palimpzest 0.7.21__py3-none-any.whl → 0.8.0__py3-none-any.whl

palimpzest/query/optimizer/cost_model.py

@@ -1,29 +1,15 @@
 from __future__ import annotations
 
 import logging
-
-# NOTE: the answer.mode() call(s) inside of _est_quality() throw a UserWarning when there are multiple
-#       answers to a convert with the same mode. This is because pandas tries to sort the answers
-#       before returning them, but since answer is a column of dicts the '<' operator fails on dicts.
-#       For now, we can simply ignore the warning b/c we pick an answer at random anyways if there are
-#       multiple w/the same count, but in the future we may want to cast the 'dict' --> 'str' or compute
-#       the mode on a per-field basis.
 import warnings
-from typing import Any
 
 import pandas as pd
 
-from palimpzest.constants import MODEL_CARDS, NAIVE_BYTES_PER_RECORD, GPT_4o_MODEL_CARD, Model
-from palimpzest.core.data.dataclasses import OperatorCostEstimates, PlanCost, RecordOpStats, SentinelPlanStats
-from palimpzest.query.operators.aggregate import ApplyGroupByOp, AverageAggregateOp, CountAggregateOp
-from palimpzest.query.operators.code_synthesis_convert import CodeSynthesisConvert
-from palimpzest.query.operators.convert import LLMConvert
-from palimpzest.query.operators.filter import LLMFilter, NonLLMFilter
-from palimpzest.query.operators.limit import LimitScanOp
+from palimpzest.constants import NAIVE_BYTES_PER_RECORD
+from palimpzest.core.models import OperatorCostEstimates, PlanCost, SentinelPlanStats
+from palimpzest.query.operators.join import JoinOp
 from palimpzest.query.operators.physical import PhysicalOperator
-from palimpzest.query.operators.rag_convert import RAGConvert
-from palimpzest.query.operators.scan import CacheScanDataOp, MarshalAndScanDataOp, ScanPhysicalOp
-from palimpzest.utils.model_helpers import get_champion_model_name, get_models
+from palimpzest.query.operators.scan import ContextScanOp, MarshalAndScanDataOp, ScanPhysicalOp
 
 warnings.simplefilter(action='ignore', category=UserWarning)
 
@@ -62,7 +48,7 @@ class SampleBasedCostModel:
     """
     def __init__(
         self,
-        sentinel_plan_stats: SentinelPlanStats,
+        sentinel_plan_stats: SentinelPlanStats | None = None,
         verbose: bool = False,
         exp_name: str | None = None,
     ):
@@ -73,25 +59,34 @@ class SampleBasedCostModel:
         self.exp_name = exp_name
 
         # construct cost, time, quality, and selectivity matrices for each operator set;
-        self.operator_to_stats = self.compute_operator_stats(sentinel_plan_stats)
-        self.costed_full_op_ids = set([
+        self.operator_to_stats = self._compute_operator_stats(sentinel_plan_stats)
+        self.costed_full_op_ids = None if self.operator_to_stats is None else set([
             full_op_id
             for _, full_op_id_to_stats in self.operator_to_stats.items()
             for full_op_id in full_op_id_to_stats
         ])
 
+        # if there is a logical operator with no samples; add all of its op ids to costed_full_op_ids;
+        # this will lead to the cost model applying the naive cost estimates for all physical op ids
+        # in this logical operator (I think?)
+        # TODO
+
         logger.info(f"Initialized SampleBasedCostModel with verbose={self.verbose}")
         logger.debug(f"Initialized SampleBasedCostModel with params: {self.__dict__}")
 
