Merge pull request #8 from mostlygeek/improve-upstream-monitoring-issue5

Improvements to handling of the upstream process so errors happen whenever one of these is first:

    the health check timeout is reached waiting for the upstream process to be ready
    the upstream process exits unexpectedly

With this change llama-swap is more compatible with use cases like containerized upstream services (#5) which pull the container before HTTP endpoints are ready.
This commit is contained in:
Benson Wong
2024-11-01 15:28:06 -07:00
committed by GitHub
2 changed files with 49 additions and 17 deletions

View File

@@ -1,6 +1,6 @@
# Seconds to wait for llama.cpp to be available to serve requests
# Default (and minimum): 15 seconds
healthCheckTimeout: 60
healthCheckTimeout: 15
models:
"llama":
@@ -35,7 +35,10 @@ models:
# until the upstream server is ready for traffic
checkEndpoint: none
# don't use this, just for testing if things are broken
# don't use these, just for testing if things are broken
"broken":
cmd: models/llama-server-osx --port 8999 -m models/doesnotexist.gguf
proxy: http://127.0.0.1:8999
proxy: http://127.0.0.1:8999
"broken_timeout":
cmd: models/llama-server-osx --port 8999 -m models/Qwen2.5-1.5B-Instruct-Q4_K_M.gguf
proxy: http://127.0.0.1:9000

View File

@@ -149,14 +149,28 @@ func (pm *ProxyManager) swapModel(requestedModel string) error {
}
pm.currentCmd = cmd
if err := pm.checkHealthEndpoint(); err != nil {
// watch for the command to exist
cmdCtx, cancel := context.WithCancelCause(context.Background())
// monitor the command's exist status
go func() {
err := cmd.Wait()
if err != nil {
cancel(fmt.Errorf("command [%s] %s", strings.Join(cmd.Args, " "), err.Error()))
} else {
cancel(nil)
}
}()
// wait for checkHealthEndpoint
if err := pm.checkHealthEndpoint(cmdCtx); err != nil {
return err
}
return nil
}
func (pm *ProxyManager) checkHealthEndpoint() error {
func (pm *ProxyManager) checkHealthEndpoint(cmdCtx context.Context) error {
if pm.currentConfig.Proxy == "" {
return fmt.Errorf("no upstream available to check /health")
@@ -179,6 +193,7 @@ func (pm *ProxyManager) checkHealthEndpoint() error {
if err != nil {
return fmt.Errorf("failed to create health url with with %s and path %s", proxyTo, checkEndpoint)
}
client := &http.Client{}
startTime := time.Now()
@@ -187,33 +202,47 @@ func (pm *ProxyManager) checkHealthEndpoint() error {
if err != nil {
return err
}
ctx, cancel := context.WithTimeout(req.Context(), 250*time.Millisecond)
ctx, cancel := context.WithTimeout(cmdCtx, 250*time.Millisecond)
defer cancel()
req = req.WithContext(ctx)
resp, err := client.Do(req)
if err != nil {
if strings.Contains(err.Error(), "connection refused") {
// if TCP dial can't connect any HTTP response after 5 seconds
// exit quickly.
if time.Since(startTime) > 5*time.Second {
return fmt.Errorf("health check endpoint took more than 5 seconds to respond")
}
ttl := (maxDuration - time.Since(startTime)).Seconds()
if err != nil {
// check if the context was cancelled
select {
case <-ctx.Done():
return context.Cause(ctx)
default:
}
if time.Since(startTime) >= maxDuration {
// wait a bit longer for TCP connection issues
if strings.Contains(err.Error(), "connection refused") {
fmt.Fprintf(pm.logMonitor, "Connection refused on %s, ttl %.0fs\n", healthURL, ttl)
time.Sleep(5 * time.Second)
} else {
time.Sleep(time.Second)
}
if ttl < 0 {
return fmt.Errorf("failed to check health from: %s", healthURL)
}
time.Sleep(time.Second)
continue
}
defer resp.Body.Close()
if resp.StatusCode == http.StatusOK {
return nil
}
if time.Since(startTime) >= maxDuration {
if ttl < 0 {
return fmt.Errorf("failed to check health from: %s", healthURL)
}
time.Sleep(time.Second)
}
}
@@ -236,7 +265,7 @@ func (pm *ProxyManager) proxyChatRequest(w http.ResponseWriter, r *http.Request)
}
if err := pm.swapModel(model); err != nil {
http.Error(w, fmt.Sprintf("unable to swap to model: %s", err.Error()), http.StatusNotFound)
http.Error(w, fmt.Sprintf("unable to swap to model, %s", err.Error()), http.StatusNotFound)
return
}