redispipe

RedisPipe

RedisPipe – is a client for redis that uses "implicit pipelining" for highest performance.

• Highlights
• Introduction
• Performance
• Limitations
• Installation
• Usage
• Contributing
• License

Highlights

• scalable: the more throughput you try to get, the more efficient it is.
• cares about redis: redis needs less CPU to perform same throughput.
• thread-safe: no need to lock around connection, no need to "return to pool", etc.
• pipelining is implicit.
• transactions are supported (but without WATCH).
• hook for custom logging.
• hook for request timing reporting.

BenchmarkParallelGetSet/radixv2-8        1000000     9245 ns/op   1268 B/op   32 allocs/op
BenchmarkParallelGetSet/redigo-8         1000000     6886 ns/op    399 B/op   13 allocs/op
BenchmarkParallelGetSet/redispipe-8      5000000     1636 ns/op    219 B/op   12 allocs/op

Introduction

https://redis.io/topics/pipelining

Pipelining improves the maximum throughput that redis can serve, and reduces CPU usage both on redis server and on the client side. Mostly it comes from saving system CPU consumption.

But it is not always possible to use pipelining explicitly: usually there are dozens of concurrent goroutines, each sends just one request at a time. To handle the usual workload, pipelining has to be implicit.

"Implicit pipelining" is used in many drivers for other languages:

• https://github.com/NodeRedis/node_redis , https://github.com/h0x91b/redis-fast-driver , and probably, other nodejs clients,
• https://github.com/andrew-bn/RedisBoost - C# connector,
• some C/C++ clients,
• all Dart clients ,
• some Erlang and Elixir clients,
• https://github.com/informatikr/hedis - Haskel client.
• http://aredis.sourceforge.net/ - Java client explicitly made for transparent pipelining,
• https://github.com/lettuce-io/lettuce-core - Java client capable for transparent pipelining,
• https://github.com/aio-libs/aioredis - Python's async connector, and some of other async python clients
• Ruby's EventMachine related connectors,
• etc

But there was no such connector for Golang. All known Golang redis connectors use a connection-per-request model with a connection pool, and provide only explicit pipelining.

This connector was created as implicitly pipelined from the ground up to achieve maximum performance in a highly concurrent environment. It writes all requests to single connection to redis, and continuously reads answers from another goroutine.

Note that it trades a bit of latency for throughput, and therefore could be not optimal for low-concurrent low-request-per-second usage. Write loop latency is configurable as WritePause parameter in connection options, and could be disabled at all, or increased to higher values (150µs is the value used in production, 50µs is default value, -1 disables write pause). Implicit runtime latency for switching goroutines still remains, however, and could not be removed.

Performance

Tests were performed on localhost Xeon(R) CPU E3-1245 v6 @ 3.70GHz (4 cores, 8 HT)

Single redis

go test -count 1 -run FooBar -bench . -benchmem -benchtime 5s ./redisconn
goos: linux
goarch: amd64
pkg: github.com/joomcode/redispipe/redisconn
BenchmarkSerialGetSet/radixv2-8             200000    32257 ns/op    256 B/op    11 allocs/op
BenchmarkSerialGetSet/redigo-8              200000    31785 ns/op     86 B/op     5 allocs/op
BenchmarkSerialGetSet/redispipe-8            30000   266490 ns/op    168 B/op     8 allocs/op
BenchmarkSerialGetSet/redispipe_pause0-8    200000    44396 ns/op    168 B/op     8 allocs/op
BenchmarkParallelGetSet/radixv2-8           500000    12756 ns/op    260 B/op    11 allocs/op
BenchmarkParallelGetSet/redigo-8           1000000    12486 ns/op    123 B/op     6 allocs/op
BenchmarkParallelGetSet/redispipe-8        5000000     1435 ns/op    168 B/op     8 allocs/op

You can see a couple of things:

• first, redispipe has highest performance in Parallel benchmarks,
• second, redispipe has lower performance for single-threaded cases.

That is true: redispipe trades latency for throughput. Every single request has additional latency for hidden batching in a connector. But thanks to batching, more requests can be sent to redis and answered by redis in an interval of time.

SerialGetSet/redispipe_pause0 shows single-threaded results with disabled additional latency for "batching" (WritePause: -1). This way redispipe is quite close to other connectors in performance, though there is still small overhead of internal design. But I would not recommend disable batching (unless your use case is single threaded), because it increases CPU usage under highly concurrent load both on client and on redis-server.

Parallel benchmark for single redis has Redis CPU usage as a bottleneck for both radix.v2 and redigo (ie Redis eats whole CPU core). But with redispipe, Redis server consumes only 75% of a core despite the fact that it could serve 8 times more requests. It clearly shows how usage of implicitly pipelined connector helps to get much more RPS from single Redis server.

