diff --git a/CMakeLists.txt b/CMakeLists.txt
index 0c3c833f4cb602a3c33e289d2bdd544f6a337a79..3ad0f39ada004eabc4b604b38e3fdc006b5b7605 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -16,6 +16,7 @@ cmake_minimum_required(VERSION 3.0)
 set(CMAKE_MODULE_PATH ${CMAKE_MODULE_PATH} "${CMAKE_CURRENT_SOURCE_DIR}/cmake")
 set(PADDLE_SERVING_SOURCE_DIR ${CMAKE_CURRENT_SOURCE_DIR})
 set(PADDLE_SERVING_BINARY_DIR ${CMAKE_CURRENT_BINARY_DIR})
+SET(PADDLE_SERVING_INSTALL_DIR ${CMAKE_BINARY_DIR}/output)
 
 include(system)
 
@@ -40,12 +41,6 @@ endif()
 set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
   "A path setting third party libraries download & build directories.")
 
-set(FLUID_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_install_dir" CACHE STRING
-  "A path setting fluid shared and static libraries")
-
-set(FLUID_INFERENCE_INSTALL_DIR "${CMAKE_BINARY_DIR}/fluid_inference_install_dir" CACHE STRING
-  "A path setting fluid inference shared and static libraries")
-
 set(THIRD_PARTY_BUILD_TYPE Release)
 
 option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
@@ -99,3 +94,6 @@ add_subdirectory(predictor)
 add_subdirectory(inferencer-fluid-cpu)
 add_subdirectory(serving)
 add_subdirectory(sdk-cpp)
+
+
+
diff --git a/configure/CMakeLists.txt b/configure/CMakeLists.txt
index 3b6e55db4f07b1b07541a2794745cd03bf48b4a2..2f5ab6e7aec7e4c91fff976547796c5f8130a359 100644
--- a/configure/CMakeLists.txt
+++ b/configure/CMakeLists.txt
@@ -18,3 +18,8 @@ target_include_directories(test_configure PUBLIC
         ${CMAKE_CURRENT_LIST_DIR}/include
         )
 target_link_libraries(test_configure configure protobuf)
+
+install(TARGETS configure 
+        ARCHIVE DESTINATION ${PADDLE_SERVING_INSTALL_DIR}/lib
+        )
+
diff --git a/inferencer-fluid-cpu/CMakeLists.txt b/inferencer-fluid-cpu/CMakeLists.txt
index 64e4cddd762a2a40aa55eb113c7eadada3019228..8c05ee14593f00cc1049d6c3c89669ae5d4873d3 100644
--- a/inferencer-fluid-cpu/CMakeLists.txt
+++ b/inferencer-fluid-cpu/CMakeLists.txt
@@ -7,3 +7,7 @@ target_include_directories(fluid_cpu_engine PUBLIC
 add_dependencies(fluid_cpu_engine pdserving paddle_fluid configure)
 target_link_libraries(fluid_cpu_engine pdserving -liomp5 -lmklml_intel -lpthread -lcrypto -lm -lrt -lssl -ldl -lz)
 
+install(TARGETS fluid_cpu_engine 
+        ARCHIVE DESTINATION ${PADDLE_SERVING_INSTALL_DIR}/lib
+        )
+
diff --git a/inferencer-fluid-cpu/include/fluid_cpu_engine.h b/inferencer-fluid-cpu/include/fluid_cpu_engine.h
index 1339e28c7baf837cec987f2f13fb08b0f2c0567b..0f1e6613b684d329b7b280abfa0e168b24082860 100644
--- a/inferencer-fluid-cpu/include/fluid_cpu_engine.h
+++ b/inferencer-fluid-cpu/include/fluid_cpu_engine.h
@@ -74,7 +74,7 @@ public:
     virtual bool Run(const void* in_data, void* out_data) {
         if (!_core->Run(*(std::vector<paddle::PaddleTensor>*)in_data, 
                     (std::vector<paddle::PaddleTensor>*)out_data)) {
-            LOG(FATAL) << "Failed call Run with paddle predictor";
+            LOG(ERROR) << "Failed call Run with paddle predictor";
             return false;
         }
 
@@ -85,13 +85,13 @@ public:
 
     virtual int clone(void* origin_core) {
         if (origin_core == NULL) {
-            LOG(FATAL) << "origin paddle Predictor is null.";
+            LOG(ERROR) << "origin paddle Predictor is null.";
             return -1;
         }
         paddle::PaddlePredictor* p_predictor = (paddle::PaddlePredictor*)origin_core;
         _core = p_predictor->Clone();
         if (_core.get() == NULL) {
-            LOG(FATAL) << "fail to clone paddle predictor: " << origin_core;
+            LOG(ERROR) << "fail to clone paddle predictor: " << origin_core;
             return -1;
         }
         return 0;
@@ -110,7 +110,7 @@ class FluidCpuAnalysisCore : public FluidFamilyCore {
 public:
     int create(const std::string& data_path) {
         if (access(data_path.c_str(), F_OK) == -1) {
-            LOG(FATAL) << "create paddle predictor failed, path not exits: "
+            LOG(ERROR) << "create paddle predictor failed, path not exits: "
                 << data_path;
             return -1;
         }
@@ -125,7 +125,7 @@ public:
         _core = paddle::CreatePaddlePredictor<
             paddle::contrib::AnalysisConfig>(analysis_config);
         if (NULL == _core.get()) {
-            LOG(FATAL) << "create paddle predictor failed, path: "
+            LOG(ERROR) << "create paddle predictor failed, path: "
                 << data_path;
             return -1;
         }
@@ -139,7 +139,7 @@ class FluidCpuNativeCore : public FluidFamilyCore {
 public:
     int create(const std::string& data_path) {
         if (access(data_path.c_str(), F_OK) == -1) {
-            LOG(FATAL) << "create paddle predictor failed, path not exits: "
+            LOG(ERROR) << "create paddle predictor failed, path not exits: "
                 << data_path;
             return -1;
         }
@@ -153,7 +153,7 @@ public:
         _core = paddle::CreatePaddlePredictor<
             paddle::NativeConfig, paddle::PaddleEngineKind::kNative>(native_config);
         if (NULL == _core.get()) {
-            LOG(FATAL) << "create paddle predictor failed, path: "
+            LOG(ERROR) << "create paddle predictor failed, path: "
                 << data_path;
             return -1;
         }
@@ -167,7 +167,7 @@ class FluidCpuAnalysisDirCore : public FluidFamilyCore {
 public:
     int create(const std::string& data_path) {
         if (access(data_path.c_str(), F_OK) == -1) {
-            LOG(FATAL) << "create paddle predictor failed, path not exits: "
+            LOG(ERROR) << "create paddle predictor failed, path not exits: "
                 << data_path;
             return -1;
         }
@@ -181,7 +181,7 @@ public:
         _core = paddle::CreatePaddlePredictor<
             paddle::contrib::AnalysisConfig>(analysis_config);
         if (NULL == _core.get()) {
-            LOG(FATAL) << "create paddle predictor failed, path: "
+            LOG(ERROR) << "create paddle predictor failed, path: "
                 << data_path;
             return -1;
         }
@@ -196,7 +196,7 @@ class FluidCpuNativeDirCore : public FluidFamilyCore {
 public:
     int create(const std::string& data_path) {
         if (access(data_path.c_str(), F_OK) == -1) {
-            LOG(FATAL) << "create paddle predictor failed, path not exits: "
+            LOG(ERROR) << "create paddle predictor failed, path not exits: "
                 << data_path;
             return -1;
         }
@@ -209,7 +209,7 @@ public:
         _core = paddle::CreatePaddlePredictor<
             paddle::NativeConfig, paddle::PaddleEngineKind::kNative>(native_config);
         if (NULL == _core.get()) {
-            LOG(FATAL) << "create paddle predictor failed, path: "
+            LOG(ERROR) << "create paddle predictor failed, path: "
                 << data_path;
             return -1;
         }
@@ -235,7 +235,7 @@ public:
         _col = col;
         _params = (float*)malloc(_row * _col * sizeof(float));
         if (_params == NULL) {
-            LOG(FATAL) << "Load " << _file_name << " malloc error.";
+            LOG(ERROR) << "Load " << _file_name << " malloc error.";
             return -1;
         }
         LOG(WARNING) << "Load parameter file[" << _file_name << "] success.";
@@ -253,20 +253,20 @@ public:
 
     int load() {
         if (_params == NULL || _row <= 0 || _col <= 0) {
-            LOG(FATAL) << "load parameter error [not inited].";
+            LOG(ERROR) << "load parameter error [not inited].";
             return -1;
         }
 
         FILE* fs = fopen(_file_name.c_str(), "rb");
         if (fs == NULL) {
-            LOG(FATAL) << "load " << _file_name << " fopen error.";
+            LOG(ERROR) << "load " << _file_name << " fopen error.";
             return -1;
         }
         static const uint32_t MODEL_FILE_HEAD_LEN = 16;
         char head[MODEL_FILE_HEAD_LEN] = {0};
         if (fread(head, 1, MODEL_FILE_HEAD_LEN, fs) != MODEL_FILE_HEAD_LEN) {
             destroy();
-            LOG(FATAL) << "Load " << _file_name << " read head error.";
+            LOG(ERROR) << "Load " << _file_name << " read head error.";
             if (fs != NULL) {
                 fclose(fs);
                 fs = NULL;
@@ -283,7 +283,7 @@ public:
             LOG(INFO) << "load " << _file_name << " read ok.";
             return 0;
         } else {
-            LOG(FATAL) << "load " << _file_name << " read error.";
+            LOG(ERROR) << "load " << _file_name << " read error.";
             destroy();
             if (fs != NULL) {
                 fclose(fs);
@@ -309,13 +309,13 @@ public:
             float exp_max, float exp_min) {
         AutoLock lock(GlobalSigmoidCreateMutex::instance());
         if (0 != _sigmoid_w.init(2, 1, sigmoid_w_file) || 0 != _sigmoid_w.load()) {
-            LOG(FATAL) << "load params sigmoid_w failed.";
+            LOG(ERROR) << "load params sigmoid_w failed.";
             return -1;
         }
         LOG(WARNING) << "load sigmoid_w [" << _sigmoid_w._params[0]
                 << "] [" << _sigmoid_w._params[1] << "].";
         if (0 != _sigmoid_b.init(2, 1, sigmoid_b_file) || 0 != _sigmoid_b.load()) {
-            LOG(FATAL) << "load params sigmoid_b failed.";
+            LOG(ERROR) << "load params sigmoid_b failed.";
             return -1;
         }
         LOG(WARNING) << "load sigmoid_b [" << _sigmoid_b._params[0]
@@ -372,7 +372,7 @@ public:
         std::string conf_file = model_path.substr(pos);
         configure::SigmoidConf conf;
         if (configure::read_proto_conf(conf_path, conf_file, &conf) != 0) {
-            LOG(FATAL) << "failed load model path: " << model_path;
+            LOG(ERROR) << "failed load model path: " << model_path;
             return -1;
         }
 
@@ -381,7 +381,7 @@ public:
         std::string fluid_model_data_path = conf.dnn_model_path();
         int ret = load_fluid_model(fluid_model_data_path);
         if (ret < 0) {
-            LOG(FATAL) << "fail to load fluid model.";
+            LOG(ERROR) << "fail to load fluid model.";
             return -1;
         }
         const char* sigmoid_w_file = conf.sigmoid_w_file().c_str();
@@ -393,7 +393,7 @@ public:
                 << "], use count[" << _core->_sigmoid_core.use_count() << "].";
         ret = _core->_sigmoid_core->load(sigmoid_w_file, sigmoid_b_file, exp_max, exp_min);
         if (ret < 0) {
-            LOG(FATAL) << "fail to load sigmoid model.";
+            LOG(ERROR) << "fail to load sigmoid model.";
             return -1;
         }
         return 0;
@@ -402,7 +402,7 @@ public:
     virtual bool Run(const void* in_data, void* out_data) {
         if (!_core->_fluid_core->Run(*(std::vector<paddle::PaddleTensor>*)in_data, 
                     (std::vector<paddle::PaddleTensor>*)out_data)) {
-            LOG(FATAL) << "Failed call Run with paddle predictor";
+            LOG(ERROR) << "Failed call Run with paddle predictor";
             return false;
         }
 
