The InferencePoolAppProtocol conformance test is failing at least against the LWEPP image.
This test verifies that the InferencePool honors appProtocol values and defaults back to HTTP/1.1 when it is omitted.
The deployment appprotocol-inference-model-server-deployment had a single container echoserver with two ports: 3000 and 3001. This is problematic because a single container cannot serve HTTP/1.1 on one port and h2c on another using the same process or configuration when using a standard echo-basic setup without either a conflict or an HTTP prober failing.
Splitting them into echoserver-http and echoserver-h2c in the same pod: the HTTP container gets probed by the HTTP/1.1 readiness probe on 3000 and the HTTP/2 container gets probed via a TCP socket probe on 3001. Now both servers can run concurrently and pass their readiness checks.
The
InferencePoolAppProtocolconformance test is failing at least against the LWEPP image.This test verifies that the InferencePool honors appProtocol values and defaults back to HTTP/1.1 when it is omitted.
The deployment
appprotocol-inference-model-server-deploymenthad a single containerechoserverwith two ports: 3000 and 3001. This is problematic because a single container cannot serve HTTP/1.1 on one port andh2con another using the same process or configuration when using a standardecho-basicsetup without either a conflict or an HTTP prober failing.Splitting them into
echoserver-httpandechoserver-h2cin the same pod: the HTTP container gets probed by the HTTP/1.1 readiness probe on 3000 and the HTTP/2 container gets probed via a TCP socket probe on 3001. Now both servers can run concurrently and pass their readiness checks.