mirror of
https://github.com/zap-proto/telemetry.git
synced 2026-10-30 15:03:28 +00:00
zeekay
8e184a0136
The three signals already shared everything that matters — the same envelope, the same port, the same node handshake, the same convention of carrying the message type in the header flags high byte. What they did not share was a way to be turned on. Each service dialed three times, configured three exporters and installed three providers: three chances to get two right and one wrong, silently, because a provider that was never installed discards every measurement while the code still reads as instrumented. That failure already happened twice in this fleet. Start does all three against one endpoint and returns one Shutdown. The zero value is a noop. An empty Endpoint installs nothing and dials nothing, so a service can call Start unconditionally at the top of main and pick up telemetry when its deployment supplies an endpoint. There is no separate disabled path to keep working — the branch not taken is the one that rots. An unreachable collector does NOT fail Start. Exporters connect lazily and retry; a service that refuses to boot because telemetry is down has made observation a dependency of serving. Signals are attempted independently, so a collector rejecting metrics does not cost the service its traces. Partial telemetry beats none, and the failure is returned rather than swallowed. Endpoints are normalized to host:port. They are widely configured as URLs from the OTLP era, and passing one through dials a host named http — a baffling failure for a config that looks correct. No OTLP, no gRPC, no protobuf in the graph.
114 lines
3.7 KiB
Go
114 lines
3.7 KiB
Go
package telemetry
|
|
|
|
import (
|
|
"context"
|
|
"testing"
|
|
)
|
|
|
|
// The whole value of the zero value is that a service can call Start
|
|
// unconditionally at the top of main. If an empty endpoint returned an error, or
|
|
// a nil Telemetry, every caller would need a branch — and the branch not taken
|
|
// is the one that rots.
|
|
func TestNoEndpointIsInertNotAnError(t *testing.T) {
|
|
tel, err := Start(Config{AppName: "svc"})
|
|
if err != nil {
|
|
t.Fatalf("empty endpoint returned an error: %v", err)
|
|
}
|
|
if tel == nil {
|
|
t.Fatal("Start returned nil; callers must be able to defer Shutdown unconditionally")
|
|
}
|
|
if tel.Enabled() {
|
|
t.Fatal("Enabled() is true with no endpoint")
|
|
}
|
|
if err := tel.Shutdown(context.Background()); err != nil {
|
|
t.Fatalf("Shutdown on a disabled Telemetry: %v", err)
|
|
}
|
|
}
|
|
|
|
// Shutdown runs on paths that may already have shut down (a failed boot, a
|
|
// second signal handler). Panicking there turns a clean exit into a crash.
|
|
func TestShutdownIsIdempotentAndNilSafe(t *testing.T) {
|
|
var nilTel *Telemetry
|
|
if err := nilTel.Shutdown(context.Background()); err != nil {
|
|
t.Fatalf("Shutdown on nil: %v", err)
|
|
}
|
|
tel, _ := Start(Config{AppName: "svc"})
|
|
for i := 0; i < 3; i++ {
|
|
if err := tel.Shutdown(context.Background()); err != nil {
|
|
t.Fatalf("Shutdown #%d: %v", i+1, err)
|
|
}
|
|
}
|
|
}
|
|
|
|
// A missing AppName with a real endpoint is a config mistake worth failing on:
|
|
// every batch would arrive unattributable, which is worse than not arriving.
|
|
func TestEndpointWithoutAppNameIsRejected(t *testing.T) {
|
|
if _, err := Start(Config{Endpoint: "127.0.0.1:4317"}); err == nil {
|
|
t.Fatal("expected an error when Endpoint is set and AppName is not")
|
|
}
|
|
}
|
|
|
|
// Endpoints are widely configured as URLs from the OTLP era. Passing one
|
|
// through dials a host named "http" — a baffling failure for a config that
|
|
// looks correct.
|
|
func TestNormalizeEndpoint(t *testing.T) {
|
|
cases := map[string]string{
|
|
"o11y.hanzo.svc:4317": "o11y.hanzo.svc:4317",
|
|
"http://o11y.hanzo.svc:4317": "o11y.hanzo.svc:4317",
|
|
"https://o11y.hanzo.svc:4317": "o11y.hanzo.svc:4317",
|
|
"http://o11y.hanzo.svc:4317/": "o11y.hanzo.svc:4317",
|
|
"https://o11y.hanzo.svc/v1/tr": "o11y.hanzo.svc",
|
|
" o11y.hanzo.svc:4317 ": "o11y.hanzo.svc:4317",
|
|
"": "",
|
|
" ": "",
|
|
}
|
|
for in, want := range cases {
|
|
if got := normalizeEndpoint(in); got != want {
|
|
t.Errorf("normalizeEndpoint(%q) = %q, want %q", in, got, want)
|
|
}
|
|
}
|
|
}
|
|
|
|
// Naming a subset must not quietly start the others: a service that asked for
|
|
// logs only should not find itself exporting metrics it never wanted.
|
|
func TestSignalSubsetIsHonored(t *testing.T) {
|
|
tel, err := Start(Config{
|
|
AppName: "svc",
|
|
Endpoint: "127.0.0.1:14317", // nothing listening; exporters connect lazily
|
|
Signals: []Signal{Logs},
|
|
})
|
|
if err != nil {
|
|
t.Fatalf("Start: %v", err)
|
|
}
|
|
defer tel.Shutdown(context.Background())
|
|
|
|
if tel.lp == nil {
|
|
t.Fatal("logs were requested but no logger provider was installed")
|
|
}
|
|
if tel.mp != nil {
|
|
t.Fatal("a metric provider was installed though only logs were requested")
|
|
}
|
|
if tel.tracer != nil {
|
|
t.Fatal("a tracer was installed though only logs were requested")
|
|
}
|
|
}
|
|
|
|
// An unreachable collector must not fail Start. Exporters connect lazily and
|
|
// retry; a service that refuses to boot because telemetry is down has made
|
|
// observation a dependency of serving.
|
|
func TestUnreachableCollectorStillStarts(t *testing.T) {
|
|
tel, err := Start(Config{
|
|
AppName: "svc",
|
|
Endpoint: "127.0.0.1:14318",
|
|
})
|
|
if err != nil {
|
|
t.Fatalf("unreachable collector failed Start: %v", err)
|
|
}
|
|
if !tel.Enabled() {
|
|
t.Fatal("Enabled() is false though providers were installed")
|
|
}
|
|
if err := tel.Shutdown(context.Background()); err != nil {
|
|
t.Fatalf("Shutdown: %v", err)
|
|
}
|
|
}
|