Why maxRequestPerConnection of istio does effect to http/1.1 requests?

I'm just learning service mesh using istio and I found a strange behavior. To understand maxRequestsPerConnection of Istio DestinationRule CRD, I write the below manifest and apply it.

apiVersion: networking.istio.io/v1alpha3
kind: DestinationRule
metadata:
  name: httpbin
spec:
  host: httpbin
  trafficPolicy:
    connectionPool:
      tcp:
        maxConnections: 1
      http:
        http1MaxPendingRequests: 1
        maxRequestsPerConnection: 1

And then, I sent requests using fortio. The result is below:

yunoMacBook-Air:labo8 yu$ kubectl exec "$FORTIO_POD" -c fortio -- /usr/bin/fortio load -c 5 -qps 0 -n 1000 -loglevel Error http://httpbin:8000/get
07:12:01 I logger.go:127> Log level is now 4 Error (was 2 Info)
Fortio 1.11.3 running at 0 queries per second, 2->2 procs, for 1000 calls: http://httpbin:8000/get
Aggregated Function Time : count 1000 avg 0.0036879818 +/- 0.004588 min 0.000379697 max 0.034176044 sum 3.68798183
# target 50% 0.00234783
# target 75% 0.0032551
# target 90% 0.008
# target 99% 0.025
# target 99.9% 0.032784
Sockets used: 876 (for perfect keepalive, would be 5)
Jitter: false
Code 200 : 126 (12.6 %)
Code 503 : 874 (87.4 %)
All done 1000 calls (plus 0 warmup) 3.688 ms avg, 1170.1 qps
yunoMacBook-Air:labo8 yu$

After that, I changed maxRequestsPerConnection value to 10:

apiVersion: networking.istio.io/v1alpha3
kind: DestinationRule
metadata:
  name: httpbin
spec:
  (...)
        maxRequestsPerConnection: 10

and I sent requests again with the same settings.

yunoMacBook-Air:labo8 yu$ kubectl exec "$FORTIO_POD" -c fortio -- /usr/bin/fortio load -c 5 -qps 0 -n 1000 -loglevel Error http://httpbin:8000/get
07:11:07 I logger.go:127> Log level is now 4 Error (was 2 Info)
Fortio 1.11.3 running at 0 queries per second, 2->2 procs, for 1000 calls: http://httpbin:8000/get
Aggregated Function Time : count 1000 avg 0.0039736575 +/- 0.004068 min 0.000404827 max 0.030141552 sum 3.97365754
# target 50% 0.00231923
# target 75% 0.00475
# target 90% 0.0104667
# target 99% 0.0192
# target 99.9% 0.025
Sockets used: 723 (for perfect keepalive, would be 5)
Jitter: false
Code 200 : 281 (28.1 %)
Code 503 : 719 (71.9 %)
All done 1000 calls (plus 0 warmup) 3.974 ms avg, 1098.3 qps
yunoMacBook-Air:labo8 yu$

200 rate increased and I cannot understand why it happened. In my understanding, fortio uses http/1.1 and only one HTTP request is in one TCP connection when I use http/1.1. So I expected that I get the same results.

Could you tell me why this happened?

Solution

First things first: HTTP/1.1 does allow multiple request per connection with Keep-Alive header. This is the default behavior (RFC 2616, Section 8.1).

The documentation is a bit unclear.

maxRequestsPerConnection description states:

Maximum number of requests per connection to a backend. Setting this parameter to 1 disables keep alive. Default 0, meaning “unlimited”, up to 2^29.

Setting maxRequestsPerConnection to 1 disables Keep-Alive. Setting it to any other value (value > 1) switches Keep-Alive back on.

Setting this field to proper value (not too high, not too low) is the hard part of configuring Istio, and is dependent on your application needs and traffic.