2960: Merge branch 'main' into 2960-keepstore-streaming
[arvados.git] / sdk / go / keepclient / support.go
1 // Copyright (C) The Arvados Authors. All rights reserved.
2 //
3 // SPDX-License-Identifier: Apache-2.0
4
5 package keepclient
6
7 import (
8         "bytes"
9         "context"
10         "crypto/md5"
11         "errors"
12         "fmt"
13         "io"
14         "io/ioutil"
15         "log"
16         "math/rand"
17         "net/http"
18         "os"
19         "strconv"
20         "strings"
21         "time"
22
23         "git.arvados.org/arvados.git/sdk/go/arvados"
24         "git.arvados.org/arvados.git/sdk/go/arvadosclient"
25         "git.arvados.org/arvados.git/sdk/go/asyncbuf"
26 )
27
28 // DebugPrintf emits debug messages. The easiest way to enable
29 // keepclient debug messages in your application is to assign
30 // log.Printf to DebugPrintf.
31 var DebugPrintf = func(string, ...interface{}) {}
32
33 func init() {
34         if arvadosclient.StringBool(os.Getenv("ARVADOS_DEBUG")) {
35                 DebugPrintf = log.Printf
36         }
37 }
38
39 type keepService struct {
40         Uuid     string `json:"uuid"`
41         Hostname string `json:"service_host"`
42         Port     int    `json:"service_port"`
43         SSL      bool   `json:"service_ssl_flag"`
44         SvcType  string `json:"service_type"`
45         ReadOnly bool   `json:"read_only"`
46 }
47
48 // Md5String returns md5 hash for the bytes in the given string
49 func Md5String(s string) string {
50         return fmt.Sprintf("%x", md5.Sum([]byte(s)))
51 }
52
53 type svcList struct {
54         Items []keepService `json:"items"`
55 }
56
57 type uploadStatus struct {
58         err            error
59         url            string
60         statusCode     int
61         replicasStored int
62         classesStored  map[string]int
63         response       string
64 }
65
66 func (kc *KeepClient) uploadToKeepServer(host string, hash string, classesTodo []string, body io.Reader,
67         uploadStatusChan chan<- uploadStatus, expectedLength int, reqid string) {
68
69         var req *http.Request
70         var err error
71         var url = fmt.Sprintf("%s/%s", host, hash)
72         if req, err = http.NewRequest("PUT", url, nil); err != nil {
73                 DebugPrintf("DEBUG: [%s] Error creating request PUT %v error: %v", reqid, url, err.Error())
74                 uploadStatusChan <- uploadStatus{err, url, 0, 0, nil, ""}
75                 return
76         }
77
78         req.ContentLength = int64(expectedLength)
79         if expectedLength > 0 {
80                 req.Body = ioutil.NopCloser(body)
81         } else {
82                 // "For client requests, a value of 0 means unknown if
83                 // Body is not nil."  In this case we do want the body
84                 // to be empty, so don't set req.Body.
85         }
86
87         req.Header.Add("X-Request-Id", reqid)
88         req.Header.Add("Authorization", "OAuth2 "+kc.Arvados.ApiToken)
89         req.Header.Add("Content-Type", "application/octet-stream")
90         req.Header.Add(XKeepDesiredReplicas, fmt.Sprint(kc.Want_replicas))
91         if len(classesTodo) > 0 {
92                 req.Header.Add(XKeepStorageClasses, strings.Join(classesTodo, ", "))
93         }
94
95         var resp *http.Response
96         if resp, err = kc.httpClient().Do(req); err != nil {
97                 DebugPrintf("DEBUG: [%s] Upload failed %v error: %v", reqid, url, err.Error())
98                 uploadStatusChan <- uploadStatus{err, url, 0, 0, nil, err.Error()}
99                 return
100         }
101
102         rep := 1
103         if xr := resp.Header.Get(XKeepReplicasStored); xr != "" {
104                 fmt.Sscanf(xr, "%d", &rep)
105         }
106         scc := resp.Header.Get(XKeepStorageClassesConfirmed)
107         classesStored, err := parseStorageClassesConfirmedHeader(scc)
108         if err != nil {
109                 DebugPrintf("DEBUG: [%s] Ignoring invalid %s header %q: %s", reqid, XKeepStorageClassesConfirmed, scc, err)
110         }
111
112         defer resp.Body.Close()
113         defer io.Copy(ioutil.Discard, resp.Body)
114
115         respbody, err2 := ioutil.ReadAll(&io.LimitedReader{R: resp.Body, N: 4096})
116         response := strings.TrimSpace(string(respbody))
117         if err2 != nil && err2 != io.EOF {
118                 DebugPrintf("DEBUG: [%s] Upload %v error: %v response: %v", reqid, url, err2.Error(), response)
119                 uploadStatusChan <- uploadStatus{err2, url, resp.StatusCode, rep, classesStored, response}
120         } else if resp.StatusCode == http.StatusOK {
121                 DebugPrintf("DEBUG: [%s] Upload %v success", reqid, url)
122                 uploadStatusChan <- uploadStatus{nil, url, resp.StatusCode, rep, classesStored, response}
123         } else {
124                 if resp.StatusCode >= 300 && response == "" {
125                         response = resp.Status
126                 }
127                 DebugPrintf("DEBUG: [%s] Upload %v error: %v response: %v", reqid, url, resp.StatusCode, response)
128                 uploadStatusChan <- uploadStatus{errors.New(resp.Status), url, resp.StatusCode, rep, classesStored, response}
129         }
130 }
131
