Files
llmgw/internal/cli/server.go
T
ilkerandClaude Opus 5 a3f71aa5ee
CI / test (push) Successful in 1m1s
Phase 1: the router serves traffic
llmgw server is real now. It fronts a static set of engine replicas with an
OpenAI-compatible API, scoring endpoints through the pipeline added in the
previous commit.

- proxy: SSE relayed event by event with a flush after each, cancellation
  propagated upstream, and separate deadlines for dial, headers, first token,
  inter-token idle and total. The idle watchdog is what distinguishes a wedged
  engine from a slow one.
- router: dispatch with retry and model fallback. Retry only happens while
  nothing has reached the client; after that, splicing two responses together
  would be worse than the short stream the client already sees.
- config: YAML via koanf overlaid on defaults, with LLMGW_ env overrides.
  Validation rejects duplicate endpoint ids and aliases that shadow a model,
  because both misroute silently at request time rather than failing loudly.
- metrics: requests, duration, TTFT, tokens, plus per-scorer values and
  per-filter drops on a separate admin listener with health and pprof.

Phase 1 exit criterion, measured in test/e2e: 1000 concurrent streams across
3 sim-engines complete at ~6800 req/s; every client disconnect frees the KV
blocks on every engine; router overhead p99 is about 0.9 ms against the plan's
2 ms budget.

The handlers live in internal/router rather than an internal/router/openai
package as the plan's tree sketched: they need the dispatch loop, and splitting
them apart would have meant an interface with exactly one implementation. Plan
§4 updated to match.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-16 01:41:13 +02:00

158 lines
4.1 KiB
Go

package cli
import (
"context"
"errors"
"fmt"
"net"
"net/http"
"os/signal"
"strings"
"syscall"
"time"
"github.com/spf13/cobra"
"golang.org/x/sync/errgroup"
"gitea.mynodes.xyz/ilker/llmgw/internal/config"
"gitea.mynodes.xyz/ilker/llmgw/internal/router"
)
func newServerCmd() *cobra.Command {
var (
configPath string
listen string
adminListen string
endpoints []string
)
cmd := &cobra.Command{
Use: "server",
Short: "Run the router (phase 1; the control plane arrives in phase 3)",
Long: "Serves the OpenAI API in front of a set of engine replicas.\n\n" +
"In this phase the replica list is static: give it in the config file, or\n" +
"pass --endpoint id=URL=model[,model] once per replica for a quick test.",
RunE: func(cmd *cobra.Command, _ []string) error {
log := logger()
cfg, err := config.Load(configPath)
if err != nil {
return err
}
if listen != "" {
cfg.Listen = listen
}
if adminListen != "" {
cfg.AdminListen = adminListen
}
for _, spec := range endpoints {
ep, err := parseEndpointFlag(spec)
if err != nil {
return err
}
cfg.Endpoints = append(cfg.Endpoints, ep)
}
if err := cfg.Validate(); err != nil {
return err
}
if len(cfg.Endpoints) == 0 {
return errors.New("no endpoints configured; use --endpoint or a config file")
}
rt, err := router.New(cfg, log)
if err != nil {
return err
}
ctx, stop := signal.NotifyContext(cmd.Context(), syscall.SIGINT, syscall.SIGTERM)
defer stop()
rt.Start(ctx)
defer rt.Stop()
adminHandler, err := rt.AdminHandler()
if err != nil {
return err
}
apiSrv := &http.Server{
Handler: rt.Handler(),
ReadHeaderTimeout: 10 * time.Second,
// No write timeout: responses stream for minutes and the proxy
// enforces its own per-phase deadlines.
}
adminSrv := &http.Server{
Handler: adminHandler,
ReadHeaderTimeout: 10 * time.Second,
}
apiLn, err := net.Listen("tcp", cfg.Listen)
if err != nil {
return fmt.Errorf("listening on %s: %w", cfg.Listen, err)
}
adminLn, err := net.Listen("tcp", cfg.AdminListen)
if err != nil {
_ = apiLn.Close()
return fmt.Errorf("listening on %s: %w", cfg.AdminListen, err)
}
log.Info("router listening",
"api", apiLn.Addr().String(),
"admin", adminLn.Addr().String(),
"endpoints", len(cfg.Endpoints),
"default_profile", cfg.Router.DefaultProfile,
)
g, gctx := errgroup.WithContext(ctx)
g.Go(func() error { return serve(apiSrv, apiLn) })
g.Go(func() error { return serve(adminSrv, adminLn) })
g.Go(func() error {
<-gctx.Done()
log.Info("shutting down")
sctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
defer cancel()
_ = adminSrv.Shutdown(sctx)
return apiSrv.Shutdown(sctx)
})
return g.Wait()
},
}
f := cmd.Flags()
f.StringVar(&configPath, "config", "", "configuration file (optional)")
f.StringVar(&listen, "listen", "", "override the client-facing address")
f.StringVar(&adminListen, "admin-listen", "", "override the metrics/pprof address")
f.StringArrayVar(&endpoints, "endpoint", nil, "static endpoint: id=URL=model[,model] (repeatable)")
return cmd
}
func serve(s *http.Server, ln net.Listener) error {
if err := s.Serve(ln); err != nil && !errors.Is(err, http.ErrServerClosed) {
return err
}
return nil
}
// parseEndpointFlag decodes id=URL=model[,model].
//
// The URL contains '=' only in pathological cases, so splitting into exactly
// three parts keeps the flag readable while staying unambiguous.
func parseEndpointFlag(spec string) (config.Endpoint, error) {
parts := strings.SplitN(spec, "=", 3)
if len(parts) != 3 {
return config.Endpoint{}, fmt.Errorf("--endpoint %q: expected id=URL=model[,model]", spec)
}
models := strings.Split(parts[2], ",")
for i := range models {
models[i] = strings.TrimSpace(models[i])
}
return config.Endpoint{
ID: strings.TrimSpace(parts[0]),
URL: strings.TrimRight(strings.TrimSpace(parts[1]), "/"),
Models: models,
Pool: "static",
Weight: 1,
}, nil
}