a response
using what is known as a conditional request: it may
If-Modified-Since header in a request for
the resource, specifying the value given in the cached response's
Last-Modified header. The server may then either
respond with a 304 (Not Modified) status code, implying that
the cache entry is valid, or it may send a normal 200 (OK) response
to replace the cache entry.
HTTP/1.0 also included a mechanism, the
header, for the client to indicate that a request should not
be satisfied from a cache.
The HTTP/1.0 caching mechanism worked moderately well, but it had many conceptual shortcomings. It did not allow either origin servers or clients to give full and explicit instructions to caches; therefore, it depended on a body of heuristics that were not well-specified. This led to two problems: incorrect caching of some responses that should not have been cached, and failure to cache some responses that could have been cached. The former causes semantic problems; the latter causes performance problems.
HTTP/1.1 attempts to clarify the concepts behind caching, and to provide explicit and extensible protocol mechanisms for caching. While it retains the basic HTTP/1.0 design, it augments that design both with new features, and with more careful specifications of the existing features.
In HTTP/1.1 terminology, a cache entry is fresh until it reaches its expiration time, at which point it becomes stale. A cache need not discard a stale entry, but it normally must revalidate it with the origin server before returning it in response to a subsequent request. However, the protocol allows both origin servers and end-user clients to override this basic rule.
In HTTP/1.0, a cache revalidated an entry using the
header. This header uses absolute timestamps with one-second resolution,
which could lead to caching errors either because of clock
synchronization errors, or because of lack of resolution. Therefore,
HTTP/1.1 introduces the more general concept of an opaque cache
validator string, known as an entity tag. If two responses
for the same resource have the same entity tag, then they must
(by specification) be identical.
entity tag is opaque, the origin server may use any information
it deems necessary to construct it (such as a fine-grained
timestamp or an internal database pointer), as long as it
meets the uniqueness requirement. Clients may compare entity
tags for equality, but cannot otherwise manipulate them.
HTTP/1.1 servers attach entity tags to responses using the
HTTP/1.1 includes a number of new conditional request-headers, in
If-Modified-Since. The most basic is
If-None-Match, which allows a client to present one or more
entity tags from its cache entries for a resource. If none of these
matches the resource's current entity tag value, the server returns a
normal response; otherwise, it may return a 304 (Not Modified)
response with an
ETag header that indicates which cache entry
is currently valid. Note that this mechanism allows the server to
cycle through a set of possible responses, while the
If-Modified-Since mechanism only generates a cache hit if the
most recent response is valid.
HTTP/1.1 also adds new conditional headers called
other forms of preconditions on requests. These
preconditions are useful in more complex situations;
in particular, see the discussion in Section 4.1 of Range
In order to make caching requirements more explicit, HTTP/1.1
adds the new
Cache-Control header, allowing an extensible
set of cache-control directives to be transmitted in both
requests and responses. The set defined by HTTP/1.1 is
quite large, so we concentrate on several notable
Because the absolute timestamps in the HTTP/1.0
can lead to failures in the presence of clock skew (and observations
suggest that serious clock skew is common), HTTP/1.1
can use relative expiration times, via the
(It also introduces an
Age header, so that caches can
indicate how long a response has been sitting in caches along
Because some users have privacy requirements that limit
caching beyond the need for semantic transparency, the
no-store directives allow servers
and clients to prevent the storage of some or all of
However, this does not guarantee privacy; only cryptographic
mechanisms can provide true privacy.
Some proxies transform responses (for example, to reduce image
complexity before transmission over a slow link [FGBA96]),
but because some responses cannot be blindly transformed without
losing information, the
no-transform directive may be used to
To support caching of negotiated responses, and for future
extensibility, HTTP/1.1 includes the
This header field carries a list of the relevant
fields that participated in the selection of the response variant.
In order to use the particular variant of the cached response in
replying to a subsequent
request, the selecting request-headers of the new request must
exactly match those of the original request.
This simple and elegant extension mechanism works for many cases of negotiation, but it does not allow for much intelligence at the cache. For example, a smart cache could, in principle, realize that one request header value is compatible with another, without being equal. The HTTP/1.1 development effort included an attempt to provide so-called ``transparent content negotiation'' that would allow caches some active participation, but ultimately no consensus developed, and this attempt [HM98b,HM98a] was separated from the HTTP/1.1 specification.
