Internet-Draft M. Toomim
Expires: Apr 20, 2024 Invisible College
Intended status: Proposed Standard Oct 21, 2024
HTTP Resource Versioning
draft-toomim-httpbis-versions-02
Abstract
HTTP resources change over time. Each change to a resource creates a
new "version" of its state. HTTP systems often need a way to
identify, read, write, navigate, and/or merge these versions, in
order to implement cache consistency, create history archives, settle
race conditions, request incremental updates to resources, interpret
incremental updates to versions, or implement distributed
collaborative editing algorithms.
This document analyzes existing methods of versioning in HTTP,
highlights limitations, and sketches a more general versioning
approach that can enable new use-cases for HTTP. An upgrade path for
legacy intermediaries is provided.
Status of this Memo
This Internet-Draft is submitted in full conformance with the
provisions of BCP 78 and BCP 79.
Internet-Drafts are working documents of the Internet Engineering
Task Force (IETF), its areas, and its working groups. Note that
other groups may also distribute working documents as
Internet-Drafts. The list of current Internet-Drafts is at
https://datatracker.ietf.org/drafts/current/.
Internet-Drafts are draft documents valid for a maximum of six months
and may be updated, replaced, or obsoleted by other documents at any
time. It is inappropriate to use Internet-Drafts as reference
material or to cite them other than as "work in progress."
The list of current Internet-Drafts can be accessed at
https://www.ietf.org/1id-abstracts.html
The list of Internet-Draft Shadow Directories can be accessed at
https://www.ietf.org/shadow.html
Table of Contents
1. Introduction ..................................................4
1.1. Existing Versioning in HTTP .................................6
1.1.1. Versioning with `Last-Modified` ...........................6
1.1.2. Versioning with `ETag` ....................................7
1.1.3. Versioning encoded within URLs ............................8
1.2. Limitations of Existing Approaches ..........................8
1.3. Design Goals for a New HTTP Versioning System ...............9
1.4. Overview of Proposed Solution ...............................9
2. HTTP Resource Versioning .....................................10
2.1. Version History ............................................10
2.2. Version Identifiers ........................................10
2.3. Version and Parents Headers ................................11
2.4. Using Versioning with HTTP Methods .........................12
2.4.1. GET the current version ..................................12
2.4.2. GET a specific version ...................................13
2.4.3. PUT a new version ........................................13
2.4.4. GET a range of historical versions .......................14
2.5. Rules for Version and Parents headers ......................16
2.6. Status 309: Version Unknown Here ...........................16
2.7. The Current-Version header .................................16
3. Version-Type Header ..........................................17
4. Versioning through Intermediaries ............................18
4.1. Detecting Legacy Intermediaries ............................19
4.2. Responding to history requests .............................20
5. Example Uses .................................................20
5.1. Incremental RSS subscription ...............................20
5.2. Hosting git via HTTP .......................................21
5.3. Resumeable uploads .........................................23
5.3.1. Version-Type: bytestream .................................23
5.3.2. Resumable Upload Protocol ................................24
5.4. Distributed collaborative editing ..........................26
5.5. Improved Header Compression for Version-Type: rle ..........28
5.6. Run-Length Compression without Header Compression ..........31
6. Acknowledgements .............................................32
7. Conventions ..................................................32
8. IANA Considerations ..........................................33
9. Copyright Notice .............................................33
10. Security Considerations ......................................33
11. Authors' Addresses ...........................................34
12. References ...................................................34
12.1. Normative References ......................................34
12.2. Informative References ....................................34
1. Introduction
From the perspective of a single computer, the version history of a
HTTP resource that is changing on that computer can be viewed in a
line of time:
o <-- oldest version
|
o
|
o
|
o <-- newest version
We call this a "linear" history.
However, if multiple computers change a resource over a network
"simultaneously" (ie. before their changes propagate to one another),
then the version history forks into a DAG, or "partial order":
o <-- oldest version
/ \
o o
\ /
o
|
o <-- newest version
HTTP systems often need a way to identify, read, write, navigate,
and/or merge versions of history in order to (1) implement better
cache consistency, (2) create history archives, (3) settle race
conditions, (4) request incremental updates to resources, (5)
interpret incremental updates to versions, or (6) implement
distributed collaborative editing algorithms.