Cluster

go test -count 1 -tags=debugredis -run FooBar -bench . -benchmem -benchtime 5s ./rediscluster
goos: linux
goarch: amd64
pkg: github.com/joomcode/redispipe/rediscluster
BenchmarkSerialGetSet/radixv2-8           200000    53585 ns/op   1007 B/op   31 allocs/op
BenchmarkSerialGetSet/redigo-8            200000    40705 ns/op    246 B/op   12 allocs/op
BenchmarkSerialGetSet/redispipe-8          30000   279838 ns/op    220 B/op   12 allocs/op
BenchmarkSerialGetSet/redispipe_pause0-8  200000    56356 ns/op    216 B/op   12 allocs/op
BenchmarkParallelGetSet/radixv2-8        1000000     9245 ns/op   1268 B/op   32 allocs/op
BenchmarkParallelGetSet/redigo-8         1000000     6886 ns/op    399 B/op   13 allocs/op
BenchmarkParallelGetSet/redispipe-8      5000000     1636 ns/op    219 B/op   12 allocs/op

With cluster configuration, internal cluster meta-info management adds additional overhead inside of the Go process. And redispipe/rediscluster attempts to provide almost lockless cluster info handling on the way of request execution.

While redigo is almost as fast in Parallel tests, it also happens to be limited by Redis's CPU usage (three redis processes eats whole 3 cpu cores). It uses a huge number of connections, and it is not trivial to recognize non-default setting that should be set to achieve this result (both KeepAlive and AliveTime should be set as high as 128). ( github.com/chasex/redis-go-cluster is used).

Each Redis uses less than 60% CPU core when redispipe is used, despite serving more requests.

Practice

In practice, performance gain is lesser, because your application does other useful work aside from sending requests to Redis. But gain is still noticeable. At our setup, we have around 10-15% less CPU usage on Redis (ie 50%CPU->35%CPU), and 5-10% improvement on the client side. WritePause is usually set to higher value (150µs) than default.

Limitations

• by default, it is not allowed to send blocking calls, because it will block the whole pipeline: BLPOP, BRPOP, BRPOPLPUSH, BZPOPMIN, BZPOPMAX, XREAD, XREADGROUP, SAVE. However, you could set ScriptMode: true option to enable these commands. ScriptMode: true also turns default WritePause to -1 (meaning it practically disables forced batching).
• WATCH is also forbidden by default: it is useless and even harmful when concurrent goroutines use the same connection. It is also allowed with ScriptMode: true, but you should be sure you use connection only from a single goroutine.
• SUBSCRIBE and PSUBSCRIBE commands are forbidden. They switch connection work mode to a completely different mode of communication, therefore it could not be combined with regular commands. This connector doesn't implement subscribing mode.

Installation

• Single connection: go get github.com/joomcode/redispipe/redisconn
• Cluster connection: go get github.com/joomcode/redispipe/rediscluster

Usage

Both redisconn.Connect and rediscluster.NewCluster creates implementations of redis.Sender. redis.Sender provides asynchronous api for sending request/requests/transactions. That api accepts redis.Future interface implementations as an argument and fullfills it asynchronously. Usually you don't need to provide your own redis.Future implementation, but rather use synchronous wrappers.

To use convenient synchronous api, one should wrap "sender" with one of wrappers:

• redis.Sync{sender} - provides simple synchronouse api
• redis.SyncCtx{sender} - provides same api, but all methods accepts context.Context, and methods returns immediately if that context is closed.
• redis.ChanFutured{sender} - provides api with future through channel closing.

Types accepted as command arguments: nil, []byte, string, int (and all other integer types), float64, float32, bool. All arguments are converted to redis bulk strings as usual (ie string and bytes - as is; numbers - in decimal notation). bool converted as "0/1", nil converted to empty string.

In difference to other redis packages, no custom types are used for request results. Results are de-serialized into plain go types and are returned as interface{}:

IO, connection, and other errors are not returned separately, but as result (and has same *errorx.Error underlying type).

package redispipe_test

import (
	"context"
	"fmt"
	"log"

