Fastly

kazuhooku@gmail.com

mnot@mnot.net https://www.mnot.net/

Applications and Real-Time HTTP Internet-Draft This specification defines a HTTP/2 frame type to allow clients to inform the server of their cache’s contents. Servers can then use this to inform their choices of what to push to clients. Discussion of this draft takes place on the HTTP working group mailing list (ietf-http-wg@w3.org), which is archived at https://lists.w3.org/Archives/Public/ietf-http-wg/. Working Group information can be found at http://httpwg.github.io/; source code and issues list for this draft can be found at https://github.com/httpwg/http-extensions/labels/cache-digest.

HTTP/2 allows a server to “push” synthetic request/response pairs into a client’s cache optimistically. While there is strong interest in using this facility to improve perceived Web browsing performance, it is sometimes counterproductive because the client might already have cached the “pushed” response. When this is the case, the bandwidth used to “push” the response is effectively wasted, and represents opportunity cost, because it could be used by other, more relevant responses. HTTP/2 allows a stream to be cancelled by a client using a RST_STREAM frame in this situation, but there is still at least one round trip of potentially wasted capacity even then. This specification defines a HTTP/2 frame type to allow clients to inform the server of their cache’s contents using a Cuckoo-filter based digest. Servers can then use this to inform their choices of what to push to clients.

The key words “MUST”, “MUST NOT”, “REQUIRED”, “SHALL”, “SHALL NOT”, “SHOULD”, “SHOULD NOT”, “RECOMMENDED”, “MAY”, and “OPTIONAL” in this document are to be interpreted as described in .

The CACHE_DIGEST frame type is 0xd (decimal 13).

+-------------------------------+-------------------------------+ | Origin-Len (16) | Origin? (\*) ... +-------------------------------+-------------------------------+ | Digest-Value? (\*) ... +---------------------------------------------------------------+

The CACHE_DIGEST frame payload has the following fields: An unsigned, 16-bit integer indicating the length, in octets, of the Origin field. A sequence of characters containing the ASCII serialization of an origin () that the Digest-Value applies to. A sequence of octets containing the digest as computed in and . The CACHE_DIGEST frame defines the following flags: RESET (0x1): When set, indicates that any and all cache digests for the applicable origin held by the recipient MUST be considered invalid. COMPLETE (0x2): When set, indicates that the currently valid set of cache digests held by the server constitutes a complete representation of the cache’s state regarding that origin, for the type of cached response indicated by the STALE flag. VALIDATORS (0x4): When set, indicates that the validators boolean in is true. STALE (0x8): When set, indicates that all cached responses represented in the digest-value are stale at the point in them that the digest was generated; otherwise, all are fresh.

A CACHE_DIGEST frame MUST be sent from a client to a server on stream 0, and conveys a digest of the contents of the client’s cache for the indicated origin. In typical use, a client will send one or more CACHE_DIGESTs immediately after the first request on a connection for a given origin, on the same stream, because there is usually a short period of inactivity then, and servers can benefit most when they understand the state of the cache before they begin pushing associated assets (e.g., CSS, JavaScript and images). Clients MAY send CACHE_DIGEST at other times. If the cache’s state is cleared, lost, or the client otherwise wishes the server to stop using previously sent CACHE_DIGESTs, it can send a CACHE_DIGEST with the RESET flag set. When generating CACHE_DIGEST, a client MUST NOT include cached responses whose URLs do not share origins with the indicated origin. Clients MUST NOT send CACHE_DIGEST frames on connections that are not authoritative (as defined in , 10.1) for the indicated origin. CACHE_DIGEST allows the client to indicate whether the set of URLs used to compute the digest represent fresh or stale stored responses, using the STALE flag. Clients MAY decide whether to only send CACHE_DIGEST frames representing their fresh stored responses, their stale stored responses, or both. Clients can choose to only send a subset of the suitable stored responses of each type (fresh or stale). However, when the CACHE_DIGEST frames sent represent the complete set of stored responses of a given type, the last such frame SHOULD have a COMPLETE flag set, to indicate to the server that it has all relevant state of that type. Note that for the purposes of COMPLETE, responses cached since the beginning of the connection or the last RESET flag on a CACHE_DIGEST frame need not be included. CACHE_DIGEST can be computed to include cached responses’ ETags, as indicated by the VALIDATORS flag. This information can be used by servers to decide what kinds of responses to push to clients; for example, a stale response that hasn’t changed could be refreshed with a 304 (Not Modified) response; one that has changed can be replaced with a 200 (OK) response, whether the cached response was fresh or stale. CACHE_DIGEST has no defined meaning when sent from servers, and SHOULD be ignored by clients.

