Files
gateway/executor.go
zeekayandGitHub 69b5e25d47 feat(gateway): HIP-0110 standalone edge-process binary (#17)
* feat(middleware): own JWT validation + identity-strip + trust helper

Moves Auth + StripIdentityHeaders from zip/middleware to
gateway/middleware. Adds gateway/trust.go exporting AssertGatewayMinted
for downstream subsystems to verify the request flowed through gateway.

The unified cloud binary's pipeline now reads:
  zip generic middleware (Recover, Logger, RequestID, Telemetry, ...)
  gateway.middleware.StripIdentityHeaders
  gateway.middleware.Auth(verifier)
  ... downstream subsystem handlers ...

* feat(gateway): HIP-0110 ZAP wire envelopes + BuildApp entry

* feat(gateway): cmd/gateway HIP-0110 edge-process main

* refactor(gateway): tighten HIP-0110 surface — 588 LOC across 4 files

* build(gateway): fix module pins, MsgType uint16 overflow, drop legacy jaeger

- go.mod: hanzoai/cloud v0.0.0 → v0.1.0, hanzoai/zip v0.0.0 → v0.1.1
  (zip v0.1.1 carries the json/v2-at-edge work merged in zip#3;
   cloud v0.1.0 carries the types-leaf + clients refactor).
- zap_wire.go: MsgTypeForward/Push/Subscribe were 0x1010/0x1020/0x1030
  (uint16) but zap.FinishWithFlags expects the type in the upper byte
  (type << 8). Anything > 0xFF overflows uint16. Repacked into 0x80/0x81/0x82
  — still upper-half of the byte (no collision with base's 100-103 IDs).
- executor.go: dropped krakend-opencensus/v2/exporter/jaeger blank import.
  The exporter's Thrift API is incompatible with current jaeger-client-go;
  inter-service tracing is migrating to OTel (kotel) + ZAP anyway.
- cmd/gateway/main.go: c.SendString → c.Fiber().SendString (zip.Ctx
  exposes Send via the underlying fiber.Ctx; matches the pattern used
  for binary responses).

Verified: go build -mod=mod ./... clean. Binary at /tmp/gateway = 108 MiB.
TestStrip+TestAuth+TestSecure all pass.

* test(gateway): per-conn memory profile — 8 KiB/conn, 1:1 goroutine

Raw-TCP-dial loadgen (not http.Client, which spawns 3 client-side
goroutines that contaminate the gateway-side measurement). Spins up
zip.App on a local listener, dials N raw TCP conns each holding
/hold, samples runtime.MemStats.HeapAlloc delta.

Apple M1 Max, Go 1.26, zip on Fiber v3 v3.2.0:

  1,000 conns: 8.05 KiB/conn, 1.00 goroutine/conn,  14 MiB heap
  8,454 conns: 8.02 KiB/conn, 1.00 goroutine/conn,  72 MiB heap (linear)

Discipline: fasthttp gives us 1 goroutine per accepted conn already.
The HIP-0110 architecture preserves that — ZAP-forward to cloud is
sync from the gateway goroutine, no fan-out. Reverse-push lookups
are O(1) on a per-replica sync.Map, no extra goroutines.

Extrapolation:
  10k  conns → ~80 MiB / replica
  100k conns → ~800 MiB (1 GiB pod ceiling)
  10 replicas behind ingress → 1M concurrent client conns

Cloud side has zero per-conn cost — worker pool dispatches N ZAP
requests per pod. This is why a stateless gateway LB in front of
cloud maximizes compute efficiency: 10-100× fewer cloud replicas
than gateway replicas required for the same client population.

* deps: bump zip v0.1.1 → v0.2.0; pin cloud at resolvable pseudo-version

cloud cannot be tagged yet due to upstream GOPROXY blockers (kms repo >500MB,
vfs missing root package at v0.3.1). Pin cloud at the pseudo-version derived
from main HEAD so consumers can resolve via proxy.golang.org without local
replace dirs.
2026-05-29 19:59:38 -07:00

