Mercurial > public > mercurial-scm > hg
diff mercurial/help/internals/wireprotocol.txt @ 37288:9bfcbe4f4745
wireproto: add streams to frame-based protocol
Previously, the frame-based protocol was just a series of frames,
with each frame associated with a request ID.
In order to scale the protocol, we'll want to enable the use of
compression. While it is possible to enable compression at the
socket/pipe level, this has its disadvantages. The big one is it
undermines the point of frames being standalone, atomic units that
can be read and written: if you add compression above the framing
protocol, you are back to having a stream-based protocol as opposed
to something frame-based.
So in order to preserve frames, compression needs to occur at
the frame payload level.
Compressing each frame's payload individually will limit compression
ratios because the window size of the compressor will be limited
by the max frame size, which is 32-64kb as currently defined. It
will also add CPU overhead, as it is more efficient for compressors
to operate on fewer, larger blocks of data than more, smaller blocks.
So compressing each frame independently is out.
This means we need to compress each frame's payload as if it is part
of a larger stream.
The simplest approach is to have 1 stream per connection. This
could certainly work. However, it has disadvantages (documented below).
We could also have 1 stream per RPC/command invocation. (This is the
model HTTP/2 goes with.) This also has disadvantages.
The main disadvantage to one global stream is that it has the very
real potential to create CPU bottlenecks doing compression. Networks
are only getting faster and the performance of single CPU cores has
been relatively flat. Newer compression formats like zstandard offer
better CPU cycle efficiency than predecessors like zlib. But it still
all too common to saturate your CPU with compression overhead long
before you saturate the network pipe.
The main disadvantage with streams per request is that you can't
reap the benefits of the compression context for multiple requests.
For example, if you send 1000 RPC requests (or HTTP/2 requests for
that matter), the response to each would have its own compression
context. The overall size of the raw responses would be larger because
compression contexts wouldn't be able to reference data from another
request or response.
The approach for streams as implemented in this commit is to support
N streams per connection and for streams to potentially span requests
and responses. As explained by the added internals docs, this
facilitates servers and clients delegating independent streams and
compression to independent threads / CPU cores. This helps alleviate
the CPU bottleneck of compression. This design also allows compression
contexts to be reused across requests/responses. This can result in
improved compression ratios and less overhead for compressors and
decompressors having to build new contexts.
Another feature that was defined was the ability for individual frames
within a stream to declare whether that individual frame's payload
uses the content encoding (read: compression) defined by the stream.
The idea here is that some servers may serve data from a combination
of caches and dynamic resolution. Data coming from caches may be
pre-compressed. We want to facilitate servers being able to essentially
stream bytes from caches to the wire with minimal overhead. Being
able to mix and match with frames are compressed within a stream
enables these types of advanced server functionality.
This commit defines the new streams mechanism. Basic code for
supporting streams in frames has been added. But that code is
seriously lacking and doesn't fully conform to the defined protocol.
For example, we don't close any streams. And support for content
encoding within streams is not yet implemented. The change was
rather invasive and I didn't think it would be reasonable to implement
the entire feature in a single commit.
For the record, I would have loved to reuse an existing multiplexing
protocol to build the new wire protocol on top of. However, I couldn't
find a protocol that offers the performance and scaling characteristics
that I desired. Namely, it should support multiple compression
contexts to facilitate scaling out to multiple CPU cores and
compression contexts should be able to live longer than single RPC
requests. HTTP/2 *almost* fits the bill. But the semantics of HTTP
message exchange state that streams can only live for a single
request-response. We /could/ tunnel on top of HTTP/2 streams and
frames with HEADER and DATA frames. But there's no guarantee that
HTTP/2 libraries and proxies would allow us to use HTTP/2 streams
and frames without the HTTP message exchange semantics defined in
RFC 7540 Section 8. Other RPC protocols like gRPC tunnel are built
on top of HTTP/2 and thus preserve its semantics of stream per
RPC invocation. Even QUIC does this. We could attempt to invent a
higher-level stream that spans HTTP/2 streams. But this would be
violating HTTP/2 because there is no guarantee that HTTP/2 streams
are routed to the same server. The best we can do - which is what
this protocol does - is shoehorn all request and response data into
a single HTTP message and create streams within. At that point, we've
defined a Content-Type in HTTP parlance. It just so happens our
media type can also work as a standalone, stream-based protocol,
without leaning on HTTP or similar protocol.