Furthermore, advanced distributed systems often devise special
formats for partially-ordered timestamps that allow inferences for
improved performance, such as lamport clocks, vector clocks, version
vectors, hash histories, and append-only-log indices.
Implementations can rely on information embedded in these timestamps
to compress history metadata, optimize partial-order computations, or
infer the value of state.
A general mechanism for versioning HTTP resources could enable a
number of new use-cases:
- RSS clients could request incremental updates when polling,
instead of re-downloading redundant unchanged feed items after
each change to any item
- Servers could accept incoming patches based on old or parallel
versions of history, and even rebase those patches for other
clients, at other points in history
- Collaborative editing could be built directly into HTTP
resources, providing the abilities of Google Docs at any URL
- Git repositories could be hosted directly over HTTP; rather than
embedding versioning information within opaque blobs that use
HTTP just as a transport
- Caches and archives could hold and serve multiple versions of a
resource, enabling audits and distributed backups
- Distributed databases could standardize network APIs to HTTP,
while retaining distributed consistency guarantees
This document analyzes existing approaches to versioning of resources
in HTTP, and sketches an approach to a more general and powerful
approach that addresses use-cases like these.
(Note that this document does NOT speak to the versioning of HTTP
APIs -- only HTTP resources, which are used within APIs.)
1.1. Existing Versioning in HTTP
Current approaches to versioning in HTTP address disparate use-cases,
but have limitations and trade-offs. The Last-Modified and ETag
headers were invented for cache consistency, but do not provide an
ordering of version history through time, nor do they handle forks
and merges in distributed time. On the other hand, a number of
forking/merging versioning systems have been proposed (WebDAV, Link
Relations) that create new resources to represent versions of
existing resources, but this approach has been more complex, and has
not seen much adoption in practice. No HTTP versioning system today
allows for articulating custom distributed timestamp formats such as
vector clocks.
1.1.1. Versioning with `Last-Modified`
The Last-Modified header specifies a clock date that caches and
clients can use to know when a change has occurred:
Last-Modified: Sat, 6 Jul 2024 07:28:00 GMT
This header is useful for caching and conditional requests (using the
If-Modified-Since header). However, it has several limitations:
1. It is limited to the precision of the wallclock. If a resource
changes within the same second, the Last-Modified date won't
change, and caches can become inconsistent.
2. It is susceptible to clock skew in distributed systems,
potentially leading to inconsistencies across different servers.
3. It doesn't work well for dynamically generated content, where the
modification time might not be meaningful or easily determined.
1.1.2. Versioning with `ETag`
The ETag header allows more precision. It specifies a version with a
string that uniquely identifies a cacheable representation:
ETag: "2u34fa7yorz0"
ETags can be strong or weak, with weak ETags prefixed by W/:
ETag: W/"2u34fa7yorz0"
ETags are used in conditional requests with If-None-Match and
If-Match headers and can be used for optimistic concurrency
control. However:
1. While helping with cache validation, ETags are not accurate
markers of time. There is no way to order versions by ETag, or
know which version came before another.
2. ETags are unique to content, not timestamps. It's possible for
the same ETag to recur over time if the resource changes back and
forth between a common state.
3. Strong ETags are sensitive to Content-Encoding. If a single
version of a resource is transmitted with different
Content-Encodings (e.g., gzip), it will be sent with different
strong ETags. Thus, one can have multiple ETags for the same
version in history, as well as a single ETag for multiple versions
of history.
1.1.3. Versioning encoded within URLs
In practice, application programmers tend to encode versions within
URLs:
https://unpkg.com/braid-text@0.0.18/index.js
This approach is common in API versioning (e.g., /api/v1/resource).
However, it has several drawbacks:
1. It loses the semantics of a "resource changing over time."
Instead, it creates multiple version resources for every single
logical resource.
2. It necessitates additional standards for version history on top of
URLs (e.g., Memento, WebDAV, Link Relations for Versioning
[RFC5829]).
3. Given a URL, we still need a standard way to extract the version
itself, get the previous and next version(s), and understand the
format of the version(s) (e.g., major.minor.patch).
4. This approach can lead to URI proliferation, potentially impacting
caching strategies and SEO.
5. It may complicate content negotiation and RESTful design
principles.
The choice to embed versions into URLs can be useful, but carries
with it additional tradeoffs. A versioning system does not need to
depend on allocating a URL for each version; but could be compatible
with doing so.
