zeta1999 / pooly

Connection pool library based on multi-armed bandit algorithms

Pooly

A package inspired by go-hostpool, to manage efficiently pools of connection across multiple hosts. Pooly uses multi-armed bandit algorithms for host selection in order to maximize service connectivity by avoiding faulty hosts.

Currently supported bandit strategies:

• Softmax
• Epsilon-greedy
• Round-robin

Basic Usage

Start two dummy netcat servers.

nc -l -p 1234 &
nc -l -p 12345 &

Get two connections from the "netcat" service using the NetDriver and the RoundRobin strategy (defaults).

package main

import "fmt"
import "github.com/3XX0/pooly"

func ping(s *pooly.Service) error {
	c, err := s.GetConn()
	if err != nil {
		return err
	}
	defer c.Release(&err, pooly.HostUp)

	_, err = c.NetConn().Write([]byte("ping\n"))
	return err
}

func main() {
	s := pooly.NewService("netcat", nil)
	defer s.Close()

	s.Add("127.0.0.1:1234")
	s.Add("127.0.0.1:12345")

	if err := ping(s); err != nil {
		fmt.Println(err)
		return
	}
	if err := ping(s); err != nil {
		fmt.Println(err)
		return
	}
}

Metrics

Pooly exports useful service metrics provided that a statsd server is running and the option ServiceConfig.StatsdAddr is set accordingly.

The following metrics are available:

metric	description
hosts.score	average score of the service hosts (in percentage)
hosts.count	number of hosts registered to the service
conns.count	number of connections spawned by the service
conns.fails	number of connection failures (temporary included)
conns.put.count	number of Release performed
conns.get.count	number of GetConn performed
conns.get.delay	average delay before receiving a connection from the service (in millisecond)
conns.get.fails	number of GetConn failures
conns.active.period	average time during which a connection was active (in millisecond)

Example of a grafana dashboard using vizu

Simulations

In order to test the behavior of different strategies, Monte Carlo simulations may be run through the Go testing framework.

go test -tags 'montecarlo_simulation' -v pooly

Softmax strategy sample output

Epsilon-greedy strategy sample output