fix: ai service race (#768)

pkg/bridge/ai/api_server.go

@@ -7,7 +7,6 @@ import (
 	"fmt"
 	"net"
 	"net/http"
-	"sync"
 	"time"
 
 	gonanoid "github.com/matoous/go-nanoid/v2"
@@ -127,14 +126,6 @@ func HandleInvoke(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	ci := &CacheItem{
-		wg:             &sync.WaitGroup{},
-		ResponseWriter: w,
-	}
-	if _, ok := service.cache[reqID]; !ok {
-		service.cache[reqID] = ci
-	}
-
 	var req ai.InvokeRequest
 	req.ReqID = reqID
 

pkg/bridge/ai/service.go

@@ -2,7 +2,6 @@ package ai
 
 import (
 	"fmt"
-	"net/http"
 	"sync"
 	"time"
 
@@ -24,27 +23,19 @@ var (
 	services *expirable.LRU[string, *Service]
 )
 
-// CacheItem cache the http.ResponseWriter, which is used for writing response from reducer.
-// TODO: http.ResponseWriter is from the SimpleRestfulServer interface, should be decoupled
-// from here.
-type CacheItem struct {
-	ResponseWriter http.ResponseWriter
-	wg             *sync.WaitGroup
-	mu             sync.Mutex
-}
-
 // Service is used to invoke LLM Provider to get the functions to be executed,
 // then, use source to send arguments which returned by llm provider to target
 // function. Finally, use reducer to aggregate all the results, and write the
 // result by the http.ResponseWriter.
 type Service struct {
-	credential   string
-	zipperAddr   string
-	md           metadata.M
-	source       yomo.Source
-	reducer      yomo.StreamFunction
-	cache        map[string]*CacheItem
+	credential string
+	zipperAddr string
+	md         metadata.M
+	source     yomo.Source
+	reducer    yomo.StreamFunction
+	// cache        map[string]*CacheItem
 	sfnCallCache map[string]*sfnAsyncCall
+	muCallCache  sync.Mutex
 	LLMProvider
 }
 
@@ -72,9 +63,9 @@ func DefaultExchangeMetadataFunc(credential string) (metadata.M, error) {
 
 func newService(credential string, zipperAddr string, aiProvider LLMProvider, exFn ExchangeMetadataFunc) (*Service, error) {
 	s := &Service{
-		credential:   credential,
-		zipperAddr:   zipperAddr,
-		cache:        make(map[string]*CacheItem),
+		credential: credential,
+		zipperAddr: zipperAddr,
+		// cache:        make(map[string]*CacheItem),
 		LLMProvider:  aiProvider,
 		sfnCallCache: make(map[string]*sfnAsyncCall),
 	}
@@ -116,7 +107,6 @@ func (s *Service) Release() {
 	if s.reducer != nil {
 		s.reducer.Close()
 	}
-	clear(s.cache)
 }
 
 func (s *Service) createSource() (yomo.Source, error) {
@@ -156,7 +146,9 @@ func (s *Service) createReducer() (yomo.StreamFunction, error) {
 		reqID := invoke.ReqID
 
 		// write parallel function calling results to cache, after all the results are written, the reducer will be done
+		s.muCallCache.Lock()
 		c, ok := s.sfnCallCache[reqID]
+		s.muCallCache.Unlock()
 		if !ok {
 			ylog.Error("[sfn-reducer] req_id not found", "req_id", reqID)
 			return
@@ -242,7 +234,9 @@ func (s *Service) runFunctionCalls(fns map[uint32][]*ai.ToolCall, reqID string)
 		wg:  &sync.WaitGroup{},
 		val: make(map[string]ai.ToolMessage),
 	}
+	s.muCallCache.Lock()
 	s.sfnCallCache[reqID] = asyncCall
+	s.muCallCache.Unlock()
 
 	for tag, tcs := range fns {
 		ylog.Debug("+++invoke toolCalls", "tag", tag, "len(toolCalls)", len(tcs), "reqID", reqID)
@@ -262,11 +256,13 @@ func (s *Service) runFunctionCalls(fns map[uint32][]*ai.ToolCall, reqID string)
 
 	arr := make([]ai.ToolMessage, 0)
 
+	asyncCall.mu.RLock()
 	for _, call := range asyncCall.val {
 		ylog.Debug("---invoke done", "id", call.ToolCallId, "content", call.Content)
 		call.Role = "tool"
 		arr = append(arr, call)
 	}
+	asyncCall.mu.RUnlock()
 
 	return arr, nil
 }
@@ -307,6 +303,6 @@ func init() {
 
 type sfnAsyncCall struct {
 	wg  *sync.WaitGroup
-	mu  sync.Mutex
+	mu  sync.RWMutex
 	val map[string]ai.ToolMessage
 }