indexify 0.4.10__tar.gz → 0.4.12__tar.gz

{indexify-0.4.10 → indexify-0.4.12}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: indexify
-Version: 0.4.10
+Version: 0.4.12
 Summary: Open Source Indexify components and helper tools
 Home-page: https://github.com/tensorlakeai/indexify
 License: Apache 2.0
@@ -17,7 +17,7 @@ Requires-Dist: aiohttp (>=3.11.0,<4.0.0)
 Requires-Dist: boto3 (>=1.37.30,<2.0.0)
 Requires-Dist: prometheus-client (>=0.21.1,<0.22.0)
 Requires-Dist: psutil (>=7.0.0,<8.0.0)
-Requires-Dist: tensorlake (==0.2.7)
+Requires-Dist: tensorlake (==0.2.8)
 Project-URL: Repository, https://github.com/tensorlakeai/indexify
 Description-Content-Type: text/markdown
 

{indexify-0.4.10 → indexify-0.4.12}/pyproject.toml

@@ -1,7 +1,7 @@
 [tool.poetry]
 name = "indexify"
 # Incremented if any of the components provided in this packages are updated.
-version = "0.4.10"
+version = "0.4.12"
 description = "Open Source Indexify components and helper tools"
 authors = ["Tensorlake Inc. <support@tensorlake.ai>"]
 license = "Apache 2.0"
@@ -25,7 +25,7 @@ prometheus-client = "^0.21.1"
 psutil = "^7.0.0"
 # Adds function-executor binary, utils lib, sdk used in indexify-cli commands.
 # We need to specify the tensorlake version exactly because pip install doesn't respect poetry.lock files.
-tensorlake = "0.2.7"
+tensorlake = "0.2.8"
 # Uncomment the next line to use local tensorlake package (only for development!)
 # tensorlake = { path = "../tensorlake", develop = true }
 # pydantic is provided by tensorlake

{indexify-0.4.10 → indexify-0.4.12}/src/indexify/executor/function_executor_controller/events.py

@@ -1,5 +1,5 @@
 from enum import Enum
-from typing import Optional
+from typing import List, Optional
 
 from indexify.executor.function_executor.function_executor import (
     FunctionExecutor,
@@ -12,7 +12,7 @@ from .task_info import TaskInfo
 
 class EventType(Enum):
     FUNCTION_EXECUTOR_CREATED = 1
-    FUNCTION_EXECUTOR_DESTROYED = 2
+    FUNCTION_EXECUTOR_TERMINATED = 2
     SHUTDOWN_INITIATED = 3
     TASK_PREPARATION_FINISHED = 4
     SCHEDULE_TASK_EXECUTION = 5
@@ -50,23 +50,32 @@ class FunctionExecutorCreated(BaseEvent):
         self.output: FunctionExecutorStartupOutput = output
 
 
-class FunctionExecutorDestroyed(BaseEvent):
+class FunctionExecutorTerminated(BaseEvent):
     """
-    Event indicating that Function Executor has been destroyed.
+    Event indicating that Function Executor has been terminated (destroyed).
     """
 
     def __init__(
-        self, is_success: bool, termination_reason: FunctionExecutorTerminationReason
+        self,
+        is_success: bool,
+        fe_termination_reason: FunctionExecutorTerminationReason,
+        allocation_ids_caused_termination: List[str],
     ):
-        super().__init__(EventType.FUNCTION_EXECUTOR_DESTROYED)
+        super().__init__(EventType.FUNCTION_EXECUTOR_TERMINATED)
         self.is_success: bool = is_success
-        self.termination_reason: FunctionExecutorTerminationReason = termination_reason
+        self.fe_termination_reason: FunctionExecutorTerminationReason = (
+            fe_termination_reason
+        )
+        self.allocation_ids_caused_termination: List[str] = (
+            allocation_ids_caused_termination
+        )
 
     def __str__(self) -> str:
         return (
             f"Event(type={self.event_type.name}, "
             f"is_success={self.is_success}, "
-            f"termination_reason={FunctionExecutorTerminationReason.Name(self.termination_reason)})"
+            f"fe_termination_reason={FunctionExecutorTerminationReason.Name(self.fe_termination_reason)}, "
+            f"allocation_ids_caused_termination={self.allocation_ids_caused_termination})"
         )
 
