skypilot-nightly 1.0.0.dev2024053101__py3-none-any.whl → 1.0.0.dev2025022801__py3-none-any.whl

sky/serve/autoscalers.py

@@ -42,11 +42,10 @@ class AutoscalerDecision:
     # TODO(MaoZiming): Add a doc to elaborate on autoscaling policies.
     def __init__(self, operator: AutoscalerDecisionOperator,
                  target: Union[Optional[Dict[str, Any]], int]):
-
-        assert (operator == AutoscalerDecisionOperator.SCALE_UP and
-                (target is None or isinstance(target, dict))) or (
-                    operator == AutoscalerDecisionOperator.SCALE_DOWN and
-                    isinstance(target, int))
+        if operator == AutoscalerDecisionOperator.SCALE_UP:
+            assert (target is None or isinstance(target, dict))
+        else:
+            assert isinstance(target, int)
         self.operator = operator
         self.target = target
 
@@ -54,9 +53,70 @@ class AutoscalerDecision:
         return f'AutoscalerDecision({self.operator}, {self.target})'
 
 
+def _generate_scale_up_decisions(
+        num: int, target: Optional[Dict[str, Any]]) -> List[AutoscalerDecision]:
+    return [
+        AutoscalerDecision(AutoscalerDecisionOperator.SCALE_UP, target)
+        for _ in range(num)
+    ]
+
+
+def _generate_scale_down_decisions(
+        replica_ids: List[int]) -> List[AutoscalerDecision]:
+    return [
+        AutoscalerDecision(AutoscalerDecisionOperator.SCALE_DOWN, replica_id)
+        for replica_id in replica_ids
+    ]
+
+
+def _select_nonterminal_replicas_to_scale_down(
+    num_replica_to_scale_down: int,
+    replica_infos: Iterable['replica_managers.ReplicaInfo'],
+) -> List[int]:
+    """Select nonterminal replicas to scale down.
+
+    We sort the replicas based on the following order:
+        1. Based on the `scale_down_decision_order` of the status. We terminate
+            the replicas that is in earlier stage first, as the replicas in
+            later stage may become ready soon.
+        2. Based on the version in ascending order, so we scale down the older
+            versions first.
+        3. Based on the replica_id in descending order, which is also the order
+            of the replicas being launched. We scale down the replicas that are
+            launched earlier first, as the replicas that are launched later may
+            become ready soon.
+
+    Args:
+        num_replica_to_scale_down: The number of replicas to scale down.
+        replica_infos: The list of replica informations to select from.
+
+    Returns:
+        The list of replica ids to scale down.
+    """
+    replicas = list(replica_infos)
+    status_order = serve_state.ReplicaStatus.scale_down_decision_order()
+    assert all(info.status in status_order for info in replicas), (
+        'All replicas to scale down should be in provisioning or launched '
+        'status.', replicas)
+    replicas = sorted(
+        replicas,
+        key=lambda info: (
+            status_order.index(info.status),
+            # version in ascending order
+            info.version,
+            # replica_id in descending order, i.e. launched order
+            -info.replica_id))
+    assert len(replicas) >= num_replica_to_scale_down, (
+        'Not enough replicas to scale down. Available replicas: ',
+        f'{replicas}, num_replica_to_scale_down: {num_replica_to_scale_down}.')
+    return [info.replica_id for info in replicas][:num_replica_to_scale_down]
+
+
 class Autoscaler:
     """Abstract class for autoscalers."""
 
+    # --------------- APIs to implement for custom autoscaler ---------------
+
     def __init__(self, service_name: str,
                  spec: 'service_spec.SkyServiceSpec') -> None:
         """Initialize the autoscaler.
@@ -67,6 +127,8 @@ class Autoscaler:
                 number of replicas, i.e. min_replicas == max_replicas.
             target_num_replicas: Target number of replicas output by autoscaler.
             latest_version: latest version of the service.
+            latest_version_ever_ready: The latest version that is ever ready.
+            update_mode: Update mode for the service.
         """
         self._service_name: str = service_name
         self.min_replicas: int = spec.min_replicas
@@ -81,6 +143,10 @@ class Autoscaler:
         self.latest_version_ever_ready: int = self.latest_version - 1
         self.update_mode = serve_utils.DEFAULT_UPDATE_MODE
 
+    def _calculate_target_num_replicas(self) -> int:
+        """Calculate target number of replicas."""
+        raise NotImplementedError
+
     def update_version(self, version: int, spec: 'service_spec.SkyServiceSpec',
                        update_mode: serve_utils.UpdateMode) -> None:
         if version <= self.latest_version:
@@ -91,9 +157,9 @@ class Autoscaler:
         self.min_replicas = spec.min_replicas
         self.max_replicas = (spec.max_replicas if spec.max_replicas is not None
                              else spec.min_replicas)
-        # Reclip self.target_num_replicas with new min and max replicas.
-        self.target_num_replicas = max(
-            self.min_replicas, min(self.max_replicas, self.target_num_replicas))
+        # Re-clip self.target_num_replicas with new min and max replicas.
+        self.target_num_replicas = self._clip_target_num_replicas(
+            self.target_num_replicas)
         self.update_mode = update_mode
 
     def collect_request_information(
@@ -101,223 +167,61 @@ class Autoscaler:
         """Collect request information from aggregator for autoscaling."""
         raise NotImplementedError
 
