wait before profiling newly created containers

gprofiler/containers_client.py

@@ -13,6 +13,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 #
+import time
 from typing import Dict, List, Optional, Set
 
 from granulate_utils.containers.client import ContainersClient
@@ -25,6 +26,8 @@
 
 logger = get_logger_adapter(__name__)
 
+NEWLY_CREATED_CONTAINER_AGE_IN_SECONDS = 3
+
 
 class ContainerNamesClient:
     def __init__(self) -> None:
@@ -73,9 +76,13 @@ def get_container_name(self, pid: int) -> str:
     def _safely_get_process_container_name(self, pid: int) -> Optional[str]:
         try:
             try:
-                container_id = get_process_container_id(Process(pid))
+                process = Process(pid)
+                container_id = get_process_container_id(process)
                 if container_id is None:
                     return None
+                # If the container is newly created, we wait a bit to make sure the container is available
+                if time.time() - process.create_time() <= NEWLY_CREATED_CONTAINER_AGE_IN_SECONDS:
+                    time.sleep(2)
             except NoSuchProcess:
                 return None
             return self._get_container_name(container_id)

gprofiler/profilers/java.py

@@ -1229,6 +1229,7 @@ def _check_async_profiler_loaded(self, process: Process) -> bool:
     def _profile_process(self, process: Process, duration: int, spawned: bool) -> ProfileData:
         comm = process_comm(process)
         exe = process_exe(process)
+        container_name = self._profiler_state.get_container_name(process.pid)
         java_version_output: Optional[str] = get_java_version_logged(process, self._profiler_state.stop_event)
 
         if self._enabled_proc_events_java:
@@ -1258,7 +1259,6 @@ def _profile_process(self, process: Process, duration: int, spawned: bool) -> Pr
             self._profiled_pids.add(process.pid)
 
         logger.info(f"Profiling{' spawned' if spawned else ''} process {process.pid} with async-profiler")
-        container_name = self._profiler_state.get_container_name(process.pid)
         app_metadata = self._metadata.get_metadata(process)
         appid = application_identifiers.get_java_app_id(process, self._collect_spark_app_name)
 

gprofiler/profilers/php.py

@@ -210,10 +210,11 @@ def extract_metadata_section(re_expr: Pattern, metadata_line: str) -> str:
             if profiler_state.processes_to_profile is not None:
                 if pid not in [process.pid for process in profiler_state.processes_to_profile]:
                     continue
+            container_name = profiler_state.get_container_name(pid)
             # TODO: appid & app metadata for php!
             appid = None
             app_metadata = None
-            profiles[pid] = ProfileData(results[pid], appid, app_metadata, profiler_state.get_container_name(pid))
+            profiles[pid] = ProfileData(results[pid], appid, app_metadata, container_name)
 
         return profiles
 

gprofiler/profilers/python_ebpf.py

@@ -262,9 +262,9 @@ def snapshot(self) -> ProcessToProfileData:
                 if self._profiler_state.processes_to_profile is not None:
                     if process not in self._profiler_state.processes_to_profile:
                         continue
+                container_name = self._profiler_state.get_container_name(pid)
                 appid = application_identifiers.get_python_app_id(process)
                 app_metadata = self._metadata.get_metadata(process)
-                container_name = self._profiler_state.get_container_name(pid)
             except NoSuchProcess:
                 appid = None
                 app_metadata = None

gprofiler/utils/__init__.py

@@ -90,45 +90,7 @@ def is_root() -> bool:
         return os.geteuid() == 0
 
 
-libc: Optional[ctypes.CDLL] = None
-
-
-def prctl(*argv: Any) -> int:
-    global libc
-    if libc is None:
-        libc = ctypes.CDLL("libc.so.6", use_errno=True)
-    return cast(int, libc.prctl(*argv))
-
-
-PR_SET_PDEATHSIG = 1
-
-
-def set_child_termination_on_parent_death() -> int:
-    ret = prctl(PR_SET_PDEATHSIG, signal.SIGTERM)
-    if ret != 0:
-        errno = ctypes.get_errno()
-        logger.warning(
-            f"Failed to set parent-death signal on child process. errno: {errno}, strerror: {os.strerror(errno)}"
-        )
-    return ret
-
-
-def wrap_callbacks(callbacks: List[Callable]) -> Callable:
-    # Expects array of callback.
-    # Returns one callback that call each one of them, and returns the retval of last callback
-    def wrapper() -> Any:
-        ret = None
-        for cb in callbacks:
-            ret = cb()
-
-        return ret
-
-    return wrapper
-
-
-def start_process(
-    cmd: Union[str, List[str]], via_staticx: bool = False, term_on_parent_death: bool = True, **kwargs: Any
-) -> Popen:
+def start_process(cmd: Union[str, List[str]], via_staticx: bool = False, **kwargs: Any) -> Popen:
     if isinstance(cmd, str):
         cmd = [cmd]
 
@@ -150,19 +112,12 @@ def start_process(
             env = env if env is not None else os.environ.copy()
             env.update({"LD_LIBRARY_PATH": ""})
 
-    if is_windows():
-        cur_preexec_fn = None  # preexec_fn is not supported on Windows platforms. subprocess.py reports this.
-    else:
-        cur_preexec_fn = kwargs.pop("preexec_fn", os.setpgrp)
-        if term_on_parent_death:
-            cur_preexec_fn = wrap_callbacks([set_child_termination_on_parent_death, cur_preexec_fn])
-
     popen = Popen(
         cmd,
         stdout=kwargs.pop("stdout", subprocess.PIPE),
         stderr=kwargs.pop("stderr", subprocess.PIPE),
         stdin=subprocess.PIPE,
-        preexec_fn=cur_preexec_fn,
+        start_new_session=is_linux(),  # TODO: change to "process_group" after upgrade to Python 3.11+
         env=env,
         **kwargs,
     )