Given the following inputs: P, an integer smaller than 256, that indicates the probability of a false positive that is acceptable, expressed as 1/2\*\*P. N, an integer that represents the number of entries - a prime number smaller than 2**32 Let f be the number of bits per fingerprint, calculated as P + 3 Let b be the bucket size, defined as 4. Let allocated be the closest power of 2 that is larger than N. Let bytes be f*allocated*b/8 rounded up to the nearest integer Add 5 to bytes Allocate memory of bytes and set it to zero. Assign it to digest-value. Set the first byte to P Set the second till fifth bytes to N in big endian form Return the digest-value. Note: allocated is necessary due to the nature of the way Cuckoo filters are creating the secondary hash, by XORing the initial hash and the fingerprint’s hash. The XOR operation means that secondary hash can pick an entry beyond the initial number of entries, up to the next power of 2. In order to avoid issues there, we allocate the table appropriately. For increased space efficiency, it is recommended that implementations pick a number of entries that’s close to the next power of 2.

Given the following inputs: URL a string corresponding to the Effective Request URI () of a cached response ETag a string corresponding to the entity-tag of a cached response (if the ETag is available; otherwise, null); maxcount - max number of cuckoo hops digest-value Let f be the value of the first byte of digest-value. Let b be the bucket size, defined as 4. Let N be the value of the second to fifth bytes of digest-value in big endian form. Let key be the return value of with URL and ETag as inputs. Let h1 be the return value of with key and N as inputs. Let dest_fingerprint be the return value of with key and f as inputs. Let h2 be the return value of with h1, dest_fingerprint and N as inputs. Let h be either h1 or h2, picked in random. While maxcount is larger than zero: Let position_start be 40 + h * f * b. Let position_end be position_start + f * b. While position_start < position_end: Let bits be f bits from digest_value starting at position_start. If bits is all zeros, set bits to dest_fingerprint and terminate these steps. Add f to position_start. Let e be a random number from 0 to b. Subtract f * (b - e) from position_start. Let bits be f bits from digest_value starting at position_start. Let fingerprint be the value of bits, read as big endian. Set bits to dest_fingerprint. Set dest_fingerprint to fingerprint. Let h be with h, dest_fingerprint and N as inputs. Subtract 1 from maxcount. Subtract f from position_start. Let fingerprint be the f bits starting at position_start. Let h1 be h Subtract 1 from maxcount. If maxcount is zero, return an error. Go to step 7.

Given the following inputs: URL a string corresponding to the Effective Request URI () of a cached response ETag a string corresponding to the entity-tag of a cached response (if the ETag is available; otherwise, null); digest-value Let f be the value of the first byte of digest-value. Let b be the bucket size, defined as 4. Let N be the value of the second to fifth bytes of digest-value in big endian form. Let key be the return value of with URL and ETag as inputs. Let h1 be the return value of with key and N as inputs. Let fingerprint be the return value of with key and f as inputs. Let h2 be the return value of with h1, fingerprint and N as inputs. Let hashes be an array containing h1 and h2. For each h in hashes: Let position_start be 40 + h * f * b. Let position_end be position_start + f * b. While position_start < position_end: Let bits be f bits from digest_value starting at position_start. If bits is fingerprint, set bits to all zeros and terminate these steps. Add f to position_start.