Network bandwidth is almost always limited. Both by intrinsically delaying the transmission of data, and through the added queueing delay caused by congestion, wasting bandwidth increases latency. HTTP/1.0 wastes bandwidth in several ways that HTTP/1.1 addresses. A typical example is a server's sending an entire (large) resource when the client only needs a small part of it. There was no way in HTTP/1.0 to request partial objects. Also, it is possible for bandwidth to be wasted in the forward direction: if a HTTP/1.0 server could not accept large requests, it would return an error code after bandwidth had already been consumed. What was missing was the ability to negotiate with a server and to ensure its ability to handle such requests before sending them.
Rangeheader in its request, specifying one or more contiguous ranges of bytes. The server can either ignore the
Rangeheader, or it can return one or more ranges in the response.
If a response contains a range, rather than the entire
resource, it carries the 206 (Partial Content) status code.
This code prevents HTTP/1.0 proxy caches from accidentally treating
the response as a full one, and then using it as a cached
response to a subsequent request.
In a range response,
Content-Range header indicates the offset and
length of the returned range, and the new
MIME type allows the transmission of multiple ranges in
Range requests can be used in a variety of ways, such as:
For example, the first kind (getting a prefix of the resource)
might be done unconditionally, or it might be done with an
If-None-Match header; the latter implies that the client only
wants the range if the underlying object has changed, and otherwise
will use its cache entry.
The second kind of request, on the other hand, is made when
the client does not have a cache entry that includes the
desired range. Therefore, the client wants the range only
if the underlying object has not changed; otherwise,
it wants the full response. This could be accomplished by
first sending a range request with an
If-Match header, and
then repeating the request without either header if the
first request fails. However, since this is an important
optimization, HTTP/1.1 includes an
which effectively performs that sequence in a single
Range requests were originally proposed by Ari Luotonen and John Franks [FL95], using an extension to the URL syntax instead of a separate header field. However, this approach proved less general than the approach ultimated used in HTTP/1.1, especially with respect to conditional requests.
Some HTTP requests (for example, the
carry request bodies, which may be arbitrarily long. If,
the server is not willing to accept the request,
perhaps because of an authentication failure,
it would be a waste of bandwidth
to transmit such a large request body.
HTTP/1.1 includes a new status code, 100 (Continue), to inform the client that the request body should be transmitted. When this mechanism is used, the client first sends its request headers, then waits for a response. If the response is an error code, such as 401 (Unauthorized), indicating that the server does not need to read the request body, the request is terminated. If the response is 100 (Continue), the client can then send the request body, knowing that the server will accept it.
However, HTTP/1.0 clients do not understand the 100 (Continue)
response. Therefore, in order to trigger the use of this
mechanism, the client sends the new
Expect header, with
a value of
could be used for other, future purposes not defined in HTTP/1.1.)
Because not all servers use this mechanism
Expect header is a relatively late addition to HTTP/1.1,
and early ``HTTP/1.1'' servers did not implement it), the client must
not wait indefinitely for a 100 (Continue) response before
sending its request body. HTTP/1.1 specifies a number of
somewhat complex rules to avoid either infinite waits or
wasted bandwidth. We lack sufficient experience based on
deployed implementations to know if this design will work
One well-known way to conserve bandwidth is through the use of data compression. While most image formats (GIF, JPEG, MPEG) are precompressed, many other data types used in the Web are not. One study showed that aggressive use of additional compression could save almost 40% of the bytes sent via HTTP [MDFK97]. While HTTP/1.0 included some support for compression, it did not provide adequate mechanisms for negotiating the use of compression, or for distinguishing between end-to-end and hop-by-hop compression.
HTTP/1.1 makes a distinction between content-codings, which are end-to-end encodings that might be inherent in the native format of a resource, and transfer-codings, which are always hop-by-hop. Compression can be done either as a content-coding or as a transfer-coding. To support this choice, and the choice between various existing and future compression codings, without breaking compatibility with the installed base, HTTP/1.1 had to carefully revise and extend the mechanisms for negotiating the use of codings.
