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         "crypto/tls"
  10         "encoding/json"
  11         "fmt"
  12         "net/http"
  13         "net/url"
  14         "sync"
  15         "time"
  16
  17         "git.arvados.org/arvados.git/sdk/go/arvados"
  18         "git.arvados.org/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         Cluster *arvados.Cluster
  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.Client{
  39                 Transport: &http.Transport{
  40                         TLSClientConfig: &tls.Config{
  41                                 InsecureSkipVerify: agg.Cluster.TLS.Insecure,
  42                         },
  43                 },
  44         }
  45         if agg.timeout == 0 {
  46                 // this is always the case, except in the test suite
  47                 agg.timeout = defaultTimeout
  48         }
  49 }
  50
  51 func (agg *Aggregator) CheckHealth() error {
  52         return nil
  53 }
  54
  55 func (agg *Aggregator) Done() <-chan struct{} {
  56         return nil
  57 }
  58
  59 func (agg *Aggregator) ServeHTTP(resp http.ResponseWriter, req *http.Request) {
  60         agg.setupOnce.Do(agg.setup)
  61         sendErr := func(statusCode int, err error) {
  62                 resp.WriteHeader(statusCode)
  63                 json.NewEncoder(resp).Encode(map[string]string{"error": err.Error()})
  64                 if agg.Log != nil {
  65                         agg.Log(req, err)
  66                 }
  67         }
  68
  69         resp.Header().Set("Content-Type", "application/json")
  70
  71         if !agg.checkAuth(req) {
  72                 sendErr(http.StatusUnauthorized, errUnauthorized)
  73                 return
  74         }
  75         if req.URL.Path == "/_health/all" {
  76                 json.NewEncoder(resp).Encode(agg.ClusterHealth())
  77         } else if req.URL.Path == "/_health/ping" {
  78                 resp.Write(healthyBody)
  79         } else {
  80                 sendErr(http.StatusNotFound, errNotFound)
  81                 return
  82         }
  83         if agg.Log != nil {
  84                 agg.Log(req, nil)
  85         }
  86 }
  87
  88 type ClusterHealthResponse struct {
  89         // "OK" if all needed services are OK, otherwise "ERROR".
  90         Health string `json:"health"`
  91
  92         // An entry for each known health check of each known instance
  93         // of each needed component: "instance of service S on node N
  94         // reports health-check C is OK."
  95         Checks map[string]CheckResult `json:"checks"`
  96
  97         // An entry for each service type: "service S is OK." This
  98         // exposes problems that can't be expressed in Checks, like
  99         // "service S is needed, but isn't configured to run
 100         // anywhere."
 101         Services map[arvados.ServiceName]ServiceHealth `json:"services"`
 102 }
 103
 104 type CheckResult struct {
 105         Health         string                 `json:"health"`
 106         Error          string                 `json:"error,omitempty"`
 107         HTTPStatusCode int                    `json:",omitempty"`
 108         HTTPStatusText string                 `json:",omitempty"`
 109         Response       map[string]interface{} `json:"response"`
 110         ResponseTime   json.Number            `json:"responseTime"`
 111 }
 112
 113 type ServiceHealth struct {
 114         Health string `json:"health"`
 115         N      int    `json:"n"`
 116 }
 117
 118 func (agg *Aggregator) ClusterHealth() ClusterHealthResponse {
 119         agg.setupOnce.Do(agg.setup)
 120         resp := ClusterHealthResponse{
 121                 Health:   "OK",
 122                 Checks:   make(map[string]CheckResult),
 123                 Services: make(map[arvados.ServiceName]ServiceHealth),
 124         }
 125
 126         mtx := sync.Mutex{}
 127         wg := sync.WaitGroup{}
 128         for svcName, svc := range agg.Cluster.Services.Map() {
 129                 // Ensure svc is listed in resp.Services.
