Boost::ASIO low-level redis client (connector), github gitee

• header only
• zero-copy (currently only for received replies from Redis)
• low-level controls (i.e. you can cancel, or do manual DNS-resolving before a connection)
• unix domain sockets support
• works on linux (clang, gcc) and windows (msvc)
• synchronous & asynchronous interface
• inspired by beast
• requirements: boost v1.69 minimum

• [feature, possible breakage] add BOOST_ASIO_NO_DEPRECATED definition for better support boost v1.74 and modernize boost API usage

• [bugfix] avoid access to protected destructor (c++17 compatibility)

• [bugfix] critical bug in protcol serialization on empty values

• relaxed c++ compiler requirements: c++11 can be used instead of c++14

• minor parsing speed improvements (upto 10% in synthetic tests)
• fix compilation issues on boost::asio 1.70
• make it possible to use DynamicBuffer_v2 (dynamic_string_buffer, dynamic_vector_buffer) from boost::asio 1.70 in addition to streambuf. DynamicBuffer_v1 was actually never supported by bredis
• [API breakage] boos::asio::use_future cannot be used with bredis and boost::asio prior v1.70 (see issue). If you need use_future then either upgrade boost::asio or use previous bredis version.

• the parsing_policy::drop_result was documented and made applicable in client code
• updated preformance results
• fixed compliation warnings (-Wall -Wextra -pedantic -Werror)
• added shortcut header include/bredis.hpp
• added redis-streams usage example
• added multi-thread example

• fixed level 4 warning in MSVC
• fixed compilation issues on archlinux
• improved documentation (numerous typos etc.)

• [bugfix] removed unneeded tx_buff.commit() on async_write which corrupted buffer

• improved protocol parser (no memory allocations during input stream validity check)
• more detailed information in protocol_error_t
• added async incr speed test example
• [small API breakage] positive_parse_result_t was enriched with parcing policy; now instead of positive_parse_result_t<Iterator> should be written:

using Policy = r::parsing_policy::keep_result;
using result_t = r::parse_result_mapper_t<Iterator, Policy>;

• [small API breakage] protocol_error_t instead of std::string what member now contains boost::system::error_code code

• added windows support
• added coroutines & futures support
• generalised (templated) buffer support
• changed return type: instead of result of parsing just result markers are returned, extraction of result can be done as separate step
• dropped queing support (queuing policy should be implemented at more higher levels)
• dropped subscription support (can be implemented at higher levels)
• dropped internal buffers (can be implemented at higher levels)
• dropped explicit cancellation (socket reference can be passed to connector, and cancellation can be done on the socket object outside of the connector)

• initial version

Results achieved with examples/speed_test_async_multi.cpp for 1 thread, Intel Core i7-8550U, void-linux, gcc 8.3.0

These results are not completely fair, because of the usage of different semantics in the APIs; however they are still interesting, as they are using different underlying event libraries (Boost::ASIO vs libev) as well as redis protocol parsing libraries (written from scratch vs hiredis)

(*) bredis with drop_result policy, i.e. replies from redis server are scanned only for formal correctness and never delivered to the caller.

The general idea is that the result of trying to parse a redis reply can be either: not enough data, protocol error (in an extreme case) or some positive parse result. The last one is just markers of the result, which is actually stored in the receive buffer (i.e. outside of markers, and outside of the bredis-connection).

The further work with markers depends on your needs: it is possible to either scan the result for the expected results (e.g. for a PONG reply on a PING command, or for OK/QUEUED replies on MULTI/EXEC commands) or to extract the results (the common redis types: nil, string, error, int or a (recursive) array of them).

When the data in the receive buffer is no longer required, it should be consumed.

Scan example:

#include "bredis/MarkerHelpers.hpp"
...
namespace r = bredis;
...
using Buffer = boost::asio::streambuf;
...
Buffer rx_buff;
auto result_markers = c.read(rx_buff);
/* check for the response */
auto eq_pong = r::marker_helpers::equality<Iterator>("PONG");
/* print true or false */
std::cout << boost::apply_visitor(eq_pong, result_markers.result) << "\n";
/* consume the buffers, after finishing work with the markers */
rx_buff.consume(result_markers.consumed);

For extraction of results it is possible to use either one of the shipped extractors or to write a custom one. Shipped extractors detach (copy / convert) the extraction results from the receive buffer.