HTTP/1.0 includes the
which indicates the end-to-end content-coding(s) used for a message;
HTTP/1.1 adds the
which indicates the hop-by-hop transfer-coding(s) used for a message.
HTTP/1.1 (unlike HTTP/1.0)
carefully specifies the
header, used by a client to indicate what content-codings
it can handle, and which ones it prefers. One tricky
issue is the need to support ``robot'' clients that
are attempting to create mirrors of the origin server's
resources; another problem is the need to interoperate
with HTTP/1.0 implementations, for which
was poorly specified.
HTTP/1.1 also includes the
TE header, which
allows the client to indicate which transfer-codings
are acceptable, and which are preferred. Note that
one important transfer-coding,
Chunked, has a
special function (not related to compression), and is
discussed further in Section 6.1.
HTTP almost always uses TCP as its transport protocol. TCP works best for long-lived connections, but the original HTTP design used a new TCP connection for each request, so each request incurred the cost of setting up a new TCP connection (at least one round-trip time across the network, plus several overhead packets). Since most Web interactions are short (the median response message size is about 4 Kbytes [MDFK97]), the TCP connections seldom get past the ``slow-start'' region [Jac88] and therefore fail to maximize their use of the available bandwidth.
Web pages frequently have embedded images, sometimes many of them, and each image is retrieved via a separate HTTP request. The use of a new TCP connection for each image retrieval serializes the display of the entire page on the connection-setup latencies for all of the requests. Netscape introduced the use of parallel TCP connections to compensate for this serialization, but the possibility of increased congestion limits the utility of this approach.
To resolve these problems, Padmanabhan and Mogul [PM95] recommended the use of persistent connections and the pipelining of requests on a persistent connection.
Before discussing persistent connections, we address a more basic issue. Given the use of intermediate proxies, HTTP makes a distinction between the end-to-end path taken by a message, and the actual hop-by-hop connection between two HTTP implementations.
HTTP/1.1 introduces the concept of hop-by-hop headers:
message headers that apply only to a given connection,
and not to the entire path. (For example, we have already described
The use of hop-by-hop headers creates a potential problem:
if such a header were to be forwarded by a naive proxy, it
might mislead the recipient.
Therefore, HTTP/1.1 includes the
This header lists all of the hop-by-hop headers in a message,
telling the recipient that these headers must be removed
from that message before it is forwarded. This extensible
the future introduction of new hop-by-hop headers; the
sender need not know whether the recipient understands
a new header in order to prevent the recipient from
forwarding the header.
Because HTTP/1.0 proxies do not understand the
header, however, HTTP/1.1 imposes an additional rule. If
Connection header is received in an HTTP/1.0 message,
then it must have been incorrectly forwarded by an
HTTP/1.0 proxy. Therefore, all of the headers it lists
were also incorrectly forwarded, and must be ignored.
Connection header may also list connection-tokens,
which are not headers but rather per-connection boolean flags. For
example, HTTP/1.1 defines the token
close to permit the peer
to indicate that it does not want to use a persistent
connection. Again, the
Connection header mechanism prevents
these tokens from being forwarded.
Keep-Aliveheader (described in [Fie95]) to request that a connection persist. This design did not interoperate with intermediate proxies (see Section 19.6.2 of [FGM+98]); HTTP/1.1 specifies a more general solution.
In recognition of their desirable properties, HTTP/1.1 makes persistent connections the default. HTTP/1.1 clients, servers, and proxies assume that a connection will be kept open after the transmission of a request and its response. The protocol does allow an implementation to close a connection at any time, in order to manage its resources, although it is best to do so only after the end of a response.
Because an implementation may prefer not to use persistent
connections if it cannot efficiently scale to large
numbers of connections or may want to cleanly terminate one for
resource-management reasons, the protocol permits it to
Connection: close header to inform
the recipient that the connection will not be reused.
