X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/7072110da82050af324cfb7a511f65a543a062d1..50128b53da4003912635b03fb27b5be2c5beaca1:/services/keepstore/azure_blob_volume.go diff --git a/services/keepstore/azure_blob_volume.go b/services/keepstore/azure_blob_volume.go index e9fda2ab76..6ca31c3832 100644 --- a/services/keepstore/azure_blob_volume.go +++ b/services/keepstore/azure_blob_volume.go @@ -2,21 +2,29 @@ package main import ( "bytes" + "context" "errors" "flag" "fmt" "io" "io/ioutil" "log" + "net/http" "os" "regexp" + "strconv" "strings" + "sync" "time" + "git.curoverse.com/arvados.git/sdk/go/arvados" "github.com/curoverse/azure-sdk-for-go/storage" ) +const azureDefaultRequestTimeout = arvados.Duration(10 * time.Minute) + var ( + azureMaxGetBytes int azureStorageAccountName string azureStorageAccountKeyFile string azureStorageReplication int @@ -37,37 +45,29 @@ func readKeyFromFile(file string) (string, error) { } type azureVolumeAdder struct { - *volumeSet + *Config +} + +// String implements flag.Value +func (s *azureVolumeAdder) String() string { + return "-" } func (s *azureVolumeAdder) Set(containerName string) error { - if containerName == "" { - return errors.New("no container name given") - } - if azureStorageAccountName == "" || azureStorageAccountKeyFile == "" { - return errors.New("-azure-storage-account-name and -azure-storage-account-key-file arguments must given before -azure-storage-container-volume") - } - accountKey, err := readKeyFromFile(azureStorageAccountKeyFile) - if err != nil { - return err - } - azClient, err := storage.NewBasicClient(azureStorageAccountName, accountKey) - if err != nil { - return errors.New("creating Azure storage client: " + err.Error()) - } - if flagSerializeIO { - log.Print("Notice: -serialize is not supported by azure-blob-container volumes.") - } - v := NewAzureBlobVolume(azClient, containerName, flagReadonly, azureStorageReplication) - if err := v.Check(); err != nil { - return err - } - *s.volumeSet = append(*s.volumeSet, v) + s.Config.Volumes = append(s.Config.Volumes, &AzureBlobVolume{ + ContainerName: containerName, + StorageAccountName: azureStorageAccountName, + StorageAccountKeyFile: azureStorageAccountKeyFile, + AzureReplication: azureStorageReplication, + ReadOnly: deprecated.flagReadonly, + }) return nil } func init() { - flag.Var(&azureVolumeAdder{&volumes}, + VolumeTypes = append(VolumeTypes, func() VolumeWithExamples { return &AzureBlobVolume{} }) + + flag.Var(&azureVolumeAdder{theConfig}, "azure-storage-container-volume", "Use the given container as a storage volume. Can be given multiple times.") flag.StringVar( @@ -79,60 +79,116 @@ func init() { &azureStorageAccountKeyFile, "azure-storage-account-key-file", "", - "File containing the account key used for subsequent --azure-storage-container-volume arguments.") + "`File` containing the account key used for subsequent --azure-storage-container-volume arguments.") flag.IntVar( &azureStorageReplication, "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 // container. type AzureBlobVolume struct { - azClient storage.Client - bsClient storage.BlobStorageClient - containerName string - readonly bool - replication int + StorageAccountName string + StorageAccountKeyFile string + ContainerName string + AzureReplication int + ReadOnly bool + RequestTimeout arvados.Duration + + azClient storage.Client + bsClient storage.BlobStorageClient } -// 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, - bsClient: client.GetBlobService(), - containerName: containerName, - readonly: readonly, - replication: replication, +// Examples implements VolumeWithExamples. +func (*AzureBlobVolume) Examples() []Volume { + return []Volume{ + &AzureBlobVolume{ + StorageAccountName: "example-account-name", + StorageAccountKeyFile: "/etc/azure_storage_account_key.txt", + ContainerName: "example-container-name", + AzureReplication: 3, + RequestTimeout: azureDefaultRequestTimeout, + }, } } -// Check returns nil if the volume is usable. -func (v *AzureBlobVolume) Check() error { - ok, err := v.bsClient.ContainerExists(v.containerName) +// Type implements Volume. +func (v *AzureBlobVolume) Type() string { + return "Azure" +} + +// Start implements Volume. +func (v *AzureBlobVolume) Start() error { + if v.ContainerName == "" { + return errors.New("no container name given") + } + if v.StorageAccountName == "" || v.StorageAccountKeyFile == "" { + return errors.New("StorageAccountName and StorageAccountKeyFile must be given") + } + accountKey, err := readKeyFromFile(v.StorageAccountKeyFile) + if err != nil { + return err + } + v.azClient, err = storage.NewBasicClient(v.StorageAccountName, accountKey) + if err != nil { + return fmt.Errorf("creating Azure storage client: %s", err) + } + + if v.RequestTimeout == 0 { + v.RequestTimeout = azureDefaultRequestTimeout + } + v.azClient.HTTPClient = &http.Client{ + Timeout: time.Duration(v.RequestTimeout), + } + v.bsClient = v.azClient.GetBlobService() + + ok, err := v.bsClient.ContainerExists(v.ContainerName) if err != nil { return err } if !ok { - return errors.New("container does not exist") + return fmt.Errorf("Azure container %q does not exist", v.ContainerName) } return nil } +// Return true if expires_at metadata attribute is found on the block +func (v *AzureBlobVolume) checkTrashed(loc string) (bool, map[string]string, error) { + metadata, err := v.bsClient.GetBlobMetadata(v.ContainerName, loc) + if err != nil { + return false, metadata, v.translateError(err) + } + if metadata["expires_at"] != "" { + return true, metadata, nil + } + return false, metadata, 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) { +func (v *AzureBlobVolume) Get(ctx context.Context, loc string, buf []byte) (int, error) { + trashed, _, err := v.checkTrashed(loc) + if err != nil { + return 0, err + } + if trashed { + return 0, os.ErrNotExist + } var deadline time.Time haveDeadline := false - buf, err := v.get(loc) - for err == nil && len(buf) == 0 && loc != "d41d8cd98f00b204e9800998ecf8427e" { + size, err := v.get(loc, buf) + for err == nil && size == 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" @@ -152,64 +208,135 @@ func (v *AzureBlobVolume) Get(loc string) ([]byte, error) { } else if time.Now().After(deadline) { break } - bufs.Put(buf) time.Sleep(azureWriteRacePollTime) - buf, err = v.get(loc) + size, err = v.get(loc, buf) } if haveDeadline { - log.Printf("Race ended with len(buf)==%d", len(buf)) + log.Printf("Race ended with size==%d", size) } - return buf, err + return size, err } -func (v *AzureBlobVolume) get(loc string) ([]byte, error) { - rdr, err := v.bsClient.GetBlob(v.containerName, loc) - if err != nil { - return nil, v.translateError(err) +func (v *AzureBlobVolume) get(loc string, buf []byte) (int, error) { + expectSize := len(buf) + 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 0, v.translateError(err) + } + if props.ContentLength > int64(BlockSize) || props.ContentLength < 0 { + return 0, fmt.Errorf("block %s invalid size %d (max %d)", loc, props.ContentLength, BlockSize) + } + expectSize = int(props.ContentLength) } - defer rdr.Close() - buf := bufs.Get(BlockSize) - n, err := io.ReadFull(rdr, buf) - switch err { - case nil, io.EOF, io.ErrUnexpectedEOF: - return buf[:n], nil - default: - bufs.Put(buf) - return nil, err + + if expectSize == 0 { + return 0, 