Fw: Action 145: 'dispatch' way of discriminating a choice for better performance

10 Apr 2012

      Updated to reflect the outcome of DFDL WG call on 5th April.

Revised details: 

A new dfdl:element property is added called dfdl:elementID of type 'DFDL 
String Literal'. This provides a single alternative name for the element. 
Only allowed on local element and global element, not on element reference 
or simple type. 

A new dfdl:choice property is added called dfdl:choiceBranchRef of type 
'DFDL Expression'. The expression must evaluate to a DFDL String Literal. 
The resultant string must match (case insensitive) the dfdl:elementID 
property values of one of the element branches of the choice, and if so 
discriminates in favour of that branch. The parser then goes straight to 
that branch, ignoring schema order. Because the branch is 'known to exist' 
no backtracking takes place if a processing error subsequently occurs.

Rules: 

- Both properties behave like dfdl:ref and dfdl:hiddenGroupRef in that it 
is not possible to set a value in scope by a dfdl:format annotation, and 
is only set at its point of use. This is because there is nothing sensible 
that could be set in scope **. Empty string is not an allowed value. 

- Both properties are only used when parsing. 

- When dfdl:choiceBranchRef is present, all choice branches must be local 
elements or element references. It is a schema definition error otherwise.

- It is a processing error if the resolved value of dfdl:choiceBranchRef 
does not match one of the branches. 

- It is a schema definition error if individual dfdl:elementID values are 
not unique across all elements that are branches of an xs:choice that 
carries dfdl:choiceBranchRef.

- It is not a schema definition error if both dfdl:initiatedContent and 
dfdl:choiceBranchRef, are provided on the same choice, nor if a 
discriminator exists on that choice branch. ++

** Consequently adding support for the property to existing DFDL 
implementations will not suddenly cause errors to appear in existing DFDL 
schemas.

++ Reasoning. If a discriminator is evaluated and there is no point of 
uncertainty, it simply behaves like an assert, specifically if it fails it 
causes a processing error without backtracking. Use of 
dfdl:choiceBranchRef is just another way to get into this situation. (The 
property dfdl:initiatedContent may be rewritten as a series of 
discriminators on the child elements which test that the (initiated) 
element exists, hence same applies).

Regards

Steve Hanson
Architect, Data Format Description Language (DFDL)
Co-Chair, OGF DFDL Working Group
IBM SWG, Hursley, UK
smh@uk.ibm.com
tel:+44-1962-815848
----- Forwarded by Steve Hanson/UK/IBM on 28/03/2012 16:32 -----

From:   Steve Hanson/UK/IBM
To:     dfdl-wg@ogf.org
Date:   27/03/2012 14:09
Subject:        Fw: [DFDL-WG] Action 145: 'dispatch' way of discriminating 
a choice for better performance

Hi Mike

I agree. When I wrote the SWIFT example expression that I realised this 
was becoming a bit too complicated.  IBM MRM has a quick way of 
dispatching a choice that uses a similar mechanism to the one you propose.
It has an optional property called 'Message Alias'. The parser first tries 
to match the tag against the element name, and if that fails it tries the 
alias.  I've modified my proposal accordingly.

Regarding wildcards, the approach you show was what we had agreed upon a 
couple of years ago when we dropped wildcards. I would prefer not to add 
them back in for 1.0, but I have had some feedback from IBMers modeling 
envelope/payload formats that they want a loose-coupling approach as that 
fits better with other tools and is scalable. Let's treat this as a 
separate issue. I'll go back to the IBMers so that I fully understand 
their concerns, and if necessary I will bring back to the WG as a new 
agenda item.  So all we need to ensure for now is that whatever mechanism 
we come up with is extensible to wildcards, if needed.

Revised details: 

A new dfdl:element property is added called dfdl:elementAlias of type 
'List of String'. This provides one or more alternative names for the 
element. Only allowed on local element and global element, not on element 
reference or simple type. 

A new dfdl:choice property is added called dfdl:choiceBranchRef of type 
DFDL Expression. The expression must evaluate to String. The string must 
match (case insensitive) either the element name or one of the 
dfdl:elementAlias property values of one of the element branches of the 
choice, and if so discriminates in favour of that branch. The parser then 
goes straight to that branch, ignoring schema order. Because the branch is 
'known to exist' no backtracking takes place if a processing error occurs.