 
@@ -75,15 +84,8 @@ class ShutdownInitiated(BaseEvent):
     Event indicating that Function Executor shutdown has been initiated.
     """
 
-    def __init__(self, termination_reason: FunctionExecutorTerminationReason):
+    def __init__(self):
         super().__init__(EventType.SHUTDOWN_INITIATED)
-        self.termination_reason: FunctionExecutorTerminationReason = termination_reason
-
-    def __str__(self) -> str:
-        return (
-            f"Event(type={self.event_type.name}, "
-            f"termination_reason={FunctionExecutorTerminationReason.Name(self.termination_reason)})"
-        )
 
 
 class TaskPreparationFinished(BaseEvent):

{indexify-0.4.10 → indexify-0.4.12}/src/indexify/executor/function_executor_controller/function_executor_controller.py

@@ -1,6 +1,7 @@
 import asyncio
 import time
 from collections.abc import Coroutine
+from enum import Enum
 from pathlib import Path
 from typing import Any, Dict, List, Optional
 
@@ -28,12 +29,11 @@ from .debug_event_loop import (
     debug_print_events,
     debug_print_processing_event,
 )
-from .destroy_function_executor import destroy_function_executor
 from .events import (
     BaseEvent,
     EventType,
     FunctionExecutorCreated,
-    FunctionExecutorDestroyed,
+    FunctionExecutorTerminated,
     ScheduleTaskExecution,
     ShutdownInitiated,
     TaskExecutionFinished,
@@ -43,12 +43,14 @@ from .events import (
 from .function_executor_startup_output import FunctionExecutorStartupOutput
 from .loggers import function_executor_logger, task_allocation_logger
 from .metrics.function_executor_controller import (
-    METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_PENDING,
-    METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_RUNNING,
-    METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_TERMINATED,
-    METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_UNKNOWN,
+    METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_NOT_STARTED,
+    METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_RUNNING,
+    METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_STARTING_UP,
+    METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_TERMINATED,
+    METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_TERMINATING,
+    METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_UNKNOWN,
     metric_control_loop_handle_event_latency,
-    metric_function_executors_with_status,
+    metric_function_executors_with_state,
     metric_runnable_tasks,
     metric_runnable_tasks_per_function_name,
     metric_schedule_task_latency,
@@ -58,9 +60,20 @@ from .prepare_task import prepare_task
 from .run_task import run_task_on_function_executor
 from .task_info import TaskInfo
 from .task_output import TaskOutput
+from .terminate_function_executor import terminate_function_executor
 from .upload_task_output import upload_task_output
 
 
+# Actual FE controller states, they are a bit different from statuses reported to the Server.
+# All the valid state transitions are forward only (can skip multiple states in a row).
+class _FE_CONTROLLER_STATE(Enum):
+    NOT_STARTED = 1
+    STARTING_UP = 2
+    RUNNING = 3
+    TERMINATING = 4
+    TERMINATED = 5
+
+
 class FunctionExecutorController:
     def __init__(
         self,
@@ -94,19 +107,17 @@ class FunctionExecutorController:
         self._logger: Any = function_executor_logger(
             function_executor_description, logger.bind(module=__name__)
         )
-        # Mutable state. No lock needed as it's modified by async tasks running in
-        # the same event loop.
+        self._destroy_lock: asyncio.Lock = asyncio.Lock()
+        # Mutable state. No lock needed as it's modified by async tasks running in the same event loop.
         self._fe: Optional[FunctionExecutor] = None
-        self._fe_termination_reason: FunctionExecutorTerminationReason = (
-            None  # Optional
-        )
-        # FE Status reported to Server.
-        self._status: FunctionExecutorStatus = (
-            FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_UNKNOWN
-        )
-        metric_function_executors_with_status.labels(
-            status=_to_fe_status_metric_label(self._status, self._logger)
+        self._internal_state = _FE_CONTROLLER_STATE.NOT_STARTED
+        metric_function_executors_with_state.labels(
+            state=_to_fe_state_metric_label(self._internal_state, self._logger)
         ).inc()
+        self._reported_state: FunctionExecutorState = FunctionExecutorState(
+            description=function_executor_description,
+            status=FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_UNKNOWN,
+        )
         # Ordered list of events to be processed by the control loop.
         self._events: List[BaseEvent] = []
         # Asyncio event used to notify the control loop that there are new events to process.
@@ -124,13 +135,6 @@ class FunctionExecutorController:
     def function_executor_id(self) -> str:
         return self._fe_description.id
 
