// CloseArtifact closes an artifact for further writes and begins process of merging and uploading
// the artifact. This operation is only valid for artifacts which are being uploaded in chunks.
// In all other cases, an error is returned.
func CloseArtifact(ctx context.Context, artifact *model.Artifact, db database.Database, s3bucket *s3.Bucket, failIfAlreadyClosed bool) error {
switch artifact.State {
case model.UPLOADED:
// Already closed. Nothing to do here.
fallthrough
case model.APPEND_COMPLETE:
// This artifact will be eventually shipped to S3. No change required.
return nil
case model.APPENDING:
artifact.State = model.APPEND_COMPLETE
if err := db.UpdateArtifact(artifact); err != nil {
return err
}
return MergeLogChunks(ctx, artifact, db, s3bucket)
case model.WAITING_FOR_UPLOAD:
// Streaming artifact was not uploaded
artifact.State = model.CLOSED_WITHOUT_DATA
if err := db.UpdateArtifact(artifact); err != nil {
return err
}
return nil
default:
return fmt.Errorf("Unexpected artifact state: %s", artifact.State)
}
}
// Merges all of the individual chunks into a single object and stores it on s3.
// The log chunks are stored in the database, while the object is uploaded to s3.
func MergeLogChunks(ctx context.Context, artifact *model.Artifact, db database.Database, s3bucket *s3.Bucket) error {
switch artifact.State {
case model.APPEND_COMPLETE:
// TODO: Reimplement using GorpDatabase
// If the file is empty, don't bother creating an object on S3.
if artifact.Size == 0 {
artifact.State = model.CLOSED_WITHOUT_DATA
artifact.S3URL = ""
// Conversion between *DatabaseEror and error is tricky. If we don't do this, a nil
// *DatabaseError can become a non-nil error.
return db.UpdateArtifact(artifact).GetError()
}
// XXX Do we need to commit here or is this handled transparently?
artifact.State = model.UPLOADING
if err := db.UpdateArtifact(artifact); err != nil {
return err
}
fileName := artifact.DefaultS3URL()
r := newLogChunkReaderWithReadahead(artifact, db)
if err := uploadArtifactToS3(s3bucket, fileName, artifact.Size, r); err != nil {
return err
}
// XXX This is a long operation and should probably be asynchronous from the
// actual HTTP request, and the client should poll to check when its uploaded.
artifact.State = model.UPLOADED
artifact.S3URL = fileName
if err := db.UpdateArtifact(artifact); err != nil {
return err
}
// From this point onwards, we will not send back any errors back to the user. If we are
// unable to delete logchunks, we log it to Sentry instead.
if _, err := db.DeleteLogChunksForArtifact(artifact.Id); err != nil {
sentry.ReportError(ctx, err)
return nil
}
return nil
case model.WAITING_FOR_UPLOAD:
fallthrough
case model.ERROR:
fallthrough
case model.APPENDING:
fallthrough
case model.UPLOADED:
fallthrough
case model.UPLOADING:
return fmt.Errorf("Artifact can only be merged when in APPEND_COMPLETE state, but state is %s", artifact.State)
default:
return fmt.Errorf("Illegal artifact state! State code is %d", artifact.State)
}
}
func AppendLogChunk(db database.Database, artifact *model.Artifact, logChunk *model.LogChunk) *HttpError {
if artifact.State != model.APPENDING {
return NewHttpError(http.StatusBadRequest, fmt.Sprintf("Unexpected artifact state: %s", artifact.State))
}
if logChunk.Size <= 0 {
return NewHttpError(http.StatusBadRequest, "Invalid chunk size %d", logChunk.Size)
}
if logChunk.Content == "" {
return NewHttpError(http.StatusBadRequest, "Empty content string")
}
if int64(len(logChunk.Content)) != logChunk.Size {
return NewHttpError(http.StatusBadRequest, "Content length does not match indicated size")
}
// Find previous chunk in DB - append only
if nextByteOffset, err := db.GetLastByteSeenForArtifact(artifact.Id); err != nil {
return NewHttpError(http.StatusInternalServerError, "Error while checking for previous byte range: %s", err)
} else if nextByteOffset != logChunk.ByteOffset {
return NewHttpError(http.StatusBadRequest, "Overlapping ranges detected, expected offset: %d, actual offset: %d", nextByteOffset, logChunk.ByteOffset)
}
logChunk.ArtifactId = artifact.Id
// Expand artifact size - redundant after above change.
