Files
llmgw/internal/config/config.go
T
Ilker ManapandClaude Opus 5 a075ff4120
CI / test (push) Successful in 1m19s
Add approximate prefix routing, and find out it did nothing
Tier 1 of plan §7.5: a radix tree of request prefixes that estimates which
replica has already prefilled one. Bytes, not tokens — needing the engine's
exact tokenizer is what makes tier 2 a separate project.

It went in inert, and the decision metric is what caught it: 144 lookups, zero
hits. The cause was a rule I had written and tested for. A node's label is the
first request that created it, so the second request for the same system
prompt diverges in its last few bytes — a new question after the same
preamble, the single most common shape there is. Requiring the whole label to
match threw away the 700 bytes they shared. What matters is the shared head,
because that is what the engine cached, and a partial label match now counts.
Two unit tests asserted the wrong rule and have been inverted.

With that fixed, measured end to end against engines whose caches are too
small for the working set: 22.5% cache hits without, 81.7% with, and the
breakdown shows the misses are exactly the first round, when the tree is
empty.

The scenario matters as much as the number. With a handful of prompts and a
large cache, scattering is free — every replica ends up holding everything —
and prefix routing can only add constraint. The first version of the test
measured that case and showed a loss, which was a true statement about a
situation the feature is not for.

Two things keep it from being the disaster prefix routing usually is. The
imbalance guard, which needs both an absolute and a relative gap before it
fires, because either alone misfires — absolute on an idle cluster, relative
on a busy one. And it is a scorer rather than a filter: affinity is a vote the
queue and inflight scorers can outvote, not a rule.

The tree records the replica the picker actually chose, not the one the scorer
proposed, and a node prefers the replica it has seen most rather than most
recently — a count is stable under a randomizing picker, where recency just
chases whatever the picker last did.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
2026-09-16 12:13:19 +02:00