1.2. Limitations of Existing Approaches
Current HTTP versioning mechanisms serve specific use cases, but have
limitations collectively and individually. Last-Modified and ETags
do not represent the order of history. URL approaches to history add
complexity to RESTful design. No approach yet enables custom
timestamp formats.
As a result, programmers today must implement multiple approaches to
versioning in their applications -- each with subtly different logic
-- and cannot implement common infrastructure for distributed
versioning, archiving, and collaborative editing that works across
HTTP systems.
1.3. Design Goals for a New HTTP Versioning System
We sketch an HTTP resource versioning system with the following
design goals:
1. Unified: A single, flexible way to identify versions across
diverse versioning needs, from simple caching to complex
distributed editing.
2. Support for non-linear history: allow branching and merging
through a partial order (DAG) of versions.
3. Extensible Version Identification: Allow for custom version ID
formats to support various timestamp schemes.
4. Optimizable for High-Performance: Supports optimizations of
advanced distributed systems.
5. Independent of additional URLs: Does not require allocation of new
URLs to represent versions; but is compatible with systems doing
so.
1.4. Overview of Proposed Solution
To meet these design goals, we propose the following:
1. Version and Parents Headers: New headers to specify the current
version of a resource and its parent versions, enabling
representation of both linear and non-linear version histories.
2. Version as Sets of Strings: Versions are represented as sets of
unique string identifiers, allowing for custom versioning schemes
and distributed timestamps.
3. Extensible Version-Type Header: Allows specification of different
timestamp formats in custom versioning schemes (e.g., git-style
hashes, bytestreams and append-only logs, vector clocks) to allow
additional computational inferences for various use cases.
4. Versioned Resource Operations: Extends standard HTTP methods (GET,
PUT, PATCH) with versioning semantics, allowing version-aware
interactions with resources.
This system provides a flexible foundation that can be adapted to
various versioning needs, from simple content distribution to complex
collaborative editing scenarios, while maintaining compatibility with
existing HTTP infrastructure.
We start by specifying how to add versioning to HTTP requests and
responses.
2. HTTP Resource Versioning
This section defines the core concepts and mechanisms for HTTP
Resource Versioning.
2.1. Version History
Each HTTP resource maintains a version history, representing its
state changes over time. This history forms a partially ordered set,
where some versions have a clear sequential relationship, while
others may occur in parallel.
2.2. Version and Event Identifiers
A "Version" is uniquely identified as a set of mutation "Event IDs",
which are formatted as a list of strings in the Structured Headers
format [RFC8941]. For example:
"ajtva12kid", "cmdpvkpll2"
Each event ID is unique, and represents a distinct change to the
resource at a specific point in time. When a resource has been
mutated by any given set of event IDs, the state of its
representations can be uniquely determined by the View() function
specified in its Merge-Type [Merge-Types]. Therefore, a Version of
the resource uniquely identifies the state of its representation as
it changes over time.
The ordering of event IDs carries no meaning. Event IDs SHOULD be
sorted lexicographically whenever received or sent, with exactly one
space after "," separators, to canonicalize the set's serialization
as unique string, e.g. as a unique cache key.
The formatting and interpretation of a resource's version and event
IDs are constrained according to its Version-Type, as defined in
Section 4.
2.3. Version and Parents Headers
This specification introduces two new HTTP headers: Version and
Parents. These headers communicate version information in requests
and responses.
The Version header specifies the current version of a resource:
Version: "dkn7ov2vwg"
The Parents header specifies the immediate predecessor version(s):
Parents: "ajtva12kid", "cmdpvkpll2"
These headers may be used in PUT, PATCH, or POST requests and GET
responses to convey the version history of a resource.
Any version can be reconstructed by first merging its parents
(according to [Merge-Types]), then applying its specific changes.
The full graph of parent relationships forms the version history DAG.
Version A is considered to precede version B if and only if A is an
ancestor of B in this DAG.
The version immediately after a mutation will contain just that
mutation's event ID. However, the version obtained by merging N
parallel mutations will contain N event IDs:
Version: "dkn7ov2vwg", "v2vwgdkn7o"
For any two event IDs A and B that are specified in a Version or
Parents header, A cannot be a descendent of B or vice versa.
If a client or server does not specify a Version for a resource it
transfers, the recipient MAY generate and assign it new event IDs.