#include "bredis/Extract.hpp"
...
auto result_markers = c.read(rx_buff);
auto extract = boost::apply_visitor(r::extractor<Iterator>(), result_markers.result);
/* safe to consume buffers now */
rx_buff.consume(result_markers.consumed);
/* we know what the type is, safe to unpack to string */
auto &reply_str = boost::get<r::extracts::string_t>(extract);
/* print "PONG" */
std::cout << reply_str.str << "\n";

Custom extractors (visitors) might be useful for performance-sensitive cases, e.g. when JSON is re-constructed in-place by using string reply markers without re-allocating the whole JSON-string reply.

The underlying reason for the decision to retrieve the final results in two steps (get markers and then scan/extract results) is that the receive buffer might be scattered (fragmented). In such cases scan and extraction can be performed without gathering receive buffers (i.e. without flattening / linearizing it) if they are separate steps.

In other words, markers have reference semantics (they refer to memory regions in the buffer, but do not own it), while extracted results have value semantics (they take ownership).

#include "bredis/Connection.hpp"
#include "bredis/MarkerHelpers.hpp"

#include <boost/variant.hpp>
...
namespace r = bredis;
namespace asio = boost::asio;
...
/* define used types */
using socket_t = asio::ip::tcp::socket;
using Buffer = boost::asio::streambuf;
using Iterator = typename r::to_iterator<Buffer>::iterator_t;
...
/* establishing connection to redis is outside of bredis */
asio::ip::tcp::endpoint end_point(
    asio::ip::make_address("127.0.0.1"), port);
socket_t socket(io_service, end_point.protocol());
socket.connect(end_point);

/* wrap socket to bredis connection */
r::Connection<socket_t> c(std::move(socket));

/* synchronously write command */
c.write("ping");

/* buffer is allocated outside of bredis connection*/
Buffer rx_buff;
/* get the result markers */
auto result_markers = c.read(rx_buff);
/* check for the response */
auto eq_pong = r::marker_helpers::equality<Iterator>("PONG");
/* print true */
std::cout << boost::apply_visitor(eq_pong, result_markers.result) << "\n";
/* consume the buffers, after finishing work with the markers */
rx_buff.consume(result_markers.consumed);

In the ping example above the PONG reply string from redis is not (re)allocated, but directly scanned from the rx_buff using a result markers. This can be useful for performance-sensitive cases, e.g. when JSON is re-constructed in-place by using string reply markers without re-allocating the whole JSON-string reply.

In cases where you need to extract the reply (i.e. to detach it from rx_buff), the following can be done:

#include "bredis/Extract.hpp"
...
auto result_markers = c.read(rx_buff);
/* extract the results */
auto extract = boost::apply_visitor(r::extractor<Iterator>(), result_markers.result);
/* safe to consume buffers now */
rx_buff.consume(result_markers.consumed);
/* we know what the type is, safe to unpack to string */
auto &reply_str = boost::get<r::extracts::string_t>(extract);
/* print "PONG" */
std::cout << reply_str.str << "\n";

The examples above throw an exception in case of I/O or protocol errors. Another way to use the API is

boost::system::error_code ec;
c.write("ping", ec);
...
parse_result = c.read(rx_buff, ec);

in case you don't want the throw-exception behaviour.

#include "bredis/Connection.hpp"
#include "bredis/MarkerHelpers.hpp"
...
namespace r = bredis;
namespace asio = boost::asio;
namespace sys = boost::system;
...
using socket_t = asio::ip::tcp::socket;
using Buffer = boost::asio::streambuf;
using Iterator = typename r::to_iterator<Buffer>::iterator_t;
using Policy = r::parsing_policy::keep_result;
using result_t = r::parse_result_mapper_t<Iterator, Policy>;