if artifact.Size < logChunk.ByteOffset+logChunk.Size {
artifact.Size = logChunk.ByteOffset + logChunk.Size
if err := db.UpdateArtifact(artifact); err != nil {
return NewHttpError(http.StatusInternalServerError, err.Error())
}
}
if err := db.InsertLogChunk(logChunk); err != nil {
return NewHttpError(http.StatusBadRequest, "Error updating log chunk: %s", err)
}
return nil
}
// CreateArtifact creates a new artifact in a open bucket.
//
// If an artifact with the same name already exists in the same bucket, we attempt to rename the
// artifact by adding a suffix.
// If the request specifies a chunked artifact, the size field is ignored and always set to zero.
// If the request is for a streamed artifact, size is mandatory.
// A relative path field may be specified to preserve the original file name and path. If no path is
// specified, the original artifact name is used by default.
func CreateArtifact(req createArtifactReq, bucket *model.Bucket, db database.Database) (*model.Artifact, *HttpError) {
if len(req.Name) == 0 {
return nil, NewHttpError(http.StatusBadRequest, "Artifact name not provided")
}
if bucket.State != model.OPEN {
return nil, NewHttpError(http.StatusBadRequest, "Bucket is already closed")
}
artifact := new(model.Artifact)
artifact.Name = req.Name
artifact.BucketId = bucket.Id
artifact.DateCreated = time.Now()
if req.DeadlineMins == 0 {
artifact.DeadlineMins = DEFAULT_DEADLINE
} else {
artifact.DeadlineMins = req.DeadlineMins
}
if req.Chunked {
artifact.State = model.APPENDING
} else {
if req.Size == 0 {
return nil, NewHttpError(http.StatusBadRequest, "Cannot create a new upload artifact without size.")
} else if req.Size > MaxArtifactSizeBytes {
return nil, NewHttpError(http.StatusRequestEntityTooLarge, fmt.Sprintf("Entity '%s' (size %d) is too large (limit %d)", req.Name, req.Size, MaxArtifactSizeBytes))
}
artifact.Size = req.Size
artifact.State = model.WAITING_FOR_UPLOAD
}
if req.RelativePath == "" {
// Use artifact name provided as default relativePath
artifact.RelativePath = req.Name
} else {
artifact.RelativePath = req.RelativePath
}
// Attempt to insert artifact and retry with a different name if it fails.
if err := db.InsertArtifact(artifact); err != nil {
for attempt := 1; attempt <= MaxDuplicateFileNameResolutionAttempts; attempt++ {
// Unable to create new artifact - if an artifact already exists, the above insert failed
// because of a collision.
if _, err := db.GetArtifactByName(bucket.Id, artifact.Name); err != nil {
// This could be a transient DB error (down/unreachable), in which case we expect the client
// to retry. There is no value in attempting alternate artifact names.