444 lines
16 KiB
Go

package gateway
import (
"context"
"fmt"
"io"
"net/http"
"os"
"github.com/gin-gonic/gin"
"golang.org/x/sync/errgroup"
krakendbf "github.com/krakend/bloomfilter/v2/krakend"
asyncamqp "github.com/krakend/krakend-amqp/v2/async"
cel "github.com/krakend/krakend-cel/v2"
cmd "github.com/krakend/krakend-cobra/v2"
_ "github.com/krakend/krakend-cors/v2/gin" // keep dep, CORS handled by hostProxyMiddleware
gelf "github.com/krakend/krakend-gelf/v2"
gologging "github.com/krakend/krakend-gologging/v2"
influxdb "github.com/krakend/krakend-influx/v2"
jose "github.com/krakend/krakend-jose/v2"
logstash "github.com/krakend/krakend-logstash/v2"
metrics "github.com/krakend/krakend-metrics/v2/gin"
opencensus "github.com/krakend/krakend-opencensus/v2"
_ "github.com/krakend/krakend-opencensus/v2/exporter/datadog"
_ "github.com/krakend/krakend-opencensus/v2/exporter/influxdb"
// jaeger exporter dropped: upstream uses an opencensus/Thrift API combo
// incompatible with current jaeger-client-go. Migration path is OTel
// (kotel below) + ZAP for inter-service.
_ "github.com/krakend/krakend-opencensus/v2/exporter/ocagent"
_ "github.com/krakend/krakend-opencensus/v2/exporter/prometheus"
_ "github.com/krakend/krakend-opencensus/v2/exporter/stackdriver"
_ "github.com/krakend/krakend-opencensus/v2/exporter/xray"
_ "github.com/krakend/krakend-opencensus/v2/exporter/zipkin"
kotel "github.com/krakend/krakend-otel"
otellura "github.com/krakend/krakend-otel/lura"
otelgin "github.com/krakend/krakend-otel/router/gin"
pubsub "github.com/krakend/krakend-pubsub/v2"
"github.com/luraproject/lura/v2/async"
"github.com/luraproject/lura/v2/config"
"github.com/luraproject/lura/v2/core"
"github.com/luraproject/lura/v2/logging"
"github.com/luraproject/lura/v2/proxy"
router "github.com/luraproject/lura/v2/router/gin"
"github.com/luraproject/lura/v2/sd/dnssrv"
serverhttp "github.com/luraproject/lura/v2/transport/http/server"
server "github.com/luraproject/lura/v2/transport/http/server/plugin"
)
// NewExecutor returns an executor for the cmd package. The executor initalizes the entire gateway by
// registering the components and composing a RouterFactory wrapping all the middlewares.
func NewExecutor(ctx context.Context) cmd.Executor {
eb := new(ExecutorBuilder)
return eb.NewCmdExecutor(ctx)
}
// PluginLoader defines the interface for the collaborator responsible of starting the plugin loaders
// Deprecated: Use PluginLoaderWithContext
type PluginLoader interface {
Load(folder, pattern string, logger logging.Logger)
}
// PluginLoaderWithContext defines the interface for the collaborator responsible of starting the plugin loaders
type PluginLoaderWithContext interface {
LoadWithContext(ctx context.Context, folder, pattern string, logger logging.Logger)
}
// SubscriberFactoriesRegister registers all the required subscriber factories from the available service
// discover components and adapters and returns a service register function.
// The service register function will register the service by the given name and port to all the available
// service discover clients
type SubscriberFactoriesRegister interface {
Register(context.Context, config.ServiceConfig, logging.Logger) func(string, int)
}
// TokenRejecterFactory returns a jose.ChainedRejecterFactory containing all the required jose.RejecterFactory.
// It also should setup and manage any service related to the management of the revocation process, if required.
type TokenRejecterFactory interface {
NewTokenRejecter(context.Context, config.ServiceConfig, logging.Logger, func(string, int)) (jose.ChainedRejecterFactory, error)
}
// MetricsAndTracesRegister registers the defined observability components and returns a metrics collector,
// if required.
type MetricsAndTracesRegister interface {
Register(context.Context, config.ServiceConfig, logging.Logger) *metrics.Metrics