If a client or server does not specify a Parents header when
transferring a new version, the recipient MAY presume that the most
recent versions it has (the frontier of time) are the parents of the
new version. It MAY also ignore or reject the update.
2.4. Using Versioning with HTTP Methods
The Version and Parents can exist in these methods:
- GET
- HEAD
- PUT
- PATCH
- POST
- DELETE
For each of these methods, if a Version or Parents header exists in
the request, it should also exist in the response, and if not, the
server may nonetheless specify it in its response.
2.4.1. GET the current version
If the Version: header is not specified, a GET request returns the
current version of the state as usual:
Request:
GET /chat
Response:
HTTP/1.1 200 OK
Version: "ej4lhb9z78"
Parents: "oakwn5b8qh", "uc9zwhw7mf"
Content-Type: application/json
Content-Length: 64
[{"text": "Hi, everyone!",
"author": {"link": "/user/tommy"}}]
The server MAY include a Version and/or Parents header in the
response, to indicate the current version and its parents.
Clients can use a HEAD request to elicit versioning history without
downloading the body:
Request:
HEAD /chat
Response:
HTTP/1.1 200 OK
Version: "ej4lhb9z78"
Parents: "oakwn5b8qh", "uc9zwhw7mf"
Content-Type: application/json
2.4.2. GET a specific version
A server can allow clients to request historical versions of a
resource in GET requests by responding to the Version and Parents
headers. A client can specify a specific version that it wants with
the Version header:
Request:
GET /chat
Version: "ej4lhb9z78"
Response:
HTTP/1.1 200 OK
Version: "ej4lhb9z78"
Parents: "oakwn5b8qh", "uc9zwhw7mf"
Content-Type: application/json
Content-Length: 64
[{"text": "Hi, everyone!",
"author": {"link": "/user/tommy"}}]
2.4.3. PUT a new version
When a PUT request changes the state of a resource, it can specify
the new version of the resource, and the parent version that it was
based on:
Request:
PUT /chat
Version: "ej4lhb9z78"
Parents: "oakwn5b8qh", "uc9zwhw7mf"
Content-Type: application/json
Content-Length: 64
[{"text": "Hi, everyone!",
"author": {"link": "/user/tommy"}}]
Response:
HTTP/1.1 200 OK
The Version and Parents headers are optional. If Version is omitted,
the recipient may assign new event IDs. If Parents is omitted, the
recipient may assume that its current version is the version's
parents.
2.4.4. GET a range of historical versions
A client can request a range of history by including a Parents and a
Version header together. The Parents marks the beginning of the
range (the oldest versions) and the Version marks the end of the
range (the newest versions) that it requests.
Request:
GET /chat
Version: "3"
Parents: "1a", "1b"
Response:
HTTP/1.1 209 Multiresponse
Current-Version: "3"
HTTP/1.1 200 OK
Version: "2"
Parents: "1a", "1b"
Content-Type: application/json
Content-Length: 64
[{"text": "Hi, everyone!",
"author": {"link": "/user/tommy"}}]
HTTP/1.1 200 OK
Version: "3"
Parents: "2"
Content-Type: application/json
Merge-Type: sync9
Content-Length: 117
[{"text": "Hi, everyone!",
"author": {"link": "/user/tommy"}}
{"text": "Yo!",
"author": {"link": "/user/yobot"}]
Note that this example uses a new "Multiresponse" code, which is
currently being drafted [Multiresponse]. See [Braid-HTTP] Section 3
for an earlier draft of the semantics.
2.5. Rules for Version and Parents headers
If a GET request contains a Version header:
- If the Parents header is absent, the server SHOULD return a
single response, containing the requested version of the resource
in its body, with the Version response header set to the same
version.
- If the server does not support historical versions, it MAY ignore
the Version header and respond as usual, but MUST NOT include the
Version header in its response.
If a GET request contains a Parents header:
- The server SHOULD send the set of versions updating the Parents
to the specified Version. If no Version is specified, then it
should update the client to the server's current version.
- If the server does not support historical versions, then it MAY
ignore the Parents header, but MUST NOT include the Parents
header in its response.
A server does not need to honor historical version requests for all
documents, for all history. If a server no longer has the historical
context needed to honor a request, it may respond with a TBD error
code.
2.6. Status 309: Version Unknown Here
Since mutation history can grow large, any peer might drop any
portion of its resources history at any time, and clients cannot rely
on any particular portion of history existing on a server or
intermediary when it makes requests.