Although HTTP/1.1 encourages the transmission of multiple requests over a single TCP connection, each request must still be sent in one contiguous message, and a server must send responses (on a given connection) in the order that it received the corresponding requests. However, a client need not wait to receive the response for one request before sending another request on the same connection. In fact, a client could send an arbitrarily large number of requests over a TCP connection before receiving any of the responses. This practice, known as pipelining, can greatly improve performance [NGBS+97]. It avoids the need to wait for network round-trips, and it makes the best possible use of the TCP protocol.
HTTP messages may carry a body of arbitrary length.
The recipient of a message needs to know where the message
The sender can use the
which gives the length of the body.
However, many responses are generated dynamically, by
CGI [CGI98] processes and similar mechanisms.
Without buffering the entire response (which would add latency),
the server cannot know how long it will be and cannot
When not using persistent connections, the solution is simple: the server closes the connection. This option is available in HTTP/1.1, but it defeats the performance advantages of persistent connections.
Chunkedtransfer-coding. The sender breaks the message body into chunks of arbitrary length, and each chunk is sent with its length prepended; it marks the end of the message with a zero-length chunk. The sender uses the
Transfer-Encoding: chunkedheader to signal the use of chunking.
This mechanism allows the sender to buffer small pieces of the message, instead of the entire message, without adding much complexity or overhead. All HTTP/1.1 implementations must be able to receive chunked messages.
Chunked transfer-coding solves another problem,
not related to performance.
In HTTP/1.0, if the sender does not include a
header, the recipient cannot tell if the message has been
truncated due to transmission problems. This ambiguity leads
to errors, especially when truncated responses are stored in caches.
Chunking solves another problem related to sender-side
message buffering. Some header fields, such as
(a cryptographic checksum over the message body), cannot be
computed until after the message body is generated. In
HTTP/1.0, the use of such header fields required the sender to buffer
the entire message.
In HTTP/1.1, a chunked message may include a trailer after the final chunk. A trailer is simply a set of one or more header fields. By placing them at the end of the message, the sender allows itself to compute them after generating the message body.
The sender alerts the recipient to the presence of message
trailers by including a
Trailer header, which lists
the set of headers deferred until the trailer. This alert,
for example, allows a browser to avoid displaying a prefix
of the response before it has received authentication
information carried in a trailer.
HTTP/1.1 imposes certain conditions on the use of trailers,
to prevent certain kinds of interoperability failure. For
example, if a server sends a lengthy message with a trailer
to an HTTP/1.1 proxy that is forwarding the response to an
HTTP/1.0 client, the proxy must either buffer the entire message
or drop the trailer. Rather than insist that proxies buffer
arbitrarily long messages, which would be infeasible, the
protocol sets rules that should prevent any critical information
in the trailer (such as authentication information) from being
lost because of this problem.
Specifically, a server cannot send a trailer unless either the
information it contains is purely optional, or the client
has sent a
TE: trailers header, indicating that it
is willing to receive trailers (and, implicitly, to buffer
the entire response if it is forwarding the message to
an HTTP/1.0 client).
Several HTTP/1.1 mechanisms, such as Digest Access
Authentication (see Section 9.1), require end-to-end
agreement on the length of the message body; this
is known as the entity-length. Hop-by-hop
transfer-codings, such as compression or chunking, can
temporarily change the transfer-length of a message.
Before this distinction was clarified,
some earlier implementations used the
Content-Length header indiscriminately.
Therefore, HTTP/1.1 gives a lengthy set of rules for
indicating and inferring the entity-length of a message.
For example, if a non-identity transfer-coding is used
(so the transfer-length and entity-length differ),
the sender is not allowed to use the
When a response contains multiple byte ranges, using
Content-Type: multipart/byteranges, then this
self-delimiting format defines the transfer-length.
Companies and organizations use URLs to advertise themselves and their products and services. When a URL appears in a medium other than the Web itself, people seem to prefer ``pure hostname'' URLs; i.e., URLs without any path syntax following the hostname. These are often known as ``vanity URLs,'' but in spite of the implied disparagement, it's unlikely that non-purist users will abandon this practice, which has led to the continuing creation of huge numbers of hostnames.
IP addresses are widely perceived as a scarce resource
(pending the uncertain transition to IPv6 [DH95]).
The Domain Name System (DNS) allows multiple host names
to be bound to the same IP address.
