package main
import (
+ "bytes"
"errors"
"flag"
"fmt"
"strings"
"time"
- "github.com/Azure/azure-sdk-for-go/storage"
+ "github.com/curoverse/azure-sdk-for-go/storage"
)
var (
azureStorageAccountName string
azureStorageAccountKeyFile string
+ azureStorageReplication int
+ azureWriteRaceInterval time.Duration = 15 * time.Second
+ azureWriteRacePollTime time.Duration = time.Second
)
func readKeyFromFile(file string) (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
if flagSerializeIO {
log.Print("Notice: -serialize is not supported by azure-blob-container volumes.")
}
- v := NewAzureBlobVolume(azClient, containerName, flagReadonly)
+ v := NewAzureBlobVolume(azClient, containerName, flagReadonly, azureStorageReplication)
if err := v.Check(); err != nil {
return err
}
"azure-storage-account-key-file",
"",
"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.")
}
// An AzureBlobVolume stores and retrieves blocks in an Azure Blob
bsClient storage.BlobStorageClient
containerName string
readonly bool
+ replication int
}
-func NewAzureBlobVolume(client storage.Client, containerName string, readonly bool) *AzureBlobVolume {
+func NewAzureBlobVolume(client storage.Client, containerName string, readonly bool, replication int) *AzureBlobVolume {
return &AzureBlobVolume{
- azClient: client,
- bsClient: client.GetBlobService(),
+ azClient: client,
+ bsClient: client.GetBlobService(),
containerName: containerName,
- readonly: readonly,
+ readonly: readonly,
+ replication: replication,
}
}
}
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))
+ }
+ return buf, 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)
}
defer rdr.Close()
buf := bufs.Get(BlockSize)
n, err := io.ReadFull(rdr, buf)
switch err {
- case io.EOF, io.ErrUnexpectedEOF:
+ case nil, io.EOF, io.ErrUnexpectedEOF:
return buf[:n], nil
default:
bufs.Put(buf)
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])
if v.readonly {
return MethodDisabledError
}
- if err := v.bsClient.CreateBlockBlob(v.containerName, loc); err != nil {
- return err
- }
- // We use the same block ID, base64("0")=="MA==", for everything.
- if err := v.bsClient.PutBlock(v.containerName, loc, "MA==", block); err != nil {
- return err
- }
- return v.bsClient.PutBlockList(v.containerName, loc, []storage.Block{{"MA==", storage.BlockStatusUncommitted}})
+ return v.bsClient.CreateBlockBlobFromReader(v.containerName, loc, uint64(len(block)), bytes.NewReader(block))
}
func (v *AzureBlobVolume) Touch(loc string) error {
if v.readonly {
return MethodDisabledError
}
- if exists, err := v.bsClient.BlobExists(v.containerName, loc); err != nil {
- return err
- } else if !exists {
- return os.ErrNotExist
- }
- return v.bsClient.PutBlockList(v.containerName, loc, []storage.Block{{"MA==", storage.BlockStatusCommitted}})
+ return v.bsClient.SetBlobMetadata(v.containerName, loc, map[string]string{
+ "touch": fmt.Sprintf("%d", time.Now()),
+ })
}
func (v *AzureBlobVolume) Mtime(loc string) (time.Time, error) {
if err != nil {
return err
}
+ 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 == "" {
}
func (v *AzureBlobVolume) Delete(loc string) error {
- // TODO: Use leases to handle races with Touch and Put.
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)
+ if err != nil {
+ return err
+ }
if t, err := v.Mtime(loc); err != nil {
return err
} else if time.Since(t) < blobSignatureTTL {
return nil
}
- return v.bsClient.DeleteBlob(v.containerName, loc)
+ return v.bsClient.DeleteBlob(v.containerName, loc, map[string]string{
+ "If-Match": props.Etag,
+ })
}
func (v *AzureBlobVolume) Status() *VolumeStatus {
func (v *AzureBlobVolume) Writable() bool {
return !v.readonly
}
+
+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
+ }
+}
projects / arvados.git / blobdiff