C[omp]ute

From: "andrew cooke" <andrew@...>

Date: Wed, 28 Jun 2006 10:05:29 -0400 (CLT)

Some simple observations / notes / reflections.

Implementing a system to load a set of database tables with data, I used a
DSL to specify the relationship between the tables (key relations) and the
data source.

I separated the language into two stages.  In the first stage you could
only define types (table structures).  The second stage extended the
first, allowing the full set of features.

That seemed a bit odd, and the implementation was rather ugly, but still
turned out to be a good idea.  The goodneess came from the possibility of
generating the first stage from the database (via postgres's "meta"
tables).

Having this information as a simple AST allowed us to combine (incomplete)
database meta-information with extra user-added text (by merging ASTs). 
That may not be clear - I am talking about the first stage only.  We could
both generate the description from the database and via text files.  The
two were complementary.

Extra checks in the second stage guaranteed that all referenced tables
were fully defined (ie non "not null" values were missing).  This check
cannot be made on the first stage because that includes all tables (even
those nto populated by this particular invocation).  In otehr words, the
first stage defines the database; the second stage selects some tables
from that for populating.

However, there were also some bad points.  First, what is the relationship
(implementation-wise) between the two stages?  I had them sharing common
base classes and copying data across (so when a table is references in the
second stage for the first time, all data from the first stage is copied
across).

It would have been better for the second stage to include a reference to
the first stage ("has a", rather than a bastardised "is a").

Another bad point was the syntax.  I argued that I didn't have time to
write a parser, and came up with what seemed like an elegant hack -
http://www.acooke.org/cute/UsefulJava0.html - which gave me the parse tree
directly (note that the language is declarative, so order of statements is
unimportant).

However, this has confused users, since the lhs of the syntax is a set of
arbitrary names to label branches in the AST and therefore has no direct
connection between the two stages.

At the same time, in another part of the system, I wrote (in less time
than the hack above, I suspect) a very simple "lazy scheme" (without,
currently, function defn / lambda abstraction, continuations, macros, or
most standard functions) (laziness allows "everything" to be a function). 
We could have used that here.

One other regret - I should have been clearer about the DSL approach. 
I've written enough language-related code to have absorbed many of the
patterns, which led to using them without a clear high-level description
(I had a design/sketch for how things would work, but not a simple,
unifying "vision").  Just thinking of the system "as a language" would
have helped clarify things, I believe.  Certainly it would have made it
easier to explain to others later (the current system, with no real
language to speak of - just this "parse trees as properties" - looks like
a monolith with arbitrarily complex configuration values, rather than an
interpreter).

Andrew