Thread Safety
This guide covers concurrent usage patterns for Fory Go, including the thread-safe wrapper and best practices for multi-goroutine environments.
Default Fory Instance
The default Fory instance is not thread-safe:
f := fory.New(fory.WithXlang(true))
// NOT SAFE: Concurrent access from multiple goroutines
go func() {
f.Serialize(value1) // Race condition!
}()
go func() {
f.Serialize(value2) // Race condition!
}()
Why Not Thread-Safe?
For performance, Fory reuses internal state:
- Buffer is cleared and reused between calls
- Reference resolvers are reset
- Context objects are recycled
This avoids allocations but requires exclusive access.
Thread-Safe Wrapper
For concurrent use, use the threadsafe package:
import "github.com/apache/fory/go/fory/threadsafe"
// Create thread-safe Fory
f := threadsafe.New()
// Safe for concurrent use
go func() {
data, _ := f.Serialize(value1)
}()
go func() {
data, _ := f.Serialize(value2)
}()
The wrapper creates instances as needed and reuses them across goroutines. Each operation exclusively borrows one instance and returns it afterward. Registered types remain available when garbage collection reclaims cached instances. Serialized output is copied before returning, so callers can retain it safely.
API
// Create thread-safe instance
f := threadsafe.New()
// Instance methods
data, err := f.Serialize(value)
err = f.Deserialize(data, &target)
// Generic functions
data, err := threadsafe.Serialize(f, &value)
err = threadsafe.Deserialize(f, data, &target)
// Global convenience functions
data, err := threadsafe.Marshal(&value)
err = threadsafe.Unmarshal(data, &target)
Type Registration
Register all types before the first serialization or deserialization operation. The first operation permanently freezes the wrapper's registrations, even if that operation fails. A later registration attempt returns an error.
f := threadsafe.New()
if err := f.RegisterStruct(User{}, 1); err != nil {
panic(err)
}
if err := f.RegisterStruct(Order{}, 2); err != nil {
panic(err)
}
// All concurrent operations use the registered types.
go func() {
data, err := f.Serialize(&User{ID: 1})
// Use data and handle err.
_, _ = data, err
}()
RegisterStructByName, RegisterEnum, and RegisterEnumByName are also
available directly on the wrapper. Every registered type is available to all
concurrent operations and remains registered across garbage collections.
For custom per-instance initialization, use NewWithFactory:
f := threadsafe.NewWithFactory(func() *fory.Fory {
inner := fory.New()
if err := inner.RegisterExtension(CustomType{}, 100, newCustomSerializer()); err != nil {
panic(err)
}
return inner
})
The factory may be called concurrently as additional instances are needed. It must return a fresh, identically configured instance on every call, with registrations completed before returning. Create stateful custom serializers separately for each instance.
Zero-Copy Considerations
Non-Thread-Safe Instance
With the default Fory, returned byte slices are views into the internal buffer:
f := fory.New(fory.WithXlang(true))
data1, _ := f.Serialize(value1)
// data1 is valid
data2, _ := f.Serialize(value2)
// data1 is NOW INVALID (buffer was reused)
Thread-Safe Instance
The thread-safe wrapper copies data automatically:
f := threadsafe.New()
data1, _ := f.Serialize(value1)
data2, _ := f.Serialize(value2)
// Both data1 and data2 are valid (independent copies)
This is safer but has allocation overhead.
Performance Comparison
| Scenario | Non-Thread-Safe | Thread-Safe |
|---|---|---|
| Single goroutine | Fastest | Slower (pool overhead) |
| Multiple goroutines | Unsafe | Safe, good scaling |
| Memory allocations | Minimal | Per-call copy |
| Buffer reuse | Yes | Per-pool-instance |
Benchmarking
func BenchmarkNonThreadSafe(b *testing.B) {
f := fory.New(fory.WithXlang(true))
f.RegisterStruct(User{}, 1)
user := &User{ID: 1, Name: "Alice"}
for i := 0; i < b.N; i++ {
data, _ := f.Serialize(user)
_ = data
}
}
func BenchmarkThreadSafe(b *testing.B) {
f := threadsafe.New()
f.RegisterStruct(User{}, 1)
user := &User{ID: 1, Name: "Alice"}
for i := 0; i < b.N; i++ {
data, _ := f.Serialize(user)
_ = data
}
}
Patterns
Per-Goroutine Instance
For maximum performance with known goroutine count:
func worker(id int) {
// Each worker has its own Fory instance
f := fory.New(fory.WithXlang(true))
f.RegisterStruct(User{}, 1)
for task := range tasks {
data, _ := f.Serialize(task)
process(data)
}
}
// Start workers
for i := 0; i < numWorkers; i++ {
go worker(i)
}
Shared Thread-Safe Instance
For dynamic goroutine count or simplicity:
// Single shared instance
var f = threadsafe.New()
func init() {
f.RegisterStruct(User{}, 1)
}
func handleRequest(user *User) []byte {
// Safe from any goroutine
data, _ := f.Serialize(user)
return data
}
HTTP Handler Example
var fory = threadsafe.New()
func init() {
fory.RegisterStruct(Response{}, 1)
}
func handler(w http.ResponseWriter, r *http.Request) {
response := &Response{
Status: "ok",
Data: getData(),
}
// Safe: threadsafe.Fory handles concurrency
data, err := fory.Serialize(response)
if err != nil {
http.Error(w, err.Error(), 500)
return
}
w.Header().Set("Content-Type", "application/octet-stream")
w.Write(data)
}
Common Mistakes
Sharing Non-Thread-Safe Instance
// WRONG: Race condition
var f = fory.New(fory.WithXlang(true))
func handler1() {
f.Serialize(value1) // Race!
}
func handler2() {
f.Serialize(value2) // Race!
}
Fix: Use threadsafe.New() or per-goroutine instances.
Keeping Reference to Buffer
// WRONG: Buffer invalidated on next call
f := fory.New(fory.WithXlang(true))
data, _ := f.Serialize(value1)
savedData := data // Just copies the slice header!
f.Serialize(value2) // Invalidates data and savedData
Fix: Clone the data or use thread-safe wrapper.
// Correct: Clone the data
data, _ := f.Serialize(value1)
savedData := make([]byte, len(data))
copy(savedData, data)
// Or use thread-safe (auto-copies)
f := threadsafe.New()
data, _ := f.Serialize(value1) // Already copied
Registering Types Concurrently
// RISKY: Concurrent registration
go func() {
f.RegisterStruct(TypeA{}, 1)
}()
go func() {
f.Serialize(value) // May not see TypeA
}()
Fix: Register all types before the first serialization or deserialization.
Best Practices
- Register types at startup: Before the first serialization or deserialization
- Clone data if keeping references: With non-thread-safe instance
- Use per-worker instances for hot paths: Eliminates pool contention
- Profile before optimizing: Thread-safe overhead may be negligible