@@ -411,12 +411,12 @@ public:
 
     virtual int clone(SigmoidFluidModel* origin_core) {
         if (origin_core == NULL) {
-            LOG(FATAL) << "origin paddle Predictor is null.";
+            LOG(ERROR) << "origin paddle Predictor is null.";
             return -1;
         }
         _core = origin_core->Clone();
         if (_core.get() == NULL) {
-            LOG(FATAL) << "fail to clone paddle predictor: " << origin_core;
+            LOG(ERROR) << "fail to clone paddle predictor: " << origin_core;
             return -1;
         }
         LOG(INFO) << "clone sigmoid core[" << _core->_sigmoid_core.get()
@@ -442,7 +442,7 @@ class FluidCpuNativeDirWithSigmoidCore : public FluidCpuWithSigmoidCore {
 public:
     int load_fluid_model(const std::string& data_path) {
         if (access(data_path.c_str(), F_OK) == -1) {
-            LOG(FATAL) << "create paddle predictor failed, path not exits: "
+            LOG(ERROR) << "create paddle predictor failed, path not exits: "
                 << data_path;
             return -1;
         }
@@ -455,7 +455,7 @@ public:
         _core->_fluid_core = paddle::CreatePaddlePredictor<
             paddle::NativeConfig, paddle::PaddleEngineKind::kNative>(native_config);
         if (NULL == _core.get()) {
-            LOG(FATAL) << "create paddle predictor failed, path: "
+            LOG(ERROR) << "create paddle predictor failed, path: "
                 << data_path;
             return -1;
         }
@@ -470,7 +470,7 @@ class FluidCpuAnalysisDirWithSigmoidCore : public FluidCpuWithSigmoidCore {
 public:
     int load_fluid_model(const std::string& data_path) {
         if (access(data_path.c_str(), F_OK) == -1) {
-            LOG(FATAL) << "create paddle predictor failed, path not exits: "
+            LOG(ERROR) << "create paddle predictor failed, path not exits: "
                 << data_path;
             return -1;
         }
@@ -484,7 +484,7 @@ public:
         _core->_fluid_core = paddle::CreatePaddlePredictor<
             paddle::contrib::AnalysisConfig>(analysis_config);
         if (NULL == _core.get()) {
-            LOG(FATAL) << "create paddle predictor failed, path: "
+            LOG(ERROR) << "create paddle predictor failed, path: "
                 << data_path;
             return -1;
         }
diff --git a/predictor/CMakeLists.txt b/predictor/CMakeLists.txt
index 9e0a9ac1259c3dee3bb0c77c68ca22da34a97ca2..307588e4d038461698445c24c0b41e50475bf717 100644
--- a/predictor/CMakeLists.txt
+++ b/predictor/CMakeLists.txt
@@ -61,3 +61,11 @@ target_include_directories(pdclient PUBLIC
         )
 target_link_libraries(pdclient protobuf boost brpc -lpthread -lcrypto -lm -lrt -lssl -ldl -lz)
 
+# install
+install(TARGETS pdclient pdserving pdcodegen
+        RUNTIME DESTINATION ${PADDLE_SERVING_INSTALL_DIR}/bin
+        ARCHIVE DESTINATION ${PADDLE_SERVING_INSTALL_DIR}/lib
+        LIBRARY DESTINATION ${PADDLE_SERVING_INSTALL_DIR}/so
+        )
+
+
diff --git a/predictor/common/macros.h b/predictor/common/macros.h
index 992dc1df201e90a2972159c244921bd00d8bebbb..a6607f60d51ed4657047b23aa445596556562b38 100644
--- a/predictor/common/macros.h
+++ b/predictor/common/macros.h
@@ -10,7 +10,7 @@ namespace predictor {
 #ifndef CATCH_ANY_AND_RET
 #define CATCH_ANY_AND_RET(errno)            \
     catch (...) {                           \
-        LOG(FATAL) << "exception catched";  \
+        LOG(ERROR) << "exception catched";  \
         return errno;                       \
     }
 #endif
diff --git a/predictor/framework/bsf-inl-tensor.h b/predictor/framework/bsf-inl-tensor.h
index 40114c9d0e542bf36781097fb7a4eaa18d4f0bb1..5568cc7aae85aa0e0f5109619036fcacd58556b7 100644
--- a/predictor/framework/bsf-inl-tensor.h
+++ b/predictor/framework/bsf-inl-tensor.h
@@ -111,23 +111,23 @@ public:
             const InArrayT& in, OutArrayT& out, bool align) {
         if (align) {
             if (out.count() <= 0 || out.size() <= 0) {
-                LOG(FATAL) << "Out tensor is empty, when aligned";
+                LOG(ERROR) << "Out tensor is empty, when aligned";
                 return false;
             }
 
             if (out.size() != in.size()) {
-                LOG(FATAL) << "In/Out tensor size not eq: " << out.size() << "!=" << in.size();
+                LOG(ERROR) << "In/Out tensor size not eq: " << out.size() << "!=" << in.size();
                 return false;
             }
 
             for (size_t fi = 0, shape0 = 0; fi < out.count(); ++fi) {
                 if (!out[fi].valid()) {
-                    LOG(FATAL) << "Out[" << fi << "] tensor not valid";
+                    LOG(ERROR) << "Out[" << fi << "] tensor not valid";
                     return false;
                 }
 
                 if (out.size() != out[fi].shape0()) {
-                    LOG(FATAL) << "Shape0 not consistency, " << out.size() << "!=" << out[fi].shape0() << ", " << fi;
+                    LOG(ERROR) << "Shape0 not consistency, " << out.size() << "!=" << out[fi].shape0() << ", " << fi;
                     return false;
                 }
             }
@@ -231,7 +231,7 @@ public:
             void* data_buf
                 = MempoolWrapper::instance().malloc(tensor_byte);
             if (!data_buf) {
-                LOG(FATAL) << "Malloc failed, size: " << tensor_byte;
+                LOG(ERROR) << "Malloc failed, size: " << tensor_byte;
                 return ;
             }
 
@@ -240,7 +240,7 @@ public:
                 TaskMetaT& tm = _tasks[ti];
                 size_t acc_byte = ins_byte * (tm.end - tm.begin);
                 if (data_byte + acc_byte > tensor_byte) {
-                    LOG(FATAL) << "Invalid bytes: " << data_byte << " + " << acc_byte << " >= " << tensor_byte;
+                    LOG(ERROR) << "Invalid bytes: " << data_byte << " + " << acc_byte << " >= " << tensor_byte;
                     return ;
                 }
 
@@ -252,7 +252,7 @@ public:
             }
 
             if (data_byte != tensor_byte) {
-                LOG(FATAL) << "Invalid tensor byte: " << data_byte << " != " << tensor_byte; 
+                LOG(ERROR) << "Invalid tensor byte: " << data_byte << " != " << tensor_byte; 
                 return ;
             }
 
@@ -270,7 +270,7 @@ public:
 
     void notify_tasks() {
         if (_batch_out.size() != _batch_in.size()) {
-            LOG(FATAL) << "batch size not consistency: " << _batch_out.size() << " != " << _batch_in.size();
+            LOG(ERROR) << "batch size not consistency: " << _batch_out.size() << " != " << _batch_in.size();
             return ;
         }
 
diff --git a/predictor/framework/bsf-inl.h b/predictor/framework/bsf-inl.h
index 1d4e88fb4621ca8cf73777e0c5a86ea9fef2e7aa..bc95886ca23e5fd795175d0edbf1a516241352af 100644
--- a/predictor/framework/bsf-inl.h
+++ b/predictor/framework/bsf-inl.h
@@ -28,7 +28,7 @@ int TaskExecutor<TaskT>::start(uint32_t thread_num, uint32_t init_timeout_sec) {
     }
 
     if (thread_num == 0) {
-        LOG(FATAL) << "cannot init BSF with zero thread";
+        LOG(ERROR) << "cannot init BSF with zero thread";
         return -1;
     }
 
@@ -42,7 +42,7 @@ int TaskExecutor<TaskT>::start(uint32_t thread_num, uint32_t init_timeout_sec) {
         int rc = THREAD_CREATE(
                 &contexts[i].tid, NULL, &TaskExecutor::thread_entry, &contexts[i]);
         if (rc != 0) {
-            LOG(FATAL) << "failed to create BSF worker thread: index=" << i << ", rc=" << rc << ", errno=" << errno << ":" << strerror(errno);
+            LOG(ERROR) << "failed to create BSF worker thread: index=" << i << ", rc=" << rc << ", errno=" << errno << ":" << strerror(errno);
             return -1;
         }
 
@@ -71,7 +71,7 @@ int TaskExecutor<TaskT>::start(uint32_t thread_num, uint32_t init_timeout_sec) {
         }
 
         if (has_error) {
-            LOG(FATAL) << "BSF thread init error";
+            LOG(ERROR) << "BSF thread init error";
             return -1;
         }
 
@@ -86,7 +86,7 @@ int TaskExecutor<TaskT>::start(uint32_t thread_num, uint32_t init_timeout_sec) {
         init_timeout -= sleep_interval;
     }
 
-    LOG(FATAL) << "BSF thread init timed out";
+    LOG(ERROR) << "BSF thread init timed out";
     return -1;
 }
 
@@ -108,19 +108,19 @@ TaskHandler<TaskT> TaskExecutor<TaskT>::schedule(
         const InArrayT& in, OutArrayT& out) {
     TaskT* task = butil::get_object<TaskT>();
     if (!task) {
-        LOG(FATAL) << "Failed get TaskT from object pool";
+        LOG(ERROR) << "Failed get TaskT from object pool";
         return TaskHandler<TaskT>::valid_handle();
     }
 
     if (!BatchTasks<TaskT>::check_valid(in, out, _batch_align)) {
-        LOG(FATAL) << "Invalid input & output";
+        LOG(ERROR) << "Invalid input & output";
         return TaskHandler<TaskT>::valid_handle();
     }
 
     int fds[2];
     int rc = pipe(fds);
     if (rc != 0) {
-        LOG(FATAL) << "call pipe() failed, errno=" << errno << ":" << strerror(errno);
+        LOG(ERROR) << "call pipe() failed, errno=" << errno << ":" << strerror(errno);
         return TaskHandler<TaskT>::valid_handle();
     }
 
@@ -149,7 +149,7 @@ bool TaskExecutor<TaskT>::fetch_batch(BatchTasks<TaskT>& batch) {
     }
 
     if (_task_queue.empty()) {
-        LOG(FATAL) << "invalid task queue!";
+        LOG(ERROR) << "invalid task queue!";
         return false;
     }
 
@@ -169,7 +169,7 @@ template<typename TaskT>
 int TaskExecutor<TaskT>::work(ThreadContext<TaskT>* context) {
     if (_thread_init_fn != NULL) {
         if (_thread_init_fn(context->user_thread_context) != 0) {
-            LOG(FATAL) << "execute thread init thunk failed, BSF thread will exit";
+            LOG(ERROR) << "execute thread init thunk failed, BSF thread will exit";
             context->init_status = -1;
             return -1;
         } else {
@@ -181,7 +181,7 @@ int TaskExecutor<TaskT>::work(ThreadContext<TaskT>* context) {
     while (!_stop) {
         if (_thread_reset_fn != NULL) {
             if (_thread_reset_fn(context->user_thread_context) != 0) {
-                LOG(FATAL) << "execute user thread reset failed";
+                LOG(ERROR) << "execute user thread reset failed";
             }
         }
 
@@ -205,7 +205,7 @@ bool TaskManager<InItemT, OutItemT>::schedule(const InArrayT& in,
         _task_owned = handler;
         return true;
     } else {
-        LOG(FATAL) << "failed to schedule task";
+        LOG(ERROR) << "failed to schedule task";
         return false;
     }
 }
diff --git a/predictor/framework/bsf.h b/predictor/framework/bsf.h
index a152236e1f4074b9aab8221c6d609ecbc9e31a7b..4b21d2651cc7a769d8fd3a534b8df5998f08eb61 100644
--- a/predictor/framework/bsf.h
+++ b/predictor/framework/bsf.h
@@ -121,7 +121,7 @@ public:
 
     void notify_tasks() {
         if (_batch_out.size() != _batch_in.size()) {
-            LOG(FATAL) << "batch size not consistency: " << _batch_out.size() << " != " << _batch_in.size();
+            LOG(ERROR) << "batch size not consistency: " << _batch_out.size() << " != " << _batch_in.size();
             return ;
         }
 
