ipfs-cluster/peer_monitor.go
Hector Sanjuan 6ee0f3bead Issue #45: Detect expired metrics and trigger re-pins
An initial, simple approach to this. The PeerMonitor will
check it's metrics, compare to the current set of peers and put
an alert in the alerts channel if the metrics for a peer have expired.

Cluster reads this channel looking for "ping" alerts. The leader
is in charge of triggering repins in all the Cids allocated to
a given peer.

Also, metrics are now broadcasted to the cluster instead of pushed only
to the leader. Since they happen every few seconds it should be okay
regarding how it scales. Main problem was that if the leader is the node
going down, the new leader will not now about it as it doesn't have any
metrics for it, so it won't trigger an alert. If it acted on that then
the component needs to know it is the leader, or cluster needs to
handle alerts in complicated ways when leadership changes. Detecting
leadership changes or letting a component know who is the leader is another
dependency from the consensus algorithm that should be avoided. Therefore
we broadcast, for the moment.

License: MIT
Signed-off-by: Hector Sanjuan <hector@protocol.ai>
2017-03-02 14:59:45 +01:00

301 lines
6.5 KiB
Go

package ipfscluster
import (
"context"
"errors"
"sync"
"time"
rpc "github.com/hsanjuan/go-libp2p-gorpc"
peer "github.com/libp2p/go-libp2p-peer"
"github.com/ipfs/ipfs-cluster/api"
)
// AlertChannelCap specifies how much buffer the alerts channel has.
var AlertChannelCap = 256
// WindowCap specifies how many metrics to keep for given host and metric type
var WindowCap = 10
// peerMetrics is just a circular queue
type peerMetrics struct {
last int
window []api.Metric
// mux sync.RWMutex
}
func newPeerMetrics(windowCap int) *peerMetrics {
w := make([]api.Metric, 0, windowCap)
return &peerMetrics{0, w}
}
func (pmets *peerMetrics) add(m api.Metric) {
// pmets.mux.Lock()
// defer pmets.mux.Unlock()
if len(pmets.window) < cap(pmets.window) {
pmets.window = append(pmets.window, m)
pmets.last = len(pmets.window) - 1
return
}
// len == cap
pmets.last = (pmets.last + 1) % cap(pmets.window)
pmets.window[pmets.last] = m
return
}
func (pmets *peerMetrics) latest() (api.Metric, error) {
// pmets.mux.RLock()
// defer pmets.mux.RUnlock()
if len(pmets.window) == 0 {
return api.Metric{}, errors.New("no metrics")
}
return pmets.window[pmets.last], nil
}
// ordered from newest to oldest
func (pmets *peerMetrics) all() []api.Metric {
// pmets.mux.RLock()
// pmets.mux.RUnlock()
wlen := len(pmets.window)
res := make([]api.Metric, 0, wlen)
if wlen == 0 {
return res
}
for i := pmets.last; i >= 0; i-- {
res = append(res, pmets.window[i])
}
for i := wlen; i > pmets.last; i-- {
res = append(res, pmets.window[i])
}
return res
}
type metricsByPeer map[peer.ID]*peerMetrics
// StdPeerMonitor is a component in charge of monitoring peers, logging
// metrics and detecting failures
type StdPeerMonitor struct {
ctx context.Context
cancel func()
rpcClient *rpc.Client
rpcReady chan struct{}
metrics map[string]metricsByPeer
metricsMux sync.RWMutex
windowCap int
alerts chan api.Alert
monitoringInterval int
shutdownLock sync.Mutex
shutdown bool
wg sync.WaitGroup
}
// NewStdPeerMonitor creates a new monitor. It receives the window capacity
// (how many metrics to keep for each peer and type of metric) and the
// monitoringInterval (interval between the checks that produce alerts)
// as parameters
func NewStdPeerMonitor(cfg *Config) *StdPeerMonitor {
if WindowCap <= 0 {
panic("windowCap too small")
}
ctx, cancel := context.WithCancel(context.Background())
mon := &StdPeerMonitor{
ctx: ctx,
cancel: cancel,
rpcReady: make(chan struct{}, 1),
metrics: make(map[string]metricsByPeer),
windowCap: WindowCap,