-    def evaluate_scaling(
+    def _generate_scaling_decisions(
         self,
         replica_infos: List['replica_managers.ReplicaInfo'],
     ) -> List[AutoscalerDecision]:
-        """Evaluate autoscale options based on replica information."""
+        """Generate Autoscaling decisions based on replica information."""
         raise NotImplementedError
 
-    @classmethod
-    def from_spec(cls, service_name: str,
-                  spec: 'service_spec.SkyServiceSpec') -> 'Autoscaler':
-        # TODO(MaoZiming): use NAME to get the class.
-        if spec.use_ondemand_fallback:
-            return FallbackRequestRateAutoscaler(service_name, spec)
-        else:
-            return RequestRateAutoscaler(service_name, spec)
-
     def _dump_dynamic_states(self) -> Dict[str, Any]:
         """Dump dynamic states from autoscaler."""
         raise NotImplementedError
 
-    def dump_dynamic_states(self) -> Dict[str, Any]:
-        """Dump dynamic states from autoscaler."""
-        states = {'latest_version_ever_ready': self.latest_version_ever_ready}
-        states.update(self._dump_dynamic_states())
-        return states
-
     def _load_dynamic_states(self, dynamic_states: Dict[str, Any]) -> None:
         """Load dynamic states to autoscaler."""
         raise NotImplementedError
 
-    def load_dynamic_states(self, dynamic_states: Dict[str, Any]) -> None:
-        """Load dynamic states to autoscaler."""
-        self.latest_version_ever_ready = dynamic_states.pop(
-            'latest_version_ever_ready', constants.INITIAL_VERSION)
-        self._load_dynamic_states(dynamic_states)
-
-
-class RequestRateAutoscaler(Autoscaler):
-    """RequestRateAutoscaler: Autoscale according to request rate.
-
-    Scales when the number of requests per replica in the given interval
-    is above or below the target qps per replica. The instance can be
-    either spot or on-demand, but not both.
-    """
-
-    def __init__(self, service_name: str,
-                 spec: 'service_spec.SkyServiceSpec') -> None:
-        """Initialize the request rate autoscaler.
+    # --------------- Utility Functions ---------------
 
-        Variables:
-            target_qps_per_replica: Target qps per replica for autoscaling.
-            qps_window_size: Window size for qps calculating.
-            request_timestamps: All request timestamps within the window.
-            upscale_counter: counter for upscale number of replicas.
-            downscale_counter: counter for downscale number of replicas.
-            scale_up_consecutive_periods: period for scaling up.
-            scale_down_consecutive_periods: period for scaling down.
+    def _clip_target_num_replicas(self, target_num_replicas: int) -> int:
+        """Clip target number of replicas with current minimal and maximum
+        number of replicas.
         """
-        super().__init__(service_name, spec)
-        self.target_qps_per_replica: Optional[
-            float] = spec.target_qps_per_replica
-        self.qps_window_size: int = constants.AUTOSCALER_QPS_WINDOW_SIZE_SECONDS
-        self.request_timestamps: List[float] = []
-        self.upscale_counter: int = 0
-        self.downscale_counter: int = 0
-        upscale_delay_seconds = (
-            spec.upscale_delay_seconds if spec.upscale_delay_seconds is not None
-            else constants.AUTOSCALER_DEFAULT_UPSCALE_DELAY_SECONDS)
-        self.scale_up_consecutive_periods: int = int(
-            upscale_delay_seconds /
-            constants.AUTOSCALER_DEFAULT_DECISION_INTERVAL_SECONDS)
-        downscale_delay_seconds = (
-            spec.downscale_delay_seconds
-            if spec.downscale_delay_seconds is not None else
-            constants.AUTOSCALER_DEFAULT_DOWNSCALE_DELAY_SECONDS)
-        self.scale_down_consecutive_periods: int = int(
-            downscale_delay_seconds /
-            constants.AUTOSCALER_DEFAULT_DECISION_INTERVAL_SECONDS)
-
-    def _cal_target_num_replicas_based_on_qps(self) -> int:
-        # Recalculate target_num_replicas based on QPS.
-        # Reclip self.target_num_replicas with new min and max replicas.
-        if self.target_qps_per_replica is None:
-            return self.min_replicas
-        target_num_replicas = math.ceil(
-            len(self.request_timestamps) / self.qps_window_size /
-            self.target_qps_per_replica)
         return max(self.min_replicas, min(self.max_replicas,
                                           target_num_replicas))
 
