host / runtime metrics broken in prometheus

[corani]

With the new Metrics API (v0.32.0), we somewhere ends up calling https://github.com/prometheus/common/blob/main/model/metric.go#L92 which rejects any metrics with dots such as the ones created by by the host and runtime instrumentation (e.g. process.cpu.time).

As a result, every time when Prometheus collects the metrics my logs are being spammed full with "xxx is not a valid metric name" errors (directly to stderr, which is another issue) and those metrics aren't returned to Prometheus.

Note: prior to the new Metrics API this worked (v0.31.0).

An example to reproduce this can be found in the comment below.

[corani]

To reproduce:

package main

import (
    "context"
    "fmt"
    "net/http"
    "net/http/httptest"

    "github.com/prometheus/client_golang/prometheus"
    "github.com/prometheus/client_golang/prometheus/promhttp"
    "go.opentelemetry.io/contrib/instrumentation/host"
    otelprom "go.opentelemetry.io/otel/exporters/prometheus"
    "go.opentelemetry.io/otel/sdk/metric"
)

func main() {
    exporter := otelprom.New()
    provider := metric.NewMeterProvider(metric.WithReader(exporter))

    registry := prometheus.NewRegistry()
    if err := registry.Register(exporter.Collector); err != nil {
        panic(err)
    }

    // this generates metrics that are rejected by the prometheus module
    if err := host.Start(host.WithMeterProvider(provider)); err != nil {
        panic(err)
    }

    // similar for `runtime` instrumentation

    // this metric does work
    counter, err := provider.Meter("my_meter").SyncInt64().Counter("my_counter")
    if err != nil {
        panic(err)
    }
    counter.Add(context.Background(), 1)

    handler := promhttp.HandlerFor(registry, promhttp.HandlerOpts{})

    w := httptest.NewRecorder()
    r := httptest.NewRequest(http.MethodGet, "/", nil)

    handler.ServeHTTP(w, r)

    fmt.Println("body:\n", w.Body.String())
}

Output:

$ go run main.go
2022/09/21 11:56:30 "system.network.io" is not a valid metric name
2022/09/21 11:56:30 "system.network.io" is not a valid metric name
2022/09/21 11:56:30 "process.cpu.time" is not a valid metric name
2022/09/21 11:56:30 "process.cpu.time" is not a valid metric name
2022/09/21 11:56:30 "system.cpu.time" is not a valid metric name
2022/09/21 11:56:30 "system.cpu.time" is not a valid metric name
2022/09/21 11:56:30 "system.cpu.time" is not a valid metric name
2022/09/21 11:56:30 "system.cpu.time" is not a valid metric name
2022/09/21 11:56:30 "system.memory.usage" is not a valid metric name
2022/09/21 11:56:30 "system.memory.usage" is not a valid metric name
2022/09/21 11:56:30 "system.memory.utilization" is not a valid metric name
2022/09/21 11:56:30 "system.memory.utilization" is not a valid metric name
body:
 # HELP my_counter
# TYPE my_counter counter
my_counter 1

[MrAlias]

Thanks for the bug report. This is currently identified and tracked by open-telemetry/opentelemetry-go#3183. Closing this as it is a duplicate. Please follow the other issue to track the resolution.