I'm sorry to bother you.
I used this extension for 400 million img, and index on signature with gist. But I found that query on signature is very low (cost more than 30 seconds for query like: select sig <-> '(0.631970, 1.085420, 0.449103, 0.885163, 0.797899, 0.571842, 0.879878, 1.050410, 0.438902, 0.540778, 0.641945, 0.395742, 0.133787, 1.013711, 0.191292, 0.326252)' as dist from img_table order by dist limit 1;. By explain, I see it's using Index Scan).
I think the root cause is that: after we "Normalize" intensiveness in the pattern, the signatures becomes too concentrated in Gist index. It hits too many blocks in memory to get the result.
Is it possible to skip the normalization step when producing Pattern & Signature? Or other solution to solve the performance issue? Thanks!
I'm sorry to bother you. I used this extension for 400 million img, and index on signature with gist. But I found that query on signature is very low (cost more than 30 seconds for query like:
select sig <-> '(0.631970, 1.085420, 0.449103, 0.885163, 0.797899, 0.571842, 0.879878, 1.050410, 0.438902, 0.540778, 0.641945, 0.395742, 0.133787, 1.013711, 0.191292, 0.326252)' as dist from img_table order by dist limit 1;. By explain, I see it's using Index Scan).I think the root cause is that: after we "Normalize" intensiveness in the pattern, the signatures becomes too concentrated in Gist index. It hits too many blocks in memory to get the result.
Is it possible to skip the normalization step when producing Pattern & Signature? Or other solution to solve the performance issue? Thanks!