@@ -133,7 +133,7 @@ public:
 
             for (size_t oi = begin; oi < end; ++oi, ++bi) {
                 if (bi >= _batch_in.size()) {
-                    LOG(FATAL) << "batch index overflow: " << bi << " > " <<_batch_in.size();
+                    LOG(ERROR) << "batch index overflow: " << bi << " > " <<_batch_in.size();
                     return ;
                 }
                 (*task->out)[oi] = _batch_out[bi];
diff --git a/predictor/framework/channel.h b/predictor/framework/channel.h
index b38ef157545cdb10fe1a6c906046180fca77744e..2327a02d9fc3424742cd07685219f35f4c449ee0 100644
--- a/predictor/framework/channel.h
+++ b/predictor/framework/channel.h
@@ -74,7 +74,7 @@ public:
 
     int share_to_bus(Bus* bus) {
         if (bus->regist(_op, this) != 0) {
-            LOG(FATAL) 
+            LOG(ERROR) 
                 << "Failed regist channel[" << _op
                 << "] to bus!";
             return -1;
@@ -128,7 +128,7 @@ public:
     }
 
     google::protobuf::Message* message_impl(derived_from_message<false>) {
-        LOG(FATAL) << "Current type: " << typeid(T).name()
+        LOG(ERROR) << "Current type: " << typeid(T).name()
             << " is not derived from protobuf.";
         return NULL; 
     }
@@ -143,7 +143,7 @@ public:
     }
 
     const google::protobuf::Message* message_impl(derived_from_message<false>) const {
-        LOG(FATAL) << "Current type: " << typeid(T).name()
+        LOG(ERROR) << "Current type: " << typeid(T).name()
             << " is not derived from protobuf.";
         return NULL; 
     }
diff --git a/predictor/framework/dag.cpp b/predictor/framework/dag.cpp
index cbfd0ea60ae856ce42a332aa698a626c30c5f4ab..4dcf4e012f7c56f22281bb97320e056ff1014b9d 100644
--- a/predictor/framework/dag.cpp
+++ b/predictor/framework/dag.cpp
@@ -34,7 +34,7 @@ int Dag::deinit() {
             if (conf != NULL) {
                 Op* op = OpRepository::instance().get_op(node->type);
                 if (op == NULL) {
-                    LOG(FATAL) << "Failed to get_op, op type[" << node->type << "]";
+                    LOG(ERROR) << "Failed to get_op, op type[" << node->type << "]";
                     return -1;
                 }
                 op->delete_config(conf);
@@ -89,7 +89,7 @@ EdgeMode Dag::parse_mode(std::string& mode) {
 int Dag::init(const char* path, const char* file, const std::string& name) {
     comcfg::Configure conf;
     if (conf.load(path, file) != 0) {
-        LOG(FATAL) << "Failed load conf from" 
+        LOG(ERROR) << "Failed load conf from" 
             << path << "/" << file << " in dag: "
             << name;
         return ERR_INTERNAL_FAILURE;
@@ -123,7 +123,7 @@ int Dag::init(const configure::Workflow& conf, const std::string& name) {
         }
         Op* op = OpRepository::instance().get_op(node->type);
         if (op == NULL) {
-            LOG(FATAL) << "Failed to get_op, op type[" << node->type << "]";
+            LOG(ERROR) << "Failed to get_op, op type[" << node->type << "]";
             return ERR_INTERNAL_FAILURE;
         }
         // node->conf could be NULL
@@ -134,7 +134,7 @@ int Dag::init(const configure::Workflow& conf, const std::string& name) {
     }
 
     if (topo_sort() != 0) {
-        LOG(FATAL) << "Topo sort dag[" << _dag_name << "] failed!";
+        LOG(ERROR) << "Topo sort dag[" << _dag_name << "] failed!";
         return ERR_INTERNAL_FAILURE;
     }
 
@@ -228,7 +228,7 @@ void Dag::regist_metric(const std::string& service_name) {
                     OP_METRIC_PREFIX + service_name + NAME_DELIMITER + node->full_name);
             Op* op = OpRepository::instance().get_op(node->type);
             if (op == NULL) {
-                LOG(FATAL) << "Failed to get_op, op type[" << node->type << "]";
+                LOG(ERROR) << "Failed to get_op, op type[" << node->type << "]";
                 return;
             }
             op->set_full_name(service_name + NAME_DELIMITER + node->full_name);
diff --git a/predictor/framework/dag_view.cpp b/predictor/framework/dag_view.cpp
index 02810a387496d699e68d5abdfbcb2786b84446ba..1d9e55ad2cff6c1a1bc8cb9b660a090b65163e49 100644
--- a/predictor/framework/dag_view.cpp
+++ b/predictor/framework/dag_view.cpp
@@ -17,12 +17,12 @@ int DagView::init(Dag* dag, const std::string& service_name) {
     for (uint32_t si = 0; si < stage_size; si++) {
         const DagStage* stage = dag->stage_by_index(si);
         if (stage == NULL) {
-            LOG(FATAL) << "Failed get stage by index:" << si;
+            LOG(ERROR) << "Failed get stage by index:" << si;
             return ERR_INTERNAL_FAILURE;
         }
         ViewStage* vstage = butil::get_object<ViewStage>();
         if (vstage == NULL) {
-            LOG(FATAL) 
+            LOG(ERROR) 
                 << "Failed get vstage from object pool" 
                 << "at:" << si;
             return ERR_MEM_ALLOC_FAILURE;
@@ -34,13 +34,13 @@ int DagView::init(Dag* dag, const std::string& service_name) {
             DagNode* node = stage->nodes[ni];
             ViewNode* vnode = butil::get_object<ViewNode>();
             if (vnode == NULL) {
-                LOG(FATAL) << "Failed get vnode at:" << ni;
+                LOG(ERROR) << "Failed get vnode at:" << ni;
                 return ERR_MEM_ALLOC_FAILURE;
             }
             // factory type
             Op* op = OpRepository::instance().get_op(node->type);
             if (op == NULL) {
-                LOG(FATAL) << "Failed get op with type:" 
+                LOG(ERROR) << "Failed get op with type:" 
                     << node->type;
                 return ERR_INTERNAL_FAILURE;
             }
@@ -91,7 +91,7 @@ int DagView::execute(butil::IOBufBuilder* debug_os) {
         int errcode = execute_one_stage(_view[si], debug_os);
         TRACEPRINTF("finish to execute stage[%u]", si);
         if (errcode < 0) {
-            LOG(FATAL) 
+            LOG(ERROR) 
                 << "failed execute stage[" 
                 << _view[si]->debug();
             return errcode;
@@ -115,7 +115,7 @@ int DagView::execute_one_stage(ViewStage* vstage,
         int errcode = op->process(debug_os != NULL);
         TRACEPRINTF("finish to execute op[%s]", op->name());
         if (errcode < 0) {
-            LOG(FATAL) 
+            LOG(ERROR) 
                 << "Execute failed, Op:" << op->debug_string();
             return errcode;
         }
@@ -156,14 +156,14 @@ const Channel* DagView::get_response_channel() const {
     // Caller obtains response channel from bus, and
     // writes it to rpc response(protbuf/json)
     if (_view.size() < 1) {
-        LOG(FATAL) << "invalid empty view stage!";
+        LOG(ERROR) << "invalid empty view stage!";
         return NULL;
     }
 
     ViewStage* last_stage = _view[_view.size() - 1];
     if (last_stage->nodes.size() != 1 
                     || last_stage->nodes[0] == NULL) {
-        LOG(FATAL) << "Invalid last stage, size[" 
+        LOG(ERROR) << "Invalid last stage, size[" 
                 << last_stage->nodes.size()
                 << "] != 1";
         return NULL;
@@ -171,7 +171,7 @@ const Channel* DagView::get_response_channel() const {
 
     Op* last_op = last_stage->nodes[0]->op;
     if (last_op == NULL) {
-        LOG(FATAL) << "Last op is NULL";
+        LOG(ERROR) << "Last op is NULL";
         return NULL;
     }
     return last_op->mutable_channel();
diff --git a/predictor/framework/infer.h b/predictor/framework/infer.h
index 634864b9d833cf337ccefe181098a044abfac590..134c5880c5e3b3c20bf01ac72e67a6018325e931 100644
--- a/predictor/framework/infer.h
+++ b/predictor/framework/infer.h
@@ -78,12 +78,12 @@ public:
         _infer_batch_size = conf.batch_infer_size();
         _infer_batch_align = conf.enable_batch_align();
         if (!check_need_reload() || load(_model_data_path) != 0) {
-            LOG(FATAL) << "Failed load model_data_path" << _model_data_path;
+            LOG(ERROR) << "Failed load model_data_path" << _model_data_path;
             return -1;
         }
 
         if (parse_version_info(conf, version) != 0) {
-            LOG(FATAL) << "Failed parse version info";
+            LOG(ERROR) << "Failed parse version info";
             return -1;
         }
 
@@ -93,7 +93,7 @@ public:
 
     int proc_initialize(const configure::EngineDesc& conf, bool version) {
         if (proc_initialize_impl(conf, version) != 0) {
-            LOG(FATAL) << "Failed proc initialize impl";
+            LOG(ERROR) << "Failed proc initialize impl";
             return -1;
         }
 
@@ -112,7 +112,7 @@ public:
         im::bsf::TaskExecutor<TaskT>::instance()->set_batch_align(_infer_batch_align);
         if (im::bsf::TaskExecutor<TaskT>::instance()->start(_infer_thread_num)
                 != 0) {
-            LOG(FATAL) << "Failed start bsf executor, threads:" << _infer_thread_num;
+            LOG(ERROR) << "Failed start bsf executor, threads:" << _infer_thread_num;
             return -1;
         }
 
@@ -152,7 +152,7 @@ public:
 
     int proc_finalize() {
         if (proc_finalize_impl() != 0) {
-            LOG(FATAL) << "Failed proc finalize impl";
+            LOG(ERROR) << "Failed proc finalize impl";
             return -1;
         }
 
@@ -198,7 +198,7 @@ private:
         } else if (_reload_mode_tag == "none") {
             return false;
         } else {
-            LOG(FATAL) << "Not support check type: "
+            LOG(ERROR) << "Not support check type: "
                 << _reload_mode_tag;
             return false;
         }
@@ -207,7 +207,7 @@ private:
     bool check_timestamp_ne() {
         struct stat st;
         if (stat(_reload_tag_file.c_str(), &st) != 0) {
-            LOG(FATAL) << "Failed stat config file:"
+            LOG(ERROR) << "Failed stat config file:"
                 << _reload_tag_file;
             return false;
         }
@@ -224,7 +224,7 @@ private:
     bool check_timestamp_gt() {
         struct stat st;
         if (stat(_reload_tag_file.c_str(), &st) != 0) {
-            LOG(FATAL) << "Failed stat config file:"
+            LOG(ERROR) << "Failed stat config file:"
                 << _reload_tag_file;
             return false;
         }
@@ -293,7 +293,7 @@ public:
 
         for (uint32_t ti = 0; ti < _reload_vec.size(); ++ti) {
             if (load_data(_reload_vec[ti], model_data_dir) != 0) {
-                LOG(FATAL) << "Failed reload engine model: " << ti;
+                LOG(ERROR) << "Failed reload engine model: " << ti;
                 return -1;
             }
         }
@@ -312,7 +312,7 @@ public:
         md->cores[next_idx] = new (std::nothrow) EngineCore;
         if (!md->cores[next_idx]
                 || md->cores[next_idx]->create(data_path) != 0) {
-            LOG(FATAL) << "Failed create model, path: " << data_path;
+            LOG(ERROR) << "Failed create model, path: " << data_path;
             return -1;
         }
         md->current_idx = next_idx;
@@ -322,13 +322,13 @@ public:
     virtual int thrd_initialize_impl() {
         // memory pool to be inited in non-serving-threads
         if (MempoolWrapper::instance().thread_initialize() != 0) {
-            LOG(FATAL) << "Failed thread initialize mempool";
+            LOG(ERROR) << "Failed thread initialize mempool";
             return -1;
         }
         
         ModelData<EngineCore>* md = new(std::nothrow) ModelData<EngineCore>;
         if (!md || load_data(md, _model_data_path) != 0) {
-            LOG(FATAL) << "Failed create thread data from " << _model_data_path;
+            LOG(ERROR) << "Failed create thread data from " << _model_data_path;
             return -1;
         }
 
