DCM cannot disable the health check dynamically. The API command exists (https://www.haproxy.com/documentation/haproxy-runtime-api/reference/disable-health/) but not used. This causes a problem with idled services whose server slot is set with the single endpoint - service's clusterIP. However a health check sent to the service's clusterIP unidles the service. In non DCM case this is mitigated by removing "check" keyword during the config template rendering (https://github.com/openshift/router/blob/b447c4d27d38a6c5f6ce2d5ceda88dbc9b90c661/images/router/haproxy/conf/haproxy-config.template#L755). Although in DCM case only the server slot is updated with the clusterIP keeping the health check enabled.

4.18+ (when TechPreview featureset is used)  

Not always but often

Origin's test "Idling with a single service and ReplicationController should idle the service and ReplicationController properly"(https://github.com/openshift/origin/blob/c22aad19d02153b206ec2efbd8e18c24b70a191a/test/extended/idling/idling.go#L230-L238).
    

  [sig-network-edge][Feature:Idling] Idling with a single service and ReplicationController [It] should idle the service and ReplicationController properly [Suite:openshift/conformance/parallel]
  github.com/openshift/origin/test/extended/idling/idling.go:235

    [FAILED] Failed after 1.209s.
    Expected
        <string>: "2"
    to contain substring
        <string>: 0
    In [It] at: github.com/openshift/origin/test/extended/idling/idling.go:129 @ 02/05/25 22:53:12.733
  ------------------------------

  Summarizing 1 Failure:
    [FAIL] [sig-network-edge][Feature:Idling] Idling with a single service and ReplicationController [It] should idle the service and ReplicationController properly [Suite:openshift/conformance/parallel]
    github.com/openshift/origin/test/extended/idling/idling.go:129

  Ran 1 of 1 Specs in 13.544 seconds
  FAIL! -- 0 Passed | 1 Failed | 0 Pending | 0 Skipped
fail [github.com/openshift/origin/test/extended/idling/idling.go:129]: Failed after 1.209s.
Expected
    <string>: "2"
to contain substring
    <string>: 0
Ginkgo exit error 1: exit with code 

Successful test.

Potential fix: fallback to the router reload for the route update event. Example: https://github.com/alebedev87/router/commit/d10e732d2a87cf22fb7e7bbfb14fb0b48d000068