...
/* establishing the connection to redis is outside of bredis */
asio::ip::tcp::endpoint end_point(
    asio::ip::make_address("127.0.0.1"), port);
socket_t socket(io_service, end_point.protocol());
socket.connect(end_point);
...
Buffer tx_buff, rx_buff;
c.async_write(
    tx_buff, r::single_command_t{"llen", "my-queue"}, [&](const sys::error_code &ec, std::size_t bytes_transferred) {
        /* tx_buff must be consumed when it is no longer needed */
        tx_buff.consume(bytes_transferred);
        c.async_read(rx_buff, [&](const sys::error_code &ec, result_t &&r) {
            /* see above how to work with the result */
            auto extract = boost::apply_visitor(r::extractor<Iterator>(), r.result);
            auto &queue_size = boost::get<r::extracts::int_t>(extract);
            std::cout << "queue size: " << queue_size << "\n";
            ...
            /* consume rx_buff when it is no longer needed */
            rx_buff.consume(r.consumed);
        });
    });

In the example above separate receive and transfer buffers are used. In theory you can use only one buffer for both operations, but you must ensure that it will not be used simultaneously for reading and writing, in other words you cannot use the pipelining redis feature.

The same as above, except the underlying socket type must be changed:

using socket_t = asio::local::stream_protocol::socket;

There is no specific support for subscriptions, but you can easily build your own like

r::single_command_t subscribe_cmd{"subscribe", "some-channel1", "some-channel2"};
c.write(subscribe_cmd);
Buffer rx_buff;

/* get the 2 confirmations, as we subscribed to 2 channels */
r::marker_helpers::check_subscription<Iterator> check_subscription{std::move(subscribe_cmd)};
for (auto i = 0; i < 2; ++i){
  auto result_markers = c.read(rx_buff);
  bool confirmed =  boost::apply_visitor(check_subscription, result_markers.result);
  if (!confirmed) {
    // do something!
    ...;
  }
  rx_buff.consume(result_markers.consumed);
}

while(true) {
  auto result_markers = c.read(rx_buff);
  auto extract = boost::apply_visitor(r::extractor<Iterator>(), result_markers.result);
  rx_buff.consume(result_markers.consumed);

  /* process the result  */
  auto& array_reply = boost::get<r::extracts::array_holder_t>(extract);
  auto* type_reply = boost::get<r::extracts::string_t>(&array_reply.elements[0]);
  if (type_reply && type_reply->str == "message") {
      auto& channel = boost::get<r::extracts::string_t>(array_reply.elements[1]);
      auto& payload = boost::get<r::extracts::string_t>(array_reply.elements[2]);
      ...
  }
}

See examples/synch-subscription.cpp for the full example.

These work similarly to the synchronous approach. However you have to provide a read callback initially and again after each successfull read

using Policy = r::parsing_policy::keep_result;
using ParseResult = r::parse_result_mapper_t<Iterator, Policy>;
using read_callback_t = std::function<void(const boost::system::error_code &error_code, ParseResult &&r)>;
using Extractor = r::extractor<Iterator>;
...
/* we can execute the subscription command synchronously, as it is easier */
c.command("subscribe", "channel-1", "channel-2");
...
Buffer rx_buff;
read_callback_t notification_callback = [&](const boost::system::error_code,
                                            ParseResult &&r) {
    auto extract = boost::apply_visitor(Extractor(), r.result);
    rx_buff.consume(r.consumed);
    /* process the result, see above */
    ...
    /* re-trigger new message processing */
    c.async_read(rx_buff, notification_callback);
};

/* initialise listening subscriptions */
c.async_read(rx_buff, notification_callback);

See examples/stream-parse.cpp for the full example.

There is no specific support for transactions in bredis, but you can easily build your own for your needs.

First, wrap your commands into a transaction:

r::command_container_t tx_commands = {
    r::single_command_t("MULTI"),
    r::single_command_t("INCR", "foo"),
    r::single_command_t("GET", "bar"),
    r::single_command_t("EXEC"),
};
r::command_wrapper_t cmd(tx_commands);
c.write(cmd);

Then, as above there were 4 redis commands, there we should receive 4 redis replies: OK, QUEUED, QUEUED followed by the array of results of the execution of the commands in the transaction (i.e. results for INCR and GET above)