@@ -341,7 +341,7 @@ public:
     int thrd_clear_impl() {
         // for non-serving-threads
         if (MempoolWrapper::instance().thread_clear() != 0) {
-            LOG(FATAL) << "Failed thread clear mempool";
+            LOG(ERROR) << "Failed thread clear mempool";
             return -1;
         }
         return 0;
@@ -360,7 +360,7 @@ public:
     EngineCore* get_core() {
         ModelData<EngineCore>* md = (ModelData<EngineCore>*)THREAD_GETSPECIFIC(_skey);
         if (!md) {
-            LOG(FATAL) << "Failed get thread specific data";
+            LOG(ERROR) << "Failed get thread specific data";
             return NULL;
         }
         return md->cores[md->current_idx];
@@ -382,7 +382,7 @@ public:
     virtual int proc_initialize(const configure::EngineDesc& conf, bool version) {
         _pd = new (std::nothrow) ModelData<EngineCore>;
         if (!_pd) {
-            LOG(FATAL) << "Failed to allocate for ProcData";
+            LOG(ERROR) << "Failed to allocate for ProcData";
             return -1;
         }
         return DBReloadableInferEngine<EngineCore>::proc_initialize(
@@ -393,7 +393,7 @@ public:
         // 加载进程级模型数据
         if (!_pd || DBReloadableInferEngine<EngineCore>::load_data(
                     _pd, model_data_dir) != 0) {
-            LOG(FATAL) 
+            LOG(ERROR) 
                 << "Failed to create common model from [" 
                 << model_data_dir << "].";
             return -1;
@@ -409,7 +409,7 @@ public:
         for (uint32_t ti = 0; ti < DBReloadableInferEngine<EngineCore>::_reload_vec.size(); ++ti) {
             if (load_data(DBReloadableInferEngine<EngineCore>::_reload_vec[ti],
                         _pd->cores[_pd->current_idx]) != 0) {
-                LOG(FATAL) << "Failed reload engine model: " << ti;
+                LOG(ERROR) << "Failed reload engine model: " << ti;
                 return -1;
             }
         }
@@ -431,7 +431,7 @@ public:
         td->cores[next_idx] = new (std::nothrow) EngineCore;
         if (!td->cores[next_idx]
                 || td->cores[next_idx]->clone(pd_core->get()) != 0) {
-            LOG(FATAL) << "Failed clone model from pd_core[ " << pd_core
+            LOG(ERROR) << "Failed clone model from pd_core[ " << pd_core
                     << "], idx[" << next_idx << "]";
             return -1;
         }
@@ -446,13 +446,13 @@ public:
     virtual int thrd_initialize_impl() {
         // memory pool to be inited in non-serving-threads
         if (MempoolWrapper::instance().thread_initialize() != 0) {
-            LOG(FATAL) << "Failed thread initialize mempool";
+            LOG(ERROR) << "Failed thread initialize mempool";
             return -1;
         }
         
         ModelData<EngineCore>* md = new(std::nothrow) ModelData<EngineCore>;
         if (!md || load_data(md, _pd->cores[_pd->current_idx]) != 0) {
-            LOG(FATAL) << "Failed clone thread data, origin_core["
+            LOG(ERROR) << "Failed clone thread data, origin_core["
                     << _pd->cores[_pd->current_idx] << "].";
             return -1;
         }
@@ -477,12 +477,12 @@ public:
         FluidFamilyCore* core
             = DBReloadableInferEngine<FluidFamilyCore>::get_core();
         if (!core || !core->get()) {
-            LOG(FATAL) << "Failed get fluid core in infer_impl()";
+            LOG(ERROR) << "Failed get fluid core in infer_impl()";
             return -1;
         }
         
         if (!core->Run(in, out)) {
-            LOG(FATAL) << "Failed run fluid family core";
+            LOG(ERROR) << "Failed run fluid family core";
             return -1;
         }
         return 0;
@@ -503,19 +503,19 @@ public:
         TensorrtFamilyCore* core
             = DBReloadableInferEngine<TensorrtFamilyCore>::get_core();
         if (!core || !core->get()) {
-            LOG(FATAL) << "Failed get fluid core in infer_impl()";
+            LOG(ERROR) << "Failed get fluid core in infer_impl()";
             return -1;
         }
 
         if (!core->Run(in, out, batch_size)) {
-            LOG(FATAL) << "Failed run fluid family core";
+            LOG(ERROR) << "Failed run fluid family core";
             return -1;
         }
         return 0;
     }
 
     int infer_impl2(const BatchTensor& in, BatchTensor& out) {
-        LOG(FATAL) << "Tensortrt donot supports infer_impl2 yet!";
+        LOG(ERROR) << "Tensortrt donot supports infer_impl2 yet!";
         return -1;
     }
 };
@@ -527,12 +527,12 @@ public:
     ~AbacusInferEngine() {}
 
     int infer_impl1(const void* in, void* out, uint32_t batch_size = -1) {
-        LOG(FATAL) << "Abacus dnn engine must use predict interface";
+        LOG(ERROR) << "Abacus dnn engine must use predict interface";
         return -1;
     }
 
     int infer_impl2(const BatchTensor& in, BatchTensor& out) {
-        LOG(FATAL) << "Abacus dnn engine must use predict interface";
+        LOG(ERROR) << "Abacus dnn engine must use predict interface";
         return -1;
     }
 
@@ -541,7 +541,7 @@ public:
         AbacusFamilyCore* core
                 = CloneDBReloadableInferEngine<AbacusFamilyCore>::get_core();
         if (!core || !core->get()) {
-            LOG(FATAL) << "Failed get abacus core in predict()";
+            LOG(ERROR) << "Failed get abacus core in predict()";
             return -1;
         }
 
@@ -551,7 +551,7 @@ public:
         AbacusFamilyCore* core
                 = CloneDBReloadableInferEngine<AbacusFamilyCore>::get_core();
         if (!core || !core->get()) {
-            LOG(FATAL) << "Failed get abacus core in predict()";
+            LOG(ERROR) << "Failed get abacus core in predict()";
             return -1;
         }
 
@@ -561,7 +561,7 @@ public:
         AbacusFamilyCore* core
                 = CloneDBReloadableInferEngine<AbacusFamilyCore>::get_core();
         if (!core || !core->get()) {
-            LOG(FATAL) << "Failed get abacus core in debug()";
+            LOG(ERROR) << "Failed get abacus core in debug()";
             return -1;
         }
         return core->debug();
@@ -571,7 +571,7 @@ public:
         AbacusFamilyCore* core
                 = CloneDBReloadableInferEngine<AbacusFamilyCore>::get_core();
         if (!core || !core->get()) {
-            LOG(FATAL) << "Failed get abacus core in set_serach_id()";
+            LOG(ERROR) << "Failed get abacus core in set_serach_id()";
             return -1;
         }
         return core->set_search_id(sid);
@@ -581,7 +581,7 @@ public:
         AbacusFamilyCore* core
                 = CloneDBReloadableInferEngine<AbacusFamilyCore>::get_core();
         if (!core || !core->get()) {
-            LOG(FATAL) << "Failed get abacus core in set_layer_dim()";
+            LOG(ERROR) << "Failed get abacus core in set_layer_dim()";
             return -1;
         }
         return core->set_hidden_layer_dim(dim);
@@ -592,7 +592,7 @@ public:
         AbacusFamilyCore* core
                 = CloneDBReloadableInferEngine<AbacusFamilyCore>::get_core();
         if (!core || !core->get()) {
-            LOG(FATAL) << "Failed get abacus core in get_input()";
+            LOG(ERROR) << "Failed get abacus core in get_input()";
             return -1;
         }
         return core->get_input(ins_idx, fea_num, in);
@@ -603,7 +603,7 @@ public:
         AbacusFamilyCore* core
                 = CloneDBReloadableInferEngine<AbacusFamilyCore>::get_core();
         if (!core || !core->get()) {
-            LOG(FATAL) << "Failed get abacus core in get_layer_value()";
+            LOG(ERROR) << "Failed get abacus core in get_layer_value()";
             return -1;
         }
         return core->get_layer_value(name, ins_num, fea_dim, out);
@@ -613,7 +613,7 @@ public:
         AbacusFamilyCore* core
                 = CloneDBReloadableInferEngine<AbacusFamilyCore>::get_core();
         if (!core || !core->get()) {
-            LOG(FATAL) << "Failed get abacus core in set_position_idx()";
+            LOG(ERROR) << "Failed get abacus core in set_position_idx()";
             return;
         }
         core->set_position_idx(input, fea, ins_idx);
@@ -628,12 +628,12 @@ public:
     ~PaddleV2InferEngine() {}
     
     int infer_impl1(const void* in, void* out, uint32_t batch_size = -1) {
-        LOG(FATAL) << "Paddle V2 engine must use predict interface";
+        LOG(ERROR) << "Paddle V2 engine must use predict interface";
         return -1;
     }
 
     int infer_impl2(const BatchTensor& in, BatchTensor& out) {
-        LOG(FATAL) << "Paddle V2 engine must use predict interface";
+        LOG(ERROR) << "Paddle V2 engine must use predict interface";
         return -1;
     }
 };
@@ -649,7 +649,7 @@ public:
 
     int proc_initialize(const configure::EngineDesc& conf) {
         if (proc_initialize(conf, false) != 0) {
-            LOG(FATAL) << "Failed proc intialize engine: " 
+            LOG(ERROR) << "Failed proc intialize engine: " 
                         << conf.name().c_str();
             return -1;
         }
@@ -665,20 +665,20 @@ public:
             = StaticInferFactory::instance().generate_object(
                     engine_type);
         if (!engine) {
-            LOG(FATAL) << "Failed generate engine with type:"
+            LOG(ERROR) << "Failed generate engine with type:"
                 << engine_type;
             return -1;
         }
 
         if (engine->proc_initialize(conf, version) != 0) {
-            LOG(FATAL) << "Failed initialize engine, type:"
+            LOG(ERROR) << "Failed initialize engine, type:"
                 << engine_type;
             return -1;
         }
 
         auto r = _versions.insert(std::make_pair(engine->version(), engine));
         if (!r.second) {
-            LOG(FATAL) << "Failed insert item: " << engine->version() 
+            LOG(ERROR) << "Failed insert item: " << engine->version() 
                 << ", type: " << engine_type;
             return -1;
         }
@@ -690,7 +690,7 @@ public:
     int proc_finalize() {
         for (auto iter = _versions.begin(); iter != _versions.end(); ++iter) {
             if (iter->second->proc_finalize() != 0) {
-                LOG(FATAL) << "Failed proc finalize version engine: " <<
+                LOG(ERROR) << "Failed proc finalize version engine: " <<
                     iter->first;
             }
             LOG(WARNING) 
@@ -702,7 +702,7 @@ public:
     int thrd_initialize() {
         for (auto iter = _versions.begin(); iter != _versions.end(); ++iter) {
             if (iter->second->thrd_initialize() != 0) {
-                LOG(FATAL) << "Failed thrd initialize version engine: " <<
+                LOG(ERROR) << "Failed thrd initialize version engine: " <<
                     iter->first;
                 return -1;
             }
@@ -715,7 +715,7 @@ public:
     int thrd_clear() {
         for (auto iter = _versions.begin(); iter != _versions.end(); ++iter) {
             if (iter->second->thrd_clear() != 0) {
-                LOG(FATAL) << "Failed thrd clear version engine: " << 
+                LOG(ERROR) << "Failed thrd clear version engine: " << 
                     iter->first;
                 return -1;
             }
@@ -727,7 +727,7 @@ public:
     int thrd_finalize() {
         for (auto iter = _versions.begin(); iter != _versions.end(); ++iter) {
             if (iter->second->thrd_finalize() != 0) {
-                LOG(FATAL) << "Failed thrd finalize version engine: " << 
+                LOG(ERROR) << "Failed thrd finalize version engine: " << 
                     iter->first;
                 return -1;
             }
@@ -739,7 +739,7 @@ public:
     int reload() {
         for (auto iter = _versions.begin(); iter != _versions.end(); ++iter) {
             if (iter->second->reload() != 0) {
-                LOG(FATAL) << "Failed reload version engine: " << 
+                LOG(ERROR) << "Failed reload version engine: " << 
                     iter->first;
                 return -1;
             }
@@ -760,7 +760,7 @@ public:
     // inference interface
     InferEngine* default_engine() const {
         if (_versions.size() != 1) {
-            LOG(FATAL) << "Ambiguous default engine version:" 
+            LOG(ERROR) << "Ambiguous default engine version:" 
                 << _versions.size();
             return NULL;
         }
@@ -797,7 +797,7 @@ public:
             const void* in, void* out, uint32_t batch_size, uint64_t version) {
         auto iter = _versions.find(version);
         if (iter == _versions.end()) {
-            LOG(FATAL) << "Not found version engine: " << version;
+            LOG(ERROR) << "Not found version engine: " << version;
             return -1;
         }
 
@@ -808,7 +808,7 @@ public:
     T* get_core(uint64_t version) {
         auto iter = _versions.find(version);
         if (iter == _versions.end()) {
-            LOG(FATAL) << "Not found version engine: "  << version;
+            LOG(ERROR) << "Not found version engine: "  << version;
             return NULL;
         }
 