395 lines
14 KiB
Go

// Package config loads llmgw's YAML configuration.
//
// Phase 1 only needs the router's half of it: the endpoint list is static here
// and will come from the control plane in phase 3 (plan §6.4). The shape is
// already the one the plan describes, so that later change stays additive.
package config
import (
"fmt"
"os"
"strings"
"time"
"github.com/knadh/koanf/parsers/yaml"
env "github.com/knadh/koanf/providers/env/v2"
"github.com/knadh/koanf/providers/file"
"github.com/knadh/koanf/v2"
"gitea.mynodes.xyz/ilker/llmgw/internal/control/catalog"
"gitea.mynodes.xyz/ilker/llmgw/internal/tlsconf"
)
// Server is the whole server configuration.
type Server struct {
// Listen is the client-facing OpenAI API address.
Listen string `koanf:"listen"`
// AdminListen serves metrics, health and pprof. Kept separate so the
// client-facing port can be exposed without exposing diagnostics.
AdminListen string `koanf:"admin_listen"`
// AgentListen is where node agents connect. Separate from the client port
// so the fleet-facing surface can be firewalled independently.
AgentListen string `koanf:"agent_listen"`
// Auth turns on API keys on the client-facing port.
Auth Auth `koanf:"auth"`
// Landing configures the page served at "/".
Landing Landing `koanf:"landing"`
// AgentTLS authenticates the fleet-facing port. It deliberately does not
// cover the client-facing one: that surface is fronted by a publicly
// trusted certificate, while this one is mutual and private.
AgentTLS tlsconf.Config `koanf:"agent_tls"`
Router Router `koanf:"router"`
Control Control `koanf:"control"`
Models []catalog.Model `koanf:"models"`
// Endpoints is the phase-1 static replica list. It is only consulted when
// no model declares a profile, which is what makes a router-only smoke test
// against sim-engines possible without a control plane.
Endpoints []Endpoint `koanf:"endpoints"`
}
// Landing configures the public page at "/".
//
// It exists for one thing the gateway cannot know: who to ask for a key. A page
// that explains the API and then leaves the reader with no route to using it is
// where onboarding stops.
type Landing struct {
// Contact is where someone without a key should go — an address, a URL, a
// channel name. Rendered as a link when it is one, as text otherwise.
Contact string `koanf:"contact"`
// Notice is an optional line of your own, for whatever the reader needs to
// know that this page cannot guess.
Notice string `koanf:"notice"`
}
// Auth configures client authentication (plan §7.9).
//
// Off by default, because a gateway on a trusted network with no keys issued
// still has to serve. Turning it on is what makes the client-facing port safe
// to expose.
type Auth struct {
Enabled bool `koanf:"enabled"`
// Store is the SQLite file holding keys, tenancy and usage.
Store string `koanf:"store"`
// CacheTTL bounds how long a revoked key keeps working, traded against a
// database read per key per TTL.
CacheTTL time.Duration `koanf:"cache_ttl"`
// NegativeTTL is how long a rejected key is remembered as rejected, so a
// client looping with a bad key cannot become a database read per request.
NegativeTTL time.Duration `koanf:"negative_ttl"`
// SignupOpen serves the signup form. Accounts it creates are pending until
// an operator approves them, so this is not a way in — but turning it off
// means accounts are made with `llmgw users create` instead.
SignupOpen bool `koanf:"signup_open"`
// Pricing is what each model costs, for budgets. A model absent from the
// table costs whatever DefaultPrice says, which is nothing unless set.
Pricing map[string]Price `koanf:"pricing"`
DefaultPrice Price `koanf:"default_price"`
// FlushInterval and BatchSize decide when usage records are written.
FlushInterval time.Duration `koanf:"flush_interval"`
BatchSize int `koanf:"batch_size"`
// BufferSize bounds the queue. Beyond it records are dropped rather than
// blocking a request: losing accounting is bad, turning a stalled database
// into an outage is worse.
BufferSize int `koanf:"buffer_size"`
}
// Price is a model's cost in micro-units per thousand tokens.
//
// What a unit means is the operator's choice — self-hosted models have no
// per-token price, so it may be GPU-seconds, a notional rate, or 1000
// everywhere, which turns a budget into a token quota (plan §7.9.4).
type Price struct {
PromptPer1k int64 `koanf:"prompt_per_1k"`
CompletionPer1k int64 `koanf:"completion_per_1k"`
}
// Control configures the control plane.
type Control struct {
// Interval is how often desired and actual state are compared.
