Not that I'm picking on you specifically Wendell, but your
reply was the most blatantly representative of a class of
responses of a particularly XSL purist/snobbish nature which
I find extremely objectionable.
I get the impression you are using "purist" and "snobbish" as synonyms,
which suggests you dislike purism per se. This is unfortunate. We need
more engineering discipline in software development, not less.
The objection to disable-output-escaping is that it violates an
architectural boundary, namely the boundary between transformation and
serialization. Violating architectural boundaries can often give you a
performance gain, often quite a significant one, but it invariably
incurs a significant cost in terms of the longevity of your software.
We don't know the requirements of your application so we don't know what
the right trade-offs are. The reason that experienced software
architects usually criticize designs that sacrifice software longevity
in favour of speed is that such a tradeoff is usually made without a
proper analysis of the lifetime cost-benefits. This is not snobbishness,
it is wisdom that comes from experience.
As for your performance problem, it appears that Xalan C++ does not
optimize the construct:
select='.|following-sibling::gnsl:Row[$RowsPerPage >
position()]'>
That is, it doesn't recognize that it only needs to look at the next
$RowsPerPage elements. It might be worth trying a different processor.
Alternatively, you could improve performance by using a key, indexing
Rows according to the value of [position() mod $RowsPerPage].
When the need comes to optimize an application, it is always best to do
this without breaking architectural boundaries if you possibly can; and
in this case, you almost certainly can.
Michael Kay
Software AG
home: Michael(_dot_)H(_dot_)Kay(_at_)ntlworld(_dot_)com
work: Michael(_dot_)Kay(_at_)softwareag(_dot_)com
There was a reply from a
Mitch Amiano which actually supplied a suggested approach
which "appeared" entirely reasonable, so tried it out. I've
included the core XSL for both approaches below: the "bad"
code which had the 'disable-output-escaping' clause and the
"good" code which generated the Page element directly.
Following are my performance numbers on a test input file
which had 22,004 Row elements and was 13,425,501 bytes large
(the time output is from the Unix time(1)
command):
For the "good" (XSLT-correct) approach:
real 2:41:32.6
user 2:31:57.0
sys 1.9
For the "bad" (d-o-e) approach:
real 1:38.4
user 1:31.8
sys 1.0
The "good" approach took hours; the "bad" approach took
minutes. For those that will care, the test environment was
a Sun Solaris platform using the interim release of the Xalan
C++ 1.4 XSLT processor.
I'm just curious, do those of you with this hard-line
"purist" attitude actually use XSL to do real work or are you
mostly academics and tool developers/vendors? I understand
staying true to a paradigm up to a point, but sooner or later
"the rubber has to hit the road".
Regards,
Ed Knoll
p.s. This is not all of the XSL, just the differences.
---- "Good" XSL ------------------------------
<xsl:variable name='PageFirstRows'
select='/gnsl:Results/gnsl:Table/gnsl:Row[
(position() mod
$RowsPerPage) = 1]' />
<xsl:template match="gnsl:Table">
<xsl:copy>
<xsl:copy-of select="@*" />
<xsl:apply-templates select="gnsl:Columns" />
<xsl:choose>
<xsl:when test="gnsl:Row">
<xsl:apply-templates select='$PageFirstRows' />
</xsl:when>
<xsl:otherwise>
<xsl:element name='Page' />
</xsl:otherwise>
</xsl:choose>
</xsl:copy>
</xsl:template>
<xsl:template match="gnsl:Row">
<xsl:element name="Page">
<xsl:for-each
select='.|following-sibling::gnsl:Row[$RowsPerPage
position()]'>
<xsl:call-template name='CopyAll' />
</xsl:for-each>
</xsl:element>
---- "Bad" XSL ------------------------------
<xsl:template match="gnsl:Table">
<xsl:copy>
<xsl:copy-of select="@*" />
<xsl:apply-templates select="gnsl:Columns" />
<xsl:choose>
<xsl:when test="gnsl:Row">
<xsl:apply-templates select="gnsl:Row" />
</xsl:when>
<xsl:otherwise>
<xsl:value-of select="$LineBreak" />
<xsl:text
disable-output-escaping="yes"><Page/></xsl:text>
<xsl:value-of select="$LineBreak" />
</xsl:otherwise>
</xsl:choose>
</xsl:copy>
</xsl:template>
<xsl:template match="gnsl:Row">
<xsl:if test="(position() mod $RowsPerPage) = 1">
<xsl:if test="position() != 1">
<xsl:value-of select="$LineBreak" />
<xsl:text
disable-output-escaping="yes"></Page></xsl:text>
<xsl:value-of select="$LineBreak" />
</xsl:if>
<xsl:value-of select="$LineBreak" />
<xsl:text disable-output-escaping="yes"><Page></xsl:text>
<xsl:value-of select="$LineBreak" />
</xsl:if>
<xsl:call-template name="CopyAll" />
<xsl:if test="position() = last()">
<xsl:value-of select="$LineBreak" />
<xsl:text disable-output-escaping="yes"></Page></xsl:text>
<xsl:value-of select="$LineBreak" />
</xsl:if>
</xsl:template>
Hey Mitch,
The horribleness of disable-output-escaping is not (to my
mind) really
an
issue of the well-formedness constraint either in the
stylesheet or in the
output -- that's something of a red herring (though it is a
risk and a sign
of the deeper problem). Rather, it's the violation of
XSLT's processing
model, in which the transformation of the node tree and the
post-transformation serialization are clearly distinguished
and kept
separate by design. *Any* solution that works by writing
markup to output
using d-o-e creates a dependency on the serialization step.
While this may
be acceptable in certain circumstances (e.g. writing SGML
entity references
to output that are not otherwise provided for, when you
*know* you're going
to write a file), it's horrible at other times, if only because the
designer has created this dependency unwittingly, and thus doesn't
understand why the transform breaks in a conformant
architecture, like
Mozilla or transformation chains in Cocoon, where no file
is getting
serialized.
The relevance of grouping is only that the "write markup"
approach is
usually resorted to by newer XSLT programmers who don't
know how else to do
grouping, and who fall back on their Perl or Javascript
experience (or just
sheer ingenuity) to suppose that writing markup is the best or only
solution to the problem (it is neither).
I doubt that any experienced XSLTer would have a problem
with either
of the
solutions you offered (or Dimitre's, or Tom's), since none of them
introduce the dependency on serialization that is the problem with
d-o-e-based techniques for "outputting open/close tags
independently".
There the distinctions are much more of coding style and
performance; but
none of them use a technique that is prone to break the
minute you move
your stylesheet into a different environment.
Cheers,
Wendell
--
Edward L. Knoll Phone (work) : (719)484-2717
e-mail (work) :
ed(_dot_)knoll(_at_)cosd(_dot_)fedex(_dot_)com
e-mail (business): eknoll(_at_)sf-inc(_dot_)com
e-mail (personal): edward(_at_)elknoll(_dot_)com
XSL-List info and archive: http://www.mulberrytech.com/xsl/xsl-list
XSL-List info and archive: http://www.mulberrytech.com/xsl/xsl-list