Given the following inputs: key, an array of characters f, an integer indicating the number of output bits Let hash-value be the SHA-256 message digest of key, expressed as an integer. Let h be the number of bits in hash-value Let fingerprint-value be 0 While fingerprint-value is 0 and h > f: Let fingerprint-value be the f least significant bits of hash-value. Let hash-value be the h-f most significant bits of hash-value. Subtract f from h. If fingerprint-value is 0, let fingerprint-value be 1. Return fingerprint-value. Note: Step 5 is to handle the extremely unlikely case where a SHA-256 digest of key is all zeros. The implications of it means that there’s an infitisimaly larger probability of getting a fingerprint-value of 1 compared to all other values. This is not a problem for any practical purpose.

Given the following inputs: URL, an array of characters ETag, an array of characters validators, a boolean indicating whether validators () are to be included in the digest Let key be URL converted to an ASCII string by percent-encoding as appropriate . If validators is true and ETag is not null: Append ETag to key as an ASCII string, including both the weak indicator (if present) and double quotes, as per . Return key TODO: Add an example of the ETag and the key calcuations.

Given the following inputs: key, an array of characters. N, an integer hash-value can be computed using the following algorithm: Let hash-value be the SHA-256 message digest of key, truncated to 32 bits, expressed as an integer. Return hash-value modulo N.

Given the following inputs: hash1, an integer indicating the previous hash. fingerprint, an integer indicating the fingerprint value. N, an integer indicating the number of entries in the digest. Let fingerprint-string be the value of fingerprint in base 10, expressed as a string. Let hash2 be the return value of with fingerprint-string and N as inputs, XORed with hash1. Return hash2.

In typical use, a server will query (as per ) the CACHE_DIGESTs received on a given connection to inform what it pushes to that client; If a given URL and ETag combination has a match in a current CACHE_DIGEST, a complete response need not be pushed; The server MAY push a 304 response for that resource, indicating the client that it hasn’t changed. If a given URL and ETag has no match in any current CACHE_DIGEST, the client does not have a cached copy, and a complete response can be pushed. Servers MAY use all CACHE_DIGESTs received for a given origin as current, as long as they do not have the RESET flag set; a CACHE_DIGEST frame with the RESET flag set MUST clear any previously stored CACHE_DIGESTs for its origin. Servers MUST treat an empty Digest-Value with a RESET flag set as effectively clearing all stored digests for that origin. Clients are not likely to send updates to CACHE_DIGEST over the lifetime of a connection; it is expected that servers will separately track what cacheable responses have been sent previously on the same connection, using that knowledge in conjunction with that provided by CACHE_DIGEST. Servers MUST ignore CACHE_DIGEST frames sent on a stream other than 0.

Given the following inputs: URL a string corresponding to the Effective Request URI () of a cached response . ETag a string corresponding to the entity-tag of a cached response (if the ETag is available; otherwise, null). validators, a boolean digest-value, an array of bits. Let f be the value of the first byte of digest-value. Let b be the bucket size, defined as 4. Let N be the value of the second to fifth bytes of digest-value in big endian form. Let key be the return value of with URL and ETag as inputs. Let h1 be the return value of with key and N as inputs. Let fingerprint be the return value of with key and f as inputs. Let h2 be the return value of with h1, fingerprint and N as inputs. Let hashes be an array containing h1 and h2. For each h in hashes: Let position_start be 40 + h * f * b. Let position_end be position_start + f * b. While position_start < position_end: Let bits be f bits from digest_value starting at position_start. If bits is fingerprint, return true Add f to position_start. Return false.

A Client SHOULD notify its support for CACHE_DIGEST frames by sending the SENDING_CACHE_DIGEST (0xXXX) SETTINGS parameter. The value of the parameter is a bit-field of which the following bits are defined: DIGEST_PENDING (0x1): When set it indicates that the client has a digest to send, and the server may choose to wait for a digest in order to make server push decisions. Rest of the bits MUST be ignored and MUST be left unset when sending. The initial value of the parameter is zero (0x0) meaning that the client has no digest to send the server.