-    def update_version(self, version: int, spec: 'service_spec.SkyServiceSpec',
-                       update_mode: serve_utils.UpdateMode) -> None:
-        super().update_version(version, spec, update_mode)
-        self.target_qps_per_replica = spec.target_qps_per_replica
-        upscale_delay_seconds = (
-            spec.upscale_delay_seconds if spec.upscale_delay_seconds is not None
-            else constants.AUTOSCALER_DEFAULT_UPSCALE_DELAY_SECONDS)
-        self.scale_up_consecutive_periods = int(
-            upscale_delay_seconds /
-            constants.AUTOSCALER_DEFAULT_DECISION_INTERVAL_SECONDS)
-        downscale_delay_seconds = (
-            spec.downscale_delay_seconds
-            if spec.downscale_delay_seconds is not None else
-            constants.AUTOSCALER_DEFAULT_DOWNSCALE_DELAY_SECONDS)
-        self.scale_down_consecutive_periods = int(
-            downscale_delay_seconds /
-            constants.AUTOSCALER_DEFAULT_DECISION_INTERVAL_SECONDS)
-
-        # We directly set the target_num_replicas here instead of
-        # calling `_set_target_num_replica_with_hysteresis` to have the replicas
-        # quickly scale after each update.
-        self.target_num_replicas = self._cal_target_num_replicas_based_on_qps()
-        # Cleanup hysteretic counters.
-        self.upscale_counter = 0
-        self.downscale_counter = 0
-
-    def collect_request_information(
-            self, request_aggregator_info: Dict[str, Any]) -> None:
-        """Collect request information from aggregator for autoscaling.
-
-        request_aggregator_info should be a dict with the following format:
-
-        {
-            'timestamps': [timestamp1 (float), timestamp2 (float), ...]
-        }
-        """
-        self.request_timestamps.extend(
-            request_aggregator_info.get('timestamps', []))
-        current_time = time.time()
-        index = bisect.bisect_left(self.request_timestamps,
-                                   current_time - self.qps_window_size)
-        self.request_timestamps = self.request_timestamps[index:]
-        logger.info(f'Num of requests in the last {self.qps_window_size} '
-                    f'seconds: {len(self.request_timestamps)}')
-
-    def _set_target_num_replica_with_hysteresis(self) -> None:
-        """Set target_num_replicas based on request rate with hysteresis."""
-        # Keep self.target_num_replicas unchange when autoscaling
-        # is not enabled, i.e. self.target_qps_per_replica is None.
-        # In this case, self.target_num_replicas will be min_replicas.
-        if self.target_qps_per_replica is None:
-            return
-
-        # Convert to requests per second.
-        target_num_replicas = self._cal_target_num_replicas_based_on_qps()
-        old_target_num_replicas = self.target_num_replicas
-
-        # Faster scale up when there is no replica.
-        if self.target_num_replicas == 0:
-            self.target_num_replicas = target_num_replicas
-        elif target_num_replicas > self.target_num_replicas:
-            self.upscale_counter += 1
-            self.downscale_counter = 0
-            if self.upscale_counter >= self.scale_up_consecutive_periods:
-                self.upscale_counter = 0
-                self.target_num_replicas = target_num_replicas
-        elif target_num_replicas < self.target_num_replicas:
-            self.downscale_counter += 1
-            self.upscale_counter = 0
-            if self.downscale_counter >= self.scale_down_consecutive_periods:
-                self.downscale_counter = 0
-                self.target_num_replicas = target_num_replicas
-        else:
-            self.upscale_counter = self.downscale_counter = 0
-
-        num_requests_per_second = len(
-            self.request_timestamps) / self.qps_window_size
-        logger.info(
-            f'Requests per second: {num_requests_per_second}. '
-            f'Current target number of replicas: {old_target_num_replicas}. '
-            f'Final target number of replicas: {self.target_num_replicas}. '
-            f'Upscale counter: {self.upscale_counter}/'
-            f'{self.scale_up_consecutive_periods}. '
-            f'Downscale counter: {self.downscale_counter}/'
-            f'{self.scale_down_consecutive_periods}')
-
     @classmethod
-    def _select_nonterminal_replicas_to_scale_down(
-            cls, num_limit: int,
-            replica_infos: Iterable['replica_managers.ReplicaInfo']
-    ) -> List[int]:
-        status_order = serve_state.ReplicaStatus.scale_down_decision_order()
-        replicas = list(replica_infos)
-        assert all(info.status in status_order for info in replicas), (
-            'All replicas to scale down should be in provisioning or launched '
-            'status.', replicas)
-        replicas = sorted(
-            replicas,
-            key=lambda info: (
-                status_order.index(info.status),
-                # Sort by version in ascending order, so we scale down the older
-                # versions first.
-                info.version,
-                # Sort `info.replica_id` in descending order so that the
-                # replicas in the same version starts to provisioning later are
-                # scaled down first.
-                -info.replica_id))
-        assert len(replicas) >= num_limit, (
-            'Not enough replicas to scale down.', replicas, num_limit)
-        return [info.replica_id for info in replicas][:num_limit]
+    def from_spec(cls, service_name: str,
+                  spec: 'service_spec.SkyServiceSpec') -> 'Autoscaler':
+        # TODO(MaoZiming): use NAME to get the class.
+        if spec.use_ondemand_fallback:
+            return FallbackRequestRateAutoscaler(service_name, spec)
+        else:
+            return RequestRateAutoscaler(service_name, spec)
 
