Replies: 2 comments 2 replies
-
@kobeyk can you try setting ref: https://developmentseed.org/titiler/advanced/performance_tuning/ |
Beta Was this translation helpful? Give feedback.
2 replies
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Excuse me, I have a problem,I converted a cog tif with rio-cogeo and put it locally, then read it with rio-tiler, and ran it with py code alone, and found that the reading process took a long time, >600ms, in fact, normal tif is very fast when read with gdal, I would like to know why this piece is slow. Here's how long it takes to run alone:
code:
In addition, I also used java to write a server to call python code, and wrote a front-end test. I found that cog as a image tile service, its performance is really a little too slow, I want to know, as cog data, whether it supports multi-user concurrent access? Is it possible to improve access efficiency by setting some parameters for tuning? In other words, in order to provide high-quality image tile services, it can only be carried out in accordance with the traditional slicing method. Below is a screenshot of the response time of the front-end call:
Beta Was this translation helpful? Give feedback.
All reactions