A server can notify its support for CACHE_DIGEST frame by sending the ACCEPT_CACHE_DIGEST (0x7) SETTINGS parameter. If the server is tempted to making optimizations based on CACHE_DIGEST frames, it SHOULD send the SETTINGS parameter immediately after the connection is established. The value of the parameter is a bit-field of which the following bits are defined: ACCEPT (0x1): When set, it indicates that the server is willing to make use of a digest of cached responses. Rest of the bits MUST be ignored and MUST be left unset when sending. The initial value of the parameter is zero (0x0) meaning that the server is not interested in seeing a CACHE_DIGEST frame. Some underlying transports allow the server’s first flight of application data to reach the client at around the same time when the client sends it’s first flight data. When such transport (e.g., TLS 1.3 in full-handshake mode) is used, a client can postpone sending the CACHE_DIGEST frame until it receives a ACCEPT_CACHE_DIGEST settings value. When the underlying transport does not have such property (e.g., TLS 1.3 in 0-RTT mode), a client can reuse the settings value found in previous connections to that origin to make assumptions.

This document registers the following entry in the Permanent Message Headers Registry, as per : Header field name: Cache-Digest Applicable protocol: http Status: experimental Author/Change controller: IESG Specification document(s): [this document] This document registers the following entry in the HTTP/2 Frame Type Registry, as per : Frame Type: CACHE_DIGEST Code: 0xd Specification: [this document] This document registers the following entry in the HTTP/2 Settings Registry, as per : Code: 0x7 Name: ACCEPT_CACHE_DIGEST Initial Value: 0x0 Reference: [this document]

The contents of a User Agent’s cache can be used to re-identify or “fingerprint” the user over time, even when other identifiers (e.g., Cookies ) are cleared. CACHE_DIGEST allows such cache-based fingerprinting to become passive, since it allows the server to discover the state of the client’s cache without any visible change in server behaviour. As a result, clients MUST mitigate for this threat when the user attempts to remove identifiers (e.g., “clearing cookies”). This could be achieved in a number of ways; for example: by clearing the cache, by changing one or both of N and P, or by adding new, synthetic entries to the digest to change its contents. TODO: discuss how effective the suggested mitigations actually would be. Additionally, User Agents SHOULD NOT send CACHE_DIGEST when in “privacy mode.”

On some web browsers that support Service Workers but not Cache Digests (yet), it is possible to achieve the benefit of using Cache Digests by emulating the frame using HTTP Headers. For the sake of interoperability with such clients, this appendix defines how a CACHE_DIGEST frame can be encoded as an HTTP header named Cache-Digest. The definition uses the Augmented Backus-Naur Form (ABNF) notation of with the list rule extension defined in .

Cache-Digest = 1#digest-entity digest-entity = digest-value *(OWS ";" OWS digest-flag) digest-value = <Digest-Value encoded using base64url> digest-flag = token

A Cache-Digest request header is defined as a list construct of cache-digest-entities. Each cache-digest-entity corresponds to a CACHE_DIGEST frame. Digest-Value is encoded using base64url . Flags that are set are encoded as digest-flags by their names that are compared case-insensitively. Origin is omitted in the header form. The value is implied from the value of the :authority pseudo header. Client MUST only send Cache-Digest headers containing digests that belong to the origin specified by the HTTP request. The example below contains one digest of fresh resource and has only the COMPLETE flag set.

Cache-Digest: AfdA; complete

Clients MUST associate Cache-Digest headers to every HTTP request, since Fetch - the HTTP API supported by Service Workers - does not define the order in which the issued requests will be sent to the server nor guarantees that all the requests will be transmitted using a single HTTP/2 connection. Also, due to the fact that any header that is supplied to Fetch is required to be end-to-end, there is an ambiguity in what a Cache-Digest header respresents when a request is transmitted through a proxy. The header may represent the cache state of a client or that of a proxy, depending on how the proxy handles the header.

Thanks to Yoav Weiss for his idea and text to use Cuckoo Filter. Thanks to Stefan Eissing for his suggestions.

Switch to Cuckoo Filter.

Added definition of the Cache-Digest header. Introduce ACCEPT_CACHE_DIGEST SETTINGS parameter. Change intended status from Standard to Experimental.

Make the scope of a digest frame explicit and shift to stream 0.