@@ -843,7 +843,7 @@ public:
     int proc_initialize(const char* path, const char* file) {
         ModelToolkitConf model_toolkit_conf;
         if (configure::read_proto_conf(path, file, &model_toolkit_conf) != 0) {
-            LOG(FATAL) << "failed load infer config, path: "
+            LOG(ERROR) << "failed load infer config, path: "
                 << path << "/" << file;
             return -1;
         }
@@ -853,19 +853,19 @@ public:
             std::string engine_name = model_toolkit_conf.engines(ei).name();
             VersionedInferEngine* engine = new (std::nothrow) VersionedInferEngine();
             if (!engine) {
-                LOG(FATAL) << "Failed generate versioned engine: " << engine_name;
+                LOG(ERROR) << "Failed generate versioned engine: " << engine_name;
                 return -1;
             }
 
             if (engine->proc_initialize(model_toolkit_conf.engines(ei)) != 0) {
-                LOG(FATAL) << "Failed initialize version engine, name:"
+                LOG(ERROR) << "Failed initialize version engine, name:"
                     << engine_name;
                 return -1;
             }
 
             auto r = _map.insert(std::make_pair(engine_name, engine));
             if (!r.second) {
-                LOG(FATAL) << "Failed insert item: " << engine_name;
+                LOG(ERROR) << "Failed insert item: " << engine_name;
                 return -1;
             }
             LOG(WARNING) << "Succ proc initialize engine: " << engine_name;
@@ -877,7 +877,7 @@ public:
     int thrd_initialize() {
         for (auto it = _map.begin(); it != _map.end(); ++it) {
             if (it->second->thrd_initialize() != 0) {
-                LOG(FATAL) << "Failed thrd initialize engine, name: "
+                LOG(ERROR) << "Failed thrd initialize engine, name: "
                     << it->first;
                 return -1;
             }
@@ -890,7 +890,7 @@ public:
     int thrd_clear() {
         for (auto it = _map.begin(); it != _map.end(); ++it) {
             if (it->second->thrd_clear() != 0) {
-                LOG(FATAL) << "Failed thrd clear engine, name: "
+                LOG(ERROR) << "Failed thrd clear engine, name: "
                     << it->first;
                 return -1;
             }
@@ -901,7 +901,7 @@ public:
     int reload() {
         for (auto it = _map.begin(); it != _map.end(); ++it) {
             if (it->second->reload() != 0) {
-                LOG(FATAL) << "Failed reload engine, name: "
+                LOG(ERROR) << "Failed reload engine, name: "
                     << it->first;
                 return -1;
             }
@@ -912,7 +912,7 @@ public:
     int thrd_finalize() {
         for (auto it = _map.begin(); it != _map.end(); ++it) {
             if (it->second->thrd_finalize() != 0) {
-                LOG(FATAL) << "Failed thrd finalize engine, name: "
+                LOG(ERROR) << "Failed thrd finalize engine, name: "
                     << it->first;
                 return -1;
             }
@@ -925,7 +925,7 @@ public:
     int proc_finalize() {
         for (auto it = _map.begin(); it != _map.end(); ++it) {
             if (it->second->proc_finalize() != 0) {
-                LOG(FATAL) << "Failed proc finalize engine, name: "
+                LOG(ERROR) << "Failed proc finalize engine, name: "
                     << it->first;
                 return -1;
             }
diff --git a/predictor/framework/infer_data.h b/predictor/framework/infer_data.h
index ad244c39e3ca3629a9d130a86c6df57f6a33a48d..cd26982af72fd9cda896c36634275a691fce79e9 100644
--- a/predictor/framework/infer_data.h
+++ b/predictor/framework/infer_data.h
@@ -86,14 +86,14 @@ struct Tensor {
     bool valid() const {
         if (shape.empty()) {
             if (data.data() || data.size()) {
-                LOG(FATAL) << "data should be empty";
+                LOG(ERROR) << "data should be empty";
                 return false;
             }
             return true;
         }
 
         if (!data.data() || !data.size()) {
-            LOG(FATAL) << "data cannot empty";
+            LOG(ERROR) << "data cannot empty";
             return false;
         }
 
@@ -103,7 +103,7 @@ struct Tensor {
         } 
 
         if (byte_size * ele_byte() != data.size()) {
-            LOG(FATAL) << "wrong data size: " << byte_size * ele_byte() << " vs. " << data.size();
+            LOG(ERROR) << "wrong data size: " << byte_size * ele_byte() << " vs. " << data.size();
             return false;
         }
 
diff --git a/predictor/framework/manager.h b/predictor/framework/manager.h
index 1cb274369d4c6e3c6b6038bd6f0cea416b3c5c0a..e280a01a45b44347bf057379329772e3545b269a 100644
--- a/predictor/framework/manager.h
+++ b/predictor/framework/manager.h
@@ -41,7 +41,7 @@ public:
     int initialize(const std::string path, const std::string file) {
         WorkflowConf workflow_conf;
         if (configure::read_proto_conf(path, file, &workflow_conf) != 0) {
-            LOG(FATAL) << "Failed load manager<" << Workflow::tag() << "> configure from " << path << "/" << file;
+            LOG(ERROR) << "Failed load manager<" << Workflow::tag() << "> configure from " << path << "/" << file;
             return -1;
         }
 
@@ -52,11 +52,11 @@ public:
             std::string name = workflow_conf.workflows(ii).name();
             Workflow* item = new (std::nothrow) Workflow();
             if (item == NULL) {
-                LOG(FATAL) << "Failed create " << Workflow::tag() << " for: " << name;
+                LOG(ERROR) << "Failed create " << Workflow::tag() << " for: " << name;
                 return -1;
             }
             if (item->init(workflow_conf.workflows(ii)) != 0) {
-                LOG(FATAL) 
+                LOG(ERROR) 
                     << "Failed init item: " << name << " at:"
                     << ii << "!";
                 return -1;
@@ -66,7 +66,7 @@ public:
                 typename boost::unordered_map<std::string, Workflow*>::iterator, bool>
                 r = _item_map.insert(std::make_pair(name, item));
             if (!r.second) {
-                LOG(FATAL) 
+                LOG(ERROR) 
                     << "Failed insert item:" << name << " at:"
                     << ii << "!";
                 return -1;
@@ -78,7 +78,7 @@ public:
         } 
         
         } catch (...) {
-            LOG(FATAL) 
+            LOG(ERROR) 
                 << "Config[" << path << "/" << file << "] format "
                 << "invalid, load failed";
             return -1;
@@ -149,7 +149,7 @@ public:
     int initialize(const std::string path, const std::string file) {
         InferServiceConf infer_service_conf;
         if (configure::read_proto_conf(path, file, &infer_service_conf) != 0) {
-            LOG(FATAL) << "Failed load manager<" << InferService::tag() << "> configure!";
+            LOG(ERROR) << "Failed load manager<" << InferService::tag() << "> configure!";
             return -1;
         }
 
@@ -160,11 +160,11 @@ public:
             std::string name = infer_service_conf.services(ii).name();
             InferService* item = new (std::nothrow) InferService();
             if (item == NULL) {
-                LOG(FATAL) << "Failed create " << InferService::tag() << " for: " << name;
+                LOG(ERROR) << "Failed create " << InferService::tag() << " for: " << name;
                 return -1;
             }
             if (item->init(infer_service_conf.services(ii)) != 0) {
-                LOG(FATAL) 
+                LOG(ERROR) 
                     << "Failed init item: " << name << " at:"
                     << ii << "!";
                 return -1;
@@ -174,7 +174,7 @@ public:
                 typename boost::unordered_map<std::string, InferService*>::iterator, bool>
                 r = _item_map.insert(std::make_pair(name, item));
             if (!r.second) {
-                LOG(FATAL) 
+                LOG(ERROR) 
                     << "Failed insert item:" << name << " at:"
                     << ii << "!";
                 return -1;
@@ -186,7 +186,7 @@ public:
         } 
         
         } catch (...) {
-            LOG(FATAL) 
+            LOG(ERROR) 
                 << "Config[" << path << "/" << file << "] format "
                 << "invalid, load failed";
             return -1;
diff --git a/predictor/framework/op_repository.cpp b/predictor/framework/op_repository.cpp
index d3effdebbe2a7a2e748933b38797883a3f3a2239..2bceff15f492048ce629fabfccac5c0c50a6049b 100644
--- a/predictor/framework/op_repository.cpp
+++ b/predictor/framework/op_repository.cpp
@@ -11,35 +11,35 @@ Op* OpRepository::get_op(std::string op_type) {
     if (iter != _repository.end()) {
         op = (iter->second)->get_op();
     } else {
-        LOG(FATAL) << "Try to create unknown op[" << op_type << "]";
+        LOG(ERROR) << "Try to create unknown op[" << op_type << "]";
     }
     return op;
 }
 
 void OpRepository::return_op(Op* op) {
     if (op == NULL) {
-        LOG(FATAL) << "Try to return NULL op";
+        LOG(ERROR) << "Try to return NULL op";
         return;
     }
     ManagerMap::iterator iter = _repository.find(op->type());
     if (iter != _repository.end()) {
         iter->second->return_op(op);
     } else {
-        LOG(FATAL) << "Try to return unknown op[" << op << "], op_type["
+        LOG(ERROR) << "Try to return unknown op[" << op << "], op_type["
                 << op->type() << "].";
     }
 }
 
 void OpRepository::return_op(const std::string& op_type, Op* op) {
     if (op == NULL) {
-        LOG(FATAL) << "Try to return NULL op";
+        LOG(ERROR) << "Try to return NULL op";
         return;
     }
     ManagerMap::iterator iter = _repository.find(op_type);
     if (iter != _repository.end()) {
         iter->second->return_op(op);
     } else {
-        LOG(FATAL) << "Try to return unknown op[" << op << "], op_type["
+        LOG(ERROR) << "Try to return unknown op[" << op << "], op_type["
                 << op_type << "].";
     }
 }
diff --git a/predictor/framework/predictor_metric.h b/predictor/framework/predictor_metric.h
index 3a3afd0b33df57f0d5dc8873f9d39f41e555f4b0..ca684504e92e35d91a9cb2a9d539a52bccd11c30 100644
--- a/predictor/framework/predictor_metric.h
+++ b/predictor/framework/predictor_metric.h
@@ -210,7 +210,7 @@ public:
         if (metric != NULL) {
             **metric << latency;
         } else {
-            LOG(FATAL) << "impossible, check if you regist[" << metric_name << "].";
+            LOG(ERROR) << "impossible, check if you regist[" << metric_name << "].";
         }
     }
     
@@ -219,7 +219,7 @@ public:
         if (metric != NULL) {
             **metric << count;
         } else {
-            LOG(FATAL) << "impossible, check if you regist[" << metric_name << "].";
+            LOG(ERROR) << "impossible, check if you regist[" << metric_name << "].";
         }
     }
     
@@ -228,7 +228,7 @@ public:
         if (metric != NULL) {
             **metric << value;
         } else {
-            LOG(FATAL) << "impossible, check if you regist[" << metric_name << "].";
+            LOG(ERROR) << "impossible, check if you regist[" << metric_name << "].";
         }
     }
 
@@ -237,7 +237,7 @@ public:
         if (metric != NULL) {
             **metric << value;
         } else {
-            LOG(FATAL) << "impossible, check if you regist[" << metric_name << "].";
+            LOG(ERROR) << "impossible, check if you regist[" << metric_name << "].";
         }
     }
 
@@ -246,7 +246,7 @@ public:
         if (metric != NULL) {
             (*metric)->update_lhs(count);
         } else {
-            LOG(FATAL) << "impossible, check if you regist[" << name << "].";
+            LOG(ERROR) << "impossible, check if you regist[" << name << "].";
         }
     }
     
@@ -255,7 +255,7 @@ public:
         if (metric != NULL) {
             (*metric)->update_rhs(count);
         } else {
-            LOG(FATAL) << "impossible, check if you regist[" << name << "].";
+            LOG(ERROR) << "impossible, check if you regist[" << name << "].";
         }
     }
     
diff --git a/predictor/framework/resource.cpp b/predictor/framework/resource.cpp
index 99c239a5facf9823f86491554d243c63e80072fd..cb35b0cf456cc7e9017d5ef816e77f425e0117ee 100644
--- a/predictor/framework/resource.cpp
+++ b/predictor/framework/resource.cpp
@@ -84,7 +84,7 @@ int Resource::thread_initialize() {
 
