sdk/go/health/aggregator.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: Apache-2.0
   4
   5 package health
   6
   7 import (
   8         "context"
   9         "encoding/json"
  10         "errors"
  11         "fmt"
  12         "net/http"
  13         "net/url"
  14         "sync"
  15         "time"
  16
  17         "git.curoverse.com/arvados.git/sdk/go/arvados"
  18         "git.curoverse.com/arvados.git/sdk/go/auth"
  19 )
  20
  21 const defaultTimeout = arvados.Duration(2 * time.Second)
  22
  23 // Aggregator implements http.Handler. It handles "GET /_health/all"
  24 // by checking the health of all configured services on the cluster
  25 // and responding 200 if everything is healthy.
  26 type Aggregator struct {
  27         setupOnce  sync.Once
  28         httpClient *http.Client
  29         timeout    arvados.Duration
  30
  31         Config *arvados.Config
  32
  33         // If non-nil, Log is called after handling each request.
  34         Log func(*http.Request, error)
  35 }
  36
  37 func (agg *Aggregator) setup() {
  38         agg.httpClient = http.DefaultClient
  39         if agg.timeout == 0 {
  40                 // this is always the case, except in the test suite
  41                 agg.timeout = defaultTimeout
  42         }
  43 }
  44
  45 func (agg *Aggregator) ServeHTTP(resp http.ResponseWriter, req *http.Request) {
  46         agg.setupOnce.Do(agg.setup)
  47         sendErr := func(statusCode int, err error) {
  48                 resp.WriteHeader(statusCode)
  49                 json.NewEncoder(resp).Encode(map[string]string{"error": err.Error()})
  50                 if agg.Log != nil {
  51                         agg.Log(req, err)
  52                 }
  53         }
  54
  55         resp.Header().Set("Content-Type", "application/json")
  56
  57         cluster, err := agg.Config.GetCluster("")
  58         if err != nil {
  59                 err = fmt.Errorf("arvados.GetCluster(): %s", err)
  60                 sendErr(http.StatusInternalServerError, err)
  61                 return
  62         }
  63         if !agg.checkAuth(req, cluster) {
  64                 sendErr(http.StatusUnauthorized, errUnauthorized)
  65                 return
  66         }
  67         if req.URL.Path != "/_health/all" {
  68                 sendErr(http.StatusNotFound, errNotFound)
  69                 return
  70         }
  71         json.NewEncoder(resp).Encode(agg.ClusterHealth(cluster))
  72         if agg.Log != nil {
  73                 agg.Log(req, nil)
  74         }
  75 }
  76
  77 type ClusterHealthResponse struct {
  78         // "OK" if all needed services are OK, otherwise "ERROR".
  79         Health string `json:"health"`
  80
  81         // An entry for each known health check of each known instance
  82         // of each needed component: "instance of service S on node N
  83         // reports health-check C is OK."
  84         Checks map[string]CheckResult `json:"checks"`
  85
  86         // An entry for each service type: "service S is OK." This
  87         // exposes problems that can't be expressed in Checks, like
  88         // "service S is needed, but isn't configured to run
  89         // anywhere."
  90         Services map[arvados.ServiceName]ServiceHealth `json:"services"`
  91 }
  92
  93 type CheckResult struct {
  94         Health         string                 `json:"health"`
  95         Error          string                 `json:"error,omitempty"`
  96         HTTPStatusCode int                    `json:",omitempty"`
  97         HTTPStatusText string                 `json:",omitempty"`
  98         Response       map[string]interface{} `json:"response"`
  99         ResponseTime   json.Number            `json:"responseTime"`
 100 }
 101
 102 type ServiceHealth struct {
 103         Health string `json:"health"`
 104         N      int    `json:"n"`
 105 }
 106
 107 func (agg *Aggregator) ClusterHealth(cluster *arvados.Cluster) ClusterHealthResponse {
 108         resp := ClusterHealthResponse{
 109                 Health:   "OK",
 110                 Checks:   make(map[string]CheckResult),
 111                 Services: make(map[arvados.ServiceName]ServiceHealth),
 112         }
 113
 114         mtx := sync.Mutex{}
 115         wg := sync.WaitGroup{}
 116         for svcName, svc := range cluster.Services.Map() {
 117                 // Ensure svc is listed in resp.Services.