     def get_decision_interval(self) -> int:
-        # Reduce autoscaler interval when target_num_replicas = 0.
-        # This will happen when min_replicas = 0 and no traffic.
+        """Get the decision interval for the autoscaler.
+
+        We reduce the decision interval when the desired number of replicas is
+        0, to make the service scale faster when the service is not running.
+        This will happen when min_replicas = 0 and no traffic.
+        """
         if self.target_num_replicas == 0:
             return constants.AUTOSCALER_NO_REPLICA_DECISION_INTERVAL_SECONDS
         else:
             return constants.AUTOSCALER_DEFAULT_DECISION_INTERVAL_SECONDS
 
-    def select_outdated_replicas_to_scale_down(
-            self,
-            replica_infos: List['replica_managers.ReplicaInfo']) -> List[int]:
+    def _select_outdated_replicas_to_scale_down(
+        self,
+        replica_infos: List['replica_managers.ReplicaInfo'],
+        active_versions: List[int],
+    ) -> List[int]:
         """Select outdated replicas to scale down."""
 
         if self.update_mode == serve_utils.UpdateMode.ROLLING:
-            latest_ready_replicas = []
-            old_nonterminal_replicas = []
+            latest_ready_replicas: List['replica_managers.ReplicaInfo'] = []
+            old_nonterminal_replicas: List['replica_managers.ReplicaInfo'] = []
             for info in replica_infos:
                 if info.version == self.latest_version:
                     if info.is_ready:
@@ -346,19 +250,12 @@ class RequestRateAutoscaler(Autoscaler):
             # `_select_replicas_to_scale_down` will make sure we scale the
             # replicas in initializing statuses first before scaling down the
             # READY old replicas.
-            return self._select_nonterminal_replicas_to_scale_down(
+            return _select_nonterminal_replicas_to_scale_down(
                 max(0,
                     len(old_nonterminal_replicas) - num_old_replicas_to_keep),
                 old_nonterminal_replicas,
             )
 
-        # Use the active versions set by replica manager to make sure we only
-        # scale down the outdated replicas that are not used by the load
-        # balancer.
-        record = serve_state.get_service_from_name(self._service_name)
-        assert record is not None, (f'No service record found for '
-                                    f'{self._service_name}')
-        active_versions = record['active_versions']
         if not active_versions:
             # active_versions can be empty when none of the replicas are ready
             # when the load balancer sync with the controller.
@@ -372,36 +269,35 @@ class RequestRateAutoscaler(Autoscaler):
         # number of ready new replicas is greater than or equal to the min
         # replicas instead of the target, to ensure the service being updated
         # to the latest version faster.
-        all_replica_ids_to_scale_down: List[int] = []
-        for info in replica_infos:
-            if info.version < latest_version_with_min_replicas:
-                all_replica_ids_to_scale_down.append(info.replica_id)
-
-        return all_replica_ids_to_scale_down
+        return [
+            info.replica_id
+            for info in replica_infos
+            if info.version < latest_version_with_min_replicas
+        ]
 
-    def evaluate_scaling(
+    def generate_scaling_decisions(
         self,
         replica_infos: List['replica_managers.ReplicaInfo'],
+        active_versions: List[int],
     ) -> List[AutoscalerDecision]:
-        """Evaluate Autoscaling decisions based on replica information.
-        If the number of launched replicas is less than the target,
-        Trigger a scale up. Else, trigger a scale down.
+        """Generate Autoscaling decisions based on replica information.
+        If the number of launched replicas is less than the target, trigger a
+        scale up. Else, trigger a scale down. This function also handles the
+        version control of the replicas.
 
         For future compatibility, we return a list of AutoscalerDecision.
         Scale-up could include both spot and on-demand, each with a resource
         override dict. Active migration could require returning both SCALE_UP
         and SCALE_DOWN.
         """
-        latest_replicas: List['replica_managers.ReplicaInfo'] = []
-        latest_nonterminal_replicas: List['replica_managers.ReplicaInfo'] = []
 
+        # Handle latest version unrecoverable failure first.
+        latest_replicas: List['replica_managers.ReplicaInfo'] = []
         for info in replica_infos:
             if info.version == self.latest_version:
                 latest_replicas.append(info)
-                if not info.is_terminal:
-                    latest_nonterminal_replicas.append(info)
-                    if info.is_ready:
-                        self.latest_version_ever_ready = self.latest_version
+                if info.is_ready:
+                    self.latest_version_ever_ready = self.latest_version
         if self.latest_version_ever_ready < self.latest_version:
             for info in latest_replicas:
                 if info.status_property.unrecoverable_failure():
@@ -411,55 +307,229 @@ class RequestRateAutoscaler(Autoscaler):
                     # and restart.
                     return []
 