     // infer manager
     if (FLAGS_enable_model_toolkit && InferManager::instance().thrd_initialize() != 0) {
-        LOG(FATAL) << "Failed thrd initialized infer manager"; 
+        LOG(ERROR) << "Failed thrd initialized infer manager"; 
         return -1;
     }
 
@@ -92,18 +92,18 @@ int Resource::thread_initialize() {
     if (p_dynamic_resource == NULL) {
         p_dynamic_resource = new (std::nothrow) DynamicResource;
         if (p_dynamic_resource == NULL) {
-            LOG(FATAL) << "failed to create tls DynamicResource";
+            LOG(ERROR) << "failed to create tls DynamicResource";
             return -1;
         }
         if (p_dynamic_resource->initialize() != 0) {
-            LOG(FATAL) << "DynamicResource initialize failed.";
+            LOG(ERROR) << "DynamicResource initialize failed.";
             delete p_dynamic_resource;
             p_dynamic_resource = NULL;
             return -1;
         }
 
         if (THREAD_SETSPECIFIC(_tls_bspec_key, p_dynamic_resource) != 0) {
-            LOG(FATAL) << "unable to set tls DynamicResource";
+            LOG(ERROR) << "unable to set tls DynamicResource";
             delete p_dynamic_resource;
             p_dynamic_resource = NULL;
             return -1;
@@ -128,21 +128,21 @@ int Resource::thread_clear() {
 
     // infer manager
     if (FLAGS_enable_model_toolkit && InferManager::instance().thrd_clear() != 0) {
-        LOG(FATAL) << "Failed thrd clear infer manager"; 
+        LOG(ERROR) << "Failed thrd clear infer manager"; 
         return -1;
     }
 
     DynamicResource* p_dynamic_resource = (DynamicResource*) THREAD_GETSPECIFIC(_tls_bspec_key);
     if (p_dynamic_resource == NULL) {
 #if 0
-        LOG(FATAL) << "tls dynamic resource shouldn't be null after thread_initialize"; 
+        LOG(ERROR) << "tls dynamic resource shouldn't be null after thread_initialize"; 
 #else
-        LOG(FATAL) << bthread_self() << ": tls dynamic resource shouldn't be null after thread_initialize"; 
+        LOG(ERROR) << bthread_self() << ": tls dynamic resource shouldn't be null after thread_initialize"; 
 #endif
         return -1;
     }
     if (p_dynamic_resource->clear() != 0) {
-        LOG(FATAL) << "Failed to invoke dynamic resource clear"; 
+        LOG(ERROR) << "Failed to invoke dynamic resource clear"; 
         return -1;
     }
 
@@ -153,7 +153,7 @@ int Resource::thread_clear() {
 
 int Resource::reload() {
     if (FLAGS_enable_model_toolkit && InferManager::instance().reload() != 0) {
-        LOG(FATAL) << "Failed reload infer manager"; 
+        LOG(ERROR) << "Failed reload infer manager"; 
         return -1;
     }
     
@@ -163,7 +163,7 @@ int Resource::reload() {
 
 int Resource::finalize() {
     if (FLAGS_enable_model_toolkit && InferManager::instance().proc_finalize() != 0) {
-        LOG(FATAL) << "Failed proc finalize infer manager";
+        LOG(ERROR) << "Failed proc finalize infer manager";
         return -1;
     }
 
diff --git a/predictor/framework/server.cpp b/predictor/framework/server.cpp
index e5b9970aed013acf91b5bc81d01703f841c0d07b..d62101918b0bee39eed88d0dbc6b2c18de941a94 100644
--- a/predictor/framework/server.cpp
+++ b/predictor/framework/server.cpp
@@ -32,12 +32,12 @@ int ServerManager::add_service_by_format(const std::string& format) {
   Service* service = 
       FormatServiceManager::instance().get_service(format);
   if (service == NULL) {
-    LOG(FATAL) << "Not found service by format:" << format << "!";
+    LOG(ERROR) << "Not found service by format:" << format << "!";
     return -1;
   }
 
   if (_format_services.find(format) != _format_services.end()) {
-    LOG(FATAL) << "Cannot insert duplicated service by format:" 
+    LOG(ERROR) << "Cannot insert duplicated service by format:" 
         << format << "!";
     return -1;
   }
@@ -45,7 +45,7 @@ int ServerManager::add_service_by_format(const std::string& format) {
   std::pair<boost::unordered_map<std::string, Service*>::iterator, bool> it 
       = _format_services.insert(std::make_pair(format, service));
   if (!it.second) {
-    LOG(FATAL) << "Failed insert service by format:"
+    LOG(ERROR) << "Failed insert service by format:"
         << format << "!";
     return -1;
   }
@@ -126,11 +126,11 @@ void* ServerManager::_reload_worker(void* args) {
     while (ServerManager::reload_starting()) {
         LOG(INFO) << "Begin reload framework...";
         if (Resource::instance().reload() != 0) {
-            LOG(FATAL) << "Failed reload resource!";
+            LOG(ERROR) << "Failed reload resource!";
         }   
 
         if (WorkflowManager::instance().reload() != 0) {
-            LOG(FATAL) << "Failed reload workflows"; 
+            LOG(ERROR) << "Failed reload workflows"; 
         }
 
         usleep(FLAGS_reload_interval_s * 1000000);
diff --git a/predictor/framework/service.cpp b/predictor/framework/service.cpp
index 5ec5f1ae4829414bca4a366ab7b1d15bfe247dfe..c00755a69646f538d76f3432c17909eac9b60274 100644
--- a/predictor/framework/service.cpp
+++ b/predictor/framework/service.cpp
@@ -30,7 +30,7 @@ int InferService::init(const configure::InferService& conf) {
 
     ServerManager& svr_mgr = ServerManager::instance();
     if (svr_mgr.add_service_by_format(_infer_service_format) != 0) {
-        LOG(FATAL) 
+        LOG(ERROR) 
             << "Not found service by format name:"
             << _infer_service_format << "!";
         return ERR_INTERNAL_FAILURE;
@@ -44,7 +44,7 @@ int InferService::init(const configure::InferService& conf) {
     if (_enable_map_request_to_workflow) {
         if (_request_to_workflow_map.init(
                 MAX_WORKFLOW_NUM_IN_ONE_SERVICE/*load_factor=80*/) != 0) {
-            LOG(FATAL) 
+            LOG(ERROR) 
                 << "init request to workflow map failed, bucket_count["
                 << MAX_WORKFLOW_NUM_IN_ONE_SERVICE << "].";
             return ERR_INTERNAL_FAILURE;
@@ -52,7 +52,7 @@ int InferService::init(const configure::InferService& conf) {
         int err = 0;
         _request_field_key = conf.request_field_key().c_str();
         if (_request_field_key == "") {
-            LOG(FATAL) 
+            LOG(ERROR) 
                 << "read request_field_key failed, request_field_key[" 
                 << _request_field_key << "].";
             return ERR_INTERNAL_FAILURE;
@@ -74,7 +74,7 @@ int InferService::init(const configure::InferService& conf) {
                 Workflow* workflow = 
                     WorkflowManager::instance().item(tokens[ti]);
                 if (workflow == NULL) {
-                    LOG(FATAL) 
+                    LOG(ERROR) 
                         << "Failed get workflow by name:" 
                         << tokens[ti] << ", ti: " << ti;
                     return ERR_INTERNAL_FAILURE;
@@ -85,7 +85,7 @@ int InferService::init(const configure::InferService& conf) {
 
             const std::string& request_field_value = conf.value_mapped_workflows(fi).request_field_value();
             if (_request_to_workflow_map.insert(request_field_value, workflows) == NULL) {
-                LOG(FATAL) 
+                LOG(ERROR) 
                     << "insert [" << request_field_value << "," 
                     << list << "] to _request_to_workflow_map failed.";
                 return ERR_INTERNAL_FAILURE;
@@ -100,7 +100,7 @@ int InferService::init(const configure::InferService& conf) {
             Workflow* workflow = 
                 WorkflowManager::instance().item(workflow_name);
             if (workflow == NULL) {
-                LOG(FATAL) 
+                LOG(ERROR) 
                     << "Failed get workflow by name:" 
                     << workflow_name;
                 return ERR_INTERNAL_FAILURE;
@@ -158,7 +158,7 @@ int InferService::inference(
             int errcode = _execute_workflow(workflow, request, response, debug_os);
             TRACEPRINTF("finish to execute workflow[%s]", workflow->name().c_str());
             if (errcode < 0) {
-                LOG(FATAL) << "Failed execute workflow[" << workflow->name()
+                LOG(ERROR) << "Failed execute workflow[" << workflow->name()
                         << "] in:" << name();
                 return errcode;
             }
@@ -171,7 +171,7 @@ int InferService::inference(
             int errcode = execute_one_workflow(fi, request, response, debug_os);
             TRACEPRINTF("finish to execute one workflow-%lu", fi);
             if (errcode < 0) {
-                LOG(FATAL) << "Failed execute 0-th workflow in:" << name();
+                LOG(ERROR) << "Failed execute 0-th workflow in:" << name();
                 return errcode;
             }
         }
@@ -192,7 +192,7 @@ int InferService::execute_one_workflow(
         google::protobuf::Message* response,
         butil::IOBufBuilder* debug_os) {
     if (index >= _flows.size()) {
-        LOG(FATAL) << "Faield execute workflow, index: "
+        LOG(ERROR) << "Faield execute workflow, index: "
             << index << " >= max:" << _flows.size();
         return ERR_OVERFLOW_FAILURE;
     } 
@@ -217,7 +217,7 @@ int InferService::_execute_workflow(
     // call actual inference interface
     int errcode = dv->execute(debug_os);
     if (errcode < 0) {
-        LOG(FATAL) << "Failed execute dag for workflow:"
+        LOG(ERROR) << "Failed execute dag for workflow:"
             << workflow->name();
         return errcode;
     }
@@ -226,7 +226,7 @@ int InferService::_execute_workflow(
     // create ender channel and copy
     const Channel* res_channel = dv->get_response_channel();
     if (!_merger || !_merger->merge(res_channel->message(), response)) {
-        LOG(FATAL) << "Failed merge channel res to response"; 
+        LOG(ERROR) << "Failed merge channel res to response"; 
         return ERR_INTERNAL_FAILURE;
     }
     TRACEPRINTF("finish to copy from");
diff --git a/predictor/framework/service_manager.h b/predictor/framework/service_manager.h
index 0267f8c2d98861d72c51a8c2f4bcc0aab5ef684c..bb1d200863190a8261435563791566bd7c46bd65 100644
--- a/predictor/framework/service_manager.h
+++ b/predictor/framework/service_manager.h
@@ -12,7 +12,7 @@ do {                                                            \
     int ret = ::baidu::paddle_serving::predictor::FormatServiceManager::instance().regist_service(\
               svr_name, svr);                                   \
     if (ret != 0) {                                             \
-        LOG(FATAL)                                              \
+        LOG(ERROR)                                              \
             << "Failed regist service["                         \
             << svr_name << "]" << "["                           \
             << typeid(svr).name() << "]"                        \
@@ -32,7 +32,7 @@ public:
 
     int regist_service(const std::string& svr_name, Service* svr) {
         if (_service_map.find(svr_name) != _service_map.end()) {
-            LOG(FATAL) 
+            LOG(ERROR) 
                 << "Service[" << svr_name << "]["
                 << typeid(svr).name() << "]"
                 << " already exist!";
@@ -42,7 +42,7 @@ public:
         std::pair<boost::unordered_map<std::string, Service*>::iterator, bool> ret;
         ret = _service_map.insert(std::make_pair(svr_name, svr));
         if (ret.second == false) {
-            LOG(FATAL)
+            LOG(ERROR)
                 << "Service[" << svr_name << "]["
                 << typeid(svr).name() << "]"
                 << " insert failed!";
diff --git a/predictor/framework/workflow.cpp b/predictor/framework/workflow.cpp
index 4a6c06539fac2bc24c58b8473522dbfe95ab2f2d..53f4a22181c97afb196798b0e78889e7b5e70ebf 100644
--- a/predictor/framework/workflow.cpp
+++ b/predictor/framework/workflow.cpp
@@ -24,12 +24,12 @@ DagView* Workflow::fetch_dag_view(const std::string& service_name) {
     } else if (_type == "Parallel") {
         view = butil::get_object<ParallelDagView>();
     } else {
-        LOG(FATAL) 
+        LOG(ERROR) 
             << "Unknown dag type:" << _type << "!";
         return NULL;
     }
     if (view == NULL) {
-        LOG(FATAL) << "create dag view from pool failed!";
+        LOG(ERROR) << "create dag view from pool failed!";
         return NULL;
     }
     view->init(&_dag, service_name);
@@ -44,7 +44,7 @@ void Workflow::return_dag_view(DagView* view) {
         butil::return_object<ParallelDagView>(
             dynamic_cast<ParallelDagView*>(view));
     } else {
-        LOG(FATAL) 
+        LOG(ERROR) 
             << "Unknown dag type:" << _type << "!";
         return ;
     }
diff --git a/predictor/op/op.cpp b/predictor/op/op.cpp
index 175fa61dcbbd87ba595adc0f9d3d4d0a61eb9c46..274c840951745c863cb916884c3df309886bb733 100644
--- a/predictor/op/op.cpp
+++ b/predictor/op/op.cpp
@@ -20,7 +20,7 @@ int Op::init(Bus* bus, Dag* dag, uint32_t id, const std::string& name,
 