Interval time.Duration `koanf:"interval"`
// BaseBackoff and MaxBackoff bound how fast a failing model is retried.
BaseBackoff time.Duration `koanf:"base_backoff"`
MaxBackoff time.Duration `koanf:"max_backoff"`
// UnreachableAfter and LostAfter stage node expiry. They are deliberately
// far apart: a network blip must not tear down replicas that still serve,
// and rebuilding a multi-node replica costs tens of minutes of loading.
UnreachableAfter time.Duration `koanf:"unreachable_after"`
LostAfter time.Duration `koanf:"lost_after"`
// PortRange is the per-node range for engine HTTP and dist-init ports.
PortRange []int `koanf:"port_range"`
// MgmtInterface is the ethernet interface NCCL bootstraps over.
MgmtInterface string `koanf:"mgmt_interface"`
}
// Router holds routing behaviour.
type Router struct {
ScrapeInterval time.Duration `koanf:"scrape_interval"`
ScrapeTimeout time.Duration `koanf:"scrape_timeout"`
// StaleAfter is how long an endpoint may go unscraped before being scored
// down (plan §7.6).
StaleAfter time.Duration `koanf:"stale_after"`
// Retries is how many *additional* endpoints may be tried after a failure,
// and only while nothing has been written to the client (plan §7.7).
Retries int `koanf:"retries"`
// Cooldown is how long a failing endpoint leaves the candidate pool.
Cooldown time.Duration `koanf:"cooldown"`
// FailureThreshold is the consecutive failure count that opens a breaker.
// 1 turns the breaker into a pure cooldown.
FailureThreshold int `koanf:"failure_threshold"`
// DefaultProfile names the routing profile used when a model picks none.
DefaultProfile string `koanf:"default_profile"`
// Prefix turns on the approximate prefix tree (plan §7.5 tier 1). Off, the
// router uses session affinity alone, which is tier 0 and already most of
// the win for multi-turn traffic.
Prefix Prefix `koanf:"prefix"`
// SessionHeader carries the session id for affinity (plan §7.5 tier 0).
SessionHeader string `koanf:"session_header"`
// InjectUsage asks engines for usage even when the client did not, so token
// accounting works regardless of client behaviour.
InjectUsage bool `koanf:"inject_usage"`
// MaxBodyBytes bounds the request body the router will read.
MaxBodyBytes int64 `koanf:"max_body_bytes"`
Timeouts Timeouts `koanf:"timeouts"`
Scoring Scoring `koanf:"scoring"`
Admission Admission `koanf:"admission"`
}
// Admission configures the flow control layer (plan §7.2).
type Admission struct {
// Enabled turns the layer on. Off, every request goes straight through,
// which is the v0.2.0 behaviour.
Enabled bool `koanf:"enabled"`
Bands []Band `koanf:"bands"`
DefaultBand string `koanf:"default_band"`
// BandHeader lets a client declare its band. Until authentication lands in
// v1.0.0 this is trusted, so the port must not be exposed to callers who
// should not be able to promote themselves.
BandHeader string `koanf:"band_header"`
QueueTTL time.Duration `koanf:"queue_ttl"`
QueueCapacityPerBand int `koanf:"queue_capacity_per_band"`
Interval time.Duration `koanf:"interval"`
}
// Prefix configures approximate prefix routing.
type Prefix struct {
Enabled bool `koanf:"enabled"`
// MaxNodes bounds the tree, MinMatch is the shortest match worth acting on
// in bytes, and TTL is how long an entry is believed — an engine evicts on
// its own schedule, so an old record may describe a cache that is gone.
MaxNodes int `koanf:"max_nodes"`
MinMatch int `koanf:"min_match"`
TTL time.Duration `koanf:"ttl"`
// MatchRatio is the share of the prompt that must match to be worth acting
// on. BalanceAbs and BalanceRel are the imbalance guard, and both must trip
// before affinity is abandoned.
MatchRatio float64 `koanf:"match_ratio"`
BalanceAbs float64 `koanf:"balance_abs"`
BalanceRel float64 `koanf:"balance_rel"`
}
// Band is one priority class.
type Band struct {
Name string `koanf:"name"`
Priority int `koanf:"priority"`
SaturationCeiling float64 `koanf:"saturation_ceiling"`
}
// Timeouts bounds each phase of an upstream request.
type Timeouts struct {
Dial time.Duration `koanf:"dial"`
ResponseHeader time.Duration `koanf:"response_header"`
FirstToken time.Duration `koanf:"first_token"`
Idle time.Duration `koanf:"idle"`
Total time.Duration `koanf:"total"`
}
// Scoring holds the thresholds the scorers normalize against. These want to be
// measured with ctl bench (plan §10), not guessed.
type Scoring struct {
TargetWaiting float64 `koanf:"target_waiting"`
TargetInflight float64 `koanf:"target_inflight"`
KVCeiling float64 `koanf:"kv_ceiling"`
SessionLoadFactor float64 `koanf:"session_load_factor"`
