Apache APISIX vs KONG

Apache apisix benchmark script: https://github.com/iresty/apisix/blob/master/benchmark/run.sh

Kong beanchmark script: 

curl -i -X POST \
  --url http://localhost:8001/services/ \
  --data 'name=example-service' \
  --data 'host=127.0.0.1'

curl -i -X POST \
  --url http://localhost:8001/services/example-service/routes \
  --data 'paths[]=/hello'

curl -i -X POST http://localhost:8001/routes/efd9d857-39bf-4154-85ec-edb7c1f53856/plugins \
  --data "name=rate-limiting" \
  --data "config.hour=999999999999" \
  --data "config.policy=local"

curl -i -X POST http://localhost:8001/routes/efd9d857-39bf-4154-85ec-edb7c1f53856/plugins \
  --data "name=prometheus"

curl -i http://127.0.0.1:8000/hello/hello

wrk -d 5 -c 16 http://127.0.0.1:8000/hello/hello

@cboitel have you used any of these after your benchmarks? what is your recommendation for someone looking to adopt any of these?
I am interested in a simple nginx use case + auth for low latency and throughput up to 3000qps in the long run.

@Dhruv-Garg79 Hi, may I ask which API gateway was ultimately chosen?

you can choose the one you like, both of them are open source, you can know everything if you want to research them.

For me: Apache APISIX is better performance and rich ecosystem, it belongs to the Apache Software Foundation, more open and safe

Configuration	Workers	Plugins	Requests/sec	Transfer/sec	Avg Latency	Max Latency	Notes
APISIX - 1 Worker, No Plugins	1	None	~16,407.59	~65.53 MB/s	~0.96 ms	~5.47 ms	Baseline performance
APISIX - 1 Worker, With Plugins	1	limit-count, prometheus	~16,129.60	~65.53 MB/s	~0.99 ms	~12.64 ms	Minor drop with plugins
APISIX - 2 Workers, No Plugins	2	None	~35,386.00	~141.33 MB/s	~0.45 ms	~9.50 ms	Strong scaling with more workers
APISIX - 2 Workers, With Plugins	2	limit-count, prometheus	~30,633.00	~124.46 MB/s	~0.52 ms	~10.34 ms	Still performs well with plugins
Fake APISIX Server - 1 Worker	1	N/A	~22,504.00	~89.55 MB/s	~0.71 ms	~4.61 ms	Bare OpenResty baseline
Fake APISIX Server - 2 Workers	2	N/A	~46,045.00	~183.25 MB/s	~0.35 ms	~10.34 ms	Max bare-metal performance
Kong - 1 Worker, No Plugins	1	None	~13,728.00	~55.76 MB/s	~1.19 ms	~6.31 ms	Kong baseline
Kong - 2 Workers, No Plugins	2	None	~26,903.00	~109.27 MB/s	~0.62 ms	~7.76 ms	Good multi-worker scaling
Kong - 1 Worker, With Plugins	1	rate-limiting, prometheus	~2,027.00	~8.66 MB/s	~7.85 ms	~21.36 ms	Significant overhead with plugins
Kong - 2 Workers, With Plugins	2	rate-limiting, prometheus	~3,768.00	~15.91 MB/s	~4.27 ms	~18.91 ms	Better than 1 worker but still slow

result table generated by AI for https://gist.github.com/membphis/137db97a4bf64d3653aa42f3e016bd01?permalink_comment_id=3351676#gistcomment-3351676

@knysfh After doing a lot of tests, I decided to have everything directly on the server side without a reverse proxy and use native K8S capabilities + changes on the main server side. Resource utilisation and performance both improved in my testing.

Also the reason was that I don't want to many components, it makes everything harder to maintain.