-        self._set_target_num_replica_with_hysteresis()
+        scaling_decisions = []
 
-        scaling_options: List[AutoscalerDecision] = []
-        all_replica_ids_to_scale_down: List[int] = []
+        # If rolling update is in progress, we scale down old replicas based on
+        # the number of ready new replicas and the traffic is directed to both
+        # old and new replicas. Or, for blue_green update, once there is
+        # min_replicas number of ready new replicas, we will direct all traffic
+        # to them, we can scale down all old replicas.
+        # TODO(MaoZiming,zhwu): corner case: We should make sure the fallback
+        # replicas are ready before scaling down the old replicas to avoid the
+        # situation that all the ready new replicas are preempted together.
+        scaling_decisions.extend(
+            _generate_scale_down_decisions(
+                self._select_outdated_replicas_to_scale_down(
+                    replica_infos, active_versions)))
+
+        # If the latest version is ever ready, we can proceed to generate
+        # decisions from the implementations in subclasses.
+        scaling_decisions.extend(
+            self._generate_scaling_decisions(replica_infos))
 
-        # Case 1. If rolling update is in progress, we scale down old replicas
-        # based on the number of ready new replicas and the traffic is directed
-        # to both old and new replicas.
-        # Or, for blue_green update, once there is min_replicas number of ready
-        # new replicas, we will direct all traffic to them, we can scale down
-        # all old replicas.
-        all_replica_ids_to_scale_down.extend(
-            self.select_outdated_replicas_to_scale_down(replica_infos))
+        if not scaling_decisions:
+            logger.info('No scaling needed.')
 