}
// EngineFactory returns a gin engine, ready to be passed to the Gateway RouterFactory
type EngineFactory interface {
NewEngine(config.ServiceConfig, router.EngineOptions) *gin.Engine
}
// ProxyFactory returns a Gateway proxy factory, ready to be passed to the Gateway RouterFactory
type ProxyFactory interface {
NewProxyFactory(logging.Logger, proxy.BackendFactory, *metrics.Metrics) proxy.Factory
}
// BackendFactory returns a Gateway backend factory, ready to be passed to the Gateway proxy factory
type BackendFactory interface {
NewBackendFactory(context.Context, logging.Logger, *metrics.Metrics) proxy.BackendFactory
}
// HandlerFactory returns a Gateway router handler factory, ready to be passed to the Gateway RouterFactory
type HandlerFactory interface {
NewHandlerFactory(logging.Logger, *metrics.Metrics, jose.RejecterFactory) router.HandlerFactory
}
// LoggerFactory returns a Gateway Logger factory, ready to be passed to the Gateway RouterFactory
type LoggerFactory interface {
NewLogger(config.ServiceConfig) (logging.Logger, io.Writer, error)
}
// RunServer defines the interface of a function used by the Gateway router to start the service
type RunServer func(context.Context, config.ServiceConfig, http.Handler) error
// RunServerFactory returns a RunServer with several wraps around the injected one
type RunServerFactory interface {
NewRunServer(logging.Logger, router.RunServerFunc) RunServer
}
// AgentStarter defines a type that starts a set of agents
type AgentStarter interface {
Start(
context.Context,
[]*config.AsyncAgent,
logging.Logger,
chan<- string,
proxy.Factory,
) func() error
}
// ExecutorBuilder is a composable builder. Every injected property is used by the NewCmdExecutor method.
type ExecutorBuilder struct {
// PluginLoader is deprecated: Use PluginLoaderWithContext
PluginLoader PluginLoader
PluginLoaderWithContext PluginLoaderWithContext
LoggerFactory LoggerFactory
SubscriberFactoriesRegister SubscriberFactoriesRegister
TokenRejecterFactory TokenRejecterFactory
MetricsAndTracesRegister MetricsAndTracesRegister
EngineFactory EngineFactory
ProxyFactory ProxyFactory
BackendFactory BackendFactory
HandlerFactory HandlerFactory
RunServerFactory RunServerFactory
AgentStarterFactory AgentStarter
Middlewares []gin.HandlerFunc
}
// NewCmdExecutor returns an executor for the cmd package. The executor initializes the entire gateway by
// delegating most of the tasks to the injected collaborators. They register the components and
// compose a RouterFactory wrapping all the middlewares.
// Every nil collaborator is replaced by the default one offered by this package.
func (e *ExecutorBuilder) NewCmdExecutor(ctx context.Context) cmd.Executor {
e.checkCollaborators()
return func(cfg config.ServiceConfig) {
cfg.Normalize()
logger, gelfWriter, gelfErr := e.LoggerFactory.NewLogger(cfg)
if gelfErr != nil {
return
}
logger.Info(fmt.Sprintf("Starting Gateway v%s", core.KrakendVersion))
startReporter(ctx, logger, cfg)
if wd, err := os.Getwd(); err == nil {
logger.Info("Working directory is", wd)
}
dnssrv.SetTTL(cfg.DNSCacheTTL)
if cfg.Plugin != nil {
e.PluginLoaderWithContext.LoadWithContext(ctx, cfg.Plugin.Folder, cfg.Plugin.Pattern, logger)
}
metricCollector := e.MetricsAndTracesRegister.Register(ctx, cfg, logger)
if metricsAndTracesCloser, ok := e.MetricsAndTracesRegister.(io.Closer); ok {
defer metricsAndTracesCloser.Close()
}
// Initializes the global cache for the JWK clients if enabled in the config
if err := jose.SetGlobalCacher(logger, cfg.ExtraConfig); err != nil && err != jose.ErrNoValidatorCfg {
logger.Error("[SERVICE: JOSE]", err.Error())
}
tokenRejecterFactory, err := e.TokenRejecterFactory.NewTokenRejecter(
ctx,
cfg,
logger,
e.SubscriberFactoriesRegister.Register(ctx, cfg, logger),
)
if err != nil && err != krakendbf.ErrNoConfig {
logger.Warning("[SERVICE: Bloomfilter]", err.Error())
