Re: Proposal: pg_confcheck - syntactic & semantic validation of postgresql configuration files

On 10/13/2015 09:20 PM, Robert Haas wrote:
> On Mon, Oct 12, 2015 at 8:01 PM, Amir Rohan <amir(dot)rohan(at)zoho(dot)com> wrote:
>> That wasn't my intention. Perhaps I'm overreacting to a long-standing
>> "Tom Lane's bucket of cold water" tradition. I'm new here.
>> I understand your point and I was only reiterating what in particular
>> makes the conf checker distinctly useful IMO, and what it could
>> provide that pg_settings doesn't.
>>
>> I've looked at parts of the pg_settings implementation and indeed some
>> of that code (and legwork) could be reused so the mundane parts
>> of writing this will be less hassle. I might have missed that if Tom and
>> you hadn't pointed that out.
>>
>> So, Fair, and thanks.
>
> No worries. I'm not upset with you, and I see where you're coming
> from. But I since I'm trying to be helpful I thought I would check
> whether it's working. Sounds like yes, which is nice.
>
> It would be spiffy if we could use the same config-file parser from
> outside postgres itself, but it seems hard. I assume the core lexer
> and parser could be adapted to work from libcommon with some
> non-enormous amount of effort, but check-functions are can and do
> assume that they are running in a backend environment; one would lose
> a lot of sanity-checking if those couldn't be executed,

I've been considering that. Reusing the parser would ensure no errors
are introduces by having a different implementation, but on the other
hand involving the pg build in installation what's intended as a
lightweight, independent tool would hurt.
Because it's dubious whether this will end up in core, I'd like
"pip install pg_confcheck" to be all that is required.

Also, anything short of building the tool in tree with an unmodified
parser amounts to forking the parser code and maintaining it along with
upstream, per version, etc'. I'm not eager to do that.

> and checking GUCs created by loadable modules seems impossible. Still, even a
> partial move toward making this code accessible in frontend code would
> be welcome from where I sit.
>

Dumping the output of the pg_settings view into json has already
provided all the type/enum information needed to type-check values and
flag unrecognized GUCs. I don't really see that as involving a heroic
amount of work, the language seems extremely simple.
There aren't that many types, type-checking them isn't that much work,
and once that's written the same checks should apply to all GUCs
registered with the server, assuming the pg_settings view is
exhaustive in that sense.

Any modules outside the main server can provide their own data
in a similar format, if it comes to that. I doubt whether such
a tool has that much appeal, especially if it isn't bundled with
the server.

So, I'll think about this some more, and write up a description of how I
think it's going to look for some feedback. Then I'll code something.

Regards,
Amir