Unfortunately, because the original designers of HTTP
did not anticipate the ``success disaster'' they were
enabling, HTTP/1.0 requests do not pass the hostname
part of the request URL. For example, if a user makes a request
for the resource at URL
the browser sends a message with the Request-Line
GET /home.html HTTP/1.0to the server at
example1.org. This prevents the binding of another HTTP server hostname, such as
exampleB.orgto the same IP address, because the server receiving such a message cannot tell which server the message is meant for. Thus, the proliferation of vanity URLs causes a proliferation of IP address allocations.
The Internet Engineering Steering Group (IESG),
which manages the IETF process, insisted that HTTP/1.1
take steps to improve conservation of IP addresses. Since
HTTP/1.1 had to interoperate with HTTP/1.0, it could not
change the format of the Request-Line to include the
server hostname. Instead, HTTP/1.1 requires requests
to include a
Host header, first proposed by John Franks [Fra94],
that carries the hostname. This converts the example above to:
GET /home.html HTTP/1.1 Host: example1.orgIf the URL references a port other than the default (TCP port 80), this is also given in the
Clearly, since HTTP/1.0 clients will not send
HTTP/1.1 servers cannot simply reject all messages without
them. However, the HTTP/1.1 specification requires that
an HTTP/1.1 server must reject any HTTP/1.1 message that does
not contain a
The intent of the
Host header mechanism, and in particular
the requirement that enforces its presence in HTTP/1.1 requests,
is to speed the transition away from assigning a new IP
address for every vanity URL. However, as long as
a substantial fraction of the users on the Internet
use browsers that do not send
no Web site operator (such as an electronic
commerce business) that depends on these users will give up
a vanity-URL IP address. The transition, therefore, may
take many years. It may be obviated by an earlier transition
to IPv6, or by the use of market mechanisms to discourage
the unnecessary consumption of IPv4 addresses.
HTTP/1.0 defined a relatively small set of sixteen status codes, including the normal 200 (OK) code. Experience revealed the need for finer resolution in error reporting.
HTTP status codes indicate the success or failure of a request. For a successful response, the status code cannot provide additional advisory information, in part because the placement of the status code in the Status-Line, instead of in a header field, prevents the use of multiple status codes.
HTTP/1.1 introduces a
Warning header, which may carry
any number of subsidiary status indications. The intent
is to allow a sender to advise the recipient that something
may be unsatisfactory about an ostensibly successful response.
HTTP/1.1 defines an initial set of
Warning codes, mostly
related to the actions of caches along the response path.
For example, a
Warning can mark a response as having been
returned by a cache during disconnected operation, when
it is not possible to validate the cache entry with the origin
Warning codes are divided into two classes, based on
the first digit of the 3-digit code. One class of warnings
must be deleted after a successful revalidation of a cache
entry; the other class must be retained with a revalidated
cache entry. Because this distinction is made based on the
first digit of the code, rather than through an exhaustive
listing of the codes, it is extensible to
defined in the future.
There are 24 new status codes in HTTP/1.1; we have discussed 100 (Continue), 206 (Partial Content), and 300 (Multiple Choices) elsewhere in this paper. A few of the more notable additions include
In recent years, the IETF has heightened its sensitivity to issues of privacy and security. One special concern has been the elimination of passwords transmitted ``in the clear.'' This increased emphasis has manifested itself in the HTTP/1.1 specification (and other closely related specifications).
WWW-Authenticateheader that identifies the authentication scheme (in this case, ``Basic'') and realm. (The realm value allows a server to partition sets of resources into ``protection spaces,'' each with its own authorization database.)
The client (user agent) typically queries the user for a username
and password for the realm, then repeats the original request,
this time including an
Authorization header that contains
the username and password.
Assuming these credentials are acceptable to it,
the origin server responds by sending the expected content.
A client may continue to send the same credentials for other
resources in the same realm on the same server, thus eliminating
the extra overhead of the challenge and response.
A serious flaw in Basic authentication is that the username and password in the credentials are unencrypted and therefore vulnerable to network snooping. The credentials also have no time dependency, so they could be collected at leisure and used long after they were collected. Digest access authentication [FHBH+97,FHBH+98] provides a simple mechanism that uses the same framework as Basic authentication while eliminating many of its flaws. (Digest access authentication, being largely separable from the HTTP/1.1 specification, has developed in parallel with it.)