Differential Revision: https://phab.mercurial-scm.org/D2907
author | Gregory Szorc <gregory.szorc@gmail.com> |
---|---|
date | Mon, 26 Mar 2018 11:00:16 -0700 |
parents | aaabd709df72 |
children | b0041036214e |
line wrap: on
line diff
--- a/mercurial/help/internals/wireprotocol.txt Mon Mar 26 13:57:22 2018 -0700 +++ b/mercurial/help/internals/wireprotocol.txt Mon Mar 26 11:00:16 2018 -0700 @@ -489,28 +489,37 @@ ordered sends and receives is required. That is, each peer has one pipe for sending data and another for receiving. +All data is read and written in atomic units called *frames*. These +are conceptually similar to TCP packets. Higher-level functionality +is built on the exchange and processing of frames. + +All frames are associated with a *stream*. A *stream* provides a +unidirectional grouping of frames. Streams facilitate two goals: +content encoding and parallelism. There is a dedicated section on +streams below. + The protocol is request-response based: the client issues requests to the server, which issues replies to those requests. Server-initiated messaging is not currently supported, but this specification carves out room to implement it. -All data is read and written in atomic units called *frames*. These -are conceptually similar to TCP packets. Higher-level functionality -is built on the exchange and processing of frames. - All frames are associated with a numbered request. Frames can thus be logically grouped by their request ID. -Frames begin with a 6 octet header followed by a variable length +Frames begin with an 8 octet header followed by a variable length payload:: - +-----------------------------------------------+ - | Length (24) | - +---------------------------------+-------------+ - | Request ID (16) | - +----------+-----------+----------+ - | Type (4) | Flags (4) | - +==========+===========+========================================| + +------------------------------------------------+ + | Length (24) | + +--------------------------------+---------------+ + | Request ID (16) | Stream ID (8) | + +------------------+-------------+---------------+ + | Stream Flags (8) | + +-----------+------+ + | Type (4) | + +-----------+ + | Flags (4) | + +===========+===================================================| | Frame Payload (0...) ... +---------------------------------------------------------------+ @@ -518,7 +527,9 @@ little endian integer. Values larger than 65535 MUST NOT be used unless given permission by the server as part of the negotiated capabilities during the handshake. The frame header is not part of the advertised -frame length. +frame length. The payload length is the over-the-wire length. If there +is content encoding applied to the payload as part of the frame's stream, +the length is the output of that content encoding, not the input. The 16-bit ``Request ID`` field denotes the integer request identifier, stored as an unsigned little endian integer. Odd numbered requests are @@ -529,7 +540,16 @@ start ordering request identifiers at ``1`` and ``0``, increment by ``2``, and wrap around if all available numbers have been exhausted. -The 4-bit ``Type`` field denotes the type of message being sent. +The 8-bit ``Stream ID`` field denotes the stream that the frame is +associated with. Frames belonging to a stream may have content +encoding applied and the receiver may need to decode the raw frame +payload to obtain the original data. Odd numbered IDs are +client-initiated. Even numbered IDs are server-initiated. + +The 8-bit ``Stream Flags`` field defines stream processing semantics. +See the section on streams below. + +The 4-bit ``Type`` field denotes the type of frame being sent. The 4-bit ``Flags`` field defines special, per-type attributes for the frame. @@ -720,6 +740,126 @@ The last atom in the frame SHOULD end with a newline (``\n``). If it doesn't, clients MAY add a newline to facilitate immediate printing. +Stream Encoding Settings (``0x08``) +----------------------------------- + +This frame type holds information defining the content encoding +settings for a *stream*. + +This frame type is likely consumed by the protocol layer and is not +passed on to applications. + +This frame type MUST ONLY occur on frames having the *Beginning of Stream* +``Stream Flag`` set. + +The payload of this frame defines what content encoding has (possibly) +been applied to the payloads of subsequent frames in this stream. + +The payload begins with an 8-bit integer defining the length of the +encoding *profile*, followed by the