When we think about communication between services/microservices, the first option that comes to mind is good old JSON. And this is not without reason, as the format has advantages, such as:
- is easily readable by both computers and humans;
- all modern programming languages can read and generate JSON;
- it is much less verbose than the previous alternative, Jurassic XML.
Using JSON is the recommendation for the vast majority of APIs developed in companies' daily lives. But in some cases, where performance is critical, we may need to consider other alternatives. This post aims to show two alternatives to JSON when it comes to communication between applications.
But what's the problem with JSON? One of its advantages is that it is "easily readable by humans," but this can be a weak point in performance. The fact is that we need to convert the JSON content to some structure that is known by the programming language we are using. An exception to this rule is if we use JavaScript, as JSON is native to it. But if you are using another language, Go, for example, we need to parse the data, as we can see in the (incomplete) code example below:
type event struct {
ID uuid.UUID
Type string `json:"type"`
Source string `json:"source"`
Subject string `json:"subject"`
Time string `json:"time"`
Data string `json:"data"`
}
var e event
err := json.NewDecoder(data).Decode(&e)
if err != nil {
http.Error(w, err.Error(), http.StatusBadRequest)
}
To solve this problem, we can test two alternatives, Protocol Buffers and Flatbuffers.
Protocol Buffers
Protobuf (Protocol Buffers), created by Google, is, according to the official website :
Protocol buffers are Google's language-neutral, platform-neutral, extensible mechanism for serializing structured data—think XML, but smaller, faster, and more straightforward. You define how you want your data to be structured once. Then, you can use specially generated source code to quickly write and read your structured data to and from various data streams using a variety of languages.
Generally used in conjunction with gRPC (but not necessarily), Protobuf is a binary protocol that significantly increases performance compared to the text format of JSON. But it "suffers" from the same problem as JSON: we need to parse it to a data structure of our language. For example, in Go:
//generated code
type Event struct {
state protoimpl.MessageState
sizeCache protoimpl.SizeCache
unknownFields protoimpl.UnknownFields
Type string `protobuf:"bytes,1,opt,name=type,proto3" json:"type,omitempty"`
Subject string `protobuf:"bytes,2,opt,name=subject,proto3" json:"subject,omitempty"`
Source string `protobuf:"bytes,3,opt,name=source,proto3" json:"source,omitempty"`
Time string `protobuf:"bytes,4,opt,name=time,proto3" json:"time,omitempty"`
Data string `protobuf:"bytes,5,opt,name=data,proto3" json:"data,omitempty"`
}
e := Event{}
err := proto.Unmarshal(data, &e)
if err != nil {
http.Error(w, err.Error(), http.StatusBadRequest)
}
Adopting a binary protocol gives us a performance gain, but we still need to solve the problem of data parsing. Our third competitor focuses on solving this problem.
Flatbuffers
According to the official website:
FlatBuffers is an efficient cross-platform serialization library for C++, C#, C, Go, Java, Kotlin, JavaScript, Lobster, Lua, TypeScript, PHP, Python, Rust, and Swift. It was initially created at Google for game development and other performance-critical applications.
Although initially created for game development, it fits perfectly into the environment we study in this post. Its advantage is that we do not need to parse the data in addition to being a binary protocol. For example, in Go:
//generated code
e := events.GetRootAsEvent(data, 0)
//we can use the data directly
saveEvent(string(e.Type()), string(e.Source()), string(e.Subject()), string(e.Time()), string(e.Data()))
But how much more performant are the two alternatives to JSON? Let's investigate...
Application
The first question that came to my mind was, "how can I apply this in a real scenario?". I imagined the following scenario:
A company with a mobile application, accessed daily by millions of customers, with an internal microservices architecture and which needs to save events generated by users and systems for auditing purposes.
This is a genuine scenario. So real that I live with it every day at the company where I work :)
Note: the scenario above is a simplification and does not represent the actual complexity of the team's application. It serves educational purposes.
The first step is defining an event in Protocol Buffers and Flatbuffers. Both have their own language for defining schemas, which we can then use to generate code in the languages we will use. I will not delve into the details of each scheme as this is easily found in the documentation.