-        # Case 2. when latest_nonterminal_replicas is less
+        return scaling_decisions
+
+    def dump_dynamic_states(self) -> Dict[str, Any]:
+        """Dump dynamic states from autoscaler."""
+        states = {'latest_version_ever_ready': self.latest_version_ever_ready}
+        states.update(self._dump_dynamic_states())
+        return states
+
+    def load_dynamic_states(self, dynamic_states: Dict[str, Any]) -> None:
+        """Load dynamic states to autoscaler."""
+        self.latest_version_ever_ready = dynamic_states.pop(
+            'latest_version_ever_ready', constants.INITIAL_VERSION)
+        self._load_dynamic_states(dynamic_states)
+
+
+class _AutoscalerWithHysteresis(Autoscaler):
+    """_AutoscalerWithHysteresis: Autoscale with hysteresis.
+
+    This is an internal class for developing autoscalers with hysteresis. It
+    only scales when the number of replicas is above or below the target number
+    of replicas for a certain number of consecutive periods.
+    """
+
+    def _setup_thresholds(self, spec: 'service_spec.SkyServiceSpec') -> None:
+        upscale_delay_seconds = (
+            spec.upscale_delay_seconds if spec.upscale_delay_seconds is not None
+            else constants.AUTOSCALER_DEFAULT_UPSCALE_DELAY_SECONDS)
+        self.scale_up_threshold: int = int(
+            upscale_delay_seconds /
+            constants.AUTOSCALER_DEFAULT_DECISION_INTERVAL_SECONDS)
+        downscale_delay_seconds = (
+            spec.downscale_delay_seconds
+            if spec.downscale_delay_seconds is not None else
+            constants.AUTOSCALER_DEFAULT_DOWNSCALE_DELAY_SECONDS)
+        self.scale_down_threshold: int = int(
+            downscale_delay_seconds /
+            constants.AUTOSCALER_DEFAULT_DECISION_INTERVAL_SECONDS)
+
+    def __init__(self, service_name: str,
+                 spec: 'service_spec.SkyServiceSpec') -> None:
+        """Initialize the hysteresis autoscaler.
+
+        Variables:
+            upscale_counter: Counter for upscale decisions of replicas.
+            downscale_counter: Counter for downscale decisions of replicas.
+            scale_up_threshold: The threshold to trigger a scale up.
+            scale_down_threshold: The threshold to trigger a scale down.
+        """
+        super().__init__(service_name, spec)
+        self.upscale_counter: int = 0
+        self.downscale_counter: int = 0
+        self._setup_thresholds(spec)
+
+    def update_version(self, version: int, spec: 'service_spec.SkyServiceSpec',
+                       update_mode: serve_utils.UpdateMode) -> None:
+        super().update_version(version, spec, update_mode)
+        # We directly set the target_num_replicas here instead of calling
+        # `_set_target_num_replicas_with_hysteresis` to have the replicas
+        # quickly scale after each update.
+        self.target_num_replicas = self._calculate_target_num_replicas()
+        # Cleanup hysteresis counters.
+        self.upscale_counter = 0
+        self.downscale_counter = 0
+        self._setup_thresholds(spec)
+
+    def _set_target_num_replicas_with_hysteresis(self) -> None:
+        """Set target_num_replicas based on request rate with hysteresis."""
+        target_num_replicas = self._calculate_target_num_replicas()
+        old_target_num_replicas = self.target_num_replicas
+
+        # Faster scale up when there is no replica.
+        if self.target_num_replicas == 0:
+            self.target_num_replicas = target_num_replicas
+        elif target_num_replicas > self.target_num_replicas:
+            self.upscale_counter += 1
+            self.downscale_counter = 0
+            if self.upscale_counter >= self.scale_up_threshold:
+                self.upscale_counter = 0
+                self.target_num_replicas = target_num_replicas
+        elif target_num_replicas < self.target_num_replicas:
+            self.downscale_counter += 1
+            self.upscale_counter = 0
+            if self.downscale_counter >= self.scale_down_threshold:
+                self.downscale_counter = 0
+                self.target_num_replicas = target_num_replicas
+        else:
+            self.upscale_counter = self.downscale_counter = 0
+
+        logger.info(
+            f'Old target number of replicas: {old_target_num_replicas}. '
+            f'Current target number of replicas: {target_num_replicas}. '
+            f'Final target number of replicas: {self.target_num_replicas}. '
+            f'Upscale counter: {self.upscale_counter}/'
+            f'{self.scale_up_threshold}. '
+            f'Downscale counter: {self.downscale_counter}/'
+            f'{self.scale_down_threshold}. ')
+
+
+class RequestRateAutoscaler(_AutoscalerWithHysteresis):
+    """RequestRateAutoscaler: Autoscale according to request rate.
+
+    Scales when the number of requests per replica in the given interval
+    is above or below the target qps per replica. The instance can be
+    either spot or on-demand, but not both.
+    """
+
+    def __init__(self, service_name: str,
+                 spec: 'service_spec.SkyServiceSpec') -> None:
+        """Initialize the request rate autoscaler.
+
+        Variables:
+            target_qps_per_replica: Target qps per replica for autoscaling.
+            qps_window_size: Window size for qps calculating.
+            request_timestamps: All request timestamps within the window.
+        """
+        super().__init__(service_name, spec)
+        self.target_qps_per_replica: Optional[
+            float] = spec.target_qps_per_replica
+        self.qps_window_size: int = constants.AUTOSCALER_QPS_WINDOW_SIZE_SECONDS
+        self.request_timestamps: List[float] = []
+
+    def _calculate_target_num_replicas(self) -> int:
+        if self.target_qps_per_replica is None:
+            return self.min_replicas
+        num_requests_per_second = len(
+            self.request_timestamps) / self.qps_window_size
+        target_num_replicas = math.ceil(num_requests_per_second /
+                                        self.target_qps_per_replica)
+        logger.info(f'Requests per second: {num_requests_per_second}. '
+                    f'Target number of replicas: {target_num_replicas}.')
+        return self._clip_target_num_replicas(target_num_replicas)
+
+    def update_version(self, version: int, spec: 'service_spec.SkyServiceSpec',
+                       update_mode: serve_utils.UpdateMode) -> None:
+        super().update_version(version, spec, update_mode)
+        self.target_qps_per_replica = spec.target_qps_per_replica
+
+    def collect_request_information(
+            self, request_aggregator_info: Dict[str, Any]) -> None:
+        """Collect request information from aggregator for autoscaling.
+
+        request_aggregator_info should be a dict with the following format:
+
+        {
+            'timestamps': [timestamp1 (float), timestamp2 (float), ...]
+        }
+        """
+        self.request_timestamps.extend(
+            request_aggregator_info.get('timestamps', []))
+        current_time = time.time()
+        index = bisect.bisect_left(self.request_timestamps,
+                                   current_time - self.qps_window_size)
+        self.request_timestamps = self.request_timestamps[index:]
+        logger.info(f'Num of requests in the last {self.qps_window_size} '
+                    f'seconds: {len(self.request_timestamps)}')
+
+    def _generate_scaling_decisions(
+        self,
+        replica_infos: List['replica_managers.ReplicaInfo'],
+    ) -> List[AutoscalerDecision]:
+        """Generate Autoscaling decisions based on request rate."""
+
+        self._set_target_num_replicas_with_hysteresis()
+
+        latest_nonterminal_replicas: List['replica_managers.ReplicaInfo'] = []
+
+        for info in replica_infos:
+            if info.version == self.latest_version:
+                if not info.is_terminal:
+                    latest_nonterminal_replicas.append(info)
+
+        scaling_decisions: List[AutoscalerDecision] = []
+
+        # Case 1. when latest_nonterminal_replicas is less
         # than num_to_provision, we always scale up new replicas.
         if len(latest_nonterminal_replicas) < self.target_num_replicas:
             num_replicas_to_scale_up = (self.target_num_replicas -
                                         len(latest_nonterminal_replicas))
             logger.info('Number of replicas to scale up: '
                         f'{num_replicas_to_scale_up}')
-            for _ in range(num_replicas_to_scale_up):
-                scaling_options.append(
-                    AutoscalerDecision(AutoscalerDecisionOperator.SCALE_UP,
-                                       target=None))
+            scaling_decisions.extend(
+                _generate_scale_up_decisions(num_replicas_to_scale_up, None))
 