     _timer = butil::get_object<TimerFlow>();
     if (!_timer) {
-        LOG(FATAL) << "Invalid timerflow in op:"
+        LOG(ERROR) << "Invalid timerflow in op:"
             << this->name();
         return -1;
     }
@@ -31,7 +31,7 @@ int Op::init(Bus* bus, Dag* dag, uint32_t id, const std::string& name,
 
     Channel* channel = mutable_channel();
     if (channel == NULL) {
-        LOG(FATAL) 
+        LOG(ERROR) 
             << "Failed mutable channel in op: " 
             << this->id() << ", " << this->name() << "!";
         return -1;
@@ -50,7 +50,7 @@ int Op::deinit() {
     _timer = NULL;
 
     if (release_channel() != 0) {
-        LOG(FATAL) << "Failed release channel in op:"  
+        LOG(ERROR) << "Failed release channel in op:"  
             << this->id() << ", " << this->name() << "!";
         return -1;
     }
@@ -60,12 +60,12 @@ int Op::deinit() {
 
 int Op::check_time(const char* tag) {
     if (!_timer) {
-        LOG(FATAL) << "Invalid timer in op";
+        LOG(ERROR) << "Invalid timer in op";
         return -1;
     }
 
     if (!_timer->check(tag)) {
-        LOG(FATAL) << "Failed check timer:" << tag;
+        LOG(ERROR) << "Failed check timer:" << tag;
         return -1;
     }
 
@@ -78,7 +78,7 @@ int Op::process(bool debug) {
         _timer->start();
     }
     if (!_has_init) {
-        LOG(FATAL) 
+        LOG(ERROR) 
             << "Make sure op has been init before inference";
         return ERR_INTERNAL_FAILURE;
     }
@@ -93,7 +93,7 @@ int Op::process(bool debug) {
     /*
     DagNode* node = _dag->node_by_name(this->name());
     if (node == NULL) {
-        LOG(FATAL) << "Failed get node of op:" << this->name();
+        LOG(ERROR) << "Failed get node of op:" << this->name();
         return -1;
     }
     boost::unordered_map<std::string, EdgeMode>& depends =
@@ -290,7 +290,7 @@ uint32_t Op::id() const {
 const std::string Op::debug_string() {
     const Channel* channel = get_channel();
     if (!channel) {
-        LOG(FATAL) << "Invalid channel!";
+        LOG(ERROR) << "Invalid channel!";
         return "Invalid channel in OP";
     }
     return channel->debug_string();
diff --git a/predictor/op/op.h b/predictor/op/op.h
index 0abbff18dfe734d9d2a986c75d1a9d9f759497f1..bf922a0a67c5ce51ef2c1fc3f1508709319ba490 100644
--- a/predictor/op/op.h
+++ b/predictor/op/op.h
@@ -43,7 +43,7 @@ public:
         OpChannel<T>* op_channel = 
             dynamic_cast<OpChannel<T>*>(channel);
         if (!op_channel) {
-            LOG(FATAL) << "Cannot dynamic cast channel of op:" 
+            LOG(ERROR) << "Cannot dynamic cast channel of op:" 
                 << this->name() << " to type: " << typeid(T).name();
             return NULL;
         }
@@ -63,7 +63,7 @@ public:
         const OpChannel<T>* op_channel = 
             dynamic_cast<const OpChannel<T>*>(channel);
         if (!op_channel) {
-            LOG(FATAL) << "Cannot dynamic cast channel of op:" 
+            LOG(ERROR) << "Cannot dynamic cast channel of op:" 
                 << this->name() << " to type: " << typeid(T).name();
             return NULL;
         }
@@ -197,7 +197,7 @@ public:
 
         _channel = butil::get_object<ChannelType>();
         if (!_channel) {
-            LOG(FATAL) 
+            LOG(ERROR) 
                 << "Failed mutable channel of type:"
                 << typeid(T).name();
             return NULL;
diff --git a/predictor/src/pdcodegen.cpp b/predictor/src/pdcodegen.cpp
index 85e5f3b69d65dec84db6bfb7efac4e3920cb04c4..1a3a7047b1ff00549aec262861c14a1ed20cfca3 100644
--- a/predictor/src/pdcodegen.cpp
+++ b/predictor/src/pdcodegen.cpp
@@ -524,7 +524,7 @@ private:
                 "   LOG(INFO) << \"Failed fetch response from stub handler, new it\";\n"
                 "   cur_res = response->New();\n"
                 "   if (cur_res == NULL) {\n"
-                "       LOG(FATAL) << \"Failed new response item!\";\n"
+                "       LOG(ERROR) << \"Failed new response item!\";\n"
                 "       _stub_handler->update_average(1, \"pack_fail\");\n"
                 "       return brpc::SubCall::Bad();\n"
                 "   }\n"
@@ -548,7 +548,7 @@ private:
                 "   response->MergeFrom(*sub_response);\n"
                 "   return brpc::ResponseMerger::MERGED;\n"
                 "} catch (const std::exception& e) {\n"
-                "   LOG(FATAL) << \"Merge failed.\";\n"
+                "   LOG(ERROR) << \"Merge failed.\";\n"
                 "   _stub_handler->update_average(1, \"pack_fail\");\n"
                 "   return brpc::ResponseMerger::FAIL;\n"
                 "}\n");
@@ -603,7 +603,7 @@ private:
                 printer->Print(
                         "sub_req = dynamic_cast<$req_type$*>(_stub_handler->fetch_request());\n"
                         "if (sub_req == NULL) {\n"
-                        "    LOG(FATAL) << \"failed fetch sub_req from stub.\";\n"
+                        "    LOG(ERROR) << \"failed fetch sub_req from stub.\";\n"
                         "    _stub_handler->update_average(1, \"pack_fail\");\n"
                         "    return brpc::SubCall::Bad();\n"
                         "}\n",
@@ -613,7 +613,7 @@ private:
             } else {
                 printer->Print(
                         "if (req->$field_name$_size() != total_size) {\n"
-                        "    LOG(FATAL) << \"pack field size not consistency: \"\n"
+                        "    LOG(ERROR) << \"pack field size not consistency: \"\n"
                         "               << total_size << \"!=\" << req->$field_name$_size()\n"
                         "               << \", field: $field_name$.\";\n"
                         "    _stub_handler->update_average(1, \"pack_fail\");\n"
@@ -643,7 +643,7 @@ private:
                     "if (sub_req == NULL) { // no packed items\n"
                     "   sub_req = dynamic_cast<$req_type$*>(_stub_handler->fetch_request());\n"
                     "   if (!sub_req) {\n"
-                    "       LOG(FATAL) << \"failed fetch sub_req from stub handler.\";\n"
+                    "       LOG(ERROR) << \"failed fetch sub_req from stub handler.\";\n"
                     "       _stub_handler->update_average(1, \"pack_fail\");\n"
                     "       return brpc::SubCall::Bad();\n"
                     "   }\n"
@@ -681,7 +681,7 @@ private:
         printer->Print(
                 "google::protobuf::Message* sub_res = _stub_handler->fetch_response();\n"
                 "if (sub_res == NULL) {\n"
-                "    LOG(FATAL) << \"failed create sub_res from res.\";\n"
+                "    LOG(ERROR) << \"failed create sub_res from res.\";\n"
                 "    _stub_handler->update_average(1, \"pack_fail\");\n"
                 "    return brpc::SubCall::Bad();\n"
                 "}\n"
diff --git a/predictor/src/pdserving.cpp b/predictor/src/pdserving.cpp
index dad683d81452535ee47f10beb046d28c0fefb332..36e2adeed4061c1dddb330e9e5b8dbaf273e1ad1 100644
--- a/predictor/src/pdserving.cpp
+++ b/predictor/src/pdserving.cpp
@@ -138,7 +138,7 @@ int main(int argc, char** argv) {
 
     int errcode = bthread_set_worker_startfn(pthread_worker_start_fn);
     if (errcode != 0) {
-        LOG(FATAL) << "Failed call pthread worker start function, error_code[" << errcode << "]";
+        LOG(ERROR) << "Failed call pthread worker start function, error_code[" << errcode << "]";
         return -1;
     }
     LOG(INFO) << "Succ call pthread worker start function";
diff --git a/sdk-cpp/CMakeLists.txt b/sdk-cpp/CMakeLists.txt
index 206b74b00ede4290743d8a093338de8de44cddbe..f3fccbc731fac26f136fe0eb75f2f449adfe0b12 100644
--- a/sdk-cpp/CMakeLists.txt
+++ b/sdk-cpp/CMakeLists.txt
@@ -40,3 +40,14 @@ target_include_directories(mapcnn_sparse PUBLIC
 target_link_libraries(mapcnn_sparse sdk-cpp -lpthread -lcrypto -lm -lrt -lssl
         -ldl -lz)
 
