On Tue, Jun 17, 2014 at 12:51 AM, Peter Geoghegan <p...@heroku.com> wrote:
> On Mon, Jun 16, 2014 at 8:56 PM, Tom Lane <t...@sss.pgh.pa.us> wrote:
>> Not having looked at the patch, but: I think the probability of
>> useless-noise HINTs could be substantially reduced if the code prints a
>> HINT only when there is a single available alternative that is clearly
>> better than the others in Levenshtein distance.  I'm not sure how much
>> better is "clearly better", but I exclude "zero" from that.  I see that
>> the original description of the patch says that it will arbitrarily
>> choose one alternative when there are several with equal Levenshtein
>> distance, and I'd say that's a bad idea.
> I disagree. I happen to think that making some guess is better than no
> guess at all here, given the fact that there aren't too many
> possibilities to choose from. I think that it might be particularly
> annoying to not show some suggestion in the event of a would-be
> ambiguous column reference where the column name is itself wrong,
> since both mistakes are common. For example, "order_id" was specified
> instead of one of either "o.orderid" or "ol.orderid", as in my
> original examples. If some correct alias was specified, that would
> make the new code prefer the appropriate Var, but it might not be, and
> that should be okay in my view.
> I'm not trying to remove the need for human judgement here. We've all
> heard stories about people who did things like input "Portland" into
> their GPS only to end up in Maine rather than Oregon, but I think in
> general you can only go so far in worrying about those cases.

Emitting a suggestion with a large distance seems like it could be
rather irritating.  If the user types in SELECT prodct_id FROM orders,
and that column does not exist, suggesting "product_id", if such a
column exists, will likely be well-received.  Suggesting a column
named, say, "price", however, will likely make at least some users say
"no I didn't mean that you stupid @%!#" - because probably the issue
there is that the user selected from the completely wrong table,
rather than getting 6 of the 9 characters they typed incorrect.

One existing tool that does something along these lines is 'git',
which seems to have some kind of a heuristic to know when to give up:

[rhaas pgsql]$ git gorp
git: 'gorp' is not a git command. See 'git --help'.

Did you mean this?
[rhaas pgsql]$ git goop
git: 'goop' is not a git command. See 'git --help'.

Did you mean this?
[rhaas pgsql]$ git good
git: 'good' is not a git command. See 'git --help'.
[rhaas pgsql]$ git puma
git: 'puma' is not a git command. See 'git --help'.

Did you mean one of these?

I suspect that the maximum useful distance is a function of the string
length.  Certainly, if the distance is greater than or equal to the
length of one of the strings involved, it's just a totally unrelated
string and thus not worth suggesting.  A useful heuristic might be
something like "distance at most 3, or at most half the string length,
whichever is less".

Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company

Sent via pgsql-hackers mailing list (pgsql-hackers@postgresql.org)
To make changes to your subscription:

Reply via email to