Buffer rx_buff;
c.async_read(rx_buff, [&](const sys::error_code &ec, result_t&& r){
    auto &replies = boost::get<r::markers::array_holder_t<Iterator>>(r.result);
    /* scan stream for OK, QUEUED, QUEUED */
    ...
    assert(replies.elements.size() == 4);
    auto eq_OK = r::marker_helpers::equality<Iterator>("OK");
    auto eq_QUEUED = r::marker_helpers::equality<Iterator>("QUEUED");
    assert(boost::apply_visitor(eq_OK, replies.elements[0]));
    assert(boost::apply_visitor(eq_QUEUED, replies.elements[1]));
    assert(boost::apply_visitor(eq_QUEUED, replies.elements[2]));

    /* get tx replies */
    auto &tx_replies = boost::get<r::markers::array_holder_t<Iterator>>(replies.elements[3]);
    ...;
    rx_buff.consume(r.consumed);
},
4); /* pay attention here */

Done in a similiar way as in Boost::ASIO (special thanks to Vinnie Falko for the suggestion)

#include <boost/asio/use_future.hpp>
...
Buffer rx_buff, tx_buff;
auto f_tx_consumed = c.async_write(tx_buff, "ping", asio::use_future);
auto f_result_markers = c.async_read(rx_buff, asio::use_future);
...
tx_buff.consume(f_tx_consumed.get());
auto result_markers = f_result_markers.get();
/* scan/extract result, and consume rx_buff as usual */

#include <boost/asio/spawn.hpp>
Buffer rx_buff, tx_buff;

boost::asio::spawn(
    io_service, [&](boost::asio::yield_context yield) mutable {
        boost::system::error_code error_code;
        auto consumed = c.async_write(tx_buff, "ping", yield[error_code]);
        tx_buff.consume(consumed);
        ...
        auto parse_result = c.async_read(rx_buff, yield[error_code], 1);
        /* scan/extract result */
        rx_buff.consume(parse_result.consumed);
    });

There is no specific support for streams (appeared in redis 5.0) in bredis, they are just usual XADD, XRANGE etc. commands and corresponding replies.

...
Buffer rx_buff;
c.write(r::single_command_t{ "XADD", "mystream", "*", "cpu-temp", "23.4", "load", "2.3" });
auto parse_result1 = c.read(rx_buff);
auto extract1 = boost::apply_visitor(Extractor(), parse_result1.result);
auto id1 = boost::get<r::extracts::string_t>(extract1);

c.write(r::single_command_t{ "XADD", "mystream", "*", "cpu-temp", "23.2", "load", "2.1" });
auto parse_result2 = c.read(rx_buff);
auto extract2 = boost::apply_visitor(Extractor(), parse_result2.result);
auto id2 = boost::get<r::extracts::string_t>(extract2);
rx_buff.consume(parse_result2.consumed);

c.write(r::single_command_t{ "XRANGE" , "mystream",  id1.str, id2.str});
auto parse_result3 = c.read(rx_buff);
auto extract3 = boost::apply_visitor(Extractor(), parse_result3.result);
rx_buff.consume(parse_result3.consumed);

auto& outer_arr = boost::get<r::extracts::array_holder_t>(extract3);
auto& inner_arr1 = boost::get<r::extracts::array_holder_t>(outer_arr.elements[0]);
auto& inner_arr2 = boost::get<r::extracts::array_holder_t>(outer_arr.elements[1]);
...

See t/SocketWithLogging.hpp for an example. The main idea is quite simple: Instead of providing a real socket implementation supplied by Boost::ASIO, provide a wrapper (proxy) which will spy on the traffic before delegating it to/from a Boost::ASIO socket.

There is nothing specific to this in bredis. If you need low-level socket operations, instead of moving socket into bredis connection, you can simply move a reference to it and keep (own) the socket somewhere outside of the bredis connection.

using socket_t = asio::ip::tcp::socket;
using next_layer_t = socket_t &;
...
asio::ip::tcp::endpoint end_point(asio::ip::make_address("127.0.0.1"), port);
socket_t socket(io_service, end_point.protocol());
socket.connect(end_point);
r::Connection<next_layer_t> c(socket);
...
socket.cancel();

bredis itself is thread-agnostic, however the underlying socket (next_layer_t) and used buffers are usually not thread-safe. To handle that in multi-thead environment the access to those objects should be sequenced via asio::io_context::strand . See the examples/multi-threads-1.cpp.