//
// We have no means of verifying there was a name collision - bail with an internal error.
return nil, NewHttpError(http.StatusInternalServerError, err.Error())
}
// File name collision - attempt to resolve
artifact.Name = fmt.Sprintf(DuplicateArtifactNameFormat, req.Name, randString(5))
if err := db.InsertArtifact(artifact); err == nil {
return artifact, nil
}
}
return nil, NewHttpError(http.StatusInternalServerError, "Exceeded retry limit avoiding duplicates")
}
return artifact, nil
}
// PutArtifact writes a streamed artifact to S3. The entire file contents are streamed directly
// through to S3. If S3 is not accessible, we don't make any attempt to buffer on disk and fail
// immediately.
func PutArtifact(ctx context.Context, artifact *model.Artifact, db database.Database, bucket *s3.Bucket, req PutArtifactReq) error {
if artifact.State != model.WAITING_FOR_UPLOAD {
return fmt.Errorf("Expected artifact to be in state WAITING_FOR_UPLOAD: %s", artifact.State)
}
// New file being inserted into DB.
// Mark status change to UPLOADING and start uploading to S3.
//
// First, verify that the size of the content being uploaded matches our expected size.
var fileSize int64
var err error
if req.ContentLength != "" {
fileSize, err = strconv.ParseInt(req.ContentLength, 10, 64) // string, base, bits
// This should never happen if a sane HTTP client is used. Nonetheless ...
if err != nil {
return fmt.Errorf("Invalid Content-Length specified")
}
} else {
// This too should never happen if a sane HTTP client is used. Nonetheless ...
return fmt.Errorf("Content-Length not specified")
}
if fileSize != artifact.Size {
return fmt.Errorf("Content length %d does not match expected file size %d", fileSize, artifact.Size)
}
artifact.State = model.UPLOADING
if err := db.UpdateArtifact(artifact); err != nil {
return err
}
cleanupAndReturn := func(err error) error {
// TODO: Is there a better way to detect and handle errors?
// Use a channel to signify upload completion. In defer, check if the channel is empty. If
// yes, mark error. Else ignore.
if err != nil {
// TODO: s/ERROR/WAITING_FOR_UPLOAD/ ?
sentry.ReportError(ctx, err)
artifact.State = model.ERROR
err2 := db.UpdateArtifact(artifact)
if err2 != nil {
log.Printf("Error while handling error: %s", err2.Error())
}
return err
}
return nil
}
b := new(bytes.Buffer)
// Note: Storing entire contents of uploaded artifact in memory can cause OOMS.
if n, err := io.CopyN(b, req.Body, artifact.Size); err != nil {
return cleanupAndReturn(fmt.Errorf("Error reading from request body (for artifact %s/%s, bytes (%d/%d) read): %s", artifact.BucketId, artifact.Name, n, artifact.Size, err))
}
fileName := artifact.DefaultS3URL()
if err := uploadArtifactToS3(bucket, fileName, artifact.Size, bytes.NewReader(b.Bytes())); err != nil {
return cleanupAndReturn(err)
}
artifact.State = model.UPLOADED
artifact.S3URL = fileName
if err := db.UpdateArtifact(artifact); err != nil {
return err
}
return nil
}
// AppendLogChunk appends a logchunk to an artifact.
// If the logchunk position does not match the current end of artifact, an error is returned.
// An exception to this is made when the last seen logchunk is repeated, which is silently ignored
// without an error.
func AppendLogChunk(ctx context.Context, db database.Database, artifact *model.Artifact, logChunkReq *createLogChunkReq) *HttpError {
if artifact.State != model.APPENDING {
return NewHttpError(http.StatusBadRequest, fmt.Sprintf("Unexpected artifact state: %s", artifact.State))
}
if logChunkReq.Size <= 0 {
return NewHttpError(http.StatusBadRequest, "Invalid chunk size %d", logChunkReq.Size)
}
var contentBytes []byte
if len(logChunkReq.Bytes) != 0 {
// If request sent Bytes, use Bytes.
if int64(len(logChunkReq.Bytes)) != logChunkReq.Size {
return NewHttpError(http.StatusBadRequest, "Content length %d does not match indicated size %d", len(logChunkReq.Bytes), logChunkReq.Size)
}
contentBytes = logChunkReq.Bytes
} else {
// Otherwise, allow Content, for now.