-    def status(self) -> FunctionExecutorStatus:
-        """Returns the current status of the Function Executor.
-
-        Not blocking.
-        """
-        return self._status
-
     def add_task_allocation(self, task_allocation: TaskAllocation) -> None:
         """Adds a task to the Function Executor.
 
@@ -205,9 +209,10 @@ class FunctionExecutorController:
         """Starts up the Function Executor and prepares it to run tasks.
 
         Not blocking. Never raises exceptions."""
-        if self._control_loop_aio_task is not None:
+        if self._internal_state != _FE_CONTROLLER_STATE.NOT_STARTED:
             self._logger.warning(
-                "ignoring startup call as the Function Executor is already started"
+                "function executor state is not NOT_STARTED, ignoring startup call",
+                internal_state=self._internal_state.name,
             )
             return
 
@@ -215,7 +220,13 @@ class FunctionExecutorController:
             self._control_loop(),
             name="function executor control loop",
         )
-        self._set_status(FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_PENDING)
+        self._update_internal_state(_FE_CONTROLLER_STATE.STARTING_UP)
+        self._update_reported_state(
+            FunctionExecutorState(
+                description=self._fe_description,
+                status=FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_PENDING,
+            )
+        )
         next_aio = create_function_executor(
             function_executor_description=self._fe_description,
             function_executor_server_factory=self._fe_server_factory,
@@ -237,17 +248,13 @@ class FunctionExecutorController:
             ),
         )
 
-    async def shutdown(
-        self, termination_reason: FunctionExecutorTerminationReason
-    ) -> None:
+    async def shutdown(self) -> None:
         """Shutsdown the Function Executor and frees all of its resources.
 
-        All the tasks are reported as failed with FE Terminated failure code.
+        No task outcomes and outputs are getting reported to Server after this call.
         Doesn't raise any exceptions. Blocks until the shutdown is complete.
         """
-        self._add_event(
-            ShutdownInitiated(termination_reason=termination_reason), source="shutdown"
-        )
+        self._add_event(ShutdownInitiated(), source="shutdown")
         try:
             await self._control_loop_aio_task
         except asyncio.CancelledError:
@@ -259,51 +266,49 @@ class FunctionExecutorController:
             )
         self._logger.info("function executor controller shutdown finished")
 
-    def _set_status(
+    def _update_internal_state(self, new_state: _FE_CONTROLLER_STATE) -> None:
+        """Updates the internal state of the Function Executor Controller.
+
+        Not blocking. Never raises exceptions."""
+        old_state: _FE_CONTROLLER_STATE = self._internal_state
+        self._internal_state = new_state
+
+        self._logger.info(
+            "function executor internal state changed",
+            old_state=old_state.name,
+            new_state=new_state.name,
+        )
+
+        metric_function_executors_with_state.labels(
+            state=_to_fe_state_metric_label(old_state, self._logger)
+        ).dec()
+        metric_function_executors_with_state.labels(
+            state=_to_fe_state_metric_label(new_state, self._logger)
+        ).inc()
+
+    def _update_reported_state(
         self,
-        status: FunctionExecutorStatus,
+        new_state: FunctionExecutorState,
     ) -> None:
-        """Sets Function Executor status and reports it to the Server.
+        """Sets new Function Executor state and reports it to the Server.
 