The performance still can be boosted if it is known beforehand that the response from redis server is not needed at all. For example, the only possible response to PING command is PONG reply, usually there is no sense it validating that PONG reply, as soon as it is known, that redis-server alredy delivered us some reply (in practice it is PONG). Another example is SET command, when redis-server usually replies with OK.

With parsing_policy::drop_result the reply result is just verified with formal compliance to redis protocol, and then it is discarded.

It should be noted, that redis can reply back with error, which aslo correct reply, but the caller side isn't able to see it when parsing_policy::drop_result is applied. So, it should be used with care, when you know what your are doing. You have been warned.

It is safe, however, to mix different parsing policies on the same connection, i.e. write SET command and read it's reply with parsing_policy::drop_result and then write GET command and read it's reply with parsing_policy::keep_result. See the examples/speed_test_async_multi.cpp.

There's a convenience header include/bredis.hpp, doing #include "bredis.hpp" will include every header under include/bredis/ .

The underlying iterator type used for the dynamic buffer type (e.g. boost::asio::streambuf)

Header: include/bredis/Markers.hpp

Namespace: bredis::markers

boost::variant for the basic types in the redis protocol , i.e. the following marker types :

• nil_t<Iterator>
• int_t<Iterator>
• string_t<Iterator> (simple string and bulk strings)
• error_t<Iterator>
• array_holder_t<Iterator>

The basic type is string_t<Iterator>, which contains from and to members (Iterator) to where the string is held. String does not contain the special redis-protocol symbols or any other metadata, i.e. it can be used to extract/flatten the whole string.

nil_t<Iterator>, int_t<Iterator>, error_t<Iterator> just have a string member to point to the underlying string in the redis protocol.

array_holder_t is recursive wrapper for the redis_result_t<Iterator>, it contains a elements member of std::array of redis_result_t<Iterator> type.

Header: include/bredis/Result.hpp

Namespace: bredis

Represents the results of a parse attempt. It is a boost::variant of the following types:

• not_enough_data_t
• protocol_error_t
• positive_parse_result_t<Iterator, Policy>

not_enough_data_t is a empty struct. It means that buffer just does not contain enough information to completely parse it.

protocol_error_t has a boost::system::error_code code member. It describes the error in the protocol (e.g. when the type in the stream is specified as an integer, but it cannot be converted to an integer). This error should never occur in production code, meaning that no (logical) errors are expected in the redis-server nor in the bredis parser. The error might occur if the buffer is corrupted.

Policy (namespace bredis::parsing_policy) specifies what to do with the result: Either drop it (bredis::parsing_policy::drop_result) or keep it (bredis::parsing_policy::keep_result). The helper parse_result_mapper_t<Iterator, Policy> helps to get the proper positive_parse_result_t<Iterator, Policy> type.

positive_parse_result_t<Iterator, Policy> contains members:

• markers::redis_result_t<Iterator> result - the result of mark-up buffer; can be used either for scanning for particular results or for extraction of results. Valid only for keep_result policy.
• size_t consumed - how many bytes of receive buffer must be consumed after using the result field.

Header: include/bredis/MarkerHelpers.hpp

Namespace: bredis::marker_helpers

Apply this boost::static_visitor<std::string>s to stringize the result (can be useful for debugging).

Apply this boost::static_visitor<bool> to find a string in the parsed results (the markup can point to integer types, but as it is transferred as a string anyway, it still can be founded as string too).

Constructor: equality<Iterator>(std::string str)

This boost::static_visitor<bool> helper is used to check whether the redis reply confirms one of the requested channels. Hence, the constructor is check_subscription(single_command_t).

Usually, the redis subscription reply is in the form:

[array] {
    [string] "subcribe",
    [string] channel_name,
    [int] reference
}

So it checks that:

1. The redis reply is a 3-element array
2. The 1st reply element is a string, and it case-insensitively matches the command, i.e. it is assumed, that command will be subscribe or psubscribe depending on the original command
3. That the 3rd reply element is a reference, and it is present among the command arguments.

It is possible to reuse the same check_subscription<Iterator> on multiple redis replies to a single subsription command for multiple channels.