alerts: make(chan api.Alert, AlertChannelCap),
monitoringInterval: cfg.MonitoringIntervalSeconds,
}
go mon.run()
return mon
}
func (mon *StdPeerMonitor) run() {
select {
case <-mon.rpcReady:
go mon.monitor()
case <-mon.ctx.Done():
}
}
// SetClient saves the given rpc.Client for later use
func (mon *StdPeerMonitor) SetClient(c *rpc.Client) {
mon.rpcClient = c
mon.rpcReady <- struct{}{}
}
// Shutdown stops the peer monitor. It particular, it will
// not deliver any alerts.
func (mon *StdPeerMonitor) Shutdown() error {
mon.shutdownLock.Lock()
defer mon.shutdownLock.Unlock()
if mon.shutdown {
logger.Warning("StdPeerMonitor already shut down")
return nil
}
logger.Info("stopping StdPeerMonitor")
close(mon.rpcReady)
mon.cancel()
mon.wg.Wait()
mon.shutdown = true
return nil
}
// LogMetric stores a metric so it can later be retrieved.
func (mon *StdPeerMonitor) LogMetric(m api.Metric) {
mon.metricsMux.Lock()
defer mon.metricsMux.Unlock()
name := m.Name
peer := m.Peer
mbyp, ok := mon.metrics[name]
if !ok {
mbyp = make(metricsByPeer)
mon.metrics[name] = mbyp
}
pmets, ok := mbyp[peer]
if !ok {
pmets = newPeerMetrics(mon.windowCap)
mbyp[peer] = pmets
}
logger.Debugf("logged '%s' metric from '%s'", name, peer)
pmets.add(m)
}
// func (mon *StdPeerMonitor) getLastMetric(name string, p peer.ID) api.Metric {
// mon.metricsMux.RLock()
// defer mon.metricsMux.RUnlock()
// emptyMetric := api.Metric{
// Name: name,
// Peer: p,
// Valid: false,
// }
// mbyp, ok := mon.metrics[name]
// if !ok {
// return emptyMetric
// }
// pmets, ok := mbyp[p]
// if !ok {
// return emptyMetric
// }
// metric, err := pmets.latest()
// if err != nil {
// return emptyMetric
// }
// return metric
// }
// LastMetrics returns last known VALID metrics of a given type
func (mon *StdPeerMonitor) LastMetrics(name string) []api.Metric {
mon.metricsMux.RLock()
defer mon.metricsMux.RUnlock()
mbyp, ok := mon.metrics[name]
if !ok {
return []api.Metric{}
}
metrics := make([]api.Metric, 0, len(mbyp))
for _, peerMetrics := range mbyp {
last, err := peerMetrics.latest()
if err != nil || last.Discard() {
continue
}
metrics = append(metrics, last)
}
return metrics
}
// Alerts returns a channel on which alerts are sent when the
// monitor detects a failure.
func (mon *StdPeerMonitor) Alerts() <-chan api.Alert {
return mon.alerts
}
func (mon *StdPeerMonitor) monitor() {
ticker := time.NewTicker(time.Second * time.Duration(mon.monitoringInterval))
for {
select {
case <-ticker.C:
logger.Debug("monitoring tick")
// Get current peers
var peers []peer.ID
err := mon.rpcClient.Call("",
"Cluster",
"PeerManagerPeers",
struct{}{},
&peers)
if err != nil {
logger.Error(err)
break
}
for k := range mon.metrics {
logger.Debug("check metrics ", k)
mon.checkMetrics(peers, k)
}
case <-mon.ctx.Done():
ticker.Stop()
return
}
}
}
// This is probably the place to implement some advanced ways of detecting down
// peers.
// Currently easy logic, just check that all peers have a valid metric.
func (mon *StdPeerMonitor) checkMetrics(peers []peer.ID, metricName string) {
mon.metricsMux.RLock()
defer mon.metricsMux.RUnlock()
// get metric windows for peers
metricsByPeer := mon.metrics[metricName]
// for each of the given current peers
for _, p := range peers {
// get metrics for that peer
pMetrics, ok := metricsByPeer[p]
if !ok { // no metrics from this peer
continue
}
last, err := pMetrics.latest()
if err != nil { // no metrics for this peer
continue
}
// send alert if metric is expired (but was valid at some point)
if last.Valid && last.Expired() {
mon.sendAlert(p, metricName)
}
}
}
func (mon *StdPeerMonitor) sendAlert(p peer.ID, metricName string) {
alrt := api.Alert{
Peer: p,
MetricName: metricName,
}
select {
case mon.alerts <- alrt:
default:
logger.Error("alert channel is full")
}
}