The message flow in Digest access authentication mirrors that of Basic and uses the same headers, but with a scheme of ``Digest.'' The server's challenge in Digest access authentication uses a nonce (one-time) value, among other information. To respond successfully, a client must compute a checksum (MD5, by default) of the username, password, nonce, HTTP method of the request, and the requested URI. Not only is the password no longer unencrypted, but the given response is correct only for a single resource and method. Thus, an attacker that can snoop on the network could only replay the request, the response for which he has already seen. Unlike with Basic authentication, obtaining these credentials does not provide access to other resources.
As with Basic authentication, the client may make further requests to the same realm and include Digest credentials, computed with the appropriate request method and request-URI. However, the origin server's nonce value may be time-dependent. The server can reject the credentials by saying the response used a stale nonce and by providing a new one. The client can then recompute its credentials without needing to ask the user for username and password again.
In addition to the straightforward authentication capability,
Digest access authentication offers two other features:
support for third-party authentication servers, and
a limited message integrity feature (through the
Some proxy servers provide service only to properly authenticated clients. This prevents, for example, other clients from stealing bandwidth from an unsuspecting proxy.
To support proxy authentication, HTTP/1.1 introduces the
headers. They play the same role as the
headers in HTTP/1.0, except that the new headers are
hop-by-hop, rather than end-to-end. Proxy authentication
may use either of the Digest or Basic authentication
schemes, but the former is preferred.
A proxy server sends the client a
Proxy-Authenticate header, containing a
challenge, in a 407 (Proxy Authentication Required)
response. The client then repeats the initial request, but
Proxy-Authorization header that contains
credentials appropriate to the challenge. After
successful proxy authentication, a client typically sends
Proxy-Authorization header to the proxy
with each subsequent request, rather than wait to be
The URI of a resource often represents information that some users may view as private. Users may prefer not to have it widely known that they have visited certain sites.
Referer [sic] header in a request
provides the server with the URI of the resource from
which the request-URI was obtained. This gives the
server information about the user's previous
page-view. To protect against unexpected privacy
violations, the HTTP/1.1 specification takes pains to
discourage sending the
inappropriately; for example, when a user enters a URL
from the keyboard, the application should not send a
Referer header describing the currently-visible
page, nor should a client send the
header in an insecure request if the referring page had
been transferred securely.
The use of a
GET-based HTML form causes the
encoding of form parameters in the request-URI. Most
proxy servers log these request-URIs. To protect
against revealing sensitive information, such as
passwords or credit-card numbers, in a URI, the
HTTP/1.1 specification strongly discourages the use of
GET-based forms for submitting such data. The
POST-based forms prevents the form
parameters from appearing in a request-URI, and
therefore from being logged inappropriately.
Content-MD5 header contains the MD5 digest of the
entity being sent[MR95]. The HTTP/1.1 specification
provides specific rules for the use of this header in the Web,
which differ slightly from its use in MIME (electronic mail). The
sender may choose to send
Content-MD5 so the recipient can detect
accidental changes to the entity during its transmission.
Content-MD5 is a good example of a header that a server might
usefully send as a trailer.
Content-MD5 value could easily be spoofed and
cannot serve as a means of security. Also, because
covers just the entity in one message, it cannot be used to
determine if a full response has been successfully reassembled from
a number of partial (range) responses, or whether response
headers have been altered.
HTTP requests are stateless. That is, from a server's perspective, each request can ordinarily be treated as independent of any other. For Web applications, however, state can sometimes be useful. For example, a shopping application would like to keep track of what is in a user's ``shopping basket,'' as the basket's contents change over the course of a sequence of HTTP requests.
Netscape introduced ``cookies'' [Net] in version 1.1 of their browser as a state management mechanism. The IETF subsequently standardized cookies in RFC2109 [KM97]. (The cookie specification is another example of how HTTP can be extended by a separate specification without affecting the core protocol. Cookie support is optional in servers and user agents, although some Web-based services will not work in their absence.)
