You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 5 Next »

(Page is work in progress.)

DFDL needs an extension that allows data much larger than memory to be manipulated.

A variety of data formats such as for image and video files, consist of fields of what is effectively metadata, surrounding large blocks of data containing compressed image or video data.

An important use case for DFDL is to expose this metadata for easy use, and to provide access to the large data via a streaming mechanism akin to opening a file.

In RDBMS systems, BLOB (Binary Large Object) and CLOB (Character large object) are the types used when the data row returned from an SQL query will not contain the actual value data, but rather a handle that can be used to open/read/write/close the BLOB or CLOB.

DFDL needs at least BLOB capability. This would enable processing of images or video of arbitrary size without the need to every hold all the data in memory.

This also eliminates the limitation on object size.

Basic Blob Requirements

The basic requirement has almost nothing to do with DFDL.

We want to represent an image file in XML, except the BLOB of compressed image data, which we want to reliably incorporate by reference.

So instead of

<?xml version="1.0" ?>
<someImage>
  <lat>44.9090</lat>
  <lon>70.2929</lon>
  <img>
   098fad0965edcab...giant.hexbinary.string...many megs or gigs in size
  </img>
</someImage>

Instead the bytes corresponding to the image data go in a separate file "img.dat", and the infoset becomes

<?xml version="1.0" ?>
<someImage>
  <lat>44.9090</lat>
  <lon>70.2929</lon>
  ... some way of saying img.dat blob goes here...
</someImage>

A few requirements:

  1. The document must still be validated relative to its DFDL schema/XML Schema - so the BLOB must be content that can be validated. That suggests it is an element. This validation does not have to touch or even verify the existence of the BLOB file.
    1. This means the element must be expressed in DFDL's subset of XML Schema. Hence, it is not an element with attributes, as attributes aren't part of the DFDL schema language.
  2. The BLOB must be able to refer to a region of bytes within a file. This is so that DFDL can be used to identify the location of the BLOB in a file being parsed, without having to copy or bring into memory the BLOB data. Rather, the Infoset can contain a BLOB that identifies the original file, and the location within it.
    1. Note: This is a special case of a general capability for any element in a DFDL schema - users may want to know its exact starting position and length, measured in bits or bytes, if only for trace/debug or verification purposes.
  3. It should not require Daffodil to be used to manipulate these XML files that contain BLOB references. Interpreting the BLOB information should not require information bases that are maintained by Daffodil libraries (e.g., mappings from GUIDs to files)
    1. We may want to provide a convenient scala/java library for this, it should not be bundled into Daffodil libraries, but be easily isolated.

One concrete suggestion is:

<?xml version="1.0" ?>
<someImage>
  <lat>44.9090</lat>
  <lon>70.2929</lon>
  <img><BLOB daf:BLOB="true">../blobs/img.dat?offset0b=0;kind=raw</BLOB></img>
</someImage>

In the above we've introduced an element named BLOB which which takes a special URI which can be absolute or relative, and it identifies the blob data. The offset0b is a zero-based byte-offset into the file where the BLOB data starts. The suffix "0b" on the name indicates that it is zero-based, to distinguish from XML normal conventions which are 1-based. The value of offset0b defaults to 0. An optional length=N attribute would constrain the length of the BLOB data, and the kind=raw gives that the data is not encoded or compressed in any way. (kind=raw would be the default.)

This URL would be parsed by conventional URL libraries. What is called the query, the part after the "?" is a list (";" separated) of pairs of "keyword=value" form.

BLOBs as Layers

A DFDL schema using a BLOB would look, for example, like this:

<element name="img" >
  <complexType>
    <sequence>
      <element name="BLOB" daf:layerBoundaryMark="[END-IMAGE]"
         type="daf:URI4BLOB" daf:layerTransform="daf:BLOB" daf:layerLengthKind="boundaryMark"/>
    </sequence>
  </complexType>
</element>

A schema containing daf:URI4BLOB would be provided and would contain roughly:

<simpleType name="URI4BLOB" dfdl:encoding="utf-8">
  <restriction base="xs:string">
     <pattern value="..regex for these URIs.."/>
  </restriction>
</simpleType>

Here we see that a BLOB is actually created by way of a layering. The BLOB layer implements isolation of the BLOB contents, and produces (when parsing),  bytes containing the URI in UTF-8 encoding. When unparsing, the layer transform takes the URI, and obtains the corresponding bytes by opening the URI to obtain a Java InputStream.