Rules: 

- Both properties behave like dfdl:ref and dfdl:hiddenGroupRef in that it 
is not possible to set a value in scope by a dfdl:format annotation, and 
is only set at its point of use. This is because there is nothing sensible 
that could be set in scope. But it has the benefit that adding support for 
the property to existing DFDL implementations will not suddenly cause 
errors to appear in existing DFDL schemas.  Empty string is not an allowed 
value. 

- Both properties are only used when parsing. 

- When dfdl:choiceBranchRef is preent, all choice branches must be local 
elements or element references. It is a schema definition error otherwise.

- It is a processing error if the resolved xs:string value of 
dfdl:choiceBranchRef does not match one of the branches. 

- It is a schema definition error if a choice has both 
dfdl:choiceBranchRef set and dfdl:initiatedContent="yes".

- It is a schema definition error if individual dfdl:elementAlias values 
are not unique a) across all elements that are branches of a given 
xs:choice and b) across all global elements in the schema.

So we now have the ability to apply a simple lookup before we start to 
process a choice. That makes the processing time for each branch of the 
choice independent of its order in the schema. 

Questions:

- What happens if we encounter a discriminator once we are processing the 
branch and its point of uncertainty is the choice ?

Regards

Steve Hanson
Architect, Data Format Description Language (DFDL)
Co-Chair, OGF DFDL Working Group
IBM SWG, Hursley, UK
smh@uk.ibm.com
tel:+44-1962-815848
----- Forwarded by Steve Hanson/UK/IBM on 27/03/2012 13:41 -----

From:   Mike Beckerle <mbeckerle.dfdl@gmail.com>
To:     Steve Hanson/UK/IBM@IBMGB
Cc:     dfdl-wg@ogf.org
Date:   21/03/2012 15:57
Subject:        Re: [DFDL-WG] Action 145: 'dispatch' way of discriminating 
a choice for better performance (updated)
Sent by:        dfdl-wg-bounces@ogf.org

The whole point of this thing is to be faster, not more general, so my
reaction is too much XPath expression complexity here.

Consider this. If the tag is some 2-character code that does NOT want
to be the same as the element names (for example because they're
digits, so they can't be the element names exactly since element names
have to begin with an alpha char. Digits also aren't useful from
readability perspective as names), then you'll need a big lookup table
in the choiceBranchRef expression that translates from the codes to
the QNames. We don't have a case statement in the expression language,
so you've just moved the big linear evaluation chain out of evaluating
choice discriminators one after another and into a big if-then-else
nest in the choiceBranchRef expression. I don't see a performance gain
here.

I suggest dropping the QName stuff, and requiring a dfdl:choiceID
property on the elements that is an NCName.  (Well, we might want
those QName functions anyway in the expression language. But I
wouldn't use them for this rapid choice dispatch feature. You could
certainly use them in discriminators. )

The expression would then have to evaluate to a value that is matched
against this choiceID. I suggest exact match, not respecting
ignoreCase for example.

That eliminates all the QName complexity and is amenable to high-speed
compact lookup table implementation.

I tend to think the element names want to be a little bit more
descriptive than these tag values would want to be so using the
element names as the tags feels undesirable to me.
Particularly because we want the tags to be conveniently computed, for
example by just grabbing a fixed-length string out of a data field.

You end up with something like this:

    <element name="tag" type="string" dfdl:length="{ 2 }" ..../>
    <choice dfdl:choiceBranchRef="{ ../tag }">
        <element name="someName"     dfdl:choiceID="02" .../>
        <element name="anotherName" dfdl:choiceID="73" .../>
        ....
   </choice>

As for the wild-card issue. I think we can finesse this. Consider this 
model:

<element name="tag" type="string" dfdl:length="{ 2 }" ..../>
<choice>
    <!-- fast dispatch for known record types -->
    <choice dfdl:choiceBranchRef="{ ../tag }">
        <element name="someName"     dfdl:choiceID="02" .../>
        <element name="anotherName" dfdl:choiceID="73" .../>
        ....
   </choice>

   <!-- wildcard -->
   <element name="extensionRecord">
      <complexType>
        <sequence>
           <!-- keep tag copy in the extension -->
           <element name="extensionType" type="string"
dfdl:inputValueCalc="../../tag" .../>
           ....
        </sequence>
     </complexType>
  </element>
</choice>

The inner choice uses the fast dispatch. The outer choice lets me also
have an alternative that absorbs a more general syntax to provide some
way for a user to model their extensions to the choice set. The
extensionType field captures the tag and stores it inside the
extension record where it won't get disassociated.