}
bpf := e.BackendFactory.NewBackendFactory(ctx, logger, metricCollector)
pf := e.ProxyFactory.NewProxyFactory(logger, bpf, metricCollector)
// we move the proxy factory out of the default proxy factory to make
// sure that is always the outer middleware and that wraps any internal
// proxy layer middleware:
pf = otellura.ProxyFactory(pf)
agentPing := make(chan string, len(cfg.AsyncAgents))
handlerF := e.HandlerFactory.NewHandlerFactory(logger, metricCollector, tokenRejecterFactory)
handlerF = otelgin.New(handlerF)
runServerChain := serverhttp.RunServerWithLoggerFactory(logger)
runServerChain = otellura.GlobalRunServer(logger, runServerChain)
runServerChain = router.RunServerFunc(e.RunServerFactory.NewRunServer(logger, runServerChain))
// Start the inbound ZAP listener (TLS 1.3+PQ) for external clients.
InitZapListenerFromEnv()
defer StopZapListener()
// The ZAP-HTTP listener is booted lazily by the RunServer chain
// (DefaultRunServerFactory.NewRunServer) once the fully-wrapped
// handler is available. Just ensure we shut it down on exit.
defer stopZapHTTPListener()
// setup the gateway router
routerFactory := router.NewFactory(router.Config{
Engine: e.EngineFactory.NewEngine(cfg, router.EngineOptions{
Logger: logger,
Writer: gelfWriter,
Health: (<-chan string)(agentPing),
}),
ProxyFactory: pf,
Middlewares: e.Middlewares,
Logger: logger,
HandlerFactory: handlerF,
RunServer: runServerChain,
})
// start the engines
logger.Info("Starting the Gateway instance")
if len(cfg.AsyncAgents) == 0 {
routerFactory.NewWithContext(ctx).Run(cfg)
return
}
// start the async agents in the same error group as the router
g, gctx := errgroup.WithContext(ctx)
gctx, closeGroupCtx := context.WithCancel(gctx)
if cfg.SequentialStart {
waitAgents := e.AgentStarterFactory.Start(gctx, cfg.AsyncAgents, logger, (chan<- string)(agentPing), pf)
g.Go(waitAgents)
} else {
g.Go(func() error {
return e.AgentStarterFactory.Start(gctx, cfg.AsyncAgents, logger, (chan<- string)(agentPing), pf)()
})
}
g.Go(func() error {
logger.Info("[SERVICE: Gin] Building the router")
routerFactory.NewWithContext(ctx).Run(cfg)
closeGroupCtx()
return nil
})
g.Wait()
}
}
func (e *ExecutorBuilder) checkCollaborators() {
if e.PluginLoader == nil {
e.PluginLoader = new(pluginLoader)
}
if e.PluginLoaderWithContext == nil {
e.PluginLoaderWithContext = new(pluginLoader)
}
if e.SubscriberFactoriesRegister == nil {
e.SubscriberFactoriesRegister = new(registerSubscriberFactories)
}
if e.TokenRejecterFactory == nil {
e.TokenRejecterFactory = new(BloomFilterJWT)
}
if e.MetricsAndTracesRegister == nil {
e.MetricsAndTracesRegister = new(MetricsAndTraces)
}
if e.EngineFactory == nil {
e.EngineFactory = new(engineFactory)
}
if e.ProxyFactory == nil {
e.ProxyFactory = new(proxyFactory)
}
if e.BackendFactory == nil {
e.BackendFactory = new(backendFactory)
}
if e.HandlerFactory == nil {
e.HandlerFactory = new(handlerFactory)
}
if e.LoggerFactory == nil {
e.LoggerFactory = new(LoggerBuilder)
}
if e.RunServerFactory == nil {
e.RunServerFactory = new(DefaultRunServerFactory)
}
if e.AgentStarterFactory == nil {
e.AgentStarterFactory = async.AgentStarter([]async.Factory{asyncamqp.StartAgent})
}
}
// DefaultRunServerFactory creates the default RunServer by wrapping the injected RunServer
// with the plugin loader and the CORS module
type DefaultRunServerFactory struct{}
func (*DefaultRunServerFactory) NewRunServer(l logging.Logger, next router.RunServerFunc) RunServer {
// CORS is handled by hostProxyMiddleware in router_engine.go.
// Skip the upstream CORS wrapper to avoid duplicate headers.
httpRun := RunServer(server.New(
l,
server.RunServer(next),
))
// Boot the ZAP-HTTP listener once per process with the same fully
// composed handler that the HTTP listener serves. Additive: HTTP
// keeps serving regardless of ZAP-HTTP state.