The basic cookie mechanism is simple. An origin server sends an arbitrary piece of (state) information to the client in its response. The client is responsible for saving the information and returning it with its next request to the origin server. RFC2109 and Netscape's original specification relax this model so that a cookie can be returned to any of a collection of related servers, rather than just to one. The specifications also restricts for which URIs on a given server the cookie may be returned. A servers may assign a lifetime to a cookie, after which it is no longer used.
Cookies have both privacy and security implications. Because their content is arbitrary, cookies may contain sensitive application-dependent information. For example, they could contain credit card numbers, user names and passwords, or other personal information. Applications that send such information over unencrypted connections leave it vulnerable to snooping, and cookies stored at a client system might reveal sensitive information to another user of (or intruder into) that client.
RFC2109 proved to be controversial, primarily because of restrictions that were introduced to protect privacy. Probably the most controversial of these has to do with ``unverifiable transactions'' and ``third-party cookies.'' Consider this scenario.
IMG(image) tag with a reference to
http://ad.example.com/adv1.gif, an advertisement.
IMGtag with a reference to
ad.example.comin the process. The response includes a new cookie from
Privacy advocates, and others, worried that:
ad.example.com, a site she didn't even know she was going to visit (an ``unverifiable transaction''); and
ad.example.comin the second image request (step 6).
Refererheader is sent with each of the image requests to
ad.example.com, then that site can begin to accumulate a profile of the user's interests from the sites she visited, here
http://www.exampleB.com/home.html. Such an advertising site could potentially select advertisements that are likely to be interesting to her. While that profiling process is relatively benign in isolation, it could become more personal if the profile can also be tied to a specific real person, not just a persona. For example, this might happen if the user goes through some kind of registration at
RFC2109 sought to limit the possible pernicious effects of cookies
by requiring user agents to reject cookies that arrive from the
responses to unverifiable transactions.
RFC2109 further stated that user agents could be configured to
accept such cookies, provided that the default was not to accept them.
This default setting was a source of concern for advertising networks
(companies that run sites like
ad.example.com in the
example) whose business model depended on cookies, and whose business
blossomed in the interval between when the specification was essentially
complete (July, 1996) and the time it appeared as an RFC (February, 1997).
RFC2109 has undergone further refinement [KM98]
in response to comments, both political and technical.
Content negotiation has proved to be a contentious and confusing area. Some aspects that appeared simple at first turned out to be quite difficult to resolve. For example, although current IETF practice is to insist on explicit character set labeling in all relevant contexts, the existing HTTP practice has been to use a default character set in most contexts, but not all implementations chose the same default. The use of unlabeled defaults greatly complicates the problem of internationalizing the Web.
HTTP/1.0 provided a few features to support content negotiation, but RFC1945 [BLFF96] never uses that term and devotes less than a page to the relevant protocol features. The HTTP/1.1 specification specifies these features with far greater care, and introduces a number of new concepts.
The goal of the content negotiation mechanism is to choose the best available representation of a resource. HTTP/1.1 provides two orthogonal forms of content negotiation, differing in where the choice is made:
Accept-Charset, etc. The server then chooses the representation that best matches the preferences expressed in these headers.
Alternates[HM98b] header name for use in agent-driven negotiation, the HTTP working group never completed a specification of this header, and server-driven negotiation remains the only usable form.
Some users may speak multiple languages, but with varying degrees of fluency. Similarly, a Web resource might be available in its original language, and in several translations of varying faithfulness. HTTP introduces the use of quality values to express the importance or degree of acceptability of various negotiable parameters. A quality value (or qvalue) is a fixed-point number between 0.0 and 1.0. For example, a native speaker of English with some fluency in French, and who can impose on a Danish-speaking office-mate, might configure a browser to generate requests including
Accept-Language: en, fr;q=0.5, da;q=0.1
Because the content-negotiation mechanism allows qvalues and wildcards, and expresses variation across many dimensions (language, character-set, content-type, and content-encoding) the automated choice of the ``best available'' variant can be complex and might generate unexpected outcomes. These choices can interact with caching in subtle ways; see the discussion in Section 3.4.
