X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/1841871773da24d4a743bf9083e1fa461a510125..bd471a9eadaf564fb4beafd7db995b7762942c1d:/sdk/go/keepclient/support.go diff --git a/sdk/go/keepclient/support.go b/sdk/go/keepclient/support.go index 8fabb308f7..8d299815b2 100644 --- a/sdk/go/keepclient/support.go +++ b/sdk/go/keepclient/support.go @@ -1,20 +1,39 @@ +// Copyright (C) The Arvados Authors. All rights reserved. +// +// SPDX-License-Identifier: Apache-2.0 + package keepclient import ( + "bytes" + "context" "crypto/md5" - "encoding/json" "errors" "fmt" - "git.curoverse.com/arvados.git/sdk/go/streamer" "io" "io/ioutil" "log" - "net" "net/http" + "os" + "strconv" "strings" - "time" + + "git.arvados.org/arvados.git/sdk/go/arvados" + "git.arvados.org/arvados.git/sdk/go/arvadosclient" + "git.arvados.org/arvados.git/sdk/go/asyncbuf" ) +// DebugPrintf emits debug messages. The easiest way to enable +// keepclient debug messages in your application is to assign +// log.Printf to DebugPrintf. +var DebugPrintf = func(string, ...interface{}) {} + +func init() { + if arvadosclient.StringBool(os.Getenv("ARVADOS_DEBUG")) { + DebugPrintf = log.Printf + } +} + type keepService struct { Uuid string `json:"uuid"` Hostname string `json:"service_host"` @@ -29,275 +48,296 @@ func Md5String(s string) string { return fmt.Sprintf("%x", md5.Sum([]byte(s))) } -// Set timeouts apply when connecting to keepproxy services (assumed to be over -// the Internet). -func (this *KeepClient) setClientSettingsProxy() { - if this.Client.Timeout == 0 { - // Maximum time to wait for a complete response - this.Client.Timeout = 300 * time.Second - - // TCP and TLS connection settings - this.Client.Transport = &http.Transport{ - Dial: (&net.Dialer{ - // The maximum time to wait to set up - // the initial TCP connection. - Timeout: 30 * time.Second, - - // The TCP keep alive heartbeat - // interval. - KeepAlive: 120 * time.Second, - }).Dial, - - TLSHandshakeTimeout: 10 * time.Second, - } - } -} - -// Set timeouts apply when connecting to keepstore services directly (assumed -// to be on the local network). -func (this *KeepClient) setClientSettingsDisk() { - if this.Client.Timeout == 0 { - // Maximum time to wait for a complete response - this.Client.Timeout = 20 * time.Second - - // TCP and TLS connection timeouts - this.Client.Transport = &http.Transport{ - Dial: (&net.Dialer{ - // The maximum time to wait to set up - // the initial TCP connection. - Timeout: 2 * time.Second, - - // The TCP keep alive heartbeat - // interval. - KeepAlive: 180 * time.Second, - }).Dial, - - TLSHandshakeTimeout: 4 * time.Second, - } - } -} - type svcList struct { Items []keepService `json:"items"` } -// DiscoverKeepServers gets list of available keep services from api server -func (this *KeepClient) DiscoverKeepServers() error { - var list svcList - - // Get keep services from api server - err := this.Arvados.Call("GET", "keep_services", "", "accessible", nil, &list) - if err != nil { - return err - } - - return this.loadKeepServers(list) -} - -// DiscoverKeepServersFromJSON gets list of available keep services from given JSON -func (this *KeepClient) DiscoverKeepServersFromJSON(services string) error { - var list svcList - - // Load keep services from given json - dec := json.NewDecoder(strings.NewReader(services)) - if err := dec.Decode(&list); err != nil { - return err - } - - return this.loadKeepServers(list) -} - -// loadKeepServers -func (this *KeepClient) loadKeepServers(list svcList) error { - listed := make(map[string]bool) - localRoots := make(map[string]string) - gatewayRoots := make(map[string]string) - writableLocalRoots := make(map[string]string) - - // replicasPerService is 1 for disks; unknown or unlimited otherwise - this.replicasPerService = 1 - this.Using_proxy = false - - for _, service := range list.Items { - scheme := "http" - if service.SSL { - scheme = "https" - } - url := fmt.Sprintf("%s://%s:%d", scheme, service.Hostname, service.Port) - - // Skip duplicates - if listed[url] { - continue - } - listed[url] = true - - localRoots[service.Uuid] = url - if service.SvcType == "proxy" { - this.Using_proxy = true - } - - if