redjet
redjet is a high-performance Go library for Redis. Its hallmark feature is a low-allocation, streaming API. See the benchmarks section for more details.
Unlike redigo and go-redis, redjet does not provide a function for every Redis command. Instead, it offers a generic interface that supports all commands and options. While this approach has less type-safety, it provides forward compatibility with new Redis features.
In the aim of both performance and ease-of-use, redjet attempts to provide
an API that closely resembles the protocol. For example, the Command method
is really a Pipeline of size 1.
Table of Contents
Basic Usage
Install:
go get github.com/coder/redjet@latestFor the most part, you can interact with Redis using a familiar interface:
package main
import (
"context"
"fmt"
"log"
"github.com/coder/redjet"
)
func main() {
client := redjet.New("localhost:6379")
ctx := context.Background()
err := client.Command(ctx, "SET", "foo", "bar").Ok()
// check error
got, err := client.Command(ctx, "GET", "foo").Bytes()
// check error
// got == []byte("bar")
}Streaming
To minimize allocations, call (*Pipeline).WriteTo instead of (*Pipeline).Bytes.
WriteTo streams the response directly to an io.Writer such as a file or HTTP response.
For example:
_, err := client.Command(ctx, "GET", "big-object").WriteTo(os.Stdout)
// check errorSimilarly, you can pass in a value that implements redjet.LenReader to
Command to stream larger values into Redis. Unfortunately, the API
cannot accept a regular io.Reader because bulk string messages in
the Redis protocol are length-prefixed.
Here's an example of streaming a large file into Redis:
bigFile, err := os.Open("bigfile.txt")
// check error
defer bigFile.Close()
stat, err := bigFile.Stat()
// check error
err = client.Command(
ctx, "SET", "bigfile",
redjet.NewLenReader(bigFile, stat.Size()),
).Ok()
// check errorIf you have no way of knowing the size of your blob in advance and still
want to avoid large allocations, you may chunk a stream into Redis using repeated APPEND commands.
Pipelining
redjet supports pipelining via the (*Client).Pipeline method. This method accepts a Pipeline, potentially that of a previous, open command.
// Set foo0, foo1, ..., foo99 to "bar", and confirm that each succeeded.
//
// This entire example only takes one round-trip to Redis!
var p *Pipeline
for i := 0; i < 100; i++ {
p = client.Pipeline(p, "SET", fmt.Sprintf("foo%d", i), "bar")
}
for r.Next() {
if err := p.Ok(); err != nil {
log.Fatal(err)
}
}
p.Close() // allow the underlying connection to be reused.PubSub
redjet suports PubSub via the NextSubMessage method. For example:
// Subscribe to a channel
sub := client.Command(ctx, "SUBSCRIBE", "my-channel")
sub.NextSubMessage() // ignore the first message, which is a confirmation of the subscription
// Publish a message to the channel
n, err := client.Command(ctx, "PUBLISH", "my-channel", "hello world").Int()
// check error
// n == 1, since there is one subscriber
// Receive the message
sub.NextSubMessage()
// sub.Payload == "hello world"
// sub.Channel == "my-channel"
// sub.Type == "message"Note that NextSubMessage will block until a message is received. To interrupt the subscription, cancel the context passed to Command.
Once a connection enters subscribe mode, the internal pool does not re-use it.
It is possible to subscribe to a channel in a performant, low-allocation way via the public API. NextSubMessage is just a convenience method.
JSON
redjet supports convenient JSON encoding and decoding via the (*Pipeline).JSON method. For example:
type Person struct {
Name string `json:"name"`
Age int `json:"age"`
}
// Set a person
// Unknown argument types are automatically encoded to JSON.
err := client.Command(ctx, "SET", "person", Person{
Name: "Alice",
Age: 30,
}).Ok()
// check error
// Get a person
var p Person
client.Command(ctx, "GET", "person").JSON(&p)
// check error
// p == Person{Name: "Alice", Age: 30}Connection Pooling
Redjet provides automatic connection pooling. Configuration knobs exist
within the Client struct that may be changed before any Commands are
issued.
If you want synchronous command execution over the same connection,
use the Pipeline method and consume the Pipeline after each call to Pipeline. Storing a long-lived Pipeline
offers the same functionality as storing a long-lived connection.
Benchmarks
On a pure throughput basis, redjet will perform similarly to redigo and go-redis. But, since redjet doesn't allocate memory for the entire response object, it consumes far less resources when handling large responses.