Example:

bredis::single_command_t subscribe_cmd{
    "subscribe", "channel-1", "channel-2"
};
...
// write the command, so the subscribe_cmd
// will be no longer be required
...;
bredis::marker_helpers::check_subscription<Iterator>
    check_subscription{std::move(subscribe_cmd)};
...;
// get the 1st reply
auto parse_result = ...;
bool channel_1_ok = boost::apply_visitor(check_subscription, parse_result.result);
...;
// get the 2nd reply
parse_result = ...;
bool channel_2_ok = boost::apply_visitor(check_subscription, parse_result.result);

Header: include/bredis/Command.hpp

Namespace: bredis

boost::variant for the basic commands:

• single_command_t
• command_container_t

single_command_t represents a single redis command with all its arguments, e.g.:

// compile-time version
r::single_command_t cmd_ping {"ping"};
r::single_command_t cmd_get {"get", "queu-name"};
...
// or runtime-version
std::vector<std::string> subscription_items { "subscribe", "channel-a", "channel-b"};
r::single_command_t cmd_subscribe {
    subscription_items.cbegin(),
    subscription_items.cend()
};

The arguments must be conversible to boost::string_ref.

command_container_t is a std::vector of single_command_t. It is useful for transactions or bulk message creation.

Header: include/bredis/Connection.hpp

Namespace: bredis

A thin wrapper around NextLayer; represents a connection to redis. NextLayer can be either asio::ip::tcp::socket or asio::ip::tcp::socket& or a custom wrapper, which follows the specification of asio::ip::tcp::socket.

The constructor template <typename... Args> Connection(Args &&... args) is used for the construction of NextLayer (stream interface).

Stream interface accessors:

• NextLayer &next_layer()
• const NextLayer &next_layer() const

return the underlying stream object.

Performs a synchonous write of a redis command:

• void write(const command_wrapper_t &command)
• void write(const command_wrapper_t &command, boost::system::error_code &ec)

Performs a synchronous read of a redis result until the buffer is parsed or some error (protocol or I/O) occurs:

• template <typename DynamicBuffer> positive_parse_result_t<Iterator, Policy = bredis::parsing_policy::keep_result> read(DynamicBuffer &rx_buff)
• template <typename DynamicBuffer> positive_parse_result_t<Iterator, Policy = bredis::parsing_policy::keep_result> read(DynamicBuffer &rx_buff, boost::system::error_code &ec);

DynamicBuffer must conform to the boost::asio::streambuf interface.

The WriteCallback template should be a callable object with the signature:

void (const boost::system::error_code&, std::size_t bytes_transferred)

The asynchronous write has the following signature:

void-or-deduced
async_write(DynamicBuffer &tx_buff, const command_wrapper_t &command,
                WriteCallback write_callback)

It writes the redis command (or commands) into a transfer buffer, sends them to the next_layer stream and invokes write_callback after completion.

tx_buff must consume bytes_transferred upon write_callback invocation.

The client must guarantee that async_write is not invoked until the previous invocation is finished.

ReadCallback template should be a callable object with the signature:

void(boost::system::error_code, r::positive_parse_result_t<Iterator, Policy = bredis::parsing_policy::keep_result>&& result)

The asynchronous read has the following signature:

void-or-deduced
async_read(DynamicBuffer &rx_buff, ReadCallback read_callback,
               std::size_t replies_count = 1, Policy = bredis::parsing_policy::keep_result{});

It reads replies_count replies from the next_layer stream, which will be stored in rx_buff, or until an error (I/O or protocol) is encountered; then read_callback will be invoked.

If replies_count is greater than 1, the result type will always be bredis::array_wrapper_t; if the replies_count is 1 then the result type depends on redis answer type.

On read_callback invocation with a successful parse result it is expected, that rx_buff will consume the amount of bytes specified in the result.

The client must guarantee that async_read is not invoked until the previous invocation is finished. If you invoke async_read from read_callback don't forget to consume rx_buff first, otherwise it leads to subtle bugs.

MIT

• Derek Colley
• Stefan Hacker
• nkochakian
• Yuval Hager
• Vinnie Falco
• Stephen Coleman
• maxtorm miximtor
• Ronny Nowak
• Stephen Chisholm
• amensel

• https://github.com/Cylix/cpp_redis
• https://github.com/hmartiro/redox
• https://github.com/nekipelov/redisclient