BLOB Use Cases

There are a few different use cases. The variations have to do with how the BLOB data is accessed, over what time span it is accessible, and when resources can be reclaimed.

Image Filtering In A Process Pipeline

Parser produces an infoset containing a durable blob handle. This blob handle provides access to the blob data even after the parser has terminated, and the process exited.

The blob handle can be opened, to get an input stream, and the bytes in it read like any Java InputStream.

The parser must be run in a BLOB='persistent' mode (API TBD for this.) which tells it to create permanent URIs, and never to release/delete the underlying resources in any automatic way.

An API provides the ability to create a blob handle for a Java OutputStream (the two can be created simultaneously), which can then be opened, written, closed/flushed, and then blob handle can be used as a replacement for an input blob handle.

The notion here is that one opens and reads-from the input blob handle, and one processes the data, and if modified, you supply, on output, a replacement blob handle.

The unparser consumes an infoset containing a blob handle, and reads from it the data, writing that as the "contents" of the corresponding element.

The parser and unparser are independent processes that do not necessarily overlap in time existence. Their only communication is through the blob handle. Hence, the blob objects are allocated at the system level, and are not part of the state of the parser nor unparser. (E.g., they could be files).

A blob handle survives a reboot of the computer system - it's state is durable, so that if you write out the infoset from a parse of data, as an XML text file, then reboot the computer, you can then read that XML text file, find the BLOB handles within it, and open them.

A blob handle is some opaque URI, supporting the openStream API.

Each BLOB must be explicitly discarded. A convenience API might walk an entire infoset (as XML), and discard each BLOB found.

A non-native attribute daf:BLOB='true' is the XML representation indicating a BLOB-valued element. The blob handle is the VALUE of the element.

The lifetime of the BLOB resources (typically files) is not controlled in any way here any more than the lifetime of the original file.

Single Process, Single Thread, SAX-style Event, Stateless

In this case,  a single process with code written in Scala/Java is performing parse, transform, and unparse of data. The code is single threaded.

The parser is generating SAX-style infoset events for the start and end of each element.

BLOBs are processed in a streaming mode (API call to set this TBD).

To process BLOB contents, the application's startElement() method would simply have to check for a blob (by calling isBLOB() method which is part of the extended API of an event handler).

(TBD: or we could require the handlers to be special blob-aware handler with a startBLOBElement() method and endBLOBElement() method. This potentially is lower overhead.)

The lifetime of this BLOB input stream is only until the SAX-style event callback returns. At that point the resources/storage can be reclaimed.

So the parser BLOB API is that the parser calls the SAX-style event handler with a BLOB method, handing it an open input stream.

The unparser BLOB API is to be such a SAX-style event handler, and implement the BLOB method, reading data from the open input stream, and unparsing it.

In this use case, the DFDL schema element corresponding to the BLOB object must carry an explicit BLOB annotation (extension to DFDL v1.0) indicating that it is to be treated as a BLOB, and that its 'value' is a BLOB handle (Which could be a BLOB URI).

However, in this case, the BLOB handle, if output as text (e.g., by printing the resulting XML instead of unparsing it), just serves to document the past skipping over of the BLOB.

It is possible to parse and unparse an arbitrarily large image file in only finite memory using this API, so long as the image file format is streamable. 

Implementation Note: For unparsing, DirectOrBufferedDataOutputStream may need to grow a special form of BufferedDataOutputStream which is a BLOB. No point in double buffering a BLOB, the BLOB object itself is very much a buffer.  We simply need to know how to recombine its data into the DirectDataOutputStream at the right point in tim


Implementation Concerns

  • Want to avoid copying the BLOB data to a file when possible.
  • Want to take advantage of the ability to store offsets+lengths for BLOBSs as locations in the original input file.
  • Ideally, a BLOB needn't be bytes of the original file, but could be bytes inside a layer, such as a compressed region
    • Ideally, such a BLOB could be accessed without having to decompress everything in advance into memory or a file. I.e., such a BLOB could be streamed from its layer, and the decompressing would happen as part of accessing it. 
    • This implies that while an offset maybe known, a length may not.
    • This implies that while an offset may be known it is not necessarily a file offset, but an offset within some layer-stream.


  • No labels