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), nil) + } + 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 { + return 0, v.translateError(err) + } + } + return 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) +func (v *AzureBlobVolume) Compare(ctx context.Context, loc string, expect []byte) error { + trashed, _, err := v.checkTrashed(loc) + if err != nil { + return err + } + if trashed { + return os.ErrNotExist + } + rdr, err := v.bsClient.GetBlob(v.ContainerName, loc) if err != nil { return v.translateError(err) } defer rdr.Close() - return compareReaderWithBuf(rdr, expect, loc[:32]) + return compareReaderWithBuf(ctx, rdr, expect, loc[:32]) } -// Put sotres a Keep block as a block blob in the container. -func (v *AzureBlobVolume) Put(loc string, block []byte) error { - if v.readonly { +// Put stores a Keep block as a block blob in the container. +func (v *AzureBlobVolume) Put(ctx context.Context, loc string, block []byte) error { + if v.ReadOnly { return MethodDisabledError } - return v.bsClient.CreateBlockBlobFromReader(v.containerName, loc, uint64(len(block)), bytes.NewReader(block)) + return v.bsClient.CreateBlockBlobFromReader(v.ContainerName, loc, uint64(len(block)), bytes.NewReader(block), nil) } // Touch updates the last-modified property of a block blob. func (v *AzureBlobVolume) Touch(loc string) error { - if v.readonly { + if v.ReadOnly { return MethodDisabledError } - return v.bsClient.SetBlobMetadata(v.containerName, loc, map[string]string{ - "touch": fmt.Sprintf("%d", time.Now()), - }) + trashed, metadata, err := v.checkTrashed(loc) + if err != nil { + return err + } + if trashed { + return os.ErrNotExist + } + + metadata["touch"] = fmt.Sprintf("%d", time.Now()) + return v.bsClient.SetBlobMetadata(v.ContainerName, loc, metadata, nil) } // 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) + trashed, _, err := v.checkTrashed(loc) + if err != nil { + return time.Time{}, err + } + if trashed { + return time.Time{}, os.ErrNotExist + } + + props, err := v.bsClient.GetBlobProperties(v.ContainerName, loc) if err != nil { return time.Time{}, err } @@ -220,10 +347,11 @@ func (v *AzureBlobVolume) Mtime(loc string) (time.Time, error) { // container. func (v *AzureBlobVolume) IndexTo(prefix string, writer io.Writer) error { params := storage.ListBlobsParameters{ - Prefix: prefix, + Prefix: prefix, + Include: "metadata", } for { - resp, err := v.bsClient.ListBlobs(v.containerName, params) + resp, err := v.bsClient.ListBlobs(v.ContainerName, params) if err != nil { return err } @@ -243,7 +371,11 @@ func (v *AzureBlobVolume) IndexTo(prefix string, writer io.Writer) error { // value. continue } - fmt.Fprintf(writer, "%s+%d %d\n", b.Name, b.Properties.ContentLength, t.Unix()) + if b.Metadata["expires_at"] != "" { + // Trashed blob; exclude it from response + continue + } + fmt.Fprintf(writer, "%s+%d %d\n", b.Name, b.Properties.ContentLength, t.UnixNano()) } if resp.NextMarker == "" { return nil @@ -252,30 +384,60 @@ func (v *AzureBlobVolume) IndexTo(prefix string, writer io.Writer) error { } } -// Delete a Keep block. -func (v *AzureBlobVolume) Delete(loc string) error { - if v.readonly { +// Trash a Keep block. +func (v *AzureBlobVolume) Trash(loc string) error { + if v.ReadOnly { return MethodDisabledError } + // Ideally we would use If-Unmodified-Since, but that // particular condition seems to be ignored by Azure. Instead, // we get the Etag before checking Mtime, and use If-Match to // ensure we don't delete data if Put() or Touch() happens // between our calls to Mtime() and DeleteBlob(). - props, err := v.bsClient.GetBlobProperties(v.containerName, loc) + props, err := v.bsClient.GetBlobProperties(v.ContainerName, loc) if