I found the low proformance of library json.
There are about 1441440 data in my single series.
When I try to query select host,last(value) from node_cpu group by host, I found it is ultra slow to parse the json body in the response. It takes me about 30s to solve the query.
When I modified the source code, using library ujson instead of json， I found it is ultra fast to parse the http response.

Here is the patchfile

use_ujson_instead_of_json.txt