Content negotiation promises to be a fertile area for additional protocol evolution. For example, the HTTP working group recognized the utility of automatic negotiation regarding client implementation features, such as screen size, resolution, and color depth. The IETF has created the Content Negotiation working group to carry forward with work in the area.
Although the HTTP working group will disband after the publication of the HTTP/1.1 specification, there are numerous pending proposals for further improvements. Here we sketch a few of the more significant ones.
The HTTP/1.1 specification placed a strong emphasis on extensibility but was not able to resolve this issue entirely. Although it has not yet achieved the status of a working group, one effort has been trying to define a general extensibility mechanism [NLL98].
As we noted in Section 10, the Content Negotiation working group is working on proposals to better define content negotiation [HM98b,HM98a] and feature tags.
Several researchers have observed that when Web resources change (thus invalidating cache entries), they usually do not change very much [DFKM97,MDFK97]. This suggests that transmitting only the differences (or delta) between the current resource value and a cached response, rather than an entire new response, could save bandwidth and time. Two of us, in conjunction with several other people, have proposed a simple extension to HTTP/1.1 to support delta encoding [MKD+99].
In today's Web, content is widely shared, but it mostly flows in one direction, from servers to clients. The Web could become a medium for distributed updates to shared content. The IETF's World Wide Web Distributed Authoring and Versioning (WEBDAV) working group is in the process of defining HTTP extensions to enable this vision [SVWD98].
HTTP/1.1 differs from HTTP/1.0 in numerous ways, both large and small. While many of these changes are clearly for the better, the protocol description has tripled in length, and many of the new features were introduced without any real experimental evaluation to back them up. The HTTP/1.1 specification also includes numerous irregularities for compatibility with the installed base of HTTP/1.0 implementations.
This increase in complexity complicates the job of client, server, and especially proxy cache implementors. It has already led to unexpected interactions between features, and will probably lead to others. We do not expect the adoption of HTTP/1.1 to go entirely without glitches. Fortunately, the numerous provisions in HTTP/1.1 for extensibility should simplify the introduction of future modifications.
We would like to thank the anonymous referees for their reviews.
http://portal.research.bell-labs.com/~dmk/cookie-3.6.txt; this is a work in progress.
ftp://ftp.ietf.org/internet-drafts/draft-mogul-http-delta-01.txt; this is a work in progress.
Balachander Krishnamurthy is a researcher at AT&T Labs-Research in Florham Park, NJ, and can be reached at email@example.com.
Jeffrey C. Mogul received an S.B. from the Massachusetts Institute of Technology in 1979, an M.S. from Stanford University in 1980, and his PhD from the Stanford University Computer Science Department in 1986. Dr. Mogul has been an active participant in the Internet community, and is the author or co-author of several Internet Standards; most recently, he has contributed extensively to the HTTP/1.1 specification. Since 1986, he has been a researcher at the Compaq (formerly Digital) Western Research Laboratory, working on network and operating systems issues for high-performance computer systems, and on improving performance of the Internet and the World Wide Web. He is a member of ACM, Sigma Xi, and CPSR, and was Programm Committee Chair for the Winter 1994 USENIX Technical Conference, and for the IEEE TCOS Sixth Workshop on Hot Topics in Operating Systems.
Address for correspondence: Compaq Computer Corporation Western Research Laboratory, 250 University Avenue, Palo Alto, California, 94301 (firstname.lastname@example.org)
David M. Kristol is currently a Member of Technical Staff at Bell Laboratories, Lucent Technologies, in the Information Sciences Research Center, where he now does research in security and electronic commerce on the Internet. Previously, he worked on formal specifications for communications protocols. For six years he was responsible for the C compilers for UNIX(r) System V as a member of Bell Labs's Unix support organization, and he was the principal developer of the System V ANSI C compiler. He joined Bell Laboratories in 1981. Earlier, Kristol worked at Massachusetts Computer Associates, where he developed a mass spectrometry data system, and at GenRad, Inc., where he developed automatic test equipment systems. He received BA and BSEE degrees from the University of Pennsylvania, Philadelphia, and MS and ME degrees (in Applied Mathematics) from Harvard University.
His email address is: email@example.com.