The file event.proto
has the Protocol Buffer definition:
syntax = "proto3";
package events;
option go_package = "./events_pb";
message Event {
string type = 1;
string subject = 2;
string source = 3;
string time = 4;
string data = 5;
}
And the file event.fbs
has the equivalent in Flatbuffers:
namespace events;
table Event {
type: string;
subject:string;
source:string;
time:string;
data:string;
}
root_type Event;
The next step is to use these definitions to generate the necessary code. The following commands install the dependencies on macOS:
go install google.golang.org/protobuf/cmd/protoc-gen-go@latest
brew install protobuf
protoc -I=. --go_out=./ event.proto
brew install flatbuffers
flatc --go event.fbs
The result is the creation of Go packages to manipulate data in each format.
With the requirements met, the next step was implementing the event API. The main.go
looked like this:
package main
import (
"fmt"
"net/http"
"os"
"github.com/go-chi/chi/v5"
"github.com/go-chi/chi/v5/middleware"
"github.com/google/uuid"
)
func main() {
r := handlers()
http.ListenAndServe(":3000", r)
}
func handlers() *chi.Mux {
r := chi.NewRouter()
if os.Getenv("DEBUG") != "false" {
r.Use(middleware.Logger)
}
r.Post("/json", processJSON())
r.Post("/fb", processFB())
r.Post("/pb", processPB())
return r
}
func saveEvent(evType, source, subject, time, data string) {
if os.Getenv("DEBUG") != "false" {
id := uuid.New()
q := fmt.Sprintf("insert into event values('%s', '%s', '%s', '%s', '%s', '%s')", id, evType, source, subject, time, data)
fmt.Println(q)
}
// save event to database
}
For better organization, I created files to separate each function, which looked as follows:
package main
import (
"encoding/json"
"net/http"
"github.com/google/uuid"
)
type event struct {
ID uuid.UUID
Type string `json:"type"`
Source string `json:"source"`
Subject string `json:"subject"`
Time string `json:"time"`
Data string `json:"data"`
}
func processJSON() http.HandlerFunc {
return func(w http.ResponseWriter, r *http.Request) {
var e event
err := json.NewDecoder(r.Body).Decode(&e)
if err != nil {
http.Error(w, err.Error(), http.StatusBadRequest)
}
saveEvent(e.Type, e.Source, e.Subject, e.Time, e.Data)
w.WriteHeader(http.StatusCreated)
w.Write([]byte("json received"))
}
}
package main
import (
"io"
"net/http"
"github.com/eminetto/post-flatbuffers/events_pb"
"google.golang.org/protobuf/proto"
)
func processPB() http.HandlerFunc {
return func(w http.ResponseWriter, r *http.Request) {
body := r.Body
data, _ := io.ReadAll(body)
e := events_pb.Event{}
err := proto.Unmarshal(data, &e)
if err != nil {
http.Error(w, err.Error(), http.StatusBadRequest)
}
saveEvent(e.GetType(), e.GetSource(), e.GetSubject(), e.GetTime(), e.GetData())
w.WriteHeader(http.StatusCreated)
w.Write([]byte("protobuf received"))
}
}
package main
import (
"io"
"net/http"
"github.com/eminetto/post-flatbuffers/events"
)
func processFB() http.HandlerFunc {
return func(w http.ResponseWriter, r *http.Request) {
body := r.Body
data, _ := io.ReadAll(body)
e := events.GetRootAsEvent(data, 0)
saveEvent(string(e.Type()), string(e.Source()), string(e.Subject()), string(e.Time()), string(e.Data()))
w.WriteHeader(http.StatusCreated)
w.Write([]byte("flatbuffer received"))
}
}
In the functions processPB()
and processFB()
, we can see how the generated packages are used to manipulate the data.
Benchmark
The last step of our proof of concept is generating the benchmark to compare the formats. I used the Go stdlib benchmark package for this.
