Request with Transfer-Encoding chunked instead of Content-Length

[olivier4576]

Which version of Elastic are you using?

[X] elastic.v7 (for Elasticsearch 7.x)
[ ] elastic.v6 (for Elasticsearch 6.x)
[ ] elastic.v5 (for Elasticsearch 5.x)
[ ] elastic.v3 (for Elasticsearch 2.x)
[ ] elastic.v2 (for Elasticsearch 1.x)

Please describe the expected behavior

ELK search sends an HTTP request with a "Content-Length: nnn" header.

Please describe the actual behavior

ELK search sends an HTTP request with a "Transfer-Encoding: chunked" header.
Transfer-Encoding chunked can be used in a HTTP request but is rarely used and this can cause problems to other tools.

Any steps to reproduce the behavior?

Reproduced in any search request.

The problem is in request.go, in func setBodyHeader(). When called in setBodyJson() or setBodyGzip(), the body is a bytes.Reader, then it must be converted to a bytes.Reader and not a bytes.Buffer to get the ContentLength.

[olivere]

I've did a few tests, especially with the PR #1441, and the benchmarks are alarming:

$ go test -test.bench=RequestSetBody -run=NONE > old.txt
...
$ go test -test.bench=RequestSetBody -run=NONE > new.txt
...
$ benchcmp old.txt new.txt
benchcmp is deprecated in favor of benchstat: https://pkg.go.dev/golang.org/x/perf/cmd/benchstat
benchmark                                              old ns/op      new ns/op      delta
BenchmarkSearchMatchAll-8                              1979627        1041992666     +52535.81%

benchmark                                              old allocs     new allocs     delta
BenchmarkSearchMatchAll-8                              173            4881           +2721.39%

benchmark                                              old bytes      new bytes      delta
BenchmarkSearchMatchAll-8                              96174          1823568        +1796.11%

We have to be very cautious. I'll have another go at this in a later release.