	"github.com/joomcode/redispipe/redis"
	"github.com/joomcode/redispipe/rediscluster"
	"github.com/joomcode/redispipe/redisconn"
)

const databaseno = 0
const password = ""

var myhandle interface{} = nil

func Example_usage() {
	ctx := context.Background()
	cluster := false

	SingleRedis := func(ctx context.Context) (redis.Sender, error) {
		opts := redisconn.Opts{
			DB:       databaseno,
			Password: password,
			Logger:   redisconn.NoopLogger{}, // shut up logging. Could be your custom implementation.
			Handle:   myhandle,               // custom data, useful for custom logging
			// Other parameters (usually, no need to change)
			// IOTimeout, DialTimeout, ReconnectTimeout, TCPKeepAlive, Concurrency, WritePause, Async
		}
		conn, err := redisconn.Connect(ctx, "127.0.0.1:6379", opts)
		return conn, err
	}

	ClusterRedis := func(ctx context.Context) (redis.Sender, error) {
		opts := rediscluster.Opts{
			HostOpts: redisconn.Opts{
				// No DB
				Password: password,
				// Usually, no need for special logger
			},
			Name:   "mycluster",               // name of a cluster
			Logger: rediscluster.NoopLogger{}, // shut up logging. Could be your custom implementation.
			Handle: myhandle,                  // custom data, useful for custom logging
			// Other parameters (usually, no need to change):
			// ConnsPerHost, ConnHostPolicy, CheckInterval, MovedRetries, WaitToMigrate, RoundRobinSeed,
		}
		addresses := []string{"127.0.0.1:20001"} // one or more of cluster addresses
		cluster, err := rediscluster.NewCluster(ctx, addresses, opts)
		return cluster, err
	}

	var sender redis.Sender
	var err error
	if cluster {
		sender, err = ClusterRedis(ctx)
	} else {
		sender, err = SingleRedis(ctx)
	}
	if err != nil {
		log.Fatal(err)
	}
	defer sender.Close()

	sync := redis.SyncCtx{sender} // wrapper for synchronous api

	res := sync.Do(ctx, "SET", "key", "ho")
	if err := redis.AsError(res); err != nil {
		log.Fatal(err)
	}
	fmt.Printf("result: %q\n", res)

	res = sync.Do(ctx, "GET", "key")
	if err := redis.AsError(res); err != nil {
		log.Fatal(err)
	}
	fmt.Printf("result: %q\n", res)

	res = sync.Send(ctx, redis.Req("HMSET", "hashkey", "field1", "val1", "field2", "val2"))
	if err := redis.AsError(res); err != nil {
		log.Fatal(err)
	}

	res = sync.Send(ctx, redis.Req("HMGET", "hashkey", "field1", "field2", "field3"))
	if err := redis.AsError(res); err != nil {
		log.Fatal(err)
	}
	for i, v := range res.([]interface{}) {
		fmt.Printf("%d: %T %q\n", i, v, v)
	}

	res = sync.Send(ctx, redis.Req("HMGET", "key", "field1"))
	if err := redis.AsError(res); err != nil {
		if rerr := redis.AsErrorx(res); rerr != nil && rerr.IsOfType(redis.ErrResult) {
			fmt.Printf("expected error: %v\n", rerr)
		} else {
			fmt.Printf("unexpected error: %v\n", err)
		}
	} else {
		fmt.Printf("unexpected missed error\n")
	}

	results := sync.SendMany(ctx, []redis.Request{
		redis.Req("GET", "key"),
		redis.Req("HMGET", "hashkey", "field1", "field3"),
	})
	// results is []interface{}, each element is result for corresponding request
	for i, res := range results {
		fmt.Printf("result[%d]: %T %q\n", i, res, res)
	}

	results, err = sync.SendTransaction(ctx, []redis.Request{
		redis.Req("SET", "a{x}", "b"),
		redis.Req("SET", "b{x}", 0),
		redis.Req("INCRBY", "b{x}", 3),
	})
	if err != nil {
		log.Fatal(err)
	}
	for i, res := range results {
		fmt.Printf("tresult[%d]: %T %q\n", i, res, res)
	}

	// Output:
	// result: "OK"
	// result: "ho"
	// 0: []uint8 "val1"
	// 1: []uint8 "val2"
	// 2: <nil> %!q(<nil>)
	// expected error: WRONGTYPE Operation against a key holding the wrong kind of value (ErrResult {connection: *redisconn.Connection{addr: 127.0.0.1:6379}})
	// result[0]: []uint8 "ho"
	// result[1]: []interface {} ["val1" <nil>]
	// tresult[0]: string "OK"
	// tresult[1]: string "OK"
	// tresult[2]: int64 '\x03'
}

Contributing

• Ask questions in Issues
• Ask questions on StackOverflow.
• Report about bugs using github Issues,
• Request new features or report about intentions to implement feature using github Issues,
• Send pull requests to fix reported bugs or to implement discussed features.
• Be kind.
• Be lenient to our misunderstanding of your problem and our unwillingness to bloat library.

License

MIT License