+// Copyright (C) The Arvados Authors. All rights reserved.
+//
+// SPDX-License-Identifier: Apache-2.0
+
package health
import (
"context"
"encoding/json"
+ "errors"
"fmt"
"net"
"net/http"
}
}
-type ServiceHealth struct {
+type ClusterHealthResponse struct {
+ // "OK" if all needed services are OK, otherwise "ERROR".
Health string `json:"health"`
- N int `json:"n"`
-}
-type ClusterHealthResponse struct {
- Health string `json:"health"`
- Checks map[string]CheckResponse `json:"checks"`
- Services map[string]ServiceHealth `json:"services"`
+ // An entry for each known health check of each known instance
+ // of each needed component: "instance of service S on node N
+ // reports health-check C is OK."
+ Checks map[string]CheckResult `json:"checks"`
+
+ // An entry for each service type: "service S is OK." This
+ // exposes problems that can't be expressed in Checks, like
+ // "service S is needed, but isn't configured to run
+ // anywhere."
+ Services map[arvados.ServiceName]ServiceHealth `json:"services"`
}
-type CheckResponse struct {
- Status int `json:"status"`
- Health string `json:"health"`
- Error string `json:"error,omitempty"`
- ResponseTime json.Number `json:"responseTime"`
+type CheckResult struct {
+ Health string `json:"health"`
+ Error string `json:"error,omitempty"`
+ HTTPStatusCode int `json:",omitempty"`
+ HTTPStatusText string `json:",omitempty"`
+ Response map[string]interface{} `json:"response"`
+ ResponseTime json.Number `json:"responseTime"`
}
-func (r *CheckResponse) OK() bool {
- return r.Health == "OK" && r.Status == http.StatusOK
+type ServiceHealth struct {
+ Health string `json:"health"`
+ N int `json:"n"`
}
func (agg *Aggregator) ClusterHealth(cluster *arvados.Cluster) ClusterHealthResponse {
resp := ClusterHealthResponse{
Health: "OK",
- Checks: make(map[string]CheckResponse),
- Services: make(map[string]ServiceHealth),
+ Checks: make(map[string]CheckResult),
+ Services: make(map[arvados.ServiceName]ServiceHealth),
}
mtx := sync.Mutex{}
wg := sync.WaitGroup{}
- for node, nodeConfig := range cluster.SystemNodes {
- for svc, addr := range map[string]string{
- "keepstore": nodeConfig.Keepstore.Listen,
- } {
+ for profileName, profile := range cluster.NodeProfiles {
+ for svc, addr := range profile.ServicePorts() {
+ // Ensure svc is listed in resp.Services.
+ mtx.Lock()
+ if _, ok := resp.Services[svc]; !ok {
+ resp.Services[svc] = ServiceHealth{Health: "ERROR"}
+ }
+ mtx.Unlock()
+
if addr == "" {
+ // svc is not expected on this node.
continue
}
+
wg.Add(1)
- go func(node string) {
+ go func(profileName string, svc arvados.ServiceName, addr string) {
defer wg.Done()
- var pingResp CheckResponse
- url, err := agg.pingURL(node, addr)
+ var result CheckResult
+ url, err := agg.pingURL(profileName, addr)
if err != nil {
- pingResp = CheckResponse{
+ result = CheckResult{
Health: "ERROR",
Error: err.Error(),
}
} else {
- pingResp = agg.ping(url, cluster)
+ result = agg.ping(url, cluster)
}
mtx.Lock()
defer mtx.Unlock()
- resp.Checks[svc+"+"+url] = pingResp
- svHealth := resp.Services[svc]
- if pingResp.OK() {
- svHealth.N++
+ resp.Checks[fmt.Sprintf("%s+%s", svc, url)] = result
+ if result.Health == "OK" {
+ h := resp.Services[svc]
+ h.N++
+ h.Health = "OK"
+ resp.Services[svc] = h
} else {
resp.Health = "ERROR"
}
- resp.Services[svc] = svHealth
- }(node)
+ }(profileName, svc, addr)
}
}
wg.Wait()
- for svc, svHealth := range resp.Services {
- if svHealth.N > 0 {
- svHealth.Health = "OK"
- } else {
- svHealth.Health = "ERROR"
+ // Report ERROR if a needed service didn't fail any checks
+ // merely because it isn't configured to run anywhere.
+ for _, sh := range resp.Services {
+ if sh.Health != "OK" {
+ resp.Health = "ERROR"
+ break
}
- resp.Services[svc] = svHealth
}
-
return resp
}
return "http://" + node + ":" + port + "/_health/ping", err
}
-func (agg *Aggregator) ping(url string, cluster *arvados.Cluster) (result CheckResponse) {
+func (agg *Aggregator) ping(url string, cluster *arvados.Cluster) (result CheckResult) {
t0 := time.Now()
var err error
result.ResponseTime = json.Number(fmt.Sprintf("%.6f", time.Since(t0).Seconds()))
if err != nil {
result.Health, result.Error = "ERROR", err.Error()
+ } else {
+ result.Health = "OK"
}
}()
}
req.Header.Set("Authorization", "Bearer "+cluster.ManagementToken)
- ctx, cancel := context.WithCancel(req.Context())
- go func() {
- select {
- case <-time.After(time.Duration(agg.timeout)):
- cancel()
- case <-ctx.Done():
- }
- }()
+ ctx, cancel := context.WithTimeout(req.Context(), time.Duration(agg.timeout))
+ defer cancel()
req = req.WithContext(ctx)
resp, err := agg.httpClient.Do(req)
if err != nil {
return
}
- result.Status = resp.StatusCode
- err = json.NewDecoder(resp.Body).Decode(&result)
+ result.HTTPStatusCode = resp.StatusCode
+ result.HTTPStatusText = resp.Status
+ err = json.NewDecoder(resp.Body).Decode(&result.Response)
if err != nil {
err = fmt.Errorf("cannot decode response: %s", err)
- return
- }
- if resp.StatusCode != http.StatusOK {
+ } else if resp.StatusCode != http.StatusOK {
err = fmt.Errorf("HTTP %d %s", resp.StatusCode, resp.Status)
- return
+ } else if h, _ := result.Response["health"].(string); h != "OK" {
+ if e, ok := result.Response["error"].(string); ok && e != "" {
+ err = errors.New(e)
+ } else {
+ err = fmt.Errorf("health=%q in ping response", h)
+ }
}
return
}
projects / arvados.git / blobdiff