132 func (kc *KeepClient) httpBlockWrite(ctx context.Context, req arvados.BlockWriteOptions) (arvados.BlockWriteResponse, error) {
133         var resp arvados.BlockWriteResponse
134         var getReader func() io.Reader
135         if req.Data == nil && req.Reader == nil {
136                 return resp, errors.New("invalid BlockWriteOptions: Data and Reader are both nil")
137         }
138         if req.DataSize < 0 {
139                 return resp, fmt.Errorf("invalid BlockWriteOptions: negative DataSize %d", req.DataSize)
140         }
141         if req.DataSize > BLOCKSIZE || len(req.Data) > BLOCKSIZE {
142                 return resp, ErrOversizeBlock
143         }
144         if req.Data != nil {
145                 if req.DataSize > len(req.Data) {
146                         return resp, errors.New("invalid BlockWriteOptions: DataSize > len(Data)")
147                 }
148                 if req.DataSize == 0 {
149                         req.DataSize = len(req.Data)
150                 }
151                 getReader = func() io.Reader { return bytes.NewReader(req.Data[:req.DataSize]) }
152         } else {
153                 buf := asyncbuf.NewBuffer(make([]byte, 0, req.DataSize))
154                 reader := req.Reader
155                 if req.Hash != "" {
156                         reader = HashCheckingReader{req.Reader, md5.New(), req.Hash}
157                 }
158                 go func() {
159                         _, err := io.Copy(buf, reader)
160                         buf.CloseWithError(err)
161                 }()
162                 getReader = buf.NewReader
163         }
164         if req.Hash == "" {
165                 m := md5.New()
166                 _, err := io.Copy(m, getReader())
167                 if err != nil {
168                         return resp, err
169                 }
170                 req.Hash = fmt.Sprintf("%x", m.Sum(nil))
171         }
172         if req.StorageClasses == nil {
173                 if len(kc.StorageClasses) > 0 {
174                         req.StorageClasses = kc.StorageClasses
175                 } else {
176                         req.StorageClasses = kc.DefaultStorageClasses
177                 }
178         }
179         if req.Replicas == 0 {
180                 req.Replicas = kc.Want_replicas
181         }
182         if req.RequestID == "" {
183                 req.RequestID = kc.getRequestID()
184         }
185         if req.Attempts == 0 {
186                 req.Attempts = 1 + kc.Retries
187         }
188
189         // Calculate the ordering for uploading to servers
190         sv := NewRootSorter(kc.WritableLocalRoots(), req.Hash).GetSortedRoots()
191
192         // The next server to try contacting
193         nextServer := 0
194
195         // The number of active writers
196         active := 0
197
198         // Used to communicate status from the upload goroutines
199         uploadStatusChan := make(chan uploadStatus)
200         defer func() {
201                 // Wait for any abandoned uploads (e.g., we started
202                 // two uploads and the first replied with replicas=2)
203                 // to finish before closing the status channel.
204                 go func() {
205                         for active > 0 {
206                                 <-uploadStatusChan
207                         }
208                         close(uploadStatusChan)
209                 }()
210         }()
211
212         replicasTodo := map[string]int{}
213         for _, c := range req.StorageClasses {
214                 replicasTodo[c] = req.Replicas
215         }
216
217         replicasPerThread := kc.replicasPerService
218         if replicasPerThread < 1 {
219                 // unlimited or unknown
220                 replicasPerThread = req.Replicas
221         }
222
223         delay := delayCalculator{InitialMaxDelay: kc.RetryDelay}
224         retriesRemaining := req.Attempts
225         var retryServers []string
226
227         lastError := make(map[string]string)
228         trackingClasses := len(replicasTodo) > 0
229
230         for retriesRemaining > 0 {
231                 retriesRemaining--
232                 nextServer = 0
233                 retryServers = []string{}
234                 for {
235                         var classesTodo []string
236                         var maxConcurrency int
237                         for sc, r := range replicasTodo {
238                                 classesTodo = append(classesTodo, sc)
239                                 if maxConcurrency == 0 || maxConcurrency > r {
240                                         // Having more than r
241                                         // writes in flight
242                                         // would overreplicate
243                                         // class sc.
244                                         maxConcurrency = r
245                                 }
246                         }
247                         if !trackingClasses {
248                                 maxConcurrency = req.Replicas - resp.Replicas
249                         }
250                         if maxConcurrency < 1 {
251                                 // If there are no non-zero entries in
252                                 // replicasTodo, we're done.
