X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/64639ed6313f01016da4e0ffd81752dedf9b052b..HEAD:/lib/service/cmd.go diff --git a/lib/service/cmd.go b/lib/service/cmd.go index cc6938cbc6..9ed0acfb8f 100644 --- a/lib/service/cmd.go +++ b/lib/service/cmd.go @@ -16,6 +16,7 @@ import ( _ "net/http/pprof" "net/url" "os" + "regexp" "strings" "time" @@ -79,9 +80,9 @@ func (c *command) RunCommand(prog string, args []string, stdin io.Reader, stdout loader := config.NewLoader(stdin, log) loader.SetupFlags(flags) - // prog is [keepstore, keep-web, git-httpd, ...] but the + // prog is [keepstore, keep-web, ...] but the // legacy config flags are [-legacy-keepstore-config, - // -legacy-keepweb-config, -legacy-git-httpd-config, ...] + // -legacy-keepweb-config, ...] legacyFlag := "-legacy-" + strings.Replace(prog, "keep-", "keep", 1) + "-config" args = loader.MungeLegacyConfigArgs(log, args, legacyFlag) @@ -154,7 +155,7 @@ func (c *command) RunCommand(prog string, args []string, stdin io.Reader, stdout httpserver.Inspect(reg, cluster.ManagementToken, httpserver.LogRequests( interceptHealthReqs(cluster.ManagementToken, handler.CheckHealth, - httpserver.NewRequestLimiter(cluster.API.MaxConcurrentRequests, handler, reg))))))) + c.requestLimiter(handler, cluster, reg))))))) srv := &httpserver.Server{ Server: http.Server{ Handler: ifCollectionInHost(instrumented, instrumented.ServeAPI(cluster.ManagementToken, instrumented)), @@ -202,9 +203,10 @@ func (c *command) RunCommand(prog string, args []string, stdin io.Reader, stdout } // If SystemLogs.RequestQueueDumpDirectory is set, monitor the -// server's incoming HTTP request queue size. When it exceeds 90% of -// API.MaxConcurrentRequests, write the /_inspect/requests data to a -// JSON file in the specified directory. +// server's incoming HTTP request limiters. When the number of +// concurrent requests in any queue ("api" or "tunnel") exceeds 90% of +// its maximum slots, write the /_inspect/requests data to a JSON file +// in the specified directory. func (c *command) requestQueueDumpCheck(cluster *arvados.Cluster, prog string, reg *prometheus.Registry, srv *http.Server, logger logrus.FieldLogger) { outdir := cluster.SystemLogs.RequestQueueDumpDirectory if outdir == "" || cluster.ManagementToken == "" { @@ -218,16 +220,29 @@ func (c *command) requestQueueDumpCheck(cluster *arvados.Cluster, prog string, r logger.WithError(err).Warn("error getting metrics") continue } - dump := false + cur := map[string]int{} // queue label => current + max := map[string]int{} // queue label => max for _, mf := range mfs { - if mf.Name != nil && *mf.Name == "arvados_concurrent_requests" && len(mf.Metric) == 1 { - n := int(mf.Metric[0].GetGauge().GetValue()) - if n > 0 && n >= cluster.API.MaxConcurrentRequests*9/10 { - dump = true - break + for _, m := range mf.GetMetric() { + for _, ml := range m.GetLabel() { + if ml.GetName() == "queue" { + n := int(m.GetGauge().GetValue()) + if name := mf.GetName(); name == "arvados_concurrent_requests" { + cur[*ml.Value] = n + } else if name == "arvados_max_concurrent_requests" { + max[*ml.Value] = n + } + } } } } + dump := false + for queue, n := range cur { + if n > 0 && max[queue] > 0 && n >= max[queue]*9/10 { + dump = true + break + } + } if dump { req, err := http.NewRequest("GET", "/_inspect/requests", nil) if err != nil { @@ -250,6 +265,90 @@ func (c *command) requestQueueDumpCheck(cluster *arvados.Cluster, prog string, r } } +// Set up a httpserver.RequestLimiter with separate queues/streams for +// API requests (obeying MaxConcurrentRequests etc) and gateway tunnel +// requests (obeying MaxGatewayTunnels). +func (c *command) requestLimiter(handler http.Handler, cluster *arvados.Cluster, reg *prometheus.Registry) http.Handler { + maxReqs := cluster.API.MaxConcurrentRequests + if maxRails := cluster.API.MaxConcurrentRailsRequests; maxRails > 0 && + (maxRails < maxReqs || maxReqs == 0) && + c.svcName == arvados.ServiceNameController { + // Ideally, we would accept up to + // MaxConcurrentRequests, and apply the + // MaxConcurrentRailsRequests limit only for requests + // that require calling upstream to RailsAPI. But for + // now we make the simplifying assumption that every + // controller request causes an upstream RailsAPI + // request. + maxReqs = maxRails + } + rqAPI := &httpserver.RequestQueue{ + Label: "api", + MaxConcurrent: maxReqs, + MaxQueue: cluster.API.MaxQueuedRequests, + MaxQueueTimeForMinPriority: cluster.API.MaxQueueTimeForLockRequests.Duration(), + } + rqTunnel := &httpserver.RequestQueue{ + Label: "tunnel", + MaxConcurrent: cluster.API.MaxGatewayTunnels, + MaxQueue: 0, + } + return &httpserver.RequestLimiter{ + Handler: handler, + Priority: c.requestPriority, + Registry: reg, + Queue: func(req *http.Request) *httpserver.RequestQueue { + if req.Method == http.MethodPost && reTunnelPath.MatchString(req.URL.Path) { + return rqTunnel + } else { + return rqAPI + } + }, + } +} + +// reTunnelPath matches paths of API endpoints that go in the "tunnel" +// queue. +var reTunnelPath = regexp.MustCompile(func() string { + rePathVar := regexp.MustCompile(`{.*?}`) + out := "" + for _, endpoint := range []arvados.APIEndpoint{ + arvados.EndpointContainerGatewayTunnel, + arvados.EndpointContainerGatewayTunnelCompat, + arvados.EndpointContainerSSH, + arvados.EndpointContainerSSHCompat, + } { + if out != "" { + out += "|" + } + out += `\Q/` + rePathVar.ReplaceAllString(endpoint.Path, `\E[^/]*\Q`) + `\E` + } + return "^(" + out + ")$" +}()) + +func (c *command) requestPriority(req *http.Request, queued time.Time) int64 { + switch { + case req.Method == http.MethodPost && strings.HasPrefix(req.URL.Path, "/arvados/v1/containers/") && strings.HasSuffix(req.URL.Path, "/lock"): + // Return 503 immediately instead of queueing. We want + // to send feedback to dispatchcloud ASAP to stop + // bringing up new containers. + return httpserver.MinPriority + case req.Method == http.MethodPost && strings.HasPrefix(req.URL.Path, "/arvados/v1/logs"): + // "Create log entry" is the most harmless kind of + // request to drop. Negative priority is called "low" + // in aggregate metrics. + return -1 + case req.Header.Get("Origin") != "": + // Handle interactive requests first. Positive + // priority is called "high" in aggregate metrics. + return 1 + default: + // Zero priority is called "normal" in aggregate + // metrics. + return 0 + } +} + // If an incoming request's target vhost has an embedded collection // UUID or PDH, handle it with hTrue, otherwise handle it with // hFalse.