package keepclient
import (
+ "bytes"
+ "context"
"crypto/md5"
"errors"
"fmt"
"math/rand"
"net/http"
"os"
+ "strconv"
"strings"
+ "time"
- "git.curoverse.com/arvados.git/sdk/go/arvadosclient"
- "git.curoverse.com/arvados.git/sdk/go/streamer"
+ "git.arvados.org/arvados.git/sdk/go/arvados"
+ "git.arvados.org/arvados.git/sdk/go/arvadosclient"
+ "git.arvados.org/arvados.git/sdk/go/asyncbuf"
)
-// Function used to emit debug messages. The easiest way to enable
+// 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{}) {}
}
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 int32) {
+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 {
- DebugPrintf("DEBUG: [%08x] 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 {
- // 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.httpClient().Do(req); err != nil {
- DebugPrintf("DEBUG: [%08x] 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{R: resp.Body, N: 4096})
response := strings.TrimSpace(string(respbody))
if err2 != nil && err2 != io.EOF {
- DebugPrintf("DEBUG: [%08x] 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 {
- DebugPrintf("DEBUG: [%08x] 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 {
if resp.StatusCode >= 300 && response == "" {
response = resp.Status
}
- DebugPrintf("DEBUG: [%08x] Upload %v error: %v response: %v", requestID, url, resp.StatusCode, response)
- upload_status <- uploadStatus{errors.New(resp.Status), url, resp.StatusCode, rep, response}
+ 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) {
-
- // Generate an arbitrary ID to identify this specific
- // transaction in debug logs.
- requestID := rand.Int31()
+func (kc *KeepClient) httpBlockWrite(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))
+ reader := req.Reader
+ if req.Hash != "" {
+ reader = HashCheckingReader{req.Reader, md5.New(), req.Hash}
+ }
+ go func() {
+ _, err := io.Copy(buf, reader)
+ 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)
}()
}()
- replicasDone := 0
- replicasTodo := 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 = replicasTodo
+ replicasPerThread = req.Replicas
}
- retriesRemaining := 1 + this.Retries
+ delay := delayCalculator{InitialMaxDelay: kc.RetryDelay}
+ retriesRemaining := req.Attempts
var retryServers []string
lastError := make(map[string]string)
+ trackingClasses := len(replicasTodo) > 0
for retriesRemaining > 0 {
- retriesRemaining -= 1
- next_server = 0
+ retriesRemaining--
+ nextServer = 0
retryServers = []string{}
- for replicasTodo > 0 {
- for active*replicasPerThread < replicasTodo {
+ 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 next_server < len(sv) {
- DebugPrintf("DEBUG: [%08x] 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
+ 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: "
msg += resp + "; "
}
msg = msg[:len(msg)-2]
- return locator, replicasDone, InsufficientReplicasError(errors.New(msg))
- } else {
- break
+ return resp, InsufficientReplicasError{error: errors.New(msg)}
}
+ break
}
}
- DebugPrintf("DEBUG: [%08x] Replicas remaining to write: %v active uploads: %v",
- requestID, replicasTodo, active)
-
- // Now wait for something to happen.
- if active > 0 {
- status := <-upload_status
- active -= 1
-
- if status.statusCode == 200 {
- // good news!
- replicasDone += status.replicas_stored
- replicasTodo -= status.replicas_stored
- locator = status.response
- delete(lastError, status.url)
- } else {
- msg := fmt.Sprintf("[%d] %s", status.statusCode, status.response)
- if len(msg) > 100 {
- msg = msg[:100]
- }
- lastError[status.url] = msg
- }
- 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, "/")])
+ 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 {
- break
+ msg := fmt.Sprintf("[%d] %s", status.statusCode, status.response)
+ if len(msg) > 100 {
+ msg = msg[:100]
+ }
+ lastError[status.url] = msg
+ }
+
+ if status.statusCode == 0 || status.statusCode == 408 || status.statusCode == 429 ||
+ (status.statusCode >= 500 && status.statusCode != http.StatusInsufficientStorage) {
+ // Timeout, too many requests, or other server side failure
+ // (do not auto-retry status 507 "full")
+ retryServers = append(retryServers, status.url[0:strings.LastIndex(status.url, "/")])
}
}
sv = retryServers
+ if len(sv) > 0 {
+ time.Sleep(delay.Next())
+ }
+ }
+
+ 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
+}
+
+// delayCalculator calculates a series of delays for implementing
+// exponential backoff with jitter. The first call to Next() returns
+// a random duration between MinimumRetryDelay and the specified
+// InitialMaxDelay (or DefaultRetryDelay if 0). The max delay is
+// doubled on each subsequent call to Next(), up to 10x the initial
+// max delay.
+type delayCalculator struct {
+ InitialMaxDelay time.Duration
+ n int // number of delays returned so far
+ nextmax time.Duration
+ limit time.Duration
+}
- return locator, replicasDone, nil
+func (dc *delayCalculator) Next() time.Duration {
+ if dc.nextmax <= MinimumRetryDelay {
+ // initialize
+ if dc.InitialMaxDelay > 0 {
+ dc.nextmax = dc.InitialMaxDelay
+ } else {
+ dc.nextmax = DefaultRetryDelay
+ }
+ dc.limit = 10 * dc.nextmax
+ }
+ d := time.Duration(rand.Float64() * float64(dc.nextmax))
+ if d < MinimumRetryDelay {
+ d = MinimumRetryDelay
+ }
+ dc.nextmax *= 2
+ if dc.nextmax > dc.limit {
+ dc.nextmax = dc.limit
+ }
+ return d
}