+# install
+install(TARGETS sdk-cpp
+        ARCHIVE DESTINATION ${PADDLE_SERVING_INSTALL_DIR}/lib
+        )
+
+install(TARGETS ximage
+        RUNTIME DESTINATION ${PADDLE_SERVING_INSTALL_DIR}/demo/client/bin)
+install(DIRECTORY ${CMAKE_CURRENT_LIST_DIR}/conf DESTINATION
+        ${PADDLE_SERVING_INSTALL_DIR}/demo/client/)
+install(DIRECTORY ${CMAKE_CURRENT_LIST_DIR}/data DESTINATION
+        ${PADDLE_SERVING_INSTALL_DIR}/demo/client/)
diff --git a/sdk-cpp/conf/predictors.conf b/sdk-cpp/conf/predictors.conf
deleted file mode 100644
index a3c315d12f2f74d9e51fe470a5f33198240045ca..0000000000000000000000000000000000000000
--- a/sdk-cpp/conf/predictors.conf
+++ /dev/null
@@ -1,33 +0,0 @@
-[DefaultVariantInfo]
-Tag : default
-[.Connection]
-ConnectTimeoutMilliSec: 2000
-RpcTimeoutMilliSec: 20000
-ConnectRetryCount : 2
-MaxConnectionPerHost : 100
-HedgeRequestTimeoutMilliSec: -1
-HedgeFetchRetryCount : 2
-BnsReloadIntervalSeconds : 10
-ConnectionType : pooled
-[.NamingInfo]
-ClusterFilterStrategy : Default
-LoadBalanceStrategy : la
-[.RpcParameter]
-# 0-NONE, 1-SNAPPY, 2-GZIP, 3-ZLIB, 4-LZ4
-CompressType : 0
-PackageSize : 20
-Protocol : baidu_std
-MaxChannelPerRequest : 3
-
-[@Predictor]
-name : ximage
-service_name : baidu.paddle_serving.predictor.image_classification.ImageClassifyService
-endpoint_router : WeightedRandomRender
-[.WeightedRandomRender]
-VariantWeightList : 50
-[.@VariantInfo]
-Tag : var1
-[..NamingInfo]
-Cluster : list://127.0.0.1:8010
-#Cluster : list://10.88.158.21:8010
-
diff --git a/sdk-cpp/conf/predictors.conf.bak b/sdk-cpp/conf/predictors.conf.bak
deleted file mode 100644
index 493e8e5cd64514c9e4d9a6127c5b75e404ca8b39..0000000000000000000000000000000000000000
--- a/sdk-cpp/conf/predictors.conf.bak
+++ /dev/null
@@ -1,71 +0,0 @@
-[DefaultVariantInfo]
-Tag : default
-[.Connection]
-ConnectTimeoutMilliSec: 2000
-RpcTimeoutMilliSec: 20000
-ConnectRetryCount : 2
-MaxConnectionPerHost : 100
-HedgeRequestTimeoutMilliSec: -1
-HedgeFetchRetryCount : 2
-BnsReloadIntervalSeconds : 10
-ConnectionType : pooled
-[.NamingInfo]
-ClusterFilterStrategy : Default
-LoadBalanceStrategy : la
-[.RpcParameter]
-# 0-NONE, 1-SNAPPY, 2-GZIP, 3-ZLIB, 4-LZ4
-CompressType : 0
-PackageSize : 20
-Protocol : baidu_std
-MaxChannelPerRequest : 3
-
-[@Predictor]
-name : ximage
-service_name : baidu.paddle_serving.predictor.image_classification.ImageClassifyService
-endpoint_router : WeightedRandomRender
-[.WeightedRandomRender]
-VariantWeightList : 50
-[.@VariantInfo]
-Tag : var1
-[..NamingInfo]
-Cluster : list://127.0.0.1:8010
-#Cluster : list://10.88.158.21:8010
-
-[@Predictor]
-name : dense_cnn
-service_name : baidu.paddle_serving.fluid_engine.DefaultDenseService
-endpoint_router : WeightedRandomRender
-[.WeightedRandomRender]
-VariantWeightList : 25
-[.@VariantInfo]
-Tag : var1
-[..NamingInfo]
-Cluster : list://10.194.83.21:8010
-#Cluster : bns://opera-ps-mapcnn-000-nj03.MAP.nj03
-[..Connection]
-
-[@Predictor]
-name : sparse_cnn
-service_name : baidu.paddle_serving.fluid_engine.DefaultSparseService
-endpoint_router : WeightedRandomRender
-[.WeightedRandomRender]
-VariantWeightList : 25
-[.@VariantInfo]
-Tag : var1
-[..NamingInfo]
-Cluster : list://10.194.83.21:8010
-#Cluster : bns://opera-ps-mapcnn-000-nj03.MAP.nj03
-[..Connection]
-
-[@Predictor]
-name : wasq
-service_name : baidu.infinite.map_rnn.MapRnnService
-endpoint_router : WeightedRandomRender
-[.WeightedRandomRender]
-VariantWeightList : 25
-[.@VariantInfo]
-Tag : var1
-[..NamingInfo]
-Cluster : list://127.0.0.1:8010
-#Cluster : bns://opera-ps-mapcnn-000-nj03.MAP.nj03
-[..Connection]
diff --git a/sdk-cpp/demo/mapcnn_dense.cpp b/sdk-cpp/demo/mapcnn_dense.cpp
index 4593dcf76d1d7e3bafac912d63a0876d100c34ac..70eec3b0dfa2f6e8b40512af2c255845645121ee 100644
--- a/sdk-cpp/demo/mapcnn_dense.cpp
+++ b/sdk-cpp/demo/mapcnn_dense.cpp
@@ -36,7 +36,7 @@ public:
         pthread_mutex_init(&_mutex, NULL);
         FILE* fp = fopen(file_name.c_str(), "r");
         if (!fp) {
-            LOG(FATAL) << "Failed open data file: " 
+            LOG(ERROR) << "Failed open data file: " 
                 << file_name;
             return -1;
         } 
@@ -58,7 +58,7 @@ public:
         for (size_t ri = 0; ri < buf_size; ri++) {
             Request* req = new Request();
             if (generate_one_req(*req, batch_size) != 0) {
-                LOG(FATAL) << "Failed generate req at: " << ri;
+                LOG(ERROR) << "Failed generate req at: " << ri;
                 fclose(fp);
                 return -1;
             }
@@ -229,7 +229,7 @@ void* work(void* p) {
     while (true) {
         Predictor* predictor = api->fetch_predictor("mapcnn");
         if (!predictor) {
-            LOG(FATAL) << "Failed fetch predictor: wasq"; 
+            LOG(ERROR) << "Failed fetch predictor: wasq"; 
             return NULL;
         }
         Request* req = input->next_req();
@@ -237,7 +237,7 @@ void* work(void* p) {
         timeval start;
         gettimeofday(&start, NULL);
         if (predictor->inference(req, &res) != 0) {
-            LOG(FATAL) << "failed call predictor with req:"
+            LOG(ERROR) << "failed call predictor with req:"
                 << req->ShortDebugString();
             return NULL;
         }
@@ -264,19 +264,19 @@ int main(int argc, char** argv) {
     int qps = atoi(argv[4]);
     PredictorApi api;
     if (api.create("./conf", "predictors.conf") != 0) {
-        LOG(FATAL) << "Failed create predictors api!"; 
+        LOG(ERROR) << "Failed create predictors api!"; 
         return -1;
     }
     InputData data;
     if (data.create(
                 "./data/pure_feature", req_buffer, batch_size, qps) != 0) {
-        LOG(FATAL) << "Failed create inputdata!";
+        LOG(ERROR) << "Failed create inputdata!";
         return -1;
     }
     Arg arg = {&api, &data};
     pthread_t* threads = new pthread_t[thread_num];
     if (!threads) {
-        LOG(FATAL) << "Failed create threads, num:" << thread_num;
+        LOG(ERROR) << "Failed create threads, num:" << thread_num;
         return -1;
     }
     for (int i = 0; i < thread_num; ++i) {
diff --git a/sdk-cpp/demo/maprnn_dense.cpp b/sdk-cpp/demo/maprnn_dense.cpp
index 964e0dbef2c09cdaa20279fae8e9036dde7e17b6..46f7b71e71f4d4e9b8ce06772c7f7fccbe3ba606 100644
--- a/sdk-cpp/demo/maprnn_dense.cpp
+++ b/sdk-cpp/demo/maprnn_dense.cpp
@@ -295,7 +295,7 @@ public:
         _data_record = data;
         /*FILE* fp = fopen(file_name.c_str(), "r");
         if (!fp) {
-            LOG(FATAL) << "Failed open data file: " 
+            LOG(ERROR) << "Failed open data file: " 
                 << file_name;
             return -1;
         } 
diff --git a/serving/CMakeLists.txt b/serving/CMakeLists.txt
index fe78975e1e25f18f1100846df53c5ff64913c1a4..47247a6a7c8ff87a39d4f5ca7584a1d1394233e0 100644
--- a/serving/CMakeLists.txt
+++ b/serving/CMakeLists.txt
@@ -12,3 +12,12 @@ target_link_libraries(image_class opencv_imgcodecs
         ${opencv_depend_libs} -Wl,--whole-archive fluid_cpu_engine
         -Wl,--no-whole-archive pdserving paddle_fluid ${paddle_depend_libs}
         ${MKLML_LIB} ${MKLML_IOMP_LIB} -lpthread -lcrypto -lm -lrt -lssl -ldl -lz)
+
+install(TARGETS image_class
+        RUNTIME DESTINATION ${PADDLE_SERVING_INSTALL_DIR}/demo/serving/bin)
+install(DIRECTORY ${CMAKE_CURRENT_LIST_DIR}/conf DESTINATION
+        ${PADDLE_SERVING_INSTALL_DIR}/demo/serving/)
+install(DIRECTORY ${CMAKE_CURRENT_LIST_DIR}/data DESTINATION
+        ${PADDLE_SERVING_INSTALL_DIR}/demo/serving/)
+
+
diff --git a/serving/conf/dense_dag.conf b/serving/conf/dense_dag.conf
deleted file mode 100644
index a10399b97ddf725e7c920a5dc7fe7e1b4ccfaf46..0000000000000000000000000000000000000000
--- a/serving/conf/dense_dag.conf
+++ /dev/null
@@ -1,4 +0,0 @@
-workflow_type: Sequence
-[@Node]
-name: dense_op
-type: DenseOp
diff --git a/serving/conf/imagec_dag.conf b/serving/conf/imagec_dag.conf
deleted file mode 100644
index 05bfd0156afb821b0b4f846dde34aadf9a2f1115..0000000000000000000000000000000000000000
--- a/serving/conf/imagec_dag.conf
+++ /dev/null
@@ -1,16 +0,0 @@
-workflow_type: Sequence
-[@Node]
-name: image_reader_op
-type: ReaderOp
-[@Node]
-name: image_classify_op
-type: ClassifyOp
-[.@Depend]
-name: image_reader_op
-mode: RO
-[@Node]
-name: write_json_op
-type: WriteOp
-[.@Depend]
-name: image_classify_op
-mode: RO
diff --git a/serving/conf/log.conf b/serving/conf/log.conf
deleted file mode 100644
index 28ed37d3d8fc039566477ac3f6d5670e7e2283f5..0000000000000000000000000000000000000000
--- a/serving/conf/log.conf
+++ /dev/null
@@ -1,32 +0,0 @@
-COMLOG_LEVEL : 16
-COMLOG_DEVICE_NUM : 2
-COMLOG_DEVICE0 : TRACE
-COMLOG_DEVICE1 : WARNING
- 
-TRACE_OPEN : 1
-TRACE_TYPE : FILE
-TRACE_PATH : ./log
-TRACE_NAME : pdserving.log
-TRACE_SYSLEVEL : 0
-TRACE_SELFLEVEL : NOTICE,TRACE,DEBUG
-TRACE_SIZE : 4000
-TRACE_SPLITE_TYPE : DATECUT
-TRACE_DATA_CUTTIME : 60
-TRACE_DATA_CRONOCUT : 1
-TRACE_RESERVED1 : %Y%m%d%H
-TRACE_LAYOUT : %L: %A %T %R
-TRACE_QUOTA_DAY : 2
- 
-WARNING_OPEN : 1
-WARNING_TYPE : FILE
-WARNING_PATH : ./log
-WARNING_NAME : pdserving.log.wf
-WARNING_SYSLEVEL : 0
-WARNING_SELFLEVEL : WARNING,FATAL
-WARNING_SIZE : 4000
-WARNING_SPLITE_TYPE : DATECUT
-WARNING_DATA_CUTTIME : 60
-WARNING_DATA_CRONOCUT : 1
-WARNING_RESERVED1 : %Y%m%d%H
-WARNING_LAYOUT : %L: %A %T %R
-WARNING_QUOTA_DAY : 2
diff --git a/serving/conf/model_toolkit.conf b/serving/conf/model_toolkit.conf
deleted file mode 100644
index 1bab437f55779236ef4b9eabdaf5cca6c288f59b..0000000000000000000000000000000000000000
--- a/serving/conf/model_toolkit.conf
+++ /dev/null
@@ -1,9 +0,0 @@
-[@Engine]
-Type : FLUID_CPU_NATIVE_DIR
-Name : image_classification_resnet
-ReloadableMeta: ./data/model/paddle/fluid_time_file
-ReloadableType: timestamp_ne
-ModelDataPath: ./data/model/paddle/fluid/SE_ResNeXt50_32x4d
-RuntimeThreadNum: 0
-BatchInferSize: 0
-EnableBatchAlign: 0
diff --git a/serving/conf/resource.conf b/serving/conf/resource.conf
deleted file mode 100644
index 7f63e42229ceae00151d72aa60802fa839adbf35..0000000000000000000000000000000000000000
--- a/serving/conf/resource.conf
+++ /dev/null
@@ -1,3 +0,0 @@
-# model toolkit conf
-model_toolkit_path: ./conf
-model_toolkit_file: model_toolkit.conf
diff --git a/serving/conf/service.conf b/serving/conf/service.conf
deleted file mode 100644
index f7c79b487f19d10ce6b3267fb579894653e490df..0000000000000000000000000000000000000000
--- a/serving/conf/service.conf
+++ /dev/null
@@ -1,7 +0,0 @@
-[@Service]
-name: ImageClassifyService
-@workflow: workflow1
-
-[@Service]
-name: BuiltinDenseFormatService
-@workflow: workflow2
diff --git a/serving/conf/workflow.conf b/serving/conf/workflow.conf
deleted file mode 100644
index 6c4c2e2e75e6e3614b855ab6060a79e63da16fbb..0000000000000000000000000000000000000000
--- a/serving/conf/workflow.conf
+++ /dev/null
@@ -1,9 +0,0 @@
-[@Workflow]
-name: workflow1
-path: ./conf
-file: imagec_dag.conf
-
-[@Workflow]
-name: workflow2
-path: ./conf
-file: dense_dag.conf
diff --git a/serving/op/reader_op.cpp b/serving/op/reader_op.cpp
index da9591d23db0470172b17f0f7c82212a4f5f5c4e..cbe69b484bb3ed904235fded8676585dc53090d2 100644
--- a/serving/op/reader_op.cpp
+++ b/serving/op/reader_op.cpp
@@ -46,7 +46,7 @@ int ReaderOp::inference() {
         const char* binary = ins.image_binary().c_str();
         size_t length = ins.image_length();
         if (length == 0) {
-            LOG(FATAL) << "Empty image, length is 0"; 
+            LOG(ERROR) << "Empty image, length is 0"; 
             return -1;
         }