Re: question on parsing postgres sql queries

On Tue, Jul 26, 2016 at 4:20 AM, Jerome Wagner <jerome.wagner@m4x.org> wrote:

> I am doing some research on postgres sql query parsing.

> I was wondering what people think of the conformance with regards to the
> real parser of the documentations on
>  - https://www.postgresql.org/docs/current/static/sql-select.html
>  - https://www.postgresql.org/docs/current/static/sql-copy.html
> ... and more generally sgmls in
> https://github.com/postgres/postgres/tree/master/doc/src/sgml/ref
>
> Would it make sense to use these sgml synopsis as some kind of source of
> truth, parse them, and automatically generate a parser for a specifc
> language ?

It might be interesting to do as an academic exercise or to audit
the accuracy of the synopses, but I don't think it's practical for
generating production-quality parsers -- at least in the short
term.  Besides issues mentioned by others (e.g., parser support for
legacy syntax we don't want to document or encourage), we sometimes
allow things through the parser so that we can code more
user-friendly messages off of the parse tree than a generated
parser would provide.

I also don't remember seeing anyone mention the problems with
forward references and metadata from system catalogs.  These either
need to be handled by a "rewind and try again" approach or (better
IMO) an additional pass or two walking the parse tree to emit a
version where generic "place-holders" are replaced by something
more specific.  See the "parse analysis" and "rewrite" steps in
PostgreSQL for how that is currently handled.  Before working in
the PostgreSQL source I had helped develop a SQL parser in ANTLR,
where the same basic parser generator is used for lexer, parser,
and tree-walker phases (using pretty much the same grammar
specifier for all of them), just taking characters, tokens, or
parse tree nodes as input -- automatic generation of "main" parser
might be feasible in such an environment (possibly with some sort
of annotations or hand-written light initial parsing phase), but I
think the later tree walkers would need to be hand-coded.

> I feel like the conformance level of the documentation is high and that the
> sgml synopis seem to be nearly programmatically sufficient to create
> parsers.
>
> what do you think ?

Nearly.

> Could the parser commiters share some lights on how the documentation
> process interacts with the parser commits ?

There is no automated interaction there -- it depends on human
attention.  On the other hand, try connecting to a database with
psql and typing:

\h create index

... (or any other command name).  The help you get there is fished
out of the docs.

--
Kevin Grittner
EDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company