string name of that profile, which +must be an ASCII string. All bytes that follow can be used by that +profile for supplemental settings definitions. See the section below +on defined encoding profiles. + +Stream States and Flags +----------------------- + +Streams can be in two states: *open* and *closed*. An *open* stream +is active and frames attached to that stream could arrive at any time. +A *closed* stream is not active. If a frame attached to a *closed* +stream arrives, that frame MUST have an appropriate stream flag +set indicating beginning of stream. All streams are in the *closed* +state by default. + +The ``Stream Flags`` field denotes a set of bit flags for defining +the relationship of this frame within a stream. The following flags +are defined: + +0x01 + Beginning of stream. The first frame in the stream MUST set this + flag. When received, the ``Stream ID`` this frame is attached to + becomes ``open``. + +0x02 + End of stream. The last frame in a stream MUST set this flag. When + received, the ``Stream ID`` this frame is attached to becomes + ``closed``. Any content encoding context associated with this stream + can be destroyed after processing the payload of this frame. + +0x04 + Apply content encoding. When set, any content encoding settings + defined by the stream should be applied when attempting to read + the frame. When not set, the frame payload isn't encoded. + +Streams +------- + +Streams - along with ``Request IDs`` - facilitate grouping of frames. +But the purpose of each is quite different and the groupings they +constitute are independent. + +A ``Request ID`` is essentially a tag. It tells you which logical +request a frame is associated with. + +A *stream* is a sequence of frames grouped for the express purpose +of applying a stateful encoding or for denoting sub-groups of frames. + +Unlike ``Request ID``s which span the request and response, a stream +is unidirectional and stream IDs are independent from client to +server. + +There is no strict hierarchical relationship between ``Request IDs`` +and *streams*. A stream can contain frames having multiple +``Request IDs``. Frames belonging to the same ``Request ID`` can +span multiple streams. + +One goal of streams is to facilitate content encoding. A stream can +define an encoding to be applied to frame payloads. For example, the +payload transmitted over the wire may contain output from a +zstandard compression operation and the receiving end may decompress +that payload to obtain the original data. + +The other goal of streams is to facilitate concurrent execution. For +example, a server could spawn 4 threads to service a request that can +be easily parallelized. Each of those 4 threads could write into its +own stream. Those streams could then in turn be delivered to 4 threads +on the receiving end, with each thread consuming its stream in near +isolation. The *main* thread on both ends merely does I/O and +encodes/decodes frame headers: the bulk of the work is done by worker +threads. + +In addition, since content encoding is defined per stream, each +*worker thread* could perform potentially CPU bound work concurrently +with other threads. This approach of applying encoding at the +sub-protocol / stream level eliminates a potential resource constraint +on the protocol stream as a whole (it is common for the throughput of +a compression engine to be smaller than the throughput of a network). + +Having multiple streams - each with their own encoding settings - also +facilitates the use of advanced data compression techniques. For +example, a transmitter could see that it is generating data faster +and slower than the receiving end is consuming it and adjust its +compression settings to trade CPU for compression ratio accordingly. + +While streams can define a content encoding, not all frames within +that stream must use that content encoding. This can be useful when +data is being served from caches and being derived dynamically. A +cache could pre-compressed data so the server doesn't have to +recompress it. The ability to pick and choose which frames are +compressed allows servers to easily send data to the wire without +involving potentially expensive encoding overhead. + +Content Encoding Profiles +------------------------- + +Streams can have named content encoding *profiles* associated with +them. A profile defines a shared understanding of content encoding +settings and behavior. + +The following profiles are defined: + +TBD + Issuing Commands ----------------