if len(logChunkReq.Content) == 0 {
return NewHttpError(http.StatusBadRequest, "Empty content string")
}
if int64(len(logChunkReq.Content)) != logChunkReq.Size {
return NewHttpError(http.StatusBadRequest, "Content length %d does not match indicated size %d", len(logChunkReq.Content), logChunkReq.Size)
}
contentBytes = []byte(logChunkReq.Content)
}
// Find previous chunk in DB - append only
nextByteOffset := artifact.Size
if nextByteOffset != logChunkReq.ByteOffset {
// There is a possibility the previous logchunk is being retried - we need to handle cases where
// a server/proxy time out caused the client not to get an ACK when it successfully uploaded the
// previous logchunk, due to which it is retrying.
//
// This is a best-effort check - if we encounter DB errors or any mismatch in the chunk
// contents, we ignore this test and claim that a range mismatch occured.
if nextByteOffset != 0 && nextByteOffset == logChunkReq.ByteOffset+logChunkReq.Size {
if prevLogChunk, err := db.GetLastLogChunkSeenForArtifact(artifact.Id); err == nil {
if prevLogChunk != nil && prevLogChunk.ByteOffset == logChunkReq.ByteOffset && prevLogChunk.Size == logChunkReq.Size && bytes.Equal(prevLogChunk.ContentBytes, contentBytes) {
sentry.ReportMessage(ctx, fmt.Sprintf("Received duplicate chunk for artifact %v of size %d at byte %d", artifact.Id, logChunkReq.Size, logChunkReq.ByteOffset))
return nil
}
}
}
return NewHttpError(http.StatusBadRequest, "Overlapping ranges detected, expected offset: %d, actual offset: %d", nextByteOffset, logChunkReq.ByteOffset)
}
// Expand artifact size - redundant after above change.
if artifact.Size < logChunkReq.ByteOffset+logChunkReq.Size {
artifact.Size = logChunkReq.ByteOffset + logChunkReq.Size
if err := db.UpdateArtifact(artifact); err != nil {
return NewHttpError(http.StatusInternalServerError, err.Error())
}
}
logChunk := &model.LogChunk{
ArtifactId: artifact.Id,
ByteOffset: logChunkReq.ByteOffset,
ContentBytes: contentBytes,
Size: logChunkReq.Size,
}
if err := db.InsertLogChunk(logChunk); err != nil {
return NewHttpError(http.StatusBadRequest, "Error updating log chunk: %s", err)
}
return nil
}
func PutArtifact(artifact *model.Artifact, db database.Database, bucket *s3.Bucket, req PutArtifactReq) error {
if artifact.State != model.WAITING_FOR_UPLOAD {
return fmt.Errorf("Expected artifact to be in state WAITING_FOR_UPLOAD: %s", artifact.State)
}
// New file being inserted into DB.
// Mark status change to UPLOADING and start uploading to S3.
//
// First, verify that the size of the content being uploaded matches our expected size.
var fileSize int64
var err error
if req.ContentLength != "" {
fileSize, err = strconv.ParseInt(req.ContentLength, 10, 64) // string, base, bits
// This should never happen if a sane HTTP client is used. Nonetheless ...
if err != nil {
return fmt.Errorf("Invalid Content-Length specified")
}
} else {
// This too should never happen if a sane HTTP client is used. Nonetheless ...
return fmt.Errorf("Content-Length not specified")
}
if fileSize != artifact.Size {
return fmt.Errorf("Content length %d does not match expected file size %d", fileSize, artifact.Size)
}
// XXX Do we need to commit here or is this handled transparently?
artifact.State = model.UPLOADING
if err := db.UpdateArtifact(artifact); err != nil {
return err
}
cleanupAndReturn := func(err error) error {
// TODO: Is there a better way to detect and handle errors?
