You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@asterixdb.apache.org by Ian Maxon <im...@apache.org> on 2025/12/02 20:40:24 UTC
Naming and format of vector distance functions
Hi fellow devs,
There is a nice patch up by Calvin
(https://asterix-gerrit.ics.uci.edu/c/asterixdb/+/20126) that adds a
variety of distance functions for vectors. The initial patch added a
function like this:
vector_distance(u,v,metric)
Which would compute the distance between two arrays of the same length
containing numerics, according to a metric, which is given as a
string. The string would be, for example "euclidean", or "manhattan".
I wasn't particularly fond of this syntax- it was inspired from
something else, and not a slight at Calvin's work. After discussing
informally with Calvin and some others familiar with the patch, I
changed the patch to instead add a separate function for each metric,
like:
euclidean_dist(u,v)
manhattan_dist(u,v)
...
and so on. To me it seems this fits better with the naming and syntax
patterns we already have. Code-wise each function continues to share
most of its implementation with the other vector distance calculation
functions.
Does anyone have any other thoughts or suggestions on the matter?
-Ian
Re: Naming and format of vector distance functions
Posted by Ian Maxon <im...@apache.org>.
To serialize some off-list discussions about this, it seemed like it
should be possible to support the 3-ary vector_distance signature in a
better way. The main thing to change is to make it so the last part is
basically an enum that is evaluated at compile-time, sort of similar
to some of our WITH clauses in DDLs. That way there can be no issues
with mistakes like "vector_distance(v1,v2,"manhatan")" causing
unintended behavior. This was one of my main qualms with this syntax
style.
On Tue, Dec 2, 2025 at 1:12 PM Mike Carey <dt...@gmail.com> wrote:
>
> This makes sense to me; it invites fewer errors (e.g., unsupported or
> misspelled metrics). We could always support the string-y syntax with a
> thin logical rewrite if we wanted/need that syntax as well for any reason.
>
> On 12/2/25 12:40 PM, Ian Maxon wrote:
> > Hi fellow devs,
> > There is a nice patch up by Calvin
> > (https://asterix-gerrit.ics.uci.edu/c/asterixdb/+/20126) that adds a
> > variety of distance functions for vectors. The initial patch added a
> > function like this:
> > vector_distance(u,v,metric)
> > Which would compute the distance between two arrays of the same length
> > containing numerics, according to a metric, which is given as a
> > string. The string would be, for example "euclidean", or "manhattan".
> >
> > I wasn't particularly fond of this syntax- it was inspired from
> > something else, and not a slight at Calvin's work. After discussing
> > informally with Calvin and some others familiar with the patch, I
> > changed the patch to instead add a separate function for each metric,
> > like:
> > euclidean_dist(u,v)
> > manhattan_dist(u,v)
> > ...
> > and so on. To me it seems this fits better with the naming and syntax
> > patterns we already have. Code-wise each function continues to share
> > most of its implementation with the other vector distance calculation
> > functions.
> >
> > Does anyone have any other thoughts or suggestions on the matter?
> >
> > -Ian
Re: Naming and format of vector distance functions
Posted by Mike Carey <dt...@gmail.com>.
This makes sense to me; it invites fewer errors (e.g., unsupported or
misspelled metrics). We could always support the string-y syntax with a
thin logical rewrite if we wanted/need that syntax as well for any reason.
On 12/2/25 12:40 PM, Ian Maxon wrote:
> Hi fellow devs,
> There is a nice patch up by Calvin
> (https://asterix-gerrit.ics.uci.edu/c/asterixdb/+/20126) that adds a
> variety of distance functions for vectors. The initial patch added a
> function like this:
> vector_distance(u,v,metric)
> Which would compute the distance between two arrays of the same length
> containing numerics, according to a metric, which is given as a
> string. The string would be, for example "euclidean", or "manhattan".
>
> I wasn't particularly fond of this syntax- it was inspired from
> something else, and not a slight at Calvin's work. After discussing
> informally with Calvin and some others familiar with the patch, I
> changed the patch to instead add a separate function for each metric,
> like:
> euclidean_dist(u,v)
> manhattan_dist(u,v)
> ...
> and so on. To me it seems this fits better with the naming and syntax
> patterns we already have. Code-wise each function continues to share
> most of its implementation with the other vector distance calculation
> functions.
>
> Does anyone have any other thoughts or suggestions on the matter?
>
> -Ian