11package static
22
33import (
4+ "context"
45 "errors"
6+ "fmt"
7+ "net"
58 "net/http"
69 "sync"
10+ "time"
11+
12+ "sigs.k8s.io/controller-runtime/pkg/manager"
13+
14+ "github.com/nginx/nginx-gateway-fabric/internal/mode/static/config"
715)
816
917// newGraphBuiltHealthChecker creates a new graphBuiltHealthChecker.
@@ -13,37 +21,94 @@ func newGraphBuiltHealthChecker() *graphBuiltHealthChecker {
1321 }
1422}
1523
16- // graphBuiltHealthChecker is used to check if the initial graph is built and the NGF Pod is ready.
24+ // graphBuiltHealthChecker is used to check if the NGF Pod is ready. The NGF Pod is ready if the initial graph has
25+ // been built and if it is leader.
1726type graphBuiltHealthChecker struct {
1827 // readyCh is a channel that is initialized in newGraphBuiltHealthChecker and represents if the NGF Pod is ready.
19- readyCh chan struct {}
20- lock sync.RWMutex
21- ready bool
28+ readyCh chan struct {}
29+ lock sync.RWMutex
30+ graphBuilt bool
31+ leader bool
32+ }
33+
34+ // createHealthProbe creates a Server runnable to serve as our health and readiness checker.
35+ func createHealthProbe (cfg config.Config , healthChecker * graphBuiltHealthChecker ) (manager.Server , error ) {
36+ // we chose to create our own health probe server instead of using the controller-runtime one because
37+ // of repetitive log which would flood our logs on non-ready non-leader NGF Pods. This health probe is
38+ // similar to the controller-runtime's health probe.
39+
40+ mux := http .NewServeMux ()
41+
42+ // copy of controller-runtime sane defaults for new http.Server
43+ s := & http.Server {
44+ Handler : mux ,
45+ MaxHeaderBytes : 1 << 20 ,
46+ IdleTimeout : 90 * time .Second , // matches http.DefaultTransport keep-alive timeout
47+ ReadHeaderTimeout : 32 * time .Second ,
48+ }
49+
50+ mux .HandleFunc (readinessEndpointName , healthChecker .readyHandler )
51+
52+ ln , err := net .Listen ("tcp" , fmt .Sprintf (":%d" , cfg .HealthConfig .Port ))
53+ if err != nil {
54+ return manager.Server {},
55+ fmt .Errorf ("error listening on %s: %w" , fmt .Sprintf (":%d" , cfg .HealthConfig .Port ), err )
56+ }
57+
58+ return manager.Server {
59+ Name : "health probe" ,
60+ Server : s ,
61+ Listener : ln ,
62+ }, nil
63+ }
64+
65+ func (h * graphBuiltHealthChecker ) readyHandler (resp http.ResponseWriter , req * http.Request ) {
66+ if err := h .readyCheck (req ); err != nil {
67+ resp .WriteHeader (http .StatusServiceUnavailable )
68+ } else {
69+ resp .WriteHeader (http .StatusOK )
70+ }
2271}
2372
2473// readyCheck returns the ready-state of the Pod. It satisfies the controller-runtime Checker type.
25- // We are considered ready after the first graph is built.
74+ // We are considered ready after the first graph is built and if the NGF Pod is leader .
2675func (h * graphBuiltHealthChecker ) readyCheck (_ * http.Request ) error {
2776 h .lock .RLock ()
2877 defer h .lock .RUnlock ()
2978
30- if ! h .ready {
31- return errors .New ("control plane is not yet ready" )
79+ if ! h .leader {
80+ return errors .New ("this Pod is not currently leader" )
81+ }
82+
83+ if ! h .graphBuilt {
84+ return errors .New ("control plane initial graph has not been built" )
3285 }
3386
3487 return nil
3588}
3689
37- // setAsReady marks the health check as ready .
38- func (h * graphBuiltHealthChecker ) setAsReady () {
90+ // setGraphBuilt marks the health check as having the initial graph built .
91+ func (h * graphBuiltHealthChecker ) setGraphBuilt () {
3992 h .lock .Lock ()
4093 defer h .lock .Unlock ()
4194
42- h .ready = true
43- close (h .readyCh )
95+ h .graphBuilt = true
4496}
4597
4698// getReadyCh returns a read-only channel, which determines if the NGF Pod is ready.
4799func (h * graphBuiltHealthChecker ) getReadyCh () <- chan struct {} {
48100 return h .readyCh
49101}
102+
103+ // setAsLeader marks the health check as leader.
104+ func (h * graphBuiltHealthChecker ) setAsLeader (_ context.Context ) {
105+ h .lock .Lock ()
106+ defer h .lock .Unlock ()
107+
108+ h .leader = true
109+
110+ // setGraphBuilt should already have been called when processing the resources on startup because the leader
111+ // election process takes longer than the initial call to HandleEventBatch. Thus, the NGF Pod should be marked as
112+ // ready and have this channel be closed.
113+ close (h .readyCh )
114+ }
0 commit comments