         Not blocking. Never raises exceptions."""
-        old_status: FunctionExecutorStatus = self._status
-        new_status: FunctionExecutorStatus = status
-        self._status: FunctionExecutorStatus = new_status
+        old_state: FunctionExecutorState = self._reported_state
+        self._reported_state = new_state
 
         self._logger.info(
-            "function executor status changed",
-            old_status=FunctionExecutorStatus.Name(old_status),
-            new_status=FunctionExecutorStatus.Name(new_status),
+            "function executor grpc status changed",
+            old_status=FunctionExecutorStatus.Name(old_state.status),
+            new_status=FunctionExecutorStatus.Name(new_state.status),
             termination_reason=_termination_reason_to_short_name(
-                self._fe_termination_reason
+                new_state.termination_reason
             ),
         )
-        metric_function_executors_with_status.labels(
-            status=_to_fe_status_metric_label(old_status, self._logger)
-        ).dec()
-        metric_function_executors_with_status.labels(
-            status=_to_fe_status_metric_label(new_status, self._logger)
-        ).inc()
 
-        self._state_reporter.update_function_executor_state(self._current_state())
+        self._state_reporter.update_function_executor_state(new_state)
         # Report the status change to the Server asap to reduce latency in the system.
         self._state_reporter.schedule_state_report()
 
-    def _current_state(self) -> FunctionExecutorState:
-        """Returns the current state of the Function Executor.
-
-        Not blocking. Never raises exceptions.
-        """
-        termination_reason: Optional[FunctionExecutorTerminationReason] = None
-        if self._fe_termination_reason is not None:
-            termination_reason = self._fe_termination_reason
-
-        return FunctionExecutorState(
-            description=self._fe_description,
-            status=self._status,
-            termination_reason=termination_reason,
-        )
-
     async def _control_loop(self) -> None:
         """Runs control loop that coordinates all the work done by the Function Executor.
 
@@ -341,8 +346,8 @@ class FunctionExecutorController:
         """
         if event.event_type == EventType.FUNCTION_EXECUTOR_CREATED:
             return self._handle_event_function_executor_created(event)
-        elif event.event_type == EventType.FUNCTION_EXECUTOR_DESTROYED:
-            return self._handle_event_function_executor_destroyed(event)
+        elif event.event_type == EventType.FUNCTION_EXECUTOR_TERMINATED:
+            return self._handle_event_function_executor_terminated(event)
         elif event.event_type == EventType.TASK_PREPARATION_FINISHED:
             return self._handle_event_task_preparation_finished(event)
         elif event.event_type == EventType.SCHEDULE_TASK_EXECUTION:
@@ -454,13 +459,25 @@ class FunctionExecutorController:
         self._state_reporter.schedule_state_report()
 
         if event.function_executor is None:
-            self._destroy_function_executor_before_termination(
-                event.output.termination_reason
+            # Server needs to increment attempts counter for all the tasks that were pending while FE was starting up.
+            # This prevents infinite retries if FEs consistently fail to start up.
+            self._start_termination(
+                fe_termination_reason=event.output.termination_reason,
+                allocation_ids_caused_termination=[
+                    task_info.allocation.allocation_id
+                    for task_info in self._tasks.values()
+                ],
             )
             return
 
         self._fe = event.function_executor
-        self._set_status(FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_RUNNING)
+        self._update_internal_state(_FE_CONTROLLER_STATE.RUNNING)
+        self._update_reported_state(
+            FunctionExecutorState(
+                description=self._fe_description,
+                status=FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_RUNNING,
+            )
+        )
         # Health checker starts after FE creation and gets automatically stopped on FE destroy.
         self._fe.health_checker().start(self._health_check_failed_callback)
         self._add_event(
@@ -468,20 +485,30 @@ class FunctionExecutorController:
             source="_handle_event_function_executor_created",
         )
 
-    def _handle_event_function_executor_destroyed(
-        self, event: FunctionExecutorDestroyed
+    def _handle_event_function_executor_terminated(
+        self, event: FunctionExecutorTerminated
     ) -> None:
-        """Handles the Function Executor destroy finished event.
+        """Handles the Function Executor terminated event.
 