 118                 mtx.Lock()
 119                 if _, ok := resp.Services[svcName]; !ok {
 120                         resp.Services[svcName] = ServiceHealth{Health: "ERROR"}
 121                 }
 122                 mtx.Unlock()
 123
 124                 for addr := range svc.InternalURLs {
 125                         wg.Add(1)
 126                         go func(svcName arvados.ServiceName, addr arvados.URL) {
 127                                 defer wg.Done()
 128                                 var result CheckResult
 129                                 pingURL, err := agg.pingURL(addr)
 130                                 if err != nil {
 131                                         result = CheckResult{
 132                                                 Health: "ERROR",
 133                                                 Error:  err.Error(),
 134                                         }
 135                                 } else {
 136                                         result = agg.ping(pingURL, cluster)
 137                                 }
 138
 139                                 mtx.Lock()
 140                                 defer mtx.Unlock()
 141                                 resp.Checks[fmt.Sprintf("%s+%s", svcName, pingURL)] = result
 142                                 if result.Health == "OK" {
 143                                         h := resp.Services[svcName]
 144                                         h.N++
 145                                         h.Health = "OK"
 146                                         resp.Services[svcName] = h
 147                                 } else {
 148                                         resp.Health = "ERROR"
 149                                 }
 150                         }(svcName, addr)
 151                 }
 152         }
 153         wg.Wait()
 154
 155         // Report ERROR if a needed service didn't fail any checks
 156         // merely because it isn't configured to run anywhere.
 157         for _, sh := range resp.Services {
 158                 if sh.Health != "OK" {
 159                         resp.Health = "ERROR"
 160                         break
 161                 }
 162         }
 163         return resp
 164 }
 165
 166 func (agg *Aggregator) pingURL(svcURL arvados.URL) (*url.URL, error) {
 167         base := url.URL(svcURL)
 168         return base.Parse("/_health/ping")
 169 }
 170
 171 func (agg *Aggregator) ping(target *url.URL, cluster *arvados.Cluster) (result CheckResult) {
 172         t0 := time.Now()
 173
 174         var err error
 175         defer func() {
 176                 result.ResponseTime = json.Number(fmt.Sprintf("%.6f", time.Since(t0).Seconds()))
 177                 if err != nil {
 178                         result.Health, result.Error = "ERROR", err.Error()
 179                 } else {
 180                         result.Health = "OK"
 181                 }
 182         }()
 183
 184         req, err := http.NewRequest("GET", target.String(), nil)
 185         if err != nil {
 186                 return
 187         }
 188         req.Header.Set("Authorization", "Bearer "+cluster.ManagementToken)
 189
 190         ctx, cancel := context.WithTimeout(req.Context(), time.Duration(agg.timeout))
 191         defer cancel()
 192         req = req.WithContext(ctx)
 193         resp, err := agg.httpClient.Do(req)
 194         if err != nil {
 195                 return
 196         }
 197         result.HTTPStatusCode = resp.StatusCode
 198         result.HTTPStatusText = resp.Status
 199         err = json.NewDecoder(resp.Body).Decode(&result.Response)
 200         if err != nil {
 201                 err = fmt.Errorf("cannot decode response: %s", err)
 202         } else if resp.StatusCode != http.StatusOK {
 203                 err = fmt.Errorf("HTTP %d %s", resp.StatusCode, resp.Status)
 204         } else if h, _ := result.Response["health"].(string); h != "OK" {
 205                 if e, ok := result.Response["error"].(string); ok && e != "" {
 206                         err = errors.New(e)
 207                 } else {
 208                         err = fmt.Errorf("health=%q in ping response", h)
 209                 }
 210         }
 211         return
 212 }
 213
 214 func (agg *Aggregator) checkAuth(req *http.Request, cluster *arvados.Cluster) bool {
 215         creds := auth.CredentialsFromRequest(req)
 216         for _, token := range creds.Tokens {
 217                 if token != "" && token == cluster.ManagementToken {
 218                         return true
 219                 }
 220         }
 221         return false
 222 }