grpc-go/xds/internal/xdsclient/load/store.go

/*
 * Copyright 2020 gRPC authors.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

// Package load provides functionality to record and maintain load data.
package load

import (
	"sync"
	"sync/atomic"
	"time"
)

const negativeOneUInt64 = ^uint64(0)

// Store keeps the loads for multiple clusters and services to be reported via
// LRS. It contains loads to reported to one LRS server. Create multiple stores
// for multiple servers.
//
// It is safe for concurrent use.
type Store struct {
	// mu only protects the map (2 layers). The read/write to *perClusterStore
	// doesn't need to hold the mu.
	mu sync.Mutex
	// clusters is a map with cluster name as the key. The second layer is a map
	// with service name as the key. Each value (perClusterStore) contains data
	// for a (cluster, service) pair.
	//
	// Note that new entries are added to this map, but never removed. This is
	// potentially a memory leak. But the memory is allocated for each new
	// (cluster,service) pair, and the memory allocated is just pointers and
	// maps. So this shouldn't get too bad.
	clusters map[string]map[string]*perClusterStore
}

// NewStore creates a Store.
func NewStore() *Store {
	return &Store{
		clusters: make(map[string]map[string]*perClusterStore),
	}
}

// Stats returns the load data for the given cluster names. Data is returned in
// a slice with no specific order.
//
// If no clusterName is given (an empty slice), all data for all known clusters
// is returned.
//
// If a cluster's Data is empty (no load to report), it's not appended to the
// returned slice.
func (s *Store) Stats(clusterNames []string) []*Data {
	var ret []*Data
	s.mu.Lock()
	defer s.mu.Unlock()

	if len(clusterNames) == 0 {
		for _, c := range s.clusters {
			ret = appendClusterStats(ret, c)
		}
		return ret
	}

	for _, n := range clusterNames {
		if c, ok := s.clusters[n]; ok {
			ret = appendClusterStats(ret, c)
		}
	}
	return ret
}

// appendClusterStats gets Data for the given cluster, append to ret, and return
// the new slice.
//
// Data is only appended to ret if it's not empty.
func appendClusterStats(ret []*Data, cluster map[string]*perClusterStore) []*Data {
	for _, d := range cluster {
		data := d.stats()
		if data == nil {
			// Skip this data if it doesn't contain any information.
			continue
		}
		ret = append(ret, data)
	}
	return ret
}

// PerCluster returns the perClusterStore for the given clusterName +
// serviceName.
func (s *Store) PerCluster(clusterName, serviceName string) PerClusterReporter {
	if s == nil {
		return nil
	}

	s.mu.Lock()
	defer s.mu.Unlock()
	c, ok := s.clusters[clusterName]
	if !ok {
		c = make(map[string]*perClusterStore)
		s.clusters[clusterName] = c
	}

	if p, ok := c[serviceName]; ok {
		return p
	}
	p := &perClusterStore{
		cluster: clusterName,
		service: serviceName,
	}
	c[serviceName] = p
	return p
}

// perClusterStore is a repository for LB policy implementations to report store
// load data. It contains load for a (cluster, edsService) pair.
//
// It is safe for concurrent use.
//
// TODO(easwars): Use regular maps with mutexes instead of sync.Map here. The
// latter is optimized for two common use cases: (1) when the entry for a given
// key is only ever written once but read many times, as in caches that only
// grow, or (2) when multiple goroutines read, write, and overwrite entries for
// disjoint sets of keys. In these two cases, use of a Map may significantly
// reduce lock contention compared to a Go map paired with a separate Mutex or
// RWMutex.
// Neither of these conditions are met here, and we should transition to a
// regular map with a mutex for better type safety.
type perClusterStore struct {
	cluster, service string
	drops            sync.Map // map[string]*uint64
	localityRPCCount sync.Map // map[string]*rpcCountData

	mu               sync.Mutex
	lastLoadReportAt time.Time
}

// Update functions are called by picker for each RPC. To avoid contention, all
// updates are done atomically.