The file main_test.go
has tests for each format:
package main
import (
"bytes"
"fmt"
"net/http"
"net/http/httptest"
"os"
"strings"
"testing"
"github.com/eminetto/post-flatbuffers/events"
"github.com/eminetto/post-flatbuffers/events_pb"
flatbuffers "github.com/google/flatbuffers/go"
"google.golang.org/protobuf/proto"
)
func benchSetup() {
os.Setenv("DEBUG", "false")
}
func BenchmarkJSON(b *testing.B) {
benchSetup()
r := handlers()
payload := fmt.Sprintf(`{
"type": "button.clicked",
"source": "Login",
"subject": "user1000",
"time": "2018-04-05T17:31:00Z",
"data": "User clicked because X"}`)
for i := 0; i < b.N; i++ {
w := httptest.NewRecorder()
req, _ := http.NewRequest("POST", "/json", strings.NewReader(payload))
r.ServeHTTP(w, req)
if w.Code != http.StatusCreated {
b.Errorf("expected status 201, got %d", w.Code)
}
}
}
func BenchmarkFlatBuffers(b *testing.B) {
benchSetup()
r := handlers()
builder := flatbuffers.NewBuilder(1024)
evtType := builder.CreateString("button.clicked")
evtSource := builder.CreateString("service-b")
evtSubject := builder.CreateString("user1000")
evtTime := builder.CreateString("2018-04-05T17:31:00Z")
evtData := builder.CreateString("User clicked because X")
events.EventStart(builder)
events.EventAddType(builder, evtType)
events.EventAddSource(builder, evtSource)
events.EventAddSubject(builder, evtSubject)
events.EventAddTime(builder, evtTime)
events.EventAddData(builder, evtData)
evt := events.EventEnd(builder)
builder.Finish(evt)
buff := builder.FinishedBytes()
for i := 0; i < b.N; i++ {
w := httptest.NewRecorder()
req, _ := http.NewRequest("POST", "/fb", bytes.NewReader(buff))
r.ServeHTTP(w, req)
if w.Code != http.StatusCreated {
b.Errorf("expected status 201, got %d", w.Code)
}
}
}
func BenchmarkProtobuffer(b *testing.B) {
benchSetup()
r := handlers()
evt := events_pb.Event{
Type: "button.clicked",
Subject: "user1000",
Source: "service-b",
Time: "2018-04-05T17:31:00Z",
Data: "User clicked because X",
}
payload, err := proto.Marshal(&evt)
if err != nil {
panic(err)
}
for i := 0; i < b.N; i++ {
w := httptest.NewRecorder()
req, _ := http.NewRequest("POST", "/pb", bytes.NewReader(payload))
r.ServeHTTP(w, req)
if w.Code != http.StatusCreated {
b.Errorf("expected status 201, got %d", w.Code)
}
}
}
It generates an event in each format and sends it to the API.
When we run the benchmark, we have the following result:
Running tool: /opt/homebrew/bin/go test -benchmem -run=^$ -coverprofile=/var/folders/vn/gff4w90d37xbfc_2tn3616h40000gn/T/vscode-gojAS4GO/go-code-cover -bench . github.com/eminetto/post-flatbuffers/cmd/api -failfast -v
goos: darwin
goarch: arm64
pkg: github.com/eminetto/post-flatbuffers/cmd/api
BenchmarkJSON
BenchmarkJSON-8 658386 1732 ns/op 2288 B/op 26 allocs/op
BenchmarkFlatBuffers
BenchmarkFlatBuffers-8 1749194 640.5 ns/op 1856 B/op 21 allocs/op
BenchmarkProtobuffer
BenchmarkProtobuffer-8 1497356 696.9 ns/op 1952 B/op 21 allocs/op
PASS
coverage: 77.5% of statements
ok github.com/eminetto/post-flatbuffers/cmd/api 5.042s
If this is the first time you have analyzed the results of a Go benchmark, I recommend reading this post, where the author describes the details of each column and its meaning.
To make it easier to visualize, I created graphs for the most critical information generated by the benchmark:
Number of iterations (higher is better)
Nanoseconds per operation (lower is better)
Number of bytes allocated per operation (lower is better)
Number of allocations per operation (lower is better)
Conclusion
The numbers show a great advantage of binary protocols over JSON, especially Flatbuffers. This advantage is that we do not need to parse the data into structures of the language we are using.
Should you refactor your applications to replace JSON with Flatbuffers? Not necessarily. Performance is just one factor that teams must consider when selecting a communication protocol between their services and applications. But if your application receives billions of requests per day, performance improvements like those presented in this post can make a big difference in terms of costs and user experience.
The codes presented here can be found in this repository. I made the examples using the Go language, but both Protocol Buffers and Flatbuffers support different programming languages, so I would love to see other versions of these comparisons. Additionally, other benchmarks can be used, such as network consumption, CPU, etc. (since we only compare memory here).
I hope this post serves as an introduction to these formats and an incentive for new tests and experiments.
Originally published at https://eltonminetto.dev on August 05, 2024