-        # Case 3: when latest_nonterminal_replicas is more
+        # Case 2: when latest_nonterminal_replicas is more
         # than self.target_num_replicas, we scale down new replicas.
+        replicas_to_scale_down = []
         if len(latest_nonterminal_replicas) > self.target_num_replicas:
             num_replicas_to_scale_down = (len(latest_nonterminal_replicas) -
                                           self.target_num_replicas)
             replicas_to_scale_down = (
-                RequestRateAutoscaler.
                 _select_nonterminal_replicas_to_scale_down(
-                    num_limit=num_replicas_to_scale_down,
-                    replica_infos=latest_nonterminal_replicas))
+                    num_replicas_to_scale_down, latest_nonterminal_replicas))
             logger.info(
                 'Number of replicas to scale down: '
                 f'{num_replicas_to_scale_down} {replicas_to_scale_down}')
-            all_replica_ids_to_scale_down.extend(replicas_to_scale_down)
 
-        for replica_id in all_replica_ids_to_scale_down:
-            scaling_options.append(
-                AutoscalerDecision(AutoscalerDecisionOperator.SCALE_DOWN,
-                                   target=replica_id))
+        scaling_decisions.extend(
+            _generate_scale_down_decisions(replicas_to_scale_down))
 
-        if not scaling_options:
-            logger.info('No scaling needed.')
-        return scaling_options
+        return scaling_decisions
 
     def _dump_dynamic_states(self) -> Dict[str, Any]:
         return {
@@ -481,16 +551,19 @@ class FallbackRequestRateAutoscaler(RequestRateAutoscaler):
 
     When spec.base_ondemand_fallback_replicas is set, we make sure
     there are at least spec.base_ondemand_fallback_replicas on-demands
-    to be always there to provide basic gurantee for the availability.
+    to be always there to provide basic guarantee for the availability.
 
     When spec.dynamic_ondemand_fallback is set, on-demand instances
     will be scheduled to provision for any preempted spot instance, i.e.,
     on-demand instance are used as dynamic fallback of spot.
     """
 
-    def __init__(self, service_name: str,
-                 spec: 'service_spec.SkyServiceSpec') -> None:
-        super().__init__(service_name, spec)
+    # job_recovery field is checked earlier in core
+    SPOT_OVERRIDE = {'use_spot': True}
+    ONDEMAND_OVERRIDE = {'use_spot': False}
+
+    def _setup_fallback_options(self,
+                                spec: 'service_spec.SkyServiceSpec') -> None:
         self.base_ondemand_fallback_replicas: int = (
             spec.base_ondemand_fallback_replicas
             if spec.base_ondemand_fallback_replicas is not None else 0)
@@ -501,37 +574,42 @@ class FallbackRequestRateAutoscaler(RequestRateAutoscaler):
             spec.dynamic_ondemand_fallback
             if spec.dynamic_ondemand_fallback is not None else False)
 
+    def __init__(self, service_name: str,
+                 spec: 'service_spec.SkyServiceSpec') -> None:
+        """Initialize the fallback request rate autoscaler.
+
+        Variables:
+            base_ondemand_fallback_replicas: Minimum number of on-demand
+                replicas to be always there.
+            dynamic_ondemand_fallback: Whether to dynamically provision
+                on-demand instances for preempted spot instances.
+        """
+        super().__init__(service_name, spec)
+        self._setup_fallback_options(spec)
+
     def update_version(self, version: int, spec: 'service_spec.SkyServiceSpec',
                        update_mode: serve_utils.UpdateMode) -> None:
         super().update_version(version, spec, update_mode=update_mode)
-        self.base_ondemand_fallback_replicas = (
-            spec.base_ondemand_fallback_replicas
-            if spec.base_ondemand_fallback_replicas is not None else 0)
-        # Assert: Either dynamic_ondemand_fallback is set
-        # or base_ondemand_fallback_replicas is greater than 0.
-        assert spec.use_ondemand_fallback
-        self.dynamic_ondemand_fallback = (spec.dynamic_ondemand_fallback
-                                          if spec.dynamic_ondemand_fallback
-                                          is not None else False)
+        self._setup_fallback_options(spec)
 
