polardbxengine/plugin/x/protocol/doc/mysqlx-protocol-implementat...

329 lines
8.0 KiB
Plaintext

/*
* Copyright (c) 2015, 2018, Oracle and/or its affiliates. All rights reserved.
*
* This program is free software; you can redistribute it and/or modify
* it under the terms of the GNU General Public License, version 2.0,
* as published by the Free Software Foundation.
*
* This program is also distributed with certain software (including
* but not limited to OpenSSL) that is licensed under separate terms,
* as designated in a particular file or component or in included license
* documentation. The authors of MySQL hereby grant you an additional
* permission to link the program and your derivative works with the
* separately licensed software that they have included with MySQL.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
* GNU General Public License, version 2.0, for more details.
*
* You should have received a copy of the GNU General Public License
* along with this program; if not, write to the Free Software
* Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
*/
/** @page mysqlx_protocol_implementation Implementation Notes
Topics in this section:
- @ref implementation_Client
- @ref implementation_Server
@ref implementation_Client "Client" and
@ref implementation_Server "Server" implementations of the
protocol should make use of the following:
- vectorized IO
- pipelining
to reduce the latency and CPU usage.
Client {#implementation_Client}
======
@par Out-of-Band Messages
The client should decode the messages it receives from the server in a
generic way and track the possible messages with a state-machine.
@code{py}
def getMessage(self, message):
## handle out-of-band message
msg = messageFactory(message.type).fromString(message.payload)
if message.type is Notification:
notification_queue.add(msg)
raise NoMessageError()
if message.type is Notice:
notice_queue.add(msg)
raise NoMessageError()
return msg
@endcode
@par Pipelining
The client may send several messages to the server without waiting for a
response for each message.
Instead of waiting for the response to a message like in:
@startuml "Client Pipeline"
Client -[#red]> Server: Sql::StmtExecute
...1 second later...
Server -[#red]-> Client: Sql::StmtExecuteOk
Client -[#blue]> Server: Sql::StmtExecute
...1 second later...
Server -[#blue]-> Client: Sql::StmtExecuteOk
Client -[#green]> Server: Sql::StmtExecute
...1 second later...
Server -[#green]-> Client: Sql::StmtExecuteOk
@enduml
the client can generate its messages and send it to the server without
waiting:
@startuml "Client Pipeline"
Client -[#red]> Server: Sql::StmtExecute
Client -[#blue]> Server: Sql::StmtExecute
Client -[#green]> Server: Sql::StmtExecute
...1 second later...
Server -[#red]-> Client: Sql::StmtExecuteOk
Server -[#blue]-> Client: Sql::StmtExecuteOk
Server -[#green]-> Client: Sql::StmtExecuteOk
@enduml
The client has to ensure that when pipeline messages that in case of an
error the following messages also error out correctly:
@startuml "Client Pipeline"
Client -[#red]> Server: Sql::StmtExecute
Client -[#blue]> Server: Sql::StmtExecute
Client -[#green]> Server: Sql::StmtExecute
...1 second later...
Server -[#red]-> Client: Sql::StmtExecuteOk
Server -[#blue]-> Client: Error
Server -[#green]-> Client: Error
@enduml
@par Vectored I/O
In network programming it is pretty common the to prefix the message
payload with the header:
- HTTP header + HTTP content
- a pipeline of messages
- message header + protobuf message
@code{py}
import struct
import socket
s = socket.create_connection(( "127.0.0.1", 33060))
msg_type = 1
msg_payload = "abc"
msg_header = struct.pack(">I", len(msg_payload)) +
struct.pack("B", msg_type)
## concat before send
s.send(msg_header + msg_payload)
## multiple syscalls
s.send(msg_header)
s.send(msg_payload)
## vectored I/O
s.sendmsg([ msg_header, msg_payload ])
@endcode
*concat before send* leads to pretty wasteful reallocations and copy
operations if the payload is huge.