The user's "extensionRecord" would not be a special DFDL wildcard
construct, just an element format they create which is general enough
to parse their extensions. Or this extension could be predefined as
part of a standard, to accept any standard-defined acceptable
extension record a user might need so long as there is some set of
rules all extension records must obey.

Given this, do we really need special wildcard constructs?

...mikeb

From:
Steve Hanson <smh@uk.ibm.com>
To:
dfdl-wg@ogf.org, 
Date:
24.03.2012 02:27
Subject:
[DFDL-WG] Action 145: 'dispatch' way of discriminating a choice for better 
performance

The enveope/payload style of data format is quite common, where the 
envelope provides control information and the payload contains the 
business data. Examples are SWIFT and SAP IDocs. Typically the envelope 
contains a tag that identifies the payload, which can be one of many 
types. For SWIFT there are 300 possible types. To model this today in DFDL 
requires an xs:choice with each type modeled as an xs:element branch of 
the choice. A discriminator on each xs:element refers back to the envelope 
tag element thus enabling the choice to be resolved. 

There are two issues with this approach. 

1) Performance. Even if the elements in the branches are ordered for 
expected frequency, there will still be cases when tens or hundreds of 
discriminators need to be evaluated before the choice is resolved. 

2) Tight coupling. When a new type is added, a new element branch needs to 
be added to the choice. 

Action 145 proposes a mechanism to solve issue #1 and which opens the door 
to a possible extension to DFDL to solve issue #2 - namely a faster way to 
resolve a choice. 

Details: 

A new dfdl:choice property is added called dfdl:choiceBranchRef of type 
DFDL Expression. The expression must evaluate to a QName which corresponds 
to one of the element branches of the choice, and asserts 'known to exist' 
for that branch.  Rules: 

- The property behaves like dfdl:ref and dfdl:hiddenGroupRef in that it is 
not possible to set a value in scope by a dfdl:format annotation, and is 
only set at its point of use. This is because there is nothing sensible 
that could be set in scope. But it has the benefit that adding support for 
the property to existing DFDL implementations will not suddenly cause 
errors to appear in existing DFDL schemas. 

- Empty string is not an allowed value. 

- The property is only used when parsing. 

- All branches must be local elements or element references. It is a 
schema definition error if any branch is a sequence, a choice or a group 
reference.   

- It is a processing error if the QName does not resolve to one of the 
branches when parsing.. 

- It is a schema definition error if a choice has the property set and 
also has dfdl:initiatedContent="yes" set locally. 

- Because the expression must return a QName, the expression language must 
provide a constructor for creating a QName from a namespace string and a 
name string. If you take SWIFT MT103 payload as an example, the tag in the 
envelope says "103" but a DFDL schema would actually model the global 
MT103 element with name "Document" and namespace 
="urn:swift:xsd:fin.103.2011". 
So the dfdl:choiceBranchRef expression would have to look like: 
{fn:QName(fn:concat(fn:concat('urn:swift:xsd:fin.', 
FinMessage/Block2/MessageType), ".2011"), 'Document')} 

So we now have the ability to derive a QName and apply it before we start 
to process a choice. That makes the processing time for each branch of the 
choice independent of its order in the schema. 

We still have issue #2 so when a new payload is added, a new branch must 
be added to the choice. A solution to this is to allows xs:any wildcard 
elements back into DFDL, then provide a property dfdl:wildcardRef which 
works in the same way as dfdl:choiceRef. So at the point of encountering 
the wildcard we know its resolution in the schema.  This obviously will 
require some further discussion, but you can see how this ability to 
evaluate an expression and return a QName can be used in multiple ways. 

Regards

Steve Hanson
Architect, Data Format Description Language (DFDL)
Co-Chair, OGF DFDL Working Group
IBM SWG, Hursley, UK
smh@uk.ibm.com
tel:+44-1962-815848

Unless stated otherwise above:
IBM United Kingdom Limited - Registered in England and Wales with number 
741598. 
Registered office: PO Box 41, North Harbour, Portsmouth, Hampshire PO6 3AU 

--
  dfdl-wg mailing list
  dfdl-wg@ogf.org
  https://www.ogf.org/mailman/listinfo/dfdl-wg

Unless stated otherwise above:
IBM United Kingdom Limited - Registered in England and Wales with number 
741598. 
Registered office: PO Box 41, North Harbour, Portsmouth, Hampshire PO6 3AU

Steve Hanson

Tim Kimber

tags

participants (2)