Two things a cluster with many nodes and many GPUs needs that were missing, and they are the same thing seen from two distances. Idle tracking. The control plane knew what was placed and whether it was healthy, never whether anyone wanted it — so on a fleet with more models than cards there was no way to answer the first question anyone asks, which is what can I stop. The router is the only thing that knows, so it now reports it: a timestamp written in Acquire, where a request already meets an endpoint under a lock already held, and collected once a tick as a map. That shape is the point. Per-request notification would be a firehose carrying one bit — the newest timestamp wins — and this costs the same on four replicas as on four hundred. It travels through an interface in the direction opposite EndpointSink, for the same reason that one exists: the control plane decides what should run, the router knows whether anyone wants it, and neither should need to be in the other's process to say so. Unknown is kept distinct from idle throughout. A router that has just restarted knows nothing, and a zero timestamp would be the oldest idle replica in every comparison — so one restart would nominate an entire fleet for eviction. Absent from the map, "-" in the table, negative in the JSON, and IdleFor returns a second value rather than a duration anything could act on by mistake. ctl status. A per-node listing is right for four nodes and wrong for two hundred: how full am I, what is broken, what is holding cards nobody is using are aggregates, and reading them off a long table is a chore, which means it gets skipped. Capacity per GPU class with usage bars, free whole nodes per fabric — the number that decides whether a multi-node replica can land, since free cards on the wrong fabric are not capacity for a model that spans nodes — replica and model counts, models that cannot be placed, and what has gone unasked-for. Derived from the same state the tables show, so it cannot disagree with them. Unhealthy GPUs count as neither used nor free, because counting them free is how a cluster looks like it has room it does not. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
346 lines
10 KiB
Go
346 lines
10 KiB
Go
package cli
|
|
|
|
import (
|
|
"context"
|
|
"crypto/tls"
|
|
"errors"
|
|
"fmt"
|
|
"log/slog"
|
|
"net"
|
|
"net/http"
|
|
"os/signal"
|
|
"strings"
|
|
"sync"
|
|
"syscall"
|
|
"time"
|
|
|
|
"github.com/spf13/cobra"
|
|
"golang.org/x/net/http2"
|
|
"golang.org/x/net/http2/h2c"
|
|
"golang.org/x/sync/errgroup"
|
|
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/config"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/control"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/control/adminapi"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/control/agentsvc"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/control/catalog"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/control/reconciler"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/router"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/version"
|
|
"gitea.mynodes.xyz/ilker/llmgw/internal/web"
|
|
)
|
|
|
|
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
|
|
}
|
|
controlMode := cfg.ControlPlaneMode()
|
|
if !controlMode && len(cfg.Endpoints) == 0 {
|
|
return errors.New("nothing to serve: declare models with profiles, " +
|
|
"or give --endpoint for a router-only run")
|
|
}
|
|
|
|
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()
|
|
|
|
// The control plane replaces the static endpoint list: agents
|
|
// register, the reconciler places replicas, and Ready ones are
|
|
// published to the router (plan §6.4).
|
|
var agentHandler http.Handler
|
|
var agentPath string
|
|
var admin *control.Controller
|
|
if controlMode {
|
|
cat, err := cfg.Catalog()
|
|
if err != nil {
|
|
return err
|
|
}
|
|
|
|
ctl, svc, err := startControlPlane(ctx, cfg, cat, rt, cfg.AgentTLS.Enabled(), log)
|
|
if err != nil {
|
|
return err
|
|
}
|
|
defer ctl.Close()
|
|
admin = ctl
|
|
agentPath, agentHandler = svc.Handler()
|
|
|
|
log.Info("control plane running",
|
|
"models", len(cfg.Models), "agent_listen", cfg.AgentListen,
|
|
"agent_mtls", cfg.AgentTLS.Enabled())
|
|
}
|
|
|
|
adminMux, err := rt.AdminHandler()
|
|
if err != nil {
|
|
return err
|
|
}
|
|
if admin != nil {
|
|
api := adminapi.New(admin)
|
|
api.Register(adminMux)
|
|
|
|
// The dashboard lives here and only here. This listener is
|
|
// bound to loopback by default and also serves pprof, which can
|
|
// dump the process heap — where API keys sit in plaintext.
|
|
// Reaching the page from elsewhere is an SSH tunnel's job:
|
|
// ssh -L 9090:127.0.0.1:9090 <host>
|
|
var usage web.UsageSource
|
|
if st := rt.Store(); st != nil {
|
|
usage = st
|
|
}
|
|
adminMux.HandleFunc("GET /", web.Dashboard(
|
|
api, usage, version.Version, 5*time.Second, true, log))
|
|
adminMux.Handle("GET /_static/", web.Static())
|
|
|
|
// The same page on the client-facing port at /admin, behind a
|
|
// session and an admin flag. Handing the router the data rather
|
|
// than the port keeps pprof where it is.
|
|
rt.SetClusterView(api, usage)
|
|
}
|
|
|
|
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: adminMux,
|
|
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,
|
|
)
|
|
|
|
var agentSrv *http.Server
|
|
var agentLn net.Listener
|
|
if agentHandler != nil {
|
|
mux := http.NewServeMux()
|
|
mux.Handle(agentPath, agentHandler)
|
|
agentSrv = &http.Server{
|
|
Handler: mux,
|
|
ReadHeaderTimeout: 10 * time.Second,
|
|
}
|
|
if !cfg.AgentTLS.Enabled() {
|
|
// Without TLS, agents speak plaintext HTTP/2, which Go will
|
|
// not do unless asked. Nothing then stops a node from
|
|
// claiming to be another one, so this belongs on a trusted
|
|
// network only.