return func(ctx context.Context, cfg config.ServiceConfig, handler http.Handler) error {
startZapHTTPListenerOnce(l, handler)
return httpRun(ctx, cfg, handler)
}
}
// LoggerBuilder is the default BuilderFactory implementation.
type LoggerBuilder struct{}
// NewLogger sets up the logging components as defined at the configuration.
func (LoggerBuilder) NewLogger(cfg config.ServiceConfig) (logging.Logger, io.Writer, error) {
var writers []io.Writer
gelfWriter, gelfErr := gelf.NewWriter(cfg.ExtraConfig)
if gelfErr == nil {
writers = append(writers, gelfWriterWrapper{gelfWriter})
gologging.SetFormatterSelector(func(w io.Writer) string {
switch w.(type) {
case gelfWriterWrapper:
return "%{message}"
default:
return gologging.DefaultPattern
}
})
} else {
gelfWriter = nil
}
logger, gologgingErr := logstash.NewLogger(cfg.ExtraConfig)
if gologgingErr != nil {
logger, gologgingErr = gologging.NewLogger(cfg.ExtraConfig, writers...)
if gologgingErr != nil {
var err error
logger, err = logging.NewLogger("DEBUG", os.Stdout, "GATEWAY")
if err != nil {
return logger, gelfWriter, err
}
if gologgingErr != gologging.ErrWrongConfig {
logger.Error("[SERVICE: Logging] Unable to create the logger:", gologgingErr.Error())
}
}
}
if gelfErr != nil && gelfErr != gelf.ErrWrongConfig {
logger.Error("[SERVICE: Logging][GELF] Unable to create the writer:", gelfErr.Error())
}
if gologgingErr == nil {
logger.Debug("[SERVICE: telemetry/logging] Improved logging started.")
}
return logger, gelfWriter, nil
}
// BloomFilterJWT is the default TokenRejecterFactory implementation.
type BloomFilterJWT struct{}
// NewTokenRejecter registers the bloomfilter component and links it to a token rejecter. Then it returns a chained
// rejecter factory with the created token rejecter and other based on the CEL component.
func (BloomFilterJWT) NewTokenRejecter(ctx context.Context, cfg config.ServiceConfig, l logging.Logger, reg func(n string, p int)) (jose.ChainedRejecterFactory, error) {
rejecter, err := krakendbf.Register(ctx, "krakend-bf", cfg, l, reg)
return jose.ChainedRejecterFactory([]jose.RejecterFactory{
jose.RejecterFactoryFunc(func(_ logging.Logger, _ *config.EndpointConfig) jose.Rejecter {
return jose.RejecterFunc(rejecter.RejectToken)
}),
jose.RejecterFactoryFunc(func(l logging.Logger, cfg *config.EndpointConfig) jose.Rejecter {
if r := cel.NewRejecter(l, cfg); r != nil {
return r
}
return jose.FixedRejecter(false)
}),
}), err
}
// MetricsAndTraces is the default implementation of the MetricsAndTracesRegister interface.
type MetricsAndTraces struct {
shutdownFn func()
}
// Register registers the metrics, influx and opencensus packages as required by the given configuration.
func (m *MetricsAndTraces) Register(ctx context.Context, cfg config.ServiceConfig, l logging.Logger) *metrics.Metrics {
metricCollector := metrics.New(ctx, cfg.ExtraConfig, l)
if err := influxdb.New(ctx, cfg.ExtraConfig, metricCollector, l); err != nil {
if err != influxdb.ErrNoConfig {
l.Warning("[SERVICE: InfluxDB]", err.Error())
}
} else {
l.Debug("[SERVICE: InfluxDB] Service correctly registered")
}
if err := opencensus.Register(ctx, cfg, append(opencensus.DefaultViews, pubsub.OpenCensusViews...)...); err != nil {
if err != opencensus.ErrNoConfig {
l.Warning("[SERVICE: OpenCensus]", err.Error())
}
} else {
l.Debug("[SERVICE: OpenCensus] Service correctly registered")
}
if shutdownFn, err := kotel.Register(ctx, l, cfg); err == nil {
m.shutdownFn = shutdownFn
} else {
l.Error(fmt.Sprintf("[SERVICE: OpenTelemetry] cannot register exporters: %s", err.Error()))
}
return metricCollector
}
func (m *MetricsAndTraces) Close() {
if m.shutdownFn != nil {
m.shutdownFn()
}
}
// startReporter is a no-op in the Hanzo fork — no telemetry is sent to upstream.
func startReporter(_ context.Context, _ logging.Logger, _ config.ServiceConfig) {}
type gelfWriterWrapper struct {
io.Writer
}