cleanup comments

Author: mattwittwer

qa/L0_vertex_ai/test.sh

@@ -699,20 +699,14 @@ else
 fi
 set -e
 
-# repository control (expect error)
+# repository control via redirect is blocked unconditionally
 rm -f ./curl.out
 set +e
 code=`curl -s -w %{http_code} -o ./curl.out -X POST -H "X-Vertex-Ai-Triton-Redirect: v2/repository/models/subadd/unload" localhost:8080/predict`
-if [ "$code" == "200" ]; then
+if [ "$code" != "403" ]; then
     cat ./curl.out
-    echo -e "\n***\n*** Test Failed\n***"
+    echo -e "\n***\n*** Failed. Expected model unload via redirect to return 403 (got $code)\n***"
     RET=1
-else
-    grep "explicit model load / unload is not allowed" ./curl.out
-    if [ $? -ne 0 ]; then
-        echo -e "\n***\n*** Failed. Expected error on model control\n***"
-        RET=1
-    fi
 fi
 set -e
 
@@ -813,8 +807,7 @@ if [ "$code" != "403" ]; then
     RET=1
 fi
 
-# Model load redirect is unconditionally blocked through the prediction
-# endpoint regardless of restricted header configuration (TRI-715 RCE vector)
+# Model load redirect is unconditionally blocked through the prediction endpoint
 rm -f ./curl.out
 code=`curl -s -w %{http_code} -o ./curl.out -X POST -H "X-Vertex-Ai-Triton-Redirect: v2/repository/models/identity_fp32/load" localhost:8080/predict`
 if [ "$code" != "403" ]; then
@@ -823,7 +816,7 @@ if [ "$code" != "403" ]; then
     RET=1
 fi
 
-# Model unload redirect is unconditionally blocked (TRI-715 DoS vector)
+# Model unload redirect is unconditionally blocked through the prediction endpoint
 rm -f ./curl.out
 code=`curl -s -w %{http_code} -o ./curl.out -X POST -H "X-Vertex-Ai-Triton-Redirect: v2/repository/models/identity_fp32/unload" localhost:8080/predict`
 if [ "$code" != "403" ]; then
@@ -941,7 +934,7 @@ export AIP_PREDICT_ROUTE="/predict"
 export AIP_HEALTH_ROUTE="/health"
 
 SERVER_LOG="vertex_max_input_size_server.log"
-SERVER_ARGS="--allow-vertex-ai=true --model-repository=restricted_single_model --vertex-ai-default-model=identity_fp32 --http-max-input-size=256"
+SERVER_ARGS="--allow-vertex-ai=true --model-repository=restricted_single_model --vertex-ai-default-model=identity_fp32 --http-max-input-size=128"
 run_server_nowait
 vertex_ai_wait_for_server_ready $SERVER_PID 10
 if [ "$WAIT_RET" != "0" ]; then

src/vertex_ai_server.cc

@@ -201,8 +201,9 @@ VertexAiAPIServer::Handle(evhtp_request_t* req)
           HandleSystemSharedMemory(req, region, action);
           return;
         }
-        // register/unregister are mutating operations that must not be
-        // reachable through the prediction endpoint.
+        // Only read-only status queries are permitted through redirect.
+        // Mutating operations (register/unregister) require the core
+        // HTTP endpoint.
         LOG_VERBOSE(1) << "Vertex AI redirect blocked: " << redirect_endpoint;
         evhtp_send_reply(req, EVHTP_RES_FORBIDDEN);
         return;
@@ -224,7 +225,8 @@ VertexAiAPIServer::Handle(evhtp_request_t* req)
           HandleRepositoryIndex(req, repo_name);
           return;
         }
-        // Model load/unload must not be reachable through the prediction endpoint.
+        // Only repository index queries are permitted through redirect.
+        // Model load/unload requires the core HTTP endpoint.
         LOG_VERBOSE(1) << "Vertex AI redirect blocked: " << redirect_endpoint;
         evhtp_send_reply(req, EVHTP_RES_FORBIDDEN);
         return;