feat: add initial support for metrics

[enocom]

This commit includes two metrics:

1. a metric that reports dial latency
2. a metric that reports open connections

Fixes #15.

[enocom]

Here's what I've been using to test these metrics:

package main

import (
	"context"
	"fmt"
	"log"
	"net"
	"os"
	"sync"
	"time"

	"cloud.google.com/go/cloudsqlconn"
	"contrib.go.opencensus.io/exporter/stackdriver"
	"github.com/jackc/pgx/v4/pgxpool"
	"go.opencensus.io/stats/view"
)

type customMetricsExporter struct{}

func (ce *customMetricsExporter) ExportView(vd *view.Data) {
	log.Printf("vd.View: %+v\n%#v\n", vd.View, vd.Rows)
	for i, row := range vd.Rows {
		log.Printf("\tRow: %#d: %#v", i, row)

		log.Printf("%#v", row.Data)
	}
}

func main() {
	ctx := context.Background()
	pgUser := os.Getenv("POSTGRES_USER")
	pgPass := os.Getenv("POSTGRES_PASS")
	pgDB := os.Getenv("POSTGRES_DB")
	connName := os.Getenv("POSTGRES_CONNECTION_NAME")

	sd, err := stackdriver.NewExporter(stackdriver.Options{
		ProjectID:         "MY_COOL_PROJECT_ID",
		ReportingInterval: 60 * time.Second,
	})
	if err != nil {
		log.Fatalf("Failed to create the Stackdriver exporter: %v", err)
	}
	// It is imperative to invoke flush before your main function exits
	defer sd.Flush()

	// Start the metrics exporter
	sd.StartMetricsExporter()
	defer sd.StopMetricsExporter()

        // uncomment the following line and comment-out the above exporter to use a console exporter
	// view.RegisterExporter(new(customMetricsExporter))

	// Configure the driver to connect to the database
	dsn := fmt.Sprintf("user=%s password=%s dbname=%s sslmode=disable", pgUser, pgPass, pgDB)
	config, err := pgxpool.ParseConfig(dsn)
	if err != nil {
		log.Fatalf("failed to parse pgx config: %v", err)
	}

	d, err := cloudsqlconn.NewDialer(context.Background())
	if err != nil {
		panic(err)
	}
	// Tell the driver to use the Cloud SQL Go Connector to create connections
	config.ConnConfig.DialFunc = func(ctx context.Context, network string, instance string) (net.Conn, error) {
		return d.Dial(ctx, connName)
	}
	defer d.Close()

	// Interact with the driver directly as you normally would
	conn, connErr := pgxpool.ConnectConfig(ctx, config)
	if connErr != nil {
		log.Fatalf("failed to connect: %s", connErr)
		return
	}

	var wg sync.WaitGroup
	for i := 0; i < 1000; i++ {
		wg.Add(1)
		go func() {
			var t time.Time
			if err := conn.QueryRow(ctx, "select now()").Scan(&t); err != nil {
				log.Println("error", err.Error())
			}
			log.Println(t)
			wg.Done()
		}()
	}
	wg.Wait()
	conn.Close()
	select {}
}

[enocom]

Here's what these metrics look like in action:

And here's a heatmap version of the dial latency:

[enocom]

Aside from the two open issues above, this is ready for another review. Notable in the recent changes: I've dropped the atomic counter management in favor of an accurate and simpler sum aggregation.

[enocom]

Here's what the unique ID (a UUID) per dialer looks like.

[enocom]

And here's trace with the dialer ID too.