Re: log bind parameter values on error

On 2019-Dec-07, Tom Lane wrote:

> 0002:

Here's a version of this part with fixes for these comments.  It applies
on top of the stringinfo_mb.c patch sent elsewhere in the thread.

(If we were to add a "log_parameters_on_error_max_length" GUC to decide
the length to log, we would get rid of the remaining magical numbers in
this code).


> Seems like BuildParamLogString's "valueLen" parameter ought to be called
> "maxlen", for consistency with 0001 and because "valueLen" is at best
> misleading about what the parameter means.
> 
> I'd toss the enlargeStringInfo call here too, as it seems overly
> complicated and underly correct or useful.
> 
> Probably doing MemoryContextReset after each parameter (even the last one!)
> is a net loss compared to just letting it go till the end.  Although
> I'd be inclined to use ALLOCSET_DEFAULT_SIZES not SMALL_SIZES if you
> do it like that.
> 
> Please do not throw away the existing comment "/* Free result of encoding
> conversion, if any */" in exec_bind_message.
> 
> As above, this risks generating partial multibyte characters.  You might
> be able to get away with letting appendStringInfoStringQuoted do the
> multibyte-aware truncation, but you probably have to copy more than just
> one more extra byte first.
> 
> I have zero faith in this:
> 
> +    params_errcxt.arg = (void *) &((ParamsErrorCbData)
> +                                   { portal->name, params });
> 
> How do you know where the compiler is putting that value, ie what
> its lifespan is going to be?  Better to use an explicit variable.
> 
> I concur with dropping testlibpq5.
> 
>             regards, tom lane


-- 
Álvaro Herrera                https://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services