         Doesn't raise any exceptions. Doesn't block.
         """
         if not event.is_success:
             self._logger.error(
-                "Function Executor destroy failed unexpectedly, this should never happen",
+                "Function Executor termination failed unexpectedly, this should never happen",
+            )
+
+        self._fe = None
+        # Set reported status only after the FE got destroyed because Server assumes that all FE resources are freed when the status changes.
+        self._update_reported_state(
+            FunctionExecutorState(
+                description=self._fe_description,
+                status=FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_TERMINATED,
+                termination_reason=event.fe_termination_reason,
+                allocation_ids_caused_termination=event.allocation_ids_caused_termination,
             )
-        # Set the status only after the FE got destroyed because Server assumes that all FE resources are freed when the status changes.
-        self._fe_termination_reason = event.termination_reason
-        self._set_status(FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_TERMINATED)
+        )
+        self._update_internal_state(_FE_CONTROLLER_STATE.TERMINATED)
+
         # Invoke the scheduler so it can fail runnable tasks with FE Terminated error.
         self._add_event(
             ScheduleTaskExecution(),
@@ -493,8 +520,14 @@ class FunctionExecutorController:
             "Function Executor health check failed, terminating Function Executor",
             reason=result.reason,
         )
-        self._destroy_function_executor_before_termination(
-            termination_reason=FunctionExecutorTerminationReason.FUNCTION_EXECUTOR_TERMINATION_REASON_UNHEALTHY
+
+        self._start_termination(
+            fe_termination_reason=FunctionExecutorTerminationReason.FUNCTION_EXECUTOR_TERMINATION_REASON_UNHEALTHY,
+            allocation_ids_caused_termination=(
+                []
+                if self._running_task is None
+                else [self._running_task.allocation.allocation_id]
+            ),
         )
 
     def _handle_event_task_preparation_finished(
@@ -532,14 +565,15 @@ class FunctionExecutorController:
         if len(self._runnable_tasks) == 0:
             return
 
-        if self._status not in [
-            FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_RUNNING,
-            FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_TERMINATED,
+        if self._internal_state not in [
+            _FE_CONTROLLER_STATE.RUNNING,
+            _FE_CONTROLLER_STATE.TERMINATING,
+            _FE_CONTROLLER_STATE.TERMINATED,
         ]:
-            return  # Can't progress pending task with the current status.
+            return  # Can't progress runnable tasks in the current state.
 
         if (
-            self._status == FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_RUNNING
+            self._internal_state == _FE_CONTROLLER_STATE.RUNNING
             and self._running_task is not None
         ):
             return
@@ -555,12 +589,15 @@ class FunctionExecutorController:
         if task_info.is_cancelled:
             task_info.output = TaskOutput.task_cancelled(task_info.allocation)
             self._start_task_output_upload(task_info)
-        elif self._status == FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_TERMINATED:
+        elif self._internal_state in [
+            _FE_CONTROLLER_STATE.TERMINATING,
+            _FE_CONTROLLER_STATE.TERMINATED,
+        ]:
             task_info.output = TaskOutput.function_executor_terminated(
                 task_info.allocation
             )
             self._start_task_output_upload(task_info)
-        elif self._status == FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_RUNNING:
+        elif self._internal_state == _FE_CONTROLLER_STATE.RUNNING:
             self._running_task = task_info
             next_aio = run_task_on_function_executor(
                 task_info=task_info,
@@ -603,8 +640,11 @@ class FunctionExecutorController:
                 ScheduleTaskExecution(), source="_handle_event_task_execution_finished"
             )
         else:
-            self._destroy_function_executor_before_termination(
-                termination_reason=event.function_executor_termination_reason
+            self._start_termination(
+                fe_termination_reason=event.function_executor_termination_reason,
+                allocation_ids_caused_termination=[
+                    event.task_info.allocation.allocation_id
+                ],
             )
 
         # Ignore is_cancelled because cancelling a task still involves uploading its output.
@@ -660,23 +700,37 @@ class FunctionExecutorController:
         )
         self._state_reporter.schedule_state_report()
 
-    def _destroy_function_executor_before_termination(
-        self, termination_reason: FunctionExecutorTerminationReason
+    def _start_termination(
+        self,
+        fe_termination_reason: FunctionExecutorTerminationReason,
+        allocation_ids_caused_termination: List[str],
     ) -> None:
-        """Destroys the Function Executor and frees all its resources to prepare for transitioning to the TERMINATED state.
+        """Starts termination of the Function Executor if it's not started yet.
 