// CallDropped adds one drop record with the given category to store.
func (ls *perClusterStore) CallDropped(category string) {
	if ls == nil {
		return
	}

	p, ok := ls.drops.Load(category)
	if !ok {
		tp := new(uint64)
		p, _ = ls.drops.LoadOrStore(category, tp)
	}
	atomic.AddUint64(p.(*uint64), 1)
}

// CallStarted adds one call started record for the given locality.
func (ls *perClusterStore) CallStarted(locality string) {
	if ls == nil {
		return
	}

	p, ok := ls.localityRPCCount.Load(locality)
	if !ok {
		tp := newRPCCountData()
		p, _ = ls.localityRPCCount.LoadOrStore(locality, tp)
	}
	p.(*rpcCountData).incrInProgress()
	p.(*rpcCountData).incrIssued()
}

// CallFinished adds one call finished record for the given locality.
// For successful calls, err needs to be nil.
func (ls *perClusterStore) CallFinished(locality string, err error) {
	if ls == nil {
		return
	}

	p, ok := ls.localityRPCCount.Load(locality)
	if !ok {
		// The map is never cleared, only values in the map are reset. So the
		// case where entry for call-finish is not found should never happen.
		return
	}
	p.(*rpcCountData).decrInProgress()
	if err == nil {
		p.(*rpcCountData).incrSucceeded()
	} else {
		p.(*rpcCountData).incrErrored()
	}
}

// CallServerLoad adds one server load record for the given locality. The
// load type is specified by desc, and its value by val.
func (ls *perClusterStore) CallServerLoad(locality, name string, d float64) {
	if ls == nil {
		return
	}

	p, ok := ls.localityRPCCount.Load(locality)
	if !ok {
		// The map is never cleared, only values in the map are reset. So the
		// case where entry for callServerLoad is not found should never happen.
		return
	}
	p.(*rpcCountData).addServerLoad(name, d)
}

// Data contains all load data reported to the Store since the most recent call
// to stats().
type Data struct {
	// Cluster is the name of the cluster this data is for.
	Cluster string
	// Service is the name of the EDS service this data is for.
	Service string
	// TotalDrops is the total number of dropped requests.
	TotalDrops uint64
	// Drops is the number of dropped requests per category.
	Drops map[string]uint64
	// LocalityStats contains load reports per locality.
	LocalityStats map[string]LocalityData
	// ReportInternal is the duration since last time load was reported (stats()
	// was called).
	ReportInterval time.Duration
}

// LocalityData contains load data for a single locality.
type LocalityData struct {
	// RequestStats contains counts of requests made to the locality.
	RequestStats RequestData
	// LoadStats contains server load data for requests made to the locality,
	// indexed by the load type.
	LoadStats map[string]ServerLoadData
}

// RequestData contains request counts.
type RequestData struct {
	// Succeeded is the number of succeeded requests.
	Succeeded uint64
	// Errored is the number of requests which ran into errors.
	Errored uint64
	// InProgress is the number of requests in flight.
	InProgress uint64
	// Issued is the total number requests that were sent.
	Issued uint64
}

// ServerLoadData contains server load data.
type ServerLoadData struct {
	// Count is the number of load reports.
	Count uint64
	// Sum is the total value of all load reports.
	Sum float64
}

func newData(cluster, service string) *Data {
	return &Data{
		Cluster:       cluster,
		Service:       service,
		Drops:         make(map[string]uint64),
		LocalityStats: make(map[string]LocalityData),
	}
}

// stats returns and resets all loads reported to the store, except inProgress
// rpc counts.
//
// It returns nil if the store doesn't contain any (new) data.
func (ls *perClusterStore) stats() *Data {
	if ls == nil {
		return nil
	}

	sd := newData(ls.cluster, ls.service)
	ls.drops.Range(func(key, val any) bool {
		d := atomic.SwapUint64(val.(*uint64), 0)
		if d == 0 {
			return true
		}
		sd.TotalDrops += d
		keyStr := key.(string)
		if keyStr != "" {
			// Skip drops without category. They are counted in total_drops, but
			// not in per category. One example is drops by circuit breaking.
			sd.Drops[keyStr] = d
		}
		return true
	})
	ls.localityRPCCount.Range(func(key, val any) bool {
		countData := val.(*rpcCountData)
		succeeded := countData.loadAndClearSucceeded()
		inProgress := countData.loadInProgress()
		errored := countData.loadAndClearErrored()
		issued := countData.loadAndClearIssued()
		if succeeded == 0 && inProgress == 0 && errored == 0 && issued == 0 {
			return true
		}