Here are some benchmarks (reproducible via make gen-bench) to illustrate:
.fullname: Get/1_B-10
β redjet β redigo β go-redis β rueidis β
β sec/op β sec/op vs base β sec/op vs base β sec/op vs base β
908.2n Β± 2% 962.4n Β± 1% +5.97% (p=0.000 n=10) 913.8n Β± 3% ~ (p=0.280 n=10) 1045.0n Β± 1% +15.06% (p=0.000 n=10)
β redjet β redigo β go-redis β rueidis β
β B/s β B/s vs base β B/s vs base β B/s vs base β
1074.2Ki Β± 2% 1015.6Ki Β± 1% -5.45% (p=0.000 n=10) 1069.3Ki Β± 2% ~ (p=0.413 n=10) 937.5Ki Β± 1% -12.73% (p=0.000 n=10)
β redjet β redigo β go-redis β rueidis β
β B/op β B/op vs base β B/op vs base β B/op vs base β
0.00 Β± 0% 41.00 Β± 0% ? (p=0.000 n=10) 275.50 Β± 2% ? (p=0.000 n=10) 249.00 Β± 0% ? (p=0.000 n=10)
β redjet β redigo β go-redis β rueidis β
β allocs/op β allocs/op vs base β allocs/op vs base β allocs/op vs base β
0.000 Β± 0% 3.000 Β± 0% ? (p=0.000 n=10) 4.000 Β± 0% ? (p=0.000 n=10) 2.000 Β± 0% ? (p=0.000 n=10)
.fullname: Get/1.0_kB-10
β redjet β redigo β go-redis β rueidis β
β sec/op β sec/op vs base β sec/op vs base β sec/op vs base β
1.302Β΅ Β± 2% 1.802Β΅ Β± 1% +38.42% (p=0.000 n=10) 1.713Β΅ Β± 3% +31.58% (p=0.000 n=10) 1.645Β΅ Β± 1% +26.35% (p=0.000 n=10)
β redjet β redigo β go-redis β rueidis β
β B/s β B/s vs base β B/s vs base β B/s vs base β
750.4Mi Β± 2% 542.1Mi Β± 1% -27.76% (p=0.000 n=10) 570.3Mi Β± 3% -24.01% (p=0.000 n=10) 593.8Mi Β± 1% -20.87% (p=0.000 n=10)
β redjet β redigo β go-redis β rueidis β
β B/op β B/op vs base β B/op vs base β B/op vs base β
0.000Ki Β± 0% 1.039Ki Β± 0% ? (p=0.000 n=10) 1.392Ki Β± 0% ? (p=0.000 n=10) 1.248Ki Β± 1% ? (p=0.000 n=10)
β redjet β redigo β go-redis β rueidis β
β allocs/op β allocs/op vs base β allocs/op vs base β allocs/op vs base β
0.000 Β± 0% 3.000 Β± 0% ? (p=0.000 n=10) 4.000 Β± 0% ? (p=0.000 n=10) 2.000 Β± 0% ? (p=0.000 n=10)
.fullname: Get/1.0_MB-10
β redjet β redigo β go-redis β rueidis β
β sec/op β sec/op vs base β sec/op vs base β sec/op vs base β
472.5Β΅ Β± 7% 477.3Β΅ Β± 2% ~ (p=0.190 n=10) 536.8Β΅ Β± 6% +13.61% (p=0.000 n=10) 475.3Β΅ Β± 6% ~ (p=0.684 n=10)
β redjet β redigo β go-redis β rueidis β
β B/s β B/s vs base β B/s vs base β B/s vs base β
2.067Gi Β± 8% 2.046Gi Β± 2% ~ (p=0.190 n=10) 1.819Gi Β± 6% -11.98% (p=0.000 n=10) 2.055Gi Β± 6% ~ (p=0.684 n=10)
β redjet β redigo β go-redis β rueidis β
β B/op β B/op vs base β B/op vs base β B/op vs base β
51.00 Β± 12% 1047849.50 Β± 0% +2054506.86% (p=0.000 n=10) 1057005.00 Β± 0% +2072458.82% (p=0.000 n=10) 1048808.50 Β± 0% +2056387.25% (p=0.000 n=10)
β redjet β redigo β go-redis β rueidis β
β allocs/op β allocs/op vs base β allocs/op vs base β allocs/op vs base β
1.000 Β± 0% 3.000 Β± 0% +200.00% (p=0.000 n=10) 4.000 Β± 0% +300.00% (p=0.000 n=10) 2.000 Β± 0% +100.00% (p=0.000 n=10)
Limitations
- redjet does not have convenient support for client side caching. But, the redjet API is flexible enough that a client could implement it themselves by following the instructions here.
- RESP3 is not supported. Practically, this means that connections aren't multiplexed, and other Redis libraries may perform better in high-concurrency scenarios.
- Certain features have not been tested but may still work:
- Redis Streams
- Monitor