Re: [O] problems with export and :cache

[Aaron Ecay]

Hi Nicolas,

2015ko azaroak 1an, Nicolas Goaziou-ek idatzi zuen:
> 
> Hello,
> 
> Aaron Ecay <address@hidden> writes:
> 
>> If this means “can it ever work?” then I think the answer is “yes it
>> can”.  But I think the current implementation is broken and likely to
>> remain so for the foreseeable future.  The issues are:
>> 
>> 1. :cache only works for code which is a pure function of its header args
>> 2. When combined with :session, the environment that the code is evaluated
>> in is not created anew each time it is run.  This makes it much easier
>> to leak references to (e.g.) variables defined in other blocks
>> 3. The proper notion of purity is not easily defined when the code does
>> things like modifying the emacs environment, touching the filesystem,
>> or accessing the language’s RNG.
>> 4. We (org devs) don’t actually understand how the semantics of cache
>> interacts with other babel features.  See:
>> <http://mid.gmane.org/address@hidden>.
>> 
>> 1-3 are likely to be extremely confusing for users, especially less
>> technically sophisticated ones (what’s a “pure function” anyway)?  The
>> inability to give a clear introductory explanation of the feature in
>> combination with 4 indicating we don’t actually understand it ourselves
>> makes me feel like we should not be advertising, let alone recommending,
>> it.
>> 
>> The only other literate programming environment that I know of that
>> implements such a feature is knitr (for R).  They address these issues
>> by providing (optional) free-variable analysis to construct a dependency
>> graph between code blocks.  There is also some handling of RNG seed
>> values.  The documentation <http://yihui.name/knitr/demo/cache/> is much
>> more comprehensive, including a prominent statement about the dangers of
>> side effect-ful code and a nuanced discussion of several issues,
>> including the RNG.
> 
> Thank you for the explanations.
> 
> Assuming the user knows what s?he is doing (so I'm not bothered by
> issues 2 and 3), :cache is still a somewhat useful feature. I don't mind
> advertising it. Though, I agree we could include big fat warnings about
> it in the manual.

See the attached patch.  I’ve tried to put all my experience and best
practices in there; comments are welcome of course.

> 
> Also, :cache might be a bit misleading as it implies more than what this
> feature offer, i.e., a dumb "don't update results if contents didn't
> change". I cannot think of a better name, tho.

There are backwards compatibility implications to renaming the header,
of course.  (And I can’t think of a better name either).

Thanks,

-- 
Aaron Ecay

>From bb0f43948384448225323abcfe7a662d110d1389 Mon Sep 17 00:00:00 2001
From: Aaron Ecay <address@hidden>
Date: Wed, 4 Nov 2015 11:57:49 +0000
Subject: [PATCH] babel: update the manual wrt :cache header arg

* doc/org.texi (cache): Update manual section.
---
 doc/org.texi | 36 +++++++++++++++++++++++++++++++-----
 1 file changed, 31 insertions(+), 5 deletions(-)

diff --git a/doc/org.texi b/doc/org.texi
index ba402bf..48ae017 100644
--- a/doc/org.texi
+++ b/doc/org.texi
@@ -16166,11 +16166,37 @@ used.
 
 The @code{:cache} header argument controls the use of in-buffer caching of
 the results of evaluating code blocks.  It can be used to avoid re-evaluating
-unchanged code blocks.  Note that the @code{:cache} header argument will not
-attempt to cache results when the @code{:session} header argument is used,
-because the results of the code block execution may be stored in the session
-outside of the Org mode buffer.  The @code{:cache} header argument can have
-one of two values: @code{yes} or @code{no}.
+unchanged code blocks.  When the cache is active, a source block is not
+re-evaluated if a result for it is present in the buffer and neither the
+header arguments (including the value of @code{:var} references) nor the text
+of the block itself has changed since the result was computed.  The feature
+helps avoid re-running long calculations.  However, there are edge cases and
+you should not rely on the cache to behave reliably in all circumstances.
+
+The caching feature works best when a babel block is a pure function of its
+arguments (see @ref{var}).  That is, the function always returns the same
+results when given the same arguments, and does not touch external resources
+(like the filesystem or the language’s RNG) in any way.
+
+The documentation of the knitr reproducible research package for the R
+language has some good discussion of issues that may arise when using the
+cache in such a context.  See @uref{http://yihui.name/knitr/demo/cache/},
+especially the sections “Even more stuff for cache?” and “Reproducibility
+with RNG”.  (Obviously, you will have to abstract away from the knitr
+implementation details which the documentation also discusses.)
+
+Note that the @code{:cache} header argument will attempt to cache results
+when the @code{:session} header argument is used, even though the results of
+the code block execution stored in the session may lead to unexpected
+results.
+
+Noweb references (see @ref{Noweb reference syntax}) are currently not
+expanded when calculating whether the text of the code block has changed.
+Perhaps in principle they ought to be, but this could introduce unexpected
+complexity.  See @uref{http://thread.gmane.org/gmane.emacs.orgmode/79046}.
+
+The @code{:cache} header argument can have one of two values: @code{yes} or
address@hidden
 
 @itemize @bullet
 @item @code{no}
-- 
2.6.2