If a server's response to a request depends on knowledge of a
particular version or span of history that it does not possess, the
server SHOULD respond with a status code of "309 Version Unknown
Here". This implies that the client should "redirect" its request to
a different peer or archival storage, outside the scope of this
specification.
2.7. The Current-Version header
While sending historical versions, a server or client can specify its
current latest version with the Current-Version header. The other
party may desire this information to know when it has caught up with
the latest version. This is also used in the resumeable uploads
example below.
3. Version-Type Header
The optional Version-Type header specifies constraints on the format
and interpretation of event IDs. This allows for various
optimizations and specialized versioning schemes.
For example:
Version-Type: git
This could indicate that event IDs will be git-style hashes,
branches, or tags. Peers could verify that the entire repository at
a given version hashes to the specified ID.
Another example:
Version-Type: dt
or
Version-Type: rle
This could specify the use of Diamond-Types or "run-length-encoded"
event IDs, which are Lamport timestamps of the form:
Version: "<agent>-<char_count>"
Diamond-Types, Automerge, and other algorithms use this format to
compress history metadata through run-length encoding of consecutive
insertions. This allows a set of 50 inserted characters to be stored
as 50 bytes plus one event ID, rather than 50 bytes plus 50 event
IDs (each of which takes up multiple bytes).
Implementers may define custom Version-Types to suit specific needs:
Version-Type: vector-clock
A vector clock event ID might take the form:
Version: "{agentid1: counter1, agentid2: counter2, ...}"
Vector clocks enable direct computation of partial order between any
two event IDs without examining the full version history graph.
(To know the order between two vector clocks A and B, one needs only
to compare each agent's counter between A and B. If A dominates
across all agents, it is newer. If B dominates, then it is newer.
Otherwise, the ordering between the two vector clocks is not known,
and we can say that they happened in parallel.)
4. Versioning through Intermediaries
Intermediaries can take advantage of versioning to uniquely
reference, store, and serve multiple states/updates across a
resource's history. To distinguish versions, intermediaries must add
the "Version" and "Parents" headers to their cache keys -- equivalent
to the Vary header:
Vary: version, parents
Intermediaries SHOULD behave as if the Version and Parents headers
have been added to the Vary header in every response passing through
them. To support legacy versioning-unaware intermediaries, the
origin server is RECOMMENDED to explicitly add or extend Vary with
"version, parents" in all its responses, unless they are certain that
no legacy intermediaries will process the response.
4.1. Detecting Legacy Intermediaries
In the case, that a legacy intermediary *does* process a versioned
response without the Vary header, it can be detected by the client by
noticing that the Version and Parents headers used in its request are
not the same as those it receives in response. Here is an example:
Presume we start with two versions:
PUT /foo
Version: "1"
Hi
PUT /foo
Version: "2"
Hi everyone!
Now, if someone GETs the old version:
GET /foo
Version: "1"
Then a versioning-aware origin server can return it:
HTTP/1.1 200 OK
Version: "1"
Hi everyone!
A versioning-unaware intermediary might cache this response,
without understanding that it depends on the version.
This causes problems for a client requesting the *newest* version:
GET /foo
Version: "2"
The versioning-unaware cache will return the most recent response
it has seen -- unaware that it is not the version requested:
HTTP/1.1 200 OK
Version: "1"
Hi everyone!
To detect this, the client SHOULD check that the Version and Parents
headers in the response contain a superset of the Event IDs specified
in the request. (A missing Version or Parents header is considered
to be an empty set for the version or parents of this superset
calculation.) The client can throw a warning to the programmer when
encountering a response that does not contain the request's Event
IDs, so that he knows to add Vary to the server's responses.
4.2. Responding to history requests
[xxx todo: In the future, an intermediary could conceivably chain the
cached responses of multiple GET responses together into a bigger
history object. We can specify the situations in which that is safe
to do here.]
5. Example Uses
[xxx to do]
- Reconnecting to feed of posts as Version-Type: arraystream
- New Cache-Control: version-immutable proposal
5.1. Incremental RSS Subscription
Traditional RSS readers inefficiently poll servers, often downloading
entire feeds when only minor changes have occurred. HTTP Resource
Versioning enables more efficient incremental updates.