-    # job_recovery field is checked earlier in core
-    def _get_spot_resources_override_dict(self) -> Dict[str, Any]:
-        return {'use_spot': True}
-
-    def _get_ondemand_resources_override_dict(self) -> Dict[str, Any]:
-        return {'use_spot': False}
-
-    def evaluate_scaling(
+    def _generate_scaling_decisions(
         self,
         replica_infos: List['replica_managers.ReplicaInfo'],
     ) -> List[AutoscalerDecision]:
+        """Generate Autoscaling decisions based on request rate, with on-demand
+        fallback.
+
+        The autoscaler will make sure there are at least
+        `base_ondemand_fallback_replicas` on-demand replicas to be always there,
+        so the service can provide basic guarantee for the availability.
+        """
+
+        self._set_target_num_replicas_with_hysteresis()
 
         latest_nonterminal_replicas = list(
             filter(
                 lambda info: not info.is_terminal and info.version == self.
                 latest_version, replica_infos))
-
-        self._set_target_num_replica_with_hysteresis()
         num_nonterminal_spot, num_ready_spot = 0, 0
         num_nonterminal_ondemand, num_ready_ondemand = 0, 0
 
@@ -546,22 +624,14 @@ class FallbackRequestRateAutoscaler(RequestRateAutoscaler):
                 num_nonterminal_ondemand += 1
 
         logger.info(
-            'Number of alive spot instances: '
-            f'{num_nonterminal_spot}, '
+            f'Number of alive spot instances: {num_nonterminal_spot}, '
             f'Number of ready spot instances: {num_ready_spot}, '
-            'Number of alive on-demand instances: '
-            f' {num_nonterminal_ondemand}, '
+            f'Number of alive on-demand instances: {num_nonterminal_ondemand}, '
             f'Number of ready on-demand instances: {num_ready_ondemand}')
 
-        scaling_options: List[AutoscalerDecision] = []
+        scaling_decisions: List[AutoscalerDecision] = []
         all_replica_ids_to_scale_down: List[int] = []
 
-        # TODO(MaoZiming,zhwu): coner case: We should make sure the fallback
-        # replicas are ready before scaling down the old replicas to avoid the
-        # situation that all the ready new replicas are preempted together.
-        all_replica_ids_to_scale_down.extend(
-            self.select_outdated_replicas_to_scale_down(replica_infos))
-
         # Decide how many spot instances to launch.
         num_spot_to_provision = (self.target_num_replicas -
                                  self.base_ondemand_fallback_replicas)
@@ -571,18 +641,15 @@ class FallbackRequestRateAutoscaler(RequestRateAutoscaler):
                                     num_nonterminal_spot)
             logger.info('Number of spot instances to scale up: '
                         f'{num_spot_to_scale_up}')
-            for _ in range(num_spot_to_scale_up):
-                scaling_options.append(
-                    AutoscalerDecision(
-                        AutoscalerDecisionOperator.SCALE_UP,
-                        target=self._get_spot_resources_override_dict()))
+            scaling_decisions.extend(
+                _generate_scale_up_decisions(num_spot_to_scale_up,
+                                             self.SPOT_OVERRIDE))
         elif num_nonterminal_spot > num_spot_to_provision:
             # Too many spot instances, scale down.
             # Get the replica to scale down with _select_replicas_to_scale_down
             num_spot_to_scale_down = (num_nonterminal_spot -
                                       num_spot_to_provision)
             replicas_to_scale_down = (
-                RequestRateAutoscaler.
                 _select_nonterminal_replicas_to_scale_down(
                     num_spot_to_scale_down,
                     filter(lambda info: info.is_spot,
@@ -606,16 +673,13 @@ class FallbackRequestRateAutoscaler(RequestRateAutoscaler):
                                         num_nonterminal_ondemand)
             logger.info('Number of on-demand instances to scale up: '
                         f'{num_ondemand_to_scale_up}')
-            for _ in range(num_ondemand_to_scale_up):
-                scaling_options.append(
-                    AutoscalerDecision(
-                        AutoscalerDecisionOperator.SCALE_UP,
-                        target=self._get_ondemand_resources_override_dict()))
+            scaling_decisions.extend(
+                _generate_scale_up_decisions(num_ondemand_to_scale_up,
+                                             self.ONDEMAND_OVERRIDE))
         else:
             num_ondemand_to_scale_down = (num_nonterminal_ondemand -
                                           num_ondemand_to_provision)
             replicas_to_scale_down = (
-                RequestRateAutoscaler.
                 _select_nonterminal_replicas_to_scale_down(
                     num_ondemand_to_scale_down,
                     filter(lambda info: not info.is_spot,
@@ -626,9 +690,7 @@ class FallbackRequestRateAutoscaler(RequestRateAutoscaler):
 
             all_replica_ids_to_scale_down.extend(replicas_to_scale_down)
 
-        for replica_id in all_replica_ids_to_scale_down:
-            scaling_options.append(
-                AutoscalerDecision(AutoscalerDecisionOperator.SCALE_DOWN,
-                                   target=replica_id))
+        scaling_decisions.extend(
+            _generate_scale_down_decisions(all_replica_ids_to_scale_down))
 
-        return scaling_options
+        return scaling_decisions