I want to create a database using any of the possible RDBMS. It will have a table with approximately 150 columns. The objective is to perform nearest neighbor search of some other objects. So it's a NNS in the 150-dimensional space.

I already tried to use some obvious methods like L1 or L2 distances but of course it takes a lot of time for tables with many rows. Also I tried to look at the KD-tree (note I did not test it) and PG-Strom but they are not a good solution for data with a many dimensions.

Can I somehow improve the speed of the described search using math methods (like KD-tree) or tech methods (like PG-Strom)?

I will try to use any RDBMS which allow to improve speed of the NNS. But MySQL and PostgreSQL are the most appropriate DBMS for me.

  • 1
    These are other problems. Simply ask another question @don-prog – Evan Carroll Aug 10 '17 at 16:02

PostgreSQL 9.6 using cube

First install the cube extension


Now we will create some n-dimensional space with 100,000 points in 50 dimensions. In addition we'll add a GIST index.

  SELECT i, cube(array_agg(random()::float)) AS c
  FROM generate_series(1,1e5) AS i
  CROSS JOIN LATERAL generate_series(1,50)
    AS x

CREATE INDEX ON space_nd USING gist ( c );
ANALYZE space_nd;

Now we will generate a single point and use the <-> operater to find the nearest point using Eucledian distance.

WITH points AS (
  SELECT cube(array_agg(random()::float)) AS c
  FROM generate_series(1,50)
    AS x
  cube_distance(space_nd.c, points.c)
FROM space_nd
ORDER BY space_nd.c <-> points.c

PostgreSQL 9.6+ supports other distance operators over cube. All of which can use the GIST index we created. Namely,

a <-> b float8  Euclidean distance between a and b.
a <#> b float8  Taxicab (L-1 metric) distance between a and b.
a <=> b float8  Chebyshev (L-inf metric) distance between a and b.

That said there is one caveat,

To make it harder for people to break things, there is a limit of 100 on the number of dimensions of cubes. This is set in cubedata.h if you need something bigger.

You ask for 150 dimensions. That may present a minor complication.

| improve this answer | |
  • 1
    The edit to cubedata.h doesn't work past 130 dimensions in my experience. Maybe you can also change all the doubles or float8s in the extension to float4, since Postgres has a limit on per-row index size that you can stay away from by halving how many bytes you use on each number. I did some testing and got more dimensions in that way, and IIRC I got past 150, but I'm not totally sure. – sudo Nov 3 '18 at 2:33
  • 1
    I had same problem with limit on dimensions and created docker image with 2048 limit: hub.docker.com/r/expert/postgresql-large-cube – expert Apr 29 '19 at 9:37

Consider performing dimension reduction first (eg. Principle Component Analysis).

Then your are doing NN in small number of dimensions with higher performance.

You can use Pl/R to perform PCA inside postgres if needed.

| improve this answer | |

Have a look at FLANN and OpenCV.

Unfortunately I am not aware of an integration of that into a RDBMS system. But there is for example integration of chemical structure information with Posgres. So in principle this can be done.

| improve this answer | |

Take a look at https://github.com/a-mma/AquilaDB it is a vector database to store Feature Vectors along with JSON Metadata. Keep it along with your RDBMS and do use metadata to maintain cross reference between data.

| improve this answer | |

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.