A client can specify its last known version using the Parents header:
Request:
GET /feed.rss
Accept: application/rss+xml
Parents: "4"
The server can then respond with only the changes since that version:
Response:
HTTP/1.1 200 OK
Content-Type: application/rss+xml+patch
Version: "5"
Parents: "4"
<?xml version="1.0" encoding="UTF-8" ?>
<rss version="2.0">
<channel>
<title>My RSS Feed</title>
<item>
<title>This is a new entry</title>
<description>Incremental update example</description>
<link>http://www.example.com/blog/post/1</link>
</item>
</channel>
</rss>
This approach significantly reduces bandwidth usage and processing
time for both client and server. The specific patch format used can
vary; see [updates] or [range-patch] for examples.
5.2. Hosting git via HTTP
We can host a git repository directly through HTTP, where each file
corresponds to a resource, and all have a version history.
Git versions are normally specified as a hash. The server can
express this with a "Version-Type: git" header:
Request:
GET /repo/readme.md
Response:
HTTP/1.1 200 OK
Content-Type: text/markdown
Version-Type: git
Version: "9531a9702af0d90dd489050ed8e25f87912a9252"
Parents: "3a4c361f8e0349fe4b25c1ff46ebec1cec66e60f"
...
Git also allows specifying a version with a short string, like
"HEAD", which works for any tag or branch. We can request the latest
"development" branch version with:
Request:
GET /repo/readme.md
Version: "development"
Response:
HTTP/1.1 200 OK
Content-Type: text/markdown
Version-Type: git
Version: "9e26e8837a4f6a4445e74eed744fe8af85efd0c2"
Parents: "1d5f89f8843b33b91d62bf95877e46b23fd86741"
...
One can also request the files from release tagged "1.3.5" using:
Request:
GET /repo/readme.md
Version: "1.3.5"
One can clone a repo by asking for all versions from the root to
HEAD:
Request:
GET /repo/readme.md
Version: "HEAD"
Parents: "ROOT"
Response:
HTTP/1.1 209 Multiresponse
HTTP/1.1 200 OK
Content-Type: text/markdown
Version-Type: git
Version: "9e26e8837a4f6a4445e74eed744fe8af85efd0c2"
Parents: "1d5f89f8843b33b91d62bf95877e46b23fd86741"
Content-Length: 190
...
HTTP/1.1 200 OK
Content-Type: text/markdown
Version-Type: git
Version: "1d5f89f8843b33b91d62bf95877e46b23fd86741"
Parents: "1cf6ab4ed836d4d7308ac93edbc6fd18a69ef88f"
Content-Length: 192
...
In fact, git itself already supports two HTTP protocols: a "dumb" and
a "smart" protocol. The dumb protocol uses plain HTTP, but doesn't
support incremental updates -- each pull re-downloads the entire pack
file. The smart protocol allows the client to specify the version it
has, and the version it wants:
0054want 31f1c37dfa1bf983e4d67e06fac28e8e6f
00093bd7884 HEAD@{1}
0032have e68fe437718c37155c7e3e5f4a3ff17c4f476940
0000
We can express this with HTTP Versioning as:
Request:
GET /repo/readme.md
Version: "31f1c37dfa1bf983e4d67e06fac28e8e6f"
Parents: "e68fe437718c37155c7e3e5f4a3ff17c4f476940"
This expresses aspects of the "smart" git protocol over plain HTTP.
5.3. Resumeable uploads
Resource Versioning semantics enable efficient implementation of
resumable uploads, providing an alternative perspective to
[Resumeable Upload].
5.3.1. Version-Type: bytestream
For uploads, we can consider the resource as an append-only
bytestream, declared with a header:
Version-Type: bytestream
Bytestream versions are represented as:
Version: "<agent>-<byte_count>"
For example, "x82ha-344" indicates "the resource state after agent
`x82ha` appended 344 bytes".
This approach creates a direct correspondence between time and
space: each version increment represents one additional byte in the
stream.
5.3.2. Resumable Upload Protocol
To initiate an upload, the client specifies the Version-Type and the
expected final version using the Current-Version header:
Request:
PUT /something
Current-Version: "abwejf-900"
Version-Type: bytestream
Content-Length: 900
<binary data of length 900>
For a successful upload, the server responds as usual:
Response:
200 OK
If the upload is interrupted, the client can query the server's
current state:
Request:
HEAD /something
Parents: "abwejf-0"
The server's response determines the client's next action:
A. Upload complete:
Response:
200 OK
Parents: "abwejf-0"
Version: "abwejf-900"
B. Partial upload:
Response:
206 Partial Content
Parents: "abwejf-0"
Version: "abwejf-400"
C. No upload progress:
Response:
416 Range Not Satisfiable
Based on the response, the client proceeds as follows:
- Case A: Upload is complete, no further action needed.