}
// Endpoint is one engine replica. Static in phase 1.
type Endpoint struct {
ID string `koanf:"id"`
URL string `koanf:"url"`
Models []string `koanf:"models"`
// Engine names the inference engine, which decides what its metrics are
// called. Empty means vLLM's names.
Engine string `koanf:"engine"`
Pool string `koanf:"pool"`
Weight float64 `koanf:"weight"`
Priority int `koanf:"priority"`
}
// Default returns the configuration used when nothing is specified.
func Default() Server {
return Server{
Listen: ":8080",
AdminListen: ":9090",
AgentListen: ":7443",
Control: Control{
Interval: 5 * time.Second,
BaseBackoff: 10 * time.Second,
MaxBackoff: 10 * time.Minute,
UnreachableAfter: 15 * time.Second,
LostAfter: 60 * time.Second,
PortRange: []int{20000, 20999},
MgmtInterface: "eth0",
},
Router: Router{
ScrapeInterval: 2 * time.Second,
ScrapeTimeout: time.Second,
StaleAfter: 10 * time.Second,
Retries: 2,
Cooldown: 15 * time.Second,
FailureThreshold: 3,
DefaultProfile: "chat",
SessionHeader: "x-session-id",
InjectUsage: true,
MaxBodyBytes: 32 << 20,
Timeouts: Timeouts{
Dial: 5 * time.Second,
ResponseHeader: 60 * time.Second,
FirstToken: 120 * time.Second,
Idle: 60 * time.Second,
Total: 30 * time.Minute,
},
Scoring: Scoring{
TargetWaiting: 4,
TargetInflight: 8,
KVCeiling: 0.9,
SessionLoadFactor: 1.25,
},
Admission: Admission{
Enabled: true,
BandHeader: "x-llmgw-band",
DefaultBand: "gold",
Bands: []Band{
{Name: "platinum", Priority: 400, SaturationCeiling: 1.00},
{Name: "gold", Priority: 300, SaturationCeiling: 0.85},
{Name: "silver", Priority: 200, SaturationCeiling: 0.60},
{Name: "bronze", Priority: 100, SaturationCeiling: 0.30},
},
QueueTTL: 30 * time.Second,
QueueCapacityPerBand: 2000,
Interval: 10 * time.Millisecond,
},
},
}
}
// Load reads configuration from path, overlaying it on the defaults, then
// applies LLMGW_-prefixed environment overrides.
//
// An empty path loads the defaults, which is what makes `llmgw server` with a
// couple of --endpoint flags usable for a smoke test.
func Load(path string) (Server, error) {
cfg := Default()
k := koanf.New(".")
if path != "" {
if _, err := os.Stat(path); err != nil {
return cfg, fmt.Errorf("config file: %w", err)
}
if err := k.Load(file.Provider(path), yaml.Parser()); err != nil {
return cfg, fmt.Errorf("parsing %s: %w", path, err)
}
}
// LLMGW_ROUTER__RETRIES=3 maps onto router.retries.
err := k.Load(env.Provider(".", env.Opt{
Prefix: "LLMGW_",
TransformFunc: func(key, value string) (string, any) {
key = strings.ToLower(strings.TrimPrefix(key, "LLMGW_"))
return strings.ReplaceAll(key, "__", "."), value
},
}), nil)
if err != nil {
return cfg, fmt.Errorf("environment overrides: %w", err)
}
if err := k.UnmarshalWithConf("", &cfg, koanf.UnmarshalConf{Tag: "koanf"}); err != nil {
return cfg, fmt.Errorf("decoding configuration: %w", err)
}
return cfg, cfg.Validate()
}
// Validate rejects configurations that would otherwise fail confusingly at
// request time.
func (s Server) Validate() error {
if s.Listen == "" {
return fmt.Errorf("listen is empty")
}
if err := s.AgentTLS.Validate(); err != nil {
return err
}
if s.Auth.Enabled && s.Auth.Store == "" {
return fmt.Errorf("auth.store is required when auth is enabled")
}
seen := make(map[string]struct{}, len(s.Endpoints))
for i, e := range s.Endpoints {
switch {
case e.ID == "":
return fmt.Errorf("endpoints[%d]: id is required", i)
case e.URL == "":
return fmt.Errorf("endpoint %q: url is required", e.ID)
case !strings.HasPrefix(e.URL, "http://") && !strings.HasPrefix(e.URL, "https://"):
return fmt.Errorf("endpoint %q: url must start with http:// or https://", e.ID)
case len(e.Models) == 0:
return fmt.Errorf("endpoint %q: at least one model is required", e.ID)
}
if _, dup := seen[e.ID]; dup {
return fmt.Errorf("endpoint id %q is used twice", e.ID)
}
seen[e.ID] = struct{}{}
}
// The catalog enforces the rest — aliases, fallbacks, parallelism
// arithmetic — so there is one place those rules live.
if _, err := s.Catalog(); err != nil {
return err
}
return nil
}
// Catalog builds the model catalog from the configuration.
func (s Server) Catalog() (*catalog.Catalog, error) {
return catalog.New(s.Models)
}
// ControlPlaneMode reports whether the server should run a control plane.
//
// A model with no profiles cannot be placed, so a configuration of bare names
// plus a static endpoint list is the router-only mode used for testing against
// sim-engines.
func (s Server) ControlPlaneMode() bool {
for _, m := range s.Models {
if len(m.Profiles) > 0 {
return true
}
}
return false
}