|
|
agentSrv.Handler = h2c.NewHandler(mux, &http2.Server{})
|
|
}
|
|
if agentLn, err = net.Listen("tcp", cfg.AgentListen); err != nil {
|
|
_ = apiLn.Close()
|
|
_ = adminLn.Close()
|
|
return fmt.Errorf("listening on %s: %w", cfg.AgentListen, err)
|
|
}
|
|
if cfg.AgentTLS.Enabled() {
|
|
tc, terr := cfg.AgentTLS.ServerConfig()
|
|
if terr != nil {
|
|
_ = apiLn.Close()
|
|
_ = adminLn.Close()
|
|
_ = agentLn.Close()
|
|
return terr
|
|
}
|
|
// Serve on an already-wrapped listener rather than
|
|
// ServeTLS: agentSrv.TLSConfig stays nil, which is exactly
|
|
// the case where net/http sets up HTTP/2 for ALPN itself.
|
|
agentLn = tls.NewListener(agentLn, tc)
|
|
}
|
|
}
|
|
|
|
g, gctx := errgroup.WithContext(ctx)
|
|
g.Go(func() error { return serve(apiSrv, apiLn) })
|
|
g.Go(func() error { return serve(adminSrv, adminLn) })
|
|
if agentSrv != nil {
|
|
g.Go(func() error { return serve(agentSrv, agentLn) })
|
|
}
|
|
g.Go(func() error {
|
|
<-gctx.Done()
|
|
log.Info("shutting down")
|
|
sctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
|
|
defer cancel()
|
|
if agentSrv != nil {
|
|
_ = agentSrv.Shutdown(sctx)
|
|
}
|
|
_ = 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
|
|
}
|
|
|
|
// startControlPlane builds the control plane and its agent service.
|
|
//
|
|
// The service is both the AgentService handler and the reconciler's Dispatcher:
|
|
// it holds the open streams, so it is the only thing that can deliver a command
|
|
// to a node.
|
|
func startControlPlane(ctx context.Context, cfg config.Server, cat *catalog.Catalog,
|
|
sink control.EndpointSink, requireIdentity bool, log *slog.Logger,
|
|
) (*control.Controller, *agentsvc.Service, error) {
|
|
ccfg := control.DefaultConfig()
|
|
ccfg.Reconciler = reconciler.Config{
|
|
Interval: cfg.Control.Interval,
|
|
BaseBackoff: cfg.Control.BaseBackoff,
|
|
MaxBackoff: cfg.Control.MaxBackoff,
|
|
UnreachableAfter: cfg.Control.UnreachableAfter,
|
|
LostAfter: cfg.Control.LostAfter,
|
|
MgmtInterface: cfg.Control.MgmtInterface,
|
|
}
|
|
if len(cfg.Control.PortRange) == 2 {
|
|
ccfg.Reconciler.PortRange = [2]int{cfg.Control.PortRange[0], cfg.Control.PortRange[1]}
|
|
}
|
|
|
|
// The service needs the inventory and the status sink, both of which the
|
|
// controller owns; the controller needs the service as its dispatcher. The
|
|
// cycle is broken by giving the controller a dispatcher that forwards to
|
|
// the service once it exists.
|
|
fwd := &dispatcherRef{}
|
|
ctl := control.New(ccfg, cat, fwd, sink, log)
|
|
// The router is both where endpoints go and where activity comes back
|
|
// from. It only reports usage when it is the thing serving requests, which
|
|
// is why this is a type assertion rather than a parameter.
|
|
if u, ok := sink.(control.UsageSource); ok {
|
|
ctl.SetUsageSource(u)
|
|
}
|
|
svc := agentsvc.New(ctl.Inventory(), ctl.Reconciler(), requireIdentity, log)
|
|
fwd.set(svc)
|
|
|
|
ctl.Start(ctx)
|
|
return ctl, svc, nil
|
|
}
|
|
|
|
// dispatcherRef lets the controller be built before the service exists.
|
|
type dispatcherRef struct {
|
|
mu sync.RWMutex
|
|
d reconciler.Dispatcher
|
|
}
|
|
|
|
func (r *dispatcherRef) set(d reconciler.Dispatcher) {
|
|
r.mu.Lock()
|
|
r.d = d
|
|
r.mu.Unlock()
|
|
}
|
|
|
|
func (r *dispatcherRef) get() (reconciler.Dispatcher, error) {
|
|
r.mu.RLock()
|
|
defer r.mu.RUnlock()
|
|
if r.d == nil {
|
|
return nil, errors.New("the agent service is not ready yet")
|
|
}
|
|
return r.d, nil
|
|
}
|
|
|
|
func (r *dispatcherRef) StartShard(ctx context.Context, cmd reconciler.ShardCommand) error {
|
|
d, err := r.get()
|
|
if err != nil {
|
|
return err
|
|
}
|
|
return d.StartShard(ctx, cmd)
|
|
}
|
|
|
|
func (r *dispatcherRef) StopShard(ctx context.Context, nodeID, shardID string) error {
|
|
d, err := r.get()
|
|
if err != nil {
|
|
return err
|
|
}
|
|
return d.StopShard(ctx, nodeID, shardID)
|
|
}
|
|
|
|
func (r *dispatcherRef) DrainShard(ctx context.Context, nodeID, shardID string) error {
|
|
d, err := r.get()
|
|
if err != nil {
|
|
return err
|
|
}
|
|
return d.DrainShard(ctx, nodeID, shardID)
|
|
}
|
|
|
|
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
|
|
}
|