service.ReadOnly == false { - writableLocalRoots[service.Uuid] = url - if service.SvcType != "disk" { - this.replicasPerService = 0 - } - } - - // Gateway services are only used when specified by - // UUID, so there's nothing to gain by filtering them - // by service type. Including all accessible services - // (gateway and otherwise) merely accommodates more - // service configurations. - gatewayRoots[service.Uuid] = url - } - - if this.Using_proxy { - this.setClientSettingsProxy() - } else { - this.setClientSettingsDisk() - } - - this.SetServiceRoots(localRoots, writableLocalRoots, gatewayRoots) - return nil -} - type uploadStatus struct { - err error - url string - statusCode int - replicas_stored int - response string + err error + url string + statusCode int + replicasStored int + classesStored map[string]int + response string } -func (this KeepClient) uploadToKeepServer(host string, hash string, body io.ReadCloser, - upload_status chan<- uploadStatus, expectedLength int64, requestId string) { +func (kc *KeepClient) uploadToKeepServer(host string, hash string, classesTodo []string, body io.Reader, + uploadStatusChan chan<- uploadStatus, expectedLength int, reqid string) { var req *http.Request var err error var url = fmt.Sprintf("%s/%s", host, hash) if req, err = http.NewRequest("PUT", url, nil); err != nil { - log.Printf("[%v] Error creating request PUT %v error: %v", requestId, url, err.Error()) - upload_status <- uploadStatus{err, url, 0, 0, ""} - body.Close() + DebugPrintf("DEBUG: [%s] Error creating request PUT %v error: %v", reqid, url, err.Error()) + uploadStatusChan <- uploadStatus{err, url, 0, 0, nil, ""} return } - req.ContentLength = expectedLength + req.ContentLength = int64(expectedLength) if expectedLength > 0 { - // http.Client.Do will close the body ReadCloser when it is - // done with it. - req.Body = body + req.Body = ioutil.NopCloser(body) } else { - // "For client requests, a value of 0 means unknown if Body is - // not nil." In this case we do want the body to be empty, so - // don't set req.Body. However, we still need to close the - // body ReadCloser. - body.Close() + // "For client requests, a value of 0 means unknown if + // Body is not nil." In this case we do want the body + // to be empty, so don't set req.Body. } - req.Header.Add("Authorization", fmt.Sprintf("OAuth2 %s", this.Arvados.ApiToken)) + req.Header.Add("X-Request-Id", reqid) + req.Header.Add("Authorization", "OAuth2 "+kc.Arvados.ApiToken) req.Header.Add("Content-Type", "application/octet-stream") - req.Header.Add(X_Keep_Desired_Replicas, fmt.Sprint(this.Want_replicas)) + req.Header.Add(XKeepDesiredReplicas, fmt.Sprint(kc.Want_replicas)) + if len(classesTodo) > 0 { + req.Header.Add(XKeepStorageClasses, strings.Join(classesTodo, ", ")) + } var resp *http.Response - if resp, err = this.Client.Do(req); err != nil { - log.Printf("[%v] Upload failed %v error: %v", requestId, url, err.Error()) - upload_status <- uploadStatus{err, url, 0, 0, ""} + if resp, err = kc.httpClient().Do(req); err != nil { + DebugPrintf("DEBUG: [%s] Upload failed %v error: %v", reqid, url, err.Error()) + uploadStatusChan <- uploadStatus{err, url, 0, 0, nil, err.Error()} return } rep := 1 - if xr := resp.Header.Get(X_Keep_Replicas_Stored); xr != "" { + if xr := resp.Header.Get(XKeepReplicasStored); xr != "" { fmt.Sscanf(xr, "%d", &rep) } + scc := resp.Header.Get(XKeepStorageClassesConfirmed) + classesStored, err := parseStorageClassesConfirmedHeader(scc) + if err != nil { + DebugPrintf("DEBUG: [%s] Ignoring invalid %s header %q: %s", reqid, XKeepStorageClassesConfirmed, scc, err) + } defer resp.Body.Close() defer io.Copy(ioutil.Discard, resp.Body) - respbody, err2 := ioutil.ReadAll(&io.LimitedReader{resp.Body, 4096}) + respbody, err2 := ioutil.ReadAll(&io.LimitedReader{R: resp.Body, N: 4096}) response := strings.TrimSpace(string(respbody)) if err2 != nil && err2 != io.EOF { - log.Printf("[%v] Upload %v error: %v response: %v", requestId, url, err2.Error(), response) - upload_status <- uploadStatus{err2, url, resp.StatusCode, rep, response} + DebugPrintf("DEBUG: [%s] Upload %v error: %v response: %v", reqid, url, err2.Error(), response) + uploadStatusChan <- uploadStatus{err2, url, resp.StatusCode, rep, classesStored, response} } else if resp.StatusCode == http.StatusOK { - log.Printf("[%v] Upload %v success", requestId, url) - upload_status <- uploadStatus{nil, url, resp.StatusCode, rep, response} + DebugPrintf("DEBUG: [%s] Upload %v success", reqid, url) + uploadStatusChan <- uploadStatus{nil, url, resp.StatusCode, rep, classesStored, response} } else { - log.Printf("[%v] Upload %v error: %v response: %v", requestId, url, resp.StatusCode, response) - upload_status <- uploadStatus{errors.New(resp.Status), url, resp.StatusCode, rep, response} + if resp.StatusCode >= 300 && response == "" { + response = resp.Status + } + DebugPrintf("DEBUG: [%s] Upload %v error: %v response: %v", reqid, url, resp.StatusCode, response) + uploadStatusChan <- uploadStatus{errors.New(resp.Status), url, resp.StatusCode, rep, classesStored, response} } } -func (this KeepClient) putReplicas( - hash string, - tr *streamer.AsyncStream, - expectedLength int64) (locator string, replicas int, err error) { - - // Take the hash of locator and timestamp in order to identify this - // specific transaction in log statements. - requestId := fmt.Sprintf("%x", md5.Sum([]byte(locator+time.Now().String())))[0:8] +func (kc *KeepClient) BlockWrite(ctx context.Context, req arvados.BlockWriteOptions) (arvados.BlockWriteResponse, error) { + var resp arvados.BlockWriteResponse + var getReader func() io.Reader + if req.Data == nil && req.Reader == nil { + return resp, errors.New("invalid BlockWriteOptions: Data and Reader are both nil") + } + if req.DataSize < 0 { + return resp, fmt.Errorf("invalid BlockWriteOptions: negative DataSize %d", req.DataSize) + } + if req.DataSize > BLOCKSIZE || len(req.Data) > BLOCKSIZE { + return resp, ErrOversizeBlock + } + if req.Data != nil { + if req.DataSize > len(req.Data) { + return resp, errors.New("invalid BlockWriteOptions: DataSize > len(Data)") + } + if req.DataSize == 0 { + req.DataSize = len(req.Data) + } + getReader = func() io.Reader { return bytes.NewReader(req.Data[:req.DataSize]) } + } else { + buf := asyncbuf.NewBuffer(make([]byte, 0, req.DataSize)) + go func() { + _, err := io.Copy(buf, HashCheckingReader{req.Reader, md5.New(), req.Hash}) + buf.CloseWithError(err) + }() + getReader = buf.NewReader + } + if req.Hash == "" { + m := md5.New() + _, err := io.Copy(m, getReader()) + if err != nil { + return resp, err + } + req.Hash = fmt.Sprintf("%x", m.Sum(nil)) + } + if req.StorageClasses == nil { + if len(kc.StorageClasses) > 0 { + req.StorageClasses = kc.StorageClasses + } else { + req.StorageClasses = kc.DefaultStorageClasses + } + } + if req.Replicas == 0 { + req.Replicas = kc.Want_replicas + } + if req.RequestID == "" { + req.RequestID = kc.getRequestID() + } + if req.Attempts == 0 { + req.Attempts = 1 + kc.Retries + } // Calculate the ordering for uploading to servers - sv := NewRootSorter(this.WritableLocalRoots(), hash).GetSortedRoots() + sv := NewRootSorter(kc.WritableLocalRoots(), req.Hash).GetSortedRoots() // The next server to try contacting - next_server := 0 + nextServer := 0 // The number of active writers active := 0 // Used to communicate status from the upload goroutines - upload_status := make(chan uploadStatus) + uploadStatusChan := make(chan uploadStatus) defer func() { // Wait for any abandoned uploads (e.g., we started // two uploads and the first replied with replicas=2) // to finish before closing the status channel. go func() { for active > 0 { - <-upload_status + <-uploadStatusChan } - close(upload_status) + close(uploadStatusChan) }() }() - // Desired number of replicas - remaining_replicas := this.Want_replicas + replicasTodo := map[string]int{} + for _, c := range req.StorageClasses { + replicasTodo[c] = req.Replicas + } - replicasPerThread := this.replicasPerService + replicasPerThread := kc.replicasPerService if replicasPerThread < 1 { // unlimited or unknown - replicasPerThread = remaining_replicas + replicasPerThread = req.Replicas } - for remaining_replicas > 0 { - for active*replicasPerThread < remaining_replicas { - // Start some upload requests - if next_server < len(sv) { - log.Printf("[%v] Begin upload %s to %s", requestId, hash, sv[next_server]) - go this.uploadToKeepServer(sv[next_server], hash, tr.MakeStreamReader(), upload_status, expectedLength, requestId) - next_server += 1 - active += 1 - } else { - if active == 0 { - return locator, (this.Want_replicas - remaining_replicas), InsufficientReplicasError + retriesRemaining := req.Attempts + var retryServers []string + + lastError := make(map[string]string) + trackingClasses := len(replicasTodo) > 0 + + for retriesRemaining > 0 { + retriesRemaining-- + nextServer = 0 + retryServers = []string{} + for { + var classesTodo []string + var maxConcurrency int + for sc, r := range replicasTodo { + classesTodo = append(classesTodo, sc) + if maxConcurrency == 0 || maxConcurrency > r { + // Having more than r + // writes in flight + // would overreplicate + // class sc. + maxConcurrency = r + } + } + if !trackingClasses { + maxConcurrency = req.Replicas - resp.Replicas + } + if maxConcurrency < 1 { + // If there are no non-zero entries in + // replicasTodo, we're done. + break + } + for active*replicasPerThread < maxConcurrency { + // Start some upload requests + if nextServer < len(sv) { + DebugPrintf("DEBUG: [%s] Begin upload %s to %s", req.RequestID, req.Hash, sv[nextServer]) + go kc.uploadToKeepServer(sv[nextServer], req.Hash, classesTodo, getReader(), uploadStatusChan, req.DataSize, req.RequestID) + nextServer++ + active++ } else { + if active == 0 && retriesRemaining == 0 { + msg := "Could not write sufficient replicas: " + for _, resp := range lastError { + msg += resp + "; " + } + msg = msg[:len(msg)-2] + return resp, InsufficientReplicasError{error: errors.New(msg)} + } break } } - } - log.Printf("[%v] Replicas remaining to write: %v active uploads: %v", - requestId, remaining_replicas, active) - // Now wait for something to happen. - status := <-upload_status - active -= 1 + DebugPrintf("DEBUG: [%s] Replicas remaining to write: %v active uploads: %v", req.RequestID, replicasTodo, active) + if active < 1 { + break + } + + // Wait for something to happen. + status := <-uploadStatusChan + active-- + + if status.statusCode == http.StatusOK { + delete(lastError, status.url) + resp.Replicas += status.replicasStored + if len(status.classesStored) == 0 { + // Server doesn't report + // storage classes. Give up + // trying to track which ones + // are satisfied; just rely on + // total # replicas. + trackingClasses = false + } + for className, replicas := range status.classesStored { + if replicasTodo[className] > replicas { + replicasTodo[className] -= replicas + } else { + delete(replicasTodo, className) + } + } + resp.Locator = status.response + } else { + msg := fmt.Sprintf("[%d] %s", status.statusCode, status.response) + if len(msg) > 100 { + msg = msg[:100] + } + lastError[status.url] = msg + } - if status.statusCode == 200 { - // good news! - remaining_replicas -= status.replicas_stored - locator = status.response + if status.statusCode == 0 || status.statusCode == 408 || status.statusCode == 429 || + (status.statusCode >= 500 && status.statusCode != 503) { + // Timeout, too many requests, or other server side failure + // Do not retry when status code is 503, which means the keep server is full + retryServers = append(retryServers, status.url[0:strings.LastIndex(status.url, "/")]) + } } + + sv = retryServers } - return locator, this.Want_replicas, nil + return resp, nil +} + +func parseStorageClassesConfirmedHeader(hdr string) (map[string]int, error) { + if hdr == "" { + return nil, nil + } + classesStored := map[string]int{} + for _, cr := range strings.Split(hdr, ",") { + cr = strings.TrimSpace(cr) + if cr == "" { + continue + } + fields := strings.SplitN(cr, "=", 2) + if len(fields) != 2 { + return nil, fmt.Errorf("expected exactly one '=' char in entry %q", cr) + } + className := fields[0] + if className == "" { + return nil, fmt.Errorf("empty class name in entry %q", cr) + } + replicas, err := strconv.Atoi(fields[1]) + if err != nil || replicas < 1 { + return nil, fmt.Errorf("invalid replica count %q", fields[1]) + } + classesStored[className] = replicas + } + return classesStored, nil }