 130                 mtx.Lock()
 131                 if _, ok := resp.Services[svcName]; !ok {
 132                         resp.Services[svcName] = ServiceHealth{Health: "ERROR"}
 133                 }
 134                 mtx.Unlock()
 135
 136                 for addr := range svc.InternalURLs {
 137                         wg.Add(1)
 138                         go func(svcName arvados.ServiceName, addr arvados.URL) {
 139                                 defer wg.Done()
 140                                 var result CheckResult
 141                                 pingURL, err := agg.pingURL(addr)
 142                                 if err != nil {
 143                                         result = CheckResult{
 144                                                 Health: "ERROR",
 145                                                 Error:  err.Error(),
 146                                         }
 147                                 } else {
 148                                         result = agg.ping(pingURL)
 149                                 }
 150
 151                                 mtx.Lock()
 152                                 defer mtx.Unlock()
 153                                 resp.Checks[fmt.Sprintf("%s+%s", svcName, pingURL)] = result
 154                                 if result.Health == "OK" {
 155                                         h := resp.Services[svcName]
 156                                         h.N++
 157                                         h.Health = "OK"
 158                                         resp.Services[svcName] = h
 159                                 } else {
 160                                         resp.Health = "ERROR"
 161                                 }
 162                         }(svcName, addr)
 163                 }
 164         }
 165         wg.Wait()
 166
 167         // Report ERROR if a needed service didn't fail any checks
 168         // merely because it isn't configured to run anywhere.
 169         for _, sh := range resp.Services {
 170                 if sh.Health != "OK" {
 171                         resp.Health = "ERROR"
 172                         break
 173                 }
 174         }
 175         return resp
 176 }
 177
 178 func (agg *Aggregator) pingURL(svcURL arvados.URL) (*url.URL, error) {
 179         base := url.URL(svcURL)
 180         return base.Parse("/_health/ping")
 181 }
 182
 183 func (agg *Aggregator) ping(target *url.URL) (result CheckResult) {
 184         t0 := time.Now()
 185         defer func() {
 186                 result.ResponseTime = json.Number(fmt.Sprintf("%.6f", time.Since(t0).Seconds()))
 187         }()
 188         result.Health = "ERROR"
 189
 190         req, err := http.NewRequest("GET", target.String(), nil)
 191         if err != nil {
 192                 result.Error = err.Error()
 193                 return
 194         }
 195         req.Header.Set("Authorization", "Bearer "+agg.Cluster.ManagementToken)
 196
 197         // Avoid workbench1's redirect-http-to-https feature
 198         req.Header.Set("X-Forwarded-Proto", "https")
 199
 200         ctx, cancel := context.WithTimeout(req.Context(), time.Duration(agg.timeout))
 201         defer cancel()
 202         req = req.WithContext(ctx)
 203         resp, err := agg.httpClient.Do(req)
 204         if err != nil {
 205                 result.Error = err.Error()
 206                 return
 207         }
 208         result.HTTPStatusCode = resp.StatusCode
 209         result.HTTPStatusText = resp.Status
 210         err = json.NewDecoder(resp.Body).Decode(&result.Response)
 211         if err != nil {
 212                 result.Error = fmt.Sprintf("cannot decode response: %s", err)
 213         } else if resp.StatusCode != http.StatusOK {
 214                 result.Error = fmt.Sprintf("HTTP %d %s", resp.StatusCode, resp.Status)
 215         } else if h, _ := result.Response["health"].(string); h != "OK" {
 216                 if e, ok := result.Response["error"].(string); ok && e != "" {
 217                         result.Error = e
 218                         return
 219                 } else {
 220                         result.Error = fmt.Sprintf("health=%q in ping response", h)
 221                         return
 222                 }
 223         }
 224         result.Health = "OK"
 225         return
 226 }
 227
 228 func (agg *Aggregator) checkAuth(req *http.Request) bool {
 229         creds := auth.CredentialsFromRequest(req)
 230         for _, token := range creds.Tokens {
 231                 if token != "" && token == agg.Cluster.ManagementToken {
 232                         return true
 233                 }
 234         }
 235         return false
 236 }