I can't understand why the /plaintext test is almost 5 times faster than the /json test.

[remittor]

Here are the results of the tests on my stand:

$ ./libreactor-server

$ wrk -t1 -c100 -d15 -H 'Connection: keep-alive' http://127.0.0.1:8080/plaintext
Running 15s test @ http://127.0.0.1:8080/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   464.05us   64.37us   3.03ms   83.36%
    Req/Sec   109.56k     3.04k  116.05k    72.67%
  1634595 requests in 15.00s, 210.45MB read
Requests/sec: 108952.50
Transfer/sec:     14.03MB

$ wrk -t1 -c100 -d15 -H 'Connection: keep-alive' http://127.0.0.1:8080/json
Running 15s test @ http://127.0.0.1:8080/json
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   488.96us   97.31us   1.84ms   93.74%
    Req/Sec   105.41k     4.17k  111.84k    74.67%
  1573334 requests in 15.00s, 232.57MB read
Requests/sec: 104875.40
Transfer/sec:     15.50MB

The executable file libreactor-server was obtained by me clearly following this instruction: https://github.com/TechEmpower/FrameworkBenchmarks/blob/master/frameworks/C/libreactor/libreactor-server.dockerfile

But I had to fix 1 line in the code to force the libreactor-server to use only 1 thread.

---

And here are the results for libreactor-server contained in database TFB:
plaintext: 7024777 req/sec (proof-link)
json: 1306654 req/sec (proof-link)

---

for 56 threads: 7024777 / 1306654 = 5.37
for  1 thread :  108952 /  104875 = 1.03

Explain to me where such a difference in the ratio of speeds comes from.

[fakeshadow]

plaintext bench uses http/1.1 pipeline feature. You can checkout the script tfb uses here

[remittor]

@fakeshadow , thanks
Now it's clearer:

$ wrk -t1 -c100 -d15 -H 'Connection: keep-alive' http://127.0.0.1:8080/plaintext -s pipeline.lua -- 16
Running 15s test @ http://127.0.0.1:8080/plaintext
  1 threads and 100 connections
  Thread Stats   Avg      Stdev     Max   +/- Stdev
    Latency   379.06us  269.40us   3.01ms   18.70%
    Req/Sec     0.93M    36.40k    0.99M    84.00%
  13863936 requests in 15.00s, 1.74GB read
Requests/sec: 924084.24
Transfer/sec:    118.97MB

924084 / 104875 = 8.81

[remittor]

Now I wanted to check the correct implementation of the pipeline on this server.
To do this, I wrote a python script that sends inseparably 10M identical requests /plaintext.
The result is deplorable. The libreactor-server simply loaded everything into RAM:

The script has not been working for 10 minutes, and the libreactor-server does not release memory.

It seems to me that this is a common situation. And it looks like you need to create a separate topic about this.