         Doesn't raise any exceptions. Doesn't block.
         """
-        next_aio = destroy_function_executor(
+        if self._internal_state in [
+            _FE_CONTROLLER_STATE.TERMINATING,
+            _FE_CONTROLLER_STATE.TERMINATED,
+        ]:
+            # _start_termination() can be called multiple times, e.g. by each failed task alloc
+            # when the FE is unhealthy. Dedup the calls to keep state machine consistent.
+            return
+
+        self._update_internal_state(_FE_CONTROLLER_STATE.TERMINATING)
+        next_aio = terminate_function_executor(
             function_executor=self._fe,
-            termination_reason=termination_reason,
+            lock=self._destroy_lock,
+            fe_termination_reason=fe_termination_reason,
+            allocation_ids_caused_termination=allocation_ids_caused_termination,
             logger=self._logger,
         )
-        self._fe = None
         self._spawn_aio_for_fe(
             aio=next_aio,
-            on_exception=FunctionExecutorDestroyed(
-                is_success=False, termination_reason=termination_reason
+            on_exception=FunctionExecutorTerminated(
+                is_success=False,
+                fe_termination_reason=fe_termination_reason,
+                allocation_ids_caused_termination=allocation_ids_caused_termination,
             ),
         )
 
@@ -716,16 +770,18 @@ class FunctionExecutorController:
                 # BaseException includes asyncio.CancelledError which is always raised here.
                 pass
 
-        if self._status != FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_TERMINATED:
-            self._handle_event_function_executor_destroyed(
-                await destroy_function_executor(
-                    function_executor=self._fe,
-                    termination_reason=event.termination_reason,
-                    logger=self._logger,
+        # Makes sure we don't run fe destroy concurrently with an event loop task.
+        # FE destroy uses asyncio.to_thread() calls so it doesn't get cancelled with all the tasks above.
+        async with self._destroy_lock:
+            if self._fe is not None:
+                self._logger.info(
+                    "destroying function executor",
                 )
-            )
-        metric_function_executors_with_status.labels(
-            status=_to_fe_status_metric_label(self._status, self._logger)
+                await self._fe.destroy()
+
+        # Cleanup the metric from this FE.
+        metric_function_executors_with_state.labels(
+            state=_to_fe_state_metric_label(self._internal_state, self._logger)
         ).dec()
 
         self._state_reporter.remove_function_executor_state(self.function_executor_id())
@@ -735,21 +791,23 @@ class FunctionExecutorController:
         debug_print_events(events=self._events, logger=self._logger)
 
 
-def _to_fe_status_metric_label(status: FunctionExecutorStatus, logger: Any) -> str:
-    if status == FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_UNKNOWN:
-        return METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_UNKNOWN
-    elif status == FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_PENDING:
-        return METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_PENDING
-    elif status == FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_RUNNING:
-        return METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_RUNNING
-    elif status == FunctionExecutorStatus.FUNCTION_EXECUTOR_STATUS_TERMINATED:
-        return METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_TERMINATED
+def _to_fe_state_metric_label(state: _FE_CONTROLLER_STATE, logger: Any) -> str:
+    if state == _FE_CONTROLLER_STATE.NOT_STARTED:
+        return METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_NOT_STARTED
+    elif state == _FE_CONTROLLER_STATE.STARTING_UP:
+        return METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_STARTING_UP
+    elif state == _FE_CONTROLLER_STATE.RUNNING:
+        return METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_RUNNING
+    elif state == _FE_CONTROLLER_STATE.TERMINATING:
+        return METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_TERMINATING
+    elif state == _FE_CONTROLLER_STATE.TERMINATED:
+        return METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_TERMINATED
     else:
         logger.error(
-            "unexpected Function Executor status",
-            status=FunctionExecutorStatus.Name(status),
+            "unexpected Function Executor internal state",
+            state=state.name,
         )
-        return METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_UNKNOWN
+        return METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_UNKNOWN
 