- Case B: Resume the upload from the last received byte:
Request:
PUT /something
Current-Version: "abwejf-900"
Parents: "abwejf-400"
Content-Range: bytes 400-900/900
Content-Length: 500
<binary data from 400-900>
- Case C: Restart the upload from the beginning.
This protocol leverages general version semantics, allowing servers
implementing HTTP Resource Versioning with the "bytestream"
Version-Type to inherently support resumable uploads.
5.4. Distributed collaborative editing
This versioning system can also support full CRDT and OT
collaborative editing features (when used with other extensions such
as [Braid-HTTP]), allowing every URL to gain the functionality of
Google Docs.
The [Braid-Text] project implements a very efficient style of this.
When you first load a resource, a server provides it as a single
version:
Request:
GET https://braid.org/test
Accept: text/plain
Subscribe: true
Response:
HTTP/1.1 209 Multiresponse
HTTP/1.1 200 OK
Version: "2agvvzgccrq-5"
Version-Type: rle
Merge-Type: simpleton
Content-Length: 12
Hello world!
Updates are expressed as a stream of patches:
Response (continued):
HTTP/1.1 200 OK
Version: "4590r8uwm63-18"
Parents: "2agvvzgccrq-5"
Content-Length: 1
Content-Range: text [12:12]
:
HTTP/1.1 200 OK
Version: "4590r8uwm63-19"
Parents: "4590r8uwm63-18"
Content-Length: 1
Content-Range: text [13:13]
)
This versioning system supports multiple [Merge-Types], and they can
even co-exist simultaneously for the same resource. For instance,
braid-text supports two merge-types simultaneously:
- The "simpleton" merge-type requires the server to rebase all
edits for the client
- The "dt" merge-type uses a fully peer-to-peer merge algorithm
called Diamond-Types
Clients can connect with either merge-type, and can even change
merge-type on-the-fly -- the version history itself can be re-used.
5.5. Improved Header Compression for Version-Type: rle
A header compression scheme can leverage Version-Type patterns to
significantly compress messages. For example, consider the following
scenario where 4 characters are inserted into a collaborative editor,
requiring 479 bytes when uncompressed:
PUT /some.txt
Version: "3f84786c-57"
Parents: "3f84786c-56"
Content-Length: 1
Content-Range: text 471:471
a
PUT /some.txt
Version: "3f84786c-58"
Parents: "3f84786c-57"
Content-Length: 1
Content-Range: text 472:472
s
PUT /some.txt
Version: "3f84786c-59"
Parents: "3f84786c-58"
Content-Length: 1
Content-Range: text 473:473
d
PUT /some.txt
Version: "3f84786c-60"
Parents: "3f84786c-59"
Content-Length: 1
Content-Range: text 474:474
f
Huffman Encoding (e.g. in HTTP's HPACK and QPACK) can compress these
headers by identifying repeated strings and assigning them codes:
U = "PUT /some.txt"
W = "Version: \"3f84786c-{ }\""
X = "Parents: \"3f84786c-{ }\""
Y = "Content-Length: 1"
Z = "Content-Range: text { }"
This compression reduces the messages to:
U
W{57}
X{56}
Y
Z{471:471}
a
U
W{58}
X{57}
Y
Z{472:472}
s
U
W{59}
X{58}
Y
Z{473:473}
d
U
W{60}
X{59}
Y
Z{474:474}
f
This initial compression reduces the 479 bytes to approximately 123
bytes.
Further compression is possible by recognizing that the numeric
parameters in each header increment by 1 from the previous header.
This allows us to take advantage of run-length encoding; specified as
Version-Type: rle. We can thus compress the entire run of inserts
using a new compression function, RUN(a, b, c):
RUN(a, b, insertions) =
for i, char in insertions
return `
U
W{a+i}
X{a+i-1}
Y
Z{b+i:b+i}
char
`
Where the parameters are:
a: Initial version number
b: Initial content range
insertions: String of characters to be inserted
We could then compress the 479 bytes down to a single 20-byte message:
RUN(57, 474, 'asdf')
For even more compact representation, sacrificing human-readability,
this can be encoded in 12 bytes:
R57,474,asdf
Further compression is possible by using variable-length integer
encoding (such as LEB128 or Protocol Buffers' varint) for the numeric
values. This approach could potentially reduce the encoding to 7
bytes for these 4 insertions, or just 1.46% of the uncompressed 479
bytes.