err != nil { return err } if t, err := v.Mtime(loc); err != nil { return err - } else if time.Since(t) < blobSignatureTTL { + } else if time.Since(t) < theConfig.BlobSignatureTTL.Duration() { return nil } - return v.bsClient.DeleteBlob(v.containerName, loc, map[string]string{ + + // If TrashLifetime == 0, just delete it + if theConfig.TrashLifetime == 0 { + return v.bsClient.DeleteBlob(v.ContainerName, loc, map[string]string{ + "If-Match": props.Etag, + }) + } + + // Otherwise, mark as trash + return v.bsClient.SetBlobMetadata(v.ContainerName, loc, map[string]string{ + "expires_at": fmt.Sprintf("%d", time.Now().Add(theConfig.TrashLifetime.Duration()).Unix()), + }, map[string]string{ "If-Match": props.Etag, }) } +// Untrash a Keep block. +// Delete the expires_at metadata attribute +func (v *AzureBlobVolume) Untrash(loc string) error { + // if expires_at does not exist, return NotFoundError + metadata, err := v.bsClient.GetBlobMetadata(v.ContainerName, loc) + if err != nil { + return v.translateError(err) + } + if metadata["expires_at"] == "" { + return os.ErrNotExist + } + + // reset expires_at metadata attribute + metadata["expires_at"] = "" + err = v.bsClient.SetBlobMetadata(v.ContainerName, loc, metadata, nil) + return v.translateError(err) +} + // Status returns a VolumeStatus struct with placeholder data. func (v *AzureBlobVolume) Status() *VolumeStatus { return &VolumeStatus{ @@ -287,19 +449,19 @@ 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) + 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 + 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 + return v.AzureReplication } // If possible, translate an Azure SDK error to a recognizable error @@ -308,7 +470,7 @@ func (v *AzureBlobVolume) translateError(err error) error { switch { case err == nil: return err - case strings.Contains(err.Error(), "404 Not Found"): + case strings.Contains(err.Error(), "Not Found"): // "storage: service returned without a response body (404 Not Found)" return os.ErrNotExist default: @@ -317,6 +479,58 @@ func (v *AzureBlobVolume) translateError(err error) error { } var keepBlockRegexp = regexp.MustCompile(`^[0-9a-f]{32}$`) + func (v *AzureBlobVolume) isKeepBlock(s string) bool { return keepBlockRegexp.MatchString(s) } + +// EmptyTrash looks for trashed blocks that exceeded TrashLifetime +// and deletes them from the volume. +func (v *AzureBlobVolume) EmptyTrash() { + var bytesDeleted, bytesInTrash int64 + var blocksDeleted, blocksInTrash int + params := storage.ListBlobsParameters{Include: "metadata"} + + for { + resp, err := v.bsClient.ListBlobs(v.ContainerName, params) + if err != nil { + log.Printf("EmptyTrash: ListBlobs: %v", err) + break + } + for _, b := range resp.Blobs { + // Check if the block is expired + if b.Metadata["expires_at"] == "" { + continue + } + + blocksInTrash++ + bytesInTrash += b.Properties.ContentLength + + expiresAt, err := strconv.ParseInt(b.Metadata["expires_at"], 10, 64) + if err != nil { + log.Printf("EmptyTrash: ParseInt(%v): %v", b.Metadata["expires_at"], err) + continue + } + + if expiresAt > time.Now().Unix() { + continue + } + + err = v.bsClient.DeleteBlob(v.ContainerName, b.Name, map[string]string{ + "If-Match": b.Properties.Etag, + }) + if err != nil { + log.Printf("EmptyTrash: DeleteBlob(%v): %v", b.Name, err) + continue + } + blocksDeleted++ + bytesDeleted += b.Properties.ContentLength + } + if resp.NextMarker == "" { + break + } + params.Marker = resp.NextMarker + } + + log.Printf("EmptyTrash stats for %v: Deleted %v bytes in %v blocks. Remaining in trash: %v bytes in %v blocks.", v.String(), bytesDeleted, blocksDeleted, bytesInTrash-bytesDeleted, blocksInTrash-blocksDeleted) +}