X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/2c07efe6ac7455059f2fccd558ea796f9c315e19..80f547e339c452e5b03be5beee00e845d56d8e18:/services/keepstore/azure_blob_volume.go diff --git a/services/keepstore/azure_blob_volume.go b/services/keepstore/azure_blob_volume.go index 79123a9747..c0033d9186 100644 --- a/services/keepstore/azure_blob_volume.go +++ b/services/keepstore/azure_blob_volume.go @@ -9,16 +9,21 @@ import ( "io/ioutil" "log" "os" + "regexp" "strings" + "sync" "time" "github.com/curoverse/azure-sdk-for-go/storage" ) var ( + azureMaxGetBytes int azureStorageAccountName string azureStorageAccountKeyFile string azureStorageReplication int + azureWriteRaceInterval = 15 * time.Second + azureWriteRacePollTime = time.Second ) func readKeyFromFile(file string) (string, error) { @@ -82,6 +87,11 @@ func init() { "azure-storage-replication", 3, "Replication level to report to clients when data is stored in an Azure container.") + flag.IntVar( + &azureMaxGetBytes, + "azure-max-get-bytes", + BlockSize, + fmt.Sprintf("Maximum bytes to request in a single GET request. If smaller than %d, use multiple concurrent range requests to retrieve a block.", BlockSize)) } // An AzureBlobVolume stores and retrieves blocks in an Azure Blob @@ -94,6 +104,9 @@ type AzureBlobVolume struct { replication int } +// NewAzureBlobVolume returns a new AzureBlobVolume using the given +// client and container name. The replication argument specifies the +// replication level to report when writing data. func NewAzureBlobVolume(client storage.Client, containerName string, readonly bool, replication int) *AzureBlobVolume { return &AzureBlobVolume{ azClient: client, @@ -116,42 +129,126 @@ func (v *AzureBlobVolume) Check() error { return nil } +// Get reads a Keep block that has been stored as a block blob in the +// container. +// +// If the block is younger than azureWriteRaceInterval and is +// unexpectedly empty, assume a PutBlob operation is in progress, and +// wait for it to finish writing. func (v *AzureBlobVolume) Get(loc string) ([]byte, error) { - rdr, err := v.bsClient.GetBlob(v.containerName, loc) - if err != nil { - if strings.Contains(err.Error(), "404 Not Found") { - // "storage: service returned without a response body (404 Not Found)" - return nil, os.ErrNotExist + var deadline time.Time + haveDeadline := false + buf, err := v.get(loc) + for err == nil && len(buf) == 0 && loc != "d41d8cd98f00b204e9800998ecf8427e" { + // Seeing a brand new empty block probably means we're + // in a race with CreateBlob, which under the hood + // (apparently) does "CreateEmpty" and "CommitData" + // with no additional transaction locking. + if !haveDeadline { + t, err := v.Mtime(loc) + if err != nil { + log.Print("Got empty block (possible race) but Mtime failed: ", err) + break + } + deadline = t.Add(azureWriteRaceInterval) + if time.Now().After(deadline) { + break + } + log.Printf("Race? Block %s is 0 bytes, %s old. Polling until %s", loc, time.Since(t), deadline) + haveDeadline = true + } else if time.Now().After(deadline) { + break } - return nil, err + bufs.Put(buf) + time.Sleep(azureWriteRacePollTime) + buf, err = v.get(loc) } - switch err := err.(type) { - case nil: - default: - log.Printf("ERROR IN Get(): %T %#v", err, err) - return nil, err + if haveDeadline { + log.Printf("Race ended with len(buf)==%d", len(buf)) } - defer rdr.Close() - buf := bufs.Get(BlockSize) - n, err := io.ReadFull(rdr, buf) - switch err { - case io.EOF, io.ErrUnexpectedEOF: - return buf[:n], nil - default: - bufs.Put(buf) - return nil, err + return buf, err +} + +func (v *AzureBlobVolume) get(loc string) ([]byte, error) { + expectSize := BlockSize + if azureMaxGetBytes < BlockSize { + // Unfortunately the handler doesn't tell us how long the blob + // is expected to be, so we have to ask Azure. + props, err := v.bsClient.GetBlobProperties(v.containerName, loc) + if err != nil { + return nil, v.translateError(err) + } + if props.ContentLength > int64(BlockSize) || props.ContentLength < 0 { + return nil, fmt.Errorf("block %s invalid size %d (max %d)", loc, props.ContentLength, BlockSize) + } + expectSize = int(props.ContentLength) } + + buf := bufs.Get(expectSize) + if expectSize == 0 { + return buf, nil + } + + // We'll update this actualSize if/when we get the last piece. + actualSize := -1 + pieces := (expectSize + azureMaxGetBytes - 1) / azureMaxGetBytes + errors := make([]error, pieces) + var wg sync.WaitGroup + wg.Add(pieces) + for p := 0; p < pieces; p++ { + go func(p int) { + defer wg.Done() + startPos := p * azureMaxGetBytes + endPos := startPos + azureMaxGetBytes + if endPos > expectSize { + endPos = expectSize + } + var rdr io.ReadCloser + var err error + if startPos == 0 && endPos == expectSize { + rdr, err = v.bsClient.GetBlob(v.containerName, loc) + } else { + rdr, err = v.bsClient.GetBlobRange(v.containerName, loc, fmt.Sprintf("%d-%d", startPos, endPos-1)) + } + if err != nil { + errors[p] = err + return + } + defer rdr.Close() + n, err := io.ReadFull(rdr, buf[startPos:endPos]) + if pieces == 1 && (err == io.ErrUnexpectedEOF || err == io.EOF) { + // If we don't know the actual size, + // and just tried reading 64 MiB, it's + // normal to encounter EOF. + } else if err != nil { + errors[p] = err + } + if p == pieces-1 { + actualSize = startPos + n + } + }(p) + } + wg.Wait() + for _, err := range errors { + if err != nil { + bufs.Put(buf) + return nil, v.translateError(err) + } + } + return buf[:actualSize], nil } +// Compare the given data with existing stored data. func (v *AzureBlobVolume) Compare(loc string, expect []byte) error { rdr, err := v.bsClient.GetBlob(v.containerName, loc) if err != nil { - return err + return v.translateError(err) } defer rdr.Close() return compareReaderWithBuf(rdr, expect, loc[:32]) } +// Put stores a Keep block as a block blob in the container. func (v *AzureBlobVolume) Put(loc string, block []byte) error { if v.readonly { return MethodDisabledError @@ -159,6 +256,7 @@ func (v *AzureBlobVolume) Put(loc string, block []byte) error { return v.bsClient.CreateBlockBlobFromReader(v.containerName, loc, uint64(len(block)), bytes.NewReader(block)) } +// Touch updates the last-modified property of a block blob. func (v *AzureBlobVolume) Touch(loc string) error { if v.readonly { return MethodDisabledError @@ -168,6 +266,7 @@ func (v *AzureBlobVolume) Touch(loc string) error { }) } +// Mtime returns the last-modified property of a block blob. func (v *AzureBlobVolume) Mtime(loc string) (time.Time, error) { props, err := v.bsClient.GetBlobProperties(v.containerName, loc) if err != nil { @@ -176,6 +275,8 @@ func (v *AzureBlobVolume) Mtime(loc string) (time.Time, error) { return time.Parse(time.RFC1123, props.LastModified) } +// IndexTo writes a list of Keep blocks that are stored in the +// container. func (v *AzureBlobVolume) IndexTo(prefix string, writer io.Writer) error { params := storage.ListBlobsParameters{ Prefix: prefix, @@ -190,6 +291,17 @@ func (v *AzureBlobVolume) IndexTo(prefix string, writer io.Writer) error { if err != nil { return err } + if !v.isKeepBlock(b.Name) { + continue + } + if b.Properties.ContentLength == 0 && t.Add(azureWriteRaceInterval).After(time.Now()) { + // A new zero-length blob is probably + // just a new non-empty blob that + // hasn't committed its data yet (see + // Get()), and in any case has no + // value. + continue + } fmt.Fprintf(writer, "%s+%d %d\n", b.Name, b.Properties.ContentLength, t.Unix()) } if resp.NextMarker == "" { @@ -199,6 +311,7 @@ func (v *AzureBlobVolume) IndexTo(prefix string, writer io.Writer) error { } } +// Delete a Keep block. func (v *AzureBlobVolume) Delete(loc string) error { if v.readonly { return MethodDisabledError @@ -222,6 +335,7 @@ func (v *AzureBlobVolume) Delete(loc string) error { }) } +// Status returns a VolumeStatus struct with placeholder data. func (v *AzureBlobVolume) Status() *VolumeStatus { return &VolumeStatus{ DeviceNum: 1, @@ -230,14 +344,39 @@ func (v *AzureBlobVolume) Status() *VolumeStatus { } } +// String returns a volume label, including the container name. func (v *AzureBlobVolume) String() string { return fmt.Sprintf("azure-storage-container:%+q", v.containerName) } +// Writable returns true, unless the -readonly flag was on when the +// volume was added. func (v *AzureBlobVolume) Writable() bool { return !v.readonly } +// Replication returns the replication level of the container, as +// specified by the -azure-storage-replication argument. func (v *AzureBlobVolume) Replication() int { return v.replication } + +// If possible, translate an Azure SDK error to a recognizable error +// like os.ErrNotExist. +func (v *AzureBlobVolume) translateError(err error) error { + switch { + case err == nil: + return err + case strings.Contains(err.Error(), "404 Not Found"): + // "storage: service returned without a response body (404 Not Found)" + return os.ErrNotExist + default: + return err + } +} + +var keepBlockRegexp = regexp.MustCompile(`^[0-9a-f]{32}$`) + +func (v *AzureBlobVolume) isKeepBlock(s string) bool { + return keepBlockRegexp.MatchString(s) +}