// Use a channel to signify upload completion. In defer, check if the channel is empty. If
// yes, mark error. Else ignore.
if err != nil {
// TODO: s/ERROR/WAITING_FOR_UPLOAD/ ?
log.Printf("Error uploading to S3: %s\n", err)
artifact.State = model.ERROR
err2 := db.UpdateArtifact(artifact)
if err2 != nil {
log.Printf("Error while handling error: %s", err2.Error())
}
return err
}
return nil
}
fileName := artifact.DefaultS3URL()
if err := bucket.PutReader(fileName, req.Body, artifact.Size, "binary/octet-stream", s3.PublicRead); err != nil {
return cleanupAndReturn(fmt.Errorf("Error uploading to S3: %s", err))
}
artifact.State = model.UPLOADED
artifact.S3URL = fileName
if err := db.UpdateArtifact(artifact); err != nil {
return err
}
return nil
}
// Merges all of the individual chunks into a single object and stores it on s3.
// The log chunks are stored in the database, while the object is uploaded to s3.
func MergeLogChunks(artifact *model.Artifact, db database.Database, s3bucket *s3.Bucket) error {
switch artifact.State {
case model.APPEND_COMPLETE:
// TODO: Reimplement using GorpDatabase
// If the file is empty, don't bother creating an object on S3.
if artifact.Size == 0 {
artifact.State = model.CLOSED_WITHOUT_DATA
artifact.S3URL = ""
// Conversion between *DatabaseEror and error is tricky. If we don't do this, a nil
// *DatabaseError can become a non-nil error.
return db.UpdateArtifact(artifact).GetError()
}
// XXX Do we need to commit here or is this handled transparently?
artifact.State = model.UPLOADING
if err := db.UpdateArtifact(artifact); err != nil {
return err
}
logChunks, err := db.ListLogChunksInArtifact(artifact.Id)
if err != nil {
return err
}
r, w := io.Pipe()
errChan := make(chan error)
uploadCompleteChan := make(chan bool)
fileName := artifact.DefaultS3URL()
// Asynchronously upload the object to s3 while reading from the r, w
// pipe. Thus anything written to "w" will be sent to S3.
go func() {
defer close(errChan)
defer close(uploadCompleteChan)
defer r.Close()
if err := s3bucket.PutReader(fileName, r, artifact.Size, "binary/octet-stream", s3.PublicRead); err != nil {
errChan <- fmt.Errorf("Error uploading to S3: %s", err)
return
}
uploadCompleteChan <- true
}()
for _, logChunk := range logChunks {
w.Write([]byte(logChunk.Content))
}
w.Close()
// Wait either for S3 upload to complete or for it to fail with an error.
// XXX This is a long operation and should probably be asynchronous from the
// actual HTTP request, and the client should poll to check when its uploaded.
select {
case _ = <-uploadCompleteChan:
artifact.State = model.UPLOADED
artifact.S3URL = fileName
if err := db.UpdateArtifact(artifact); err != nil {
return err
}
// From this point onwards, we will not send back any errors back to the user. If we are
// unable to delete logchunks, we log it to Sentry instead.
if n, err := db.DeleteLogChunksForArtifact(artifact.Id); err != nil {
// TODO: Send this error to Sentry
log.Printf("Error deleting logchunks for artifact %d: %v\n", artifact.Id, err)
return nil
} else if n != int64(len(logChunks)) {
// TODO: Send this error to Sentry
log.Printf("Mismatch in number of logchunks while deleting logchunks for artifact %d:"+
"Expected: %d Actual: %d\n", artifact.Id, len(logChunks), n)
}
return nil
case err := <-errChan:
return err
}
case model.WAITING_FOR_UPLOAD:
fallthrough
case model.ERROR:
fallthrough
case model.APPENDING:
fallthrough
case model.UPLOADED:
fallthrough
case model.UPLOADING:
return fmt.Errorf("Artifact can only be merged when in APPEND_COMPLETE state, but state is %s", artifact.State)
default:
return fmt.Errorf("Illegal artifact state! State code is %d", artifact.State)
}
}