		ld := LocalityData{
			RequestStats: RequestData{
				Succeeded:  succeeded,
				Errored:    errored,
				InProgress: inProgress,
				Issued:     issued,
			},
			LoadStats: make(map[string]ServerLoadData),
		}
		countData.serverLoads.Range(func(key, val any) bool {
			sum, count := val.(*rpcLoadData).loadAndClear()
			if count == 0 {
				return true
			}
			ld.LoadStats[key.(string)] = ServerLoadData{
				Count: count,
				Sum:   sum,
			}
			return true
		})
		sd.LocalityStats[key.(string)] = ld
		return true
	})

	ls.mu.Lock()
	sd.ReportInterval = time.Since(ls.lastLoadReportAt)
	ls.lastLoadReportAt = time.Now()
	ls.mu.Unlock()

	if sd.TotalDrops == 0 && len(sd.Drops) == 0 && len(sd.LocalityStats) == 0 {
		return nil
	}
	return sd
}

type rpcCountData struct {
	// Only atomic accesses are allowed for the fields.
	succeeded  *uint64
	errored    *uint64
	inProgress *uint64
	issued     *uint64

	// Map from load desc to load data (sum+count). Loading data from map is
	// atomic, but updating data takes a lock, which could cause contention when
	// multiple RPCs try to report loads for the same desc.
	//
	// To fix the contention, shard this map.
	serverLoads sync.Map // map[string]*rpcLoadData
}

func newRPCCountData() *rpcCountData {
	return &rpcCountData{
		succeeded:  new(uint64),
		errored:    new(uint64),
		inProgress: new(uint64),
		issued:     new(uint64),
	}
}

func (rcd *rpcCountData) incrSucceeded() {
	atomic.AddUint64(rcd.succeeded, 1)
}

func (rcd *rpcCountData) loadAndClearSucceeded() uint64 {
	return atomic.SwapUint64(rcd.succeeded, 0)
}

func (rcd *rpcCountData) incrErrored() {
	atomic.AddUint64(rcd.errored, 1)
}

func (rcd *rpcCountData) loadAndClearErrored() uint64 {
	return atomic.SwapUint64(rcd.errored, 0)
}

func (rcd *rpcCountData) incrInProgress() {
	atomic.AddUint64(rcd.inProgress, 1)
}

func (rcd *rpcCountData) decrInProgress() {
	atomic.AddUint64(rcd.inProgress, negativeOneUInt64) // atomic.Add(x, -1)
}

func (rcd *rpcCountData) loadInProgress() uint64 {
	return atomic.LoadUint64(rcd.inProgress) // InProgress count is not clear when reading.
}

func (rcd *rpcCountData) incrIssued() {
	atomic.AddUint64(rcd.issued, 1)
}

func (rcd *rpcCountData) loadAndClearIssued() uint64 {
	return atomic.SwapUint64(rcd.issued, 0)
}

func (rcd *rpcCountData) addServerLoad(name string, d float64) {
	loads, ok := rcd.serverLoads.Load(name)
	if !ok {
		tl := newRPCLoadData()
		loads, _ = rcd.serverLoads.LoadOrStore(name, tl)
	}
	loads.(*rpcLoadData).add(d)
}

// Data for server loads (from trailers or oob). Fields in this struct must be
// updated consistently.
//
// The current solution is to hold a lock, which could cause contention. To fix,
// shard serverLoads map in rpcCountData.
type rpcLoadData struct {
	mu    sync.Mutex
	sum   float64
	count uint64
}

func newRPCLoadData() *rpcLoadData {
	return &rpcLoadData{}
}

func (rld *rpcLoadData) add(v float64) {
	rld.mu.Lock()
	rld.sum += v
	rld.count++
	rld.mu.Unlock()
}

func (rld *rpcLoadData) loadAndClear() (s float64, c uint64) {
	rld.mu.Lock()
	s = rld.sum
	rld.sum = 0
	c = rld.count
	rld.count = 0
	rld.mu.Unlock()
	return
}