 
 _termination_reason_to_short_name_map = {
@@ -757,8 +815,6 @@ _termination_reason_to_short_name_map = {
     FunctionExecutorTerminationReason.FUNCTION_EXECUTOR_TERMINATION_REASON_STARTUP_FAILED_INTERNAL_ERROR: "STARTUP_FAILED_INTERNAL_ERROR",
     FunctionExecutorTerminationReason.FUNCTION_EXECUTOR_TERMINATION_REASON_STARTUP_FAILED_FUNCTION_ERROR: "STARTUP_FAILED_FUNCTION_ERROR",
     FunctionExecutorTerminationReason.FUNCTION_EXECUTOR_TERMINATION_REASON_STARTUP_FAILED_FUNCTION_TIMEOUT: "STARTUP_FAILED_FUNCTION_TIMEOUT",
-    FunctionExecutorTerminationReason.FUNCTION_EXECUTOR_TERMINATION_REASON_EXECUTOR_SHUTDOWN: "EXECUTOR_SHUTDOWN",
-    FunctionExecutorTerminationReason.FUNCTION_EXECUTOR_TERMINATION_REASON_REMOVED_FROM_DESIRED_STATE: "REMOVED_FROM_DESIRED_STATE",
     FunctionExecutorTerminationReason.FUNCTION_EXECUTOR_TERMINATION_REASON_UNHEALTHY: "UNHEALTHY",
     FunctionExecutorTerminationReason.FUNCTION_EXECUTOR_TERMINATION_REASON_INTERNAL_ERROR: "INTERNAL_ERROR",
     FunctionExecutorTerminationReason.FUNCTION_EXECUTOR_TERMINATION_REASON_FUNCTION_TIMEOUT: "FUNCTION_TIMEOUT",

{indexify-0.4.10 → indexify-0.4.12}/src/indexify/executor/function_executor_controller/metrics/function_executor_controller.py

@@ -34,27 +34,34 @@ metric_runnable_tasks_per_function_name: prometheus_client.Gauge = (
     )
 )
 
-metric_function_executors_with_status: prometheus_client.Gauge = (
-    prometheus_client.Gauge(
-        "function_executors_with_status",
-        "Number of Function Executors with a particular status",
-        ["status"],
-    )
+metric_function_executors_with_state: prometheus_client.Gauge = prometheus_client.Gauge(
+    "function_executors_with_state",
+    "Number of Function Executors with a particular internal state",
+    ["state"],
 )
-METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_UNKNOWN = "unknown"
-METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_PENDING = "pending"
-METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_RUNNING = "running"
-METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_TERMINATED = "terminated"
+METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_UNKNOWN = "unknown"
+METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_NOT_STARTED = "not_started"
+METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_STARTING_UP = "starting_up"
+METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_RUNNING = "running"
+METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_TERMINATING = "terminating"
+METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_TERMINATED = "terminated"
+
 
-metric_function_executors_with_status.labels(
-    status=METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_UNKNOWN
+metric_function_executors_with_state.labels(
+    state=METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_UNKNOWN
+)
+metric_function_executors_with_state.labels(
+    state=METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_NOT_STARTED
+)
+metric_function_executors_with_state.labels(
+    state=METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_STARTING_UP
 )
-metric_function_executors_with_status.labels(
-    status=METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_PENDING
+metric_function_executors_with_state.labels(
+    state=METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_RUNNING
 )
-metric_function_executors_with_status.labels(
-    status=METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_RUNNING
+metric_function_executors_with_state.labels(
+    state=METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_TERMINATING
 )
-metric_function_executors_with_status.labels(
-    status=METRIC_FUNCTION_EXECUTORS_WITH_STATUS_LABEL_TERMINATED
+metric_function_executors_with_state.labels(
+    state=METRIC_FUNCTION_EXECUTORS_WITH_STATE_LABEL_TERMINATED
 )