5.6. Run-Length Compression without Header Compression
In practice, run-length compression can also be applied without
relying on header compression. By specifying Version-Type: rle, any
peer can receive a series of N insertions as a single update and
infer that the string in the update was composed of N separate
insertions. For instance, the previous example of four PUT requests
can be compressed into a single update without any header
compression:
PUT /some.txt
Version: "3f84786c-60"
Parents: "3f84786c-56"
Version-Type: rle
Content-Length: 4
Content-Range: text 471:474
asdf
This single request expresses the change from version 3f84786c-56 to
3f84786c-60, implicitly including the three intermediate versions
(3f84786c-57, 3f84786c-58, and 3f84786c-59). Any client or server
that understands "Version-Type: rle" can infer these intermediate
versions when necessary by slicing the content "asdf" to the
appropriate length for each version.
For example:
- Version 3f84786c-57 would contain "a"
- Version 3f84786c-58 would contain "as"
- Version 3f84786c-59 would contain "asd"
- Version 3f84786c-60 would contain the full "asdf"
This approach significantly reduces the number of HTTP requests and
amount of data required for a series of small, consecutive inserts,
while still allowing for precise version control and the ability to
reconstruct any intermediate state.
6. Acknowledgements
This is derived from prior draft [Braid-HTTP] with authors:
- Michael Toomim
- Greg Little
- Raphael Walker
- Bryn Bellomy
- Joseph Gentle
And incorporates additional ideas from:
- Rahul Gupta
- Duane Johnson
- Mitar Milutinovic
- Paul Kuchenko
7. Conventions
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 [RFC2119].
8. IANA Considerations
This document defines the following new HTTP headers, which should be
added to the "Permanent Message Header Field Names" registry at
<https://www.iana.org/assignments/message-headers>:
Header Field Name: Version
Protocol: http
Status: standard
Reference: This document
Header Field Name: Parents
Protocol: http
Status: standard
Reference: This document
Header Field Name: Version-Type
Protocol: http
Status: standard
Reference: This document
Header Field Name: Current-Version
Protocol: http
Status: standard
Reference: This document
9. Copyright Notice
Copyright (c) 2023 IETF Trust and the persons identified as the
document authors. All rights reserved.
This document is subject to BCP 78 and the IETF Trust's Legal
Provisions Relating to IETF Documents
(http://trustee.ietf.org/license-info) in effect on the date of
publication of this document. Please review these documents
carefully, as they describe your rights and restrictions with respect
to this document. Code Components extracted from this document must
include Simplified BSD License text as described in Section 4.e of
the Trust Legal Provisions and are provided without warranty as
described in the Simplified BSD License.
10. Security Considerations
XXX Todo
11. Authors' Addresses
For more information, the authors of this document are best contacted
via Internet mail:
Michael Toomim
Invisible College, Berkeley
2053 Berkeley Way
Berkeley, CA 94704
EMail: toomim@gmail.com
Web: https://invisible.college/@toomim
12. References
12.1. Normative References
[RFC5789] "PATCH Method for HTTP", RFC 5789.
[RFC9110] "HTTP Semantics", RFC 9110.
[Braid-HTTP] draft-toomim-httpbis-braid-http-04
[Merge-Types] draft-toomim-httpbis-merge-types-00
[Updates] draft-toomim-httpbis-updates-[TBD]
[Multiresponse] draft-toomim-httpbis-multiresponse-[TBD]
[Range-Patch] draft-toomim-httpbis-range-patch-00
[Resumeable Upload] draft-ietf-httpbis-resumable-upload
12.2. Informative References
[XHR] Van Kestern, A., Aubourg, J., Song, J., and R. M.
Steen, H. "XMLHttpRequest", September 2019.
<https://xhr.spec.whatwg.org/>
[SSE] Hickson, I. "Server-Sent Events", W3C Recommendation,
February 2015.
<https://www.w3.org/TR/2015/REC-eventsource-20150203/>