*multiple syscalls* is pretty wasteful for small messages as a few bytes
only the whole machinery of copying data between user land and kernel
land has to be started.
*vectored io* combines the best of both approaches and sends multiple
buffers to the OS in one syscall and OS can optimize sending multiple
buffers in on TCP packet.
On Unix this is handled by ``writev(2)``, on Windows exists
``WSASend()``
@note
Any good buffered iostream implementation should already make use of
vectored I/O.
Known good implementation:
- Boost::ASIO
- GIO's GBufferedIOStream
@par Corking
Further control about how when to actual send data to the other endpoint
can be achieved with "corking":
- linux: ``TCP_CORK`` http://linux.die.net/man/7/tcp
- freebsd/macosx: ``TCP_NOPUSH``
https://www.freebsd.org/cgi/man.cgi?query=tcp&sektion=4&manpath=FreeBSD+9.0-RELEASE
They work in combination with ``TCP_NODELAY`` (aka Nagle's Algorithm).
- http://stackoverflow.com/questions/3761276/when-should-i-use-tcp-nodelay-and-when-tcp-cork?rq=1
Server {#implementation_Server}
======
@par Pipelining
The protocol is structured in a way that the messages can be decoded
completely without of knowing the state of the message sequence.
If data is available on the network, the server has to:
- read the message
- decode the message
- execute the message
Instead of a synchronous read-execution cycle:
@startuml "Server Pipeline"
participant Network
participant Reader
participant Executor
[-> Reader: message ready
Reader -> Network: receive
activate Reader
activate Network
Network --> Reader: data
deactivate Network
Reader -> Reader: decode(data)
Reader -> Executor: start_execute(msg)
deactivate Reader
activate Executor
Executor -> Executor: execute(msg)
Executor -> Executor: encode(response_msg)
[-> Reader: message ready
Executor -> Network: send(data)
activate Network
Network --> Executor: ok
deactivate Network
deactivate Executor
Reader -> Network: receive
activate Reader
activate Network
Network --> Reader: data
deactivate Network
Reader -> Reader: decode(data)
Reader -> Executor: start_execute(msg)
deactivate Reader
activate Executor
Executor -> Executor: execute(msg)
Executor -> Executor: encode(response_msg)
Executor -> Network: send(data)
activate Network
Network --> Executor: ok
deactivate Network
deactivate Executor
@enduml
the Reader and the Executor can be decoupled into separate threads:
@startuml "Separate Threads"
participant Network
participant Reader
box "Executor Thread"
participant ExecQueue
participant Executor
end box
[-> Reader: message ready
Executor -> ExecQueue: wait_for_msg
activate Executor
Reader -> Network: receive
activate Reader
activate Network
Network --> Reader: data
deactivate Network
Reader -> Reader: decode(data)
Reader -> ExecQueue: start_execute(msg)
ExecQueue --> Reader: ok
deactivate Reader
ExecQueue --> Executor: msg
Executor -> Executor: execute(msg)
Executor -> Executor: encode(response_msg)
[-> Reader: message ready
Reader -> Network: receive
activate Reader
activate Network
Network --> Reader: data
deactivate Network
Reader -> Reader: decode(data)
Executor -> Network: send(data)
activate Network
Network --> Executor: ok
deactivate Network
deactivate Executor
Reader -> ExecQueue: start_execute(msg)
Executor -> ExecQueue: wait_for_msg
activate Executor
ExecQueue --> Reader: ok
deactivate Reader
ExecQueue --> Executor: msg
Executor -> Executor: execute(msg)
Executor -> Executor: encode(response_msg)
Executor -> Network: send(data)
activate Network
Network --> Executor: ok
deactivate Network
deactivate Executor
@enduml
which allows to hide cost of decoding the message behind the execution
of the previous message.
The amount of messages that are prefetched this way should be
configurable to allow a trade-off between:
- resource usage
- parallelism
*/