253                                 break
254                         }
255                         for active*replicasPerThread < maxConcurrency {
256                                 // Start some upload requests
257                                 if nextServer < len(sv) {
258                                         DebugPrintf("DEBUG: [%s] Begin upload %s to %s", req.RequestID, req.Hash, sv[nextServer])
259                                         go kc.uploadToKeepServer(sv[nextServer], req.Hash, classesTodo, getReader(), uploadStatusChan, req.DataSize, req.RequestID)
260                                         nextServer++
261                                         active++
262                                 } else {
263                                         if active == 0 && retriesRemaining == 0 {
264                                                 msg := "Could not write sufficient replicas: "
265                                                 for _, resp := range lastError {
266                                                         msg += resp + "; "
267                                                 }
268                                                 msg = msg[:len(msg)-2]
269                                                 return resp, InsufficientReplicasError{error: errors.New(msg)}
270                                         }
271                                         break
272                                 }
273                         }
274
275                         DebugPrintf("DEBUG: [%s] Replicas remaining to write: %v active uploads: %v", req.RequestID, replicasTodo, active)
276                         if active < 1 {
277                                 break
278                         }
279
280                         // Wait for something to happen.
281                         status := <-uploadStatusChan
282                         active--
283
284                         if status.statusCode == http.StatusOK {
285                                 delete(lastError, status.url)
286                                 resp.Replicas += status.replicasStored
287                                 if len(status.classesStored) == 0 {
288                                         // Server doesn't report
289                                         // storage classes. Give up
290                                         // trying to track which ones
291                                         // are satisfied; just rely on
292                                         // total # replicas.
293                                         trackingClasses = false
294                                 }
295                                 for className, replicas := range status.classesStored {
296                                         if replicasTodo[className] > replicas {
297                                                 replicasTodo[className] -= replicas
298                                         } else {
299                                                 delete(replicasTodo, className)
300                                         }
301                                 }
302                                 resp.Locator = status.response
303                         } else {
304                                 msg := fmt.Sprintf("[%d] %s", status.statusCode, status.response)
305                                 if len(msg) > 100 {
306                                         msg = msg[:100]
307                                 }
308                                 lastError[status.url] = msg
309                         }
310
311                         if status.statusCode == 0 || status.statusCode == 408 || status.statusCode == 429 ||
312                                 (status.statusCode >= 500 && status.statusCode != http.StatusInsufficientStorage) {
313                                 // Timeout, too many requests, or other server side failure
314                                 // (do not auto-retry status 507 "full")
315                                 retryServers = append(retryServers, status.url[0:strings.LastIndex(status.url, "/")])
316                         }
317                 }
318
319                 sv = retryServers
320                 if len(sv) > 0 {
321                         time.Sleep(delay.Next())
322                 }
323         }
324
325         return resp, nil
326 }
327
328 func parseStorageClassesConfirmedHeader(hdr string) (map[string]int, error) {
329         if hdr == "" {
330                 return nil, nil
331         }
332         classesStored := map[string]int{}
333         for _, cr := range strings.Split(hdr, ",") {
334                 cr = strings.TrimSpace(cr)
335                 if cr == "" {
336                         continue
337                 }
338                 fields := strings.SplitN(cr, "=", 2)
339                 if len(fields) != 2 {
340                         return nil, fmt.Errorf("expected exactly one '=' char in entry %q", cr)
341                 }
342                 className := fields[0]
343                 if className == "" {
344                         return nil, fmt.Errorf("empty class name in entry %q", cr)
345                 }
346                 replicas, err := strconv.Atoi(fields[1])
347                 if err != nil || replicas < 1 {
348                         return nil, fmt.Errorf("invalid replica count %q", fields[1])
349                 }
350                 classesStored[className] = replicas
351         }
352         return classesStored, nil
353 }
354
355 // delayCalculator calculates a series of delays for implementing
356 // exponential backoff with jitter.  The first call to Next() returns
357 // a random duration between MinimumRetryDelay and the specified
358 // InitialMaxDelay (or DefaultRetryDelay if 0).  The max delay is
359 // doubled on each subsequent call to Next(), up to 10x the initial
360 // max delay.
361 type delayCalculator struct {
362         InitialMaxDelay time.Duration
363         n               int // number of delays returned so far
364         nextmax         time.Duration
365         limit           time.Duration
366 }
367
368 func (dc *delayCalculator) Next() time.Duration {
369         if dc.nextmax <= MinimumRetryDelay {
370                 // initialize
371                 if dc.InitialMaxDelay > 0 {
372                         dc.nextmax = dc.InitialMaxDelay
373                 } else {
374                         dc.nextmax = DefaultRetryDelay
375                 }
376                 dc.limit = 10 * dc.nextmax
377         }
378         d := time.Duration(rand.Float64() * float64(dc.nextmax))
379         if d < MinimumRetryDelay {
380                 d = MinimumRetryDelay
381         }
382         dc.nextmax *= 2
383         if dc.nextmax > dc.limit {
384                 dc.nextmax = dc.limit
385         }
386         return d
387 }