     def get_costed_full_op_ids(self):
         return self.costed_full_op_ids
 
-    def compute_operator_stats(self, sentinel_plan_stats: SentinelPlanStats) -> dict:
+    def _compute_operator_stats(self, sentinel_plan_stats: SentinelPlanStats | None) -> dict:
         logger.debug("Computing operator statistics")
+        # if no stats are provided, simply return None
+        if sentinel_plan_stats is None:
+            return None
+
         # flatten the nested dictionary of execution data and pull out fields relevant to cost estimation
         execution_record_op_stats = []
-        for logical_op_id, full_op_id_to_op_stats in sentinel_plan_stats.operator_stats.items():
-            logger.debug(f"Computing operator statistics for logical_op_id: {logical_op_id}")
+        for unique_logical_op_id, full_op_id_to_op_stats in sentinel_plan_stats.operator_stats.items():
+            logger.debug(f"Computing operator statistics for logical_op_id: {unique_logical_op_id}")
             # flatten the execution data into a list of RecordOpStats
             op_set_execution_data = [
                 record_op_stats
@@ -102,17 +97,17 @@ class SampleBasedCostModel:
             # add entries from execution data into matrices
             for record_op_stats in op_set_execution_data:
                 record_op_stats_dict = {
-                    "logical_op_id": logical_op_id,
+                    "unique_logical_op_id": unique_logical_op_id,
                     "full_op_id": record_op_stats.full_op_id,
                     "record_id": record_op_stats.record_id,
-                    "record_parent_id": record_op_stats.record_parent_id,
+                    "record_parent_ids": record_op_stats.record_parent_ids,
                     "cost_per_record": record_op_stats.cost_per_record,
                     "time_per_record": record_op_stats.time_per_record,
                     "quality": record_op_stats.quality,
                     "passed_operator": record_op_stats.passed_operator,
-                    "source_idx": record_op_stats.record_source_idx,  # TODO: remove
-                    "op_details": record_op_stats.op_details,         # TODO: remove
-                    "answer": record_op_stats.answer,                 # TODO: remove
+                    "source_indices": record_op_stats.record_source_indices,  # TODO: remove
+                    "op_details": record_op_stats.op_details,                 # TODO: remove
+                    "answer": record_op_stats.answer,                         # TODO: remove
                 }
                 execution_record_op_stats.append(record_op_stats_dict)
 
@@ -121,22 +116,22 @@ class SampleBasedCostModel:
 
         # for each full_op_id, compute its average cost_per_record, time_per_record, selectivity, and quality
         operator_to_stats = {}
-        for logical_op_id, logical_op_df in operator_stats_df.groupby("logical_op_id"):
-            logger.debug(f"Computing operator statistics for logical_op_id: {logical_op_id}")
-            operator_to_stats[logical_op_id] = {}
+        for unique_logical_op_id, logical_op_df in operator_stats_df.groupby("unique_logical_op_id"):
+            logger.debug(f"Computing operator statistics for unique_logical_op_id: {unique_logical_op_id}")
+            operator_to_stats[unique_logical_op_id] = {}
 
             for full_op_id, physical_op_df in logical_op_df.groupby("full_op_id"):
-                # compute the number of input records processed by this operator; use source_idx for scan operator(s)
+                # compute the number of input records processed by this operator; use source_indices for scan operator(s)
                 num_source_records = (
-                    len(physical_op_df.record_parent_id.unique())
-                    if not physical_op_df.record_parent_id.isna().all()
-                    else len(physical_op_df.source_idx.unique())
+                    physical_op_df.record_parent_ids.apply(tuple).nunique()
+                    if not physical_op_df.record_parent_ids.isna().all()
+                    else physical_op_df.source_indices.apply(tuple).nunique()
                 )
 
                 # compute selectivity
                 selectivity = physical_op_df.passed_operator.sum() / num_source_records
 
-                operator_to_stats[logical_op_id][full_op_id] = {
+                operator_to_stats[unique_logical_op_id][full_op_id] = {
                     "cost": physical_op_df.cost_per_record.mean(),
                     "time": physical_op_df.time_per_record.mean(),
                     "quality": physical_op_df.quality.mean(),
@@ -150,347 +145,18 @@ class SampleBasedCostModel:
         logger.debug(f"Done computing operator statistics for {len(operator_to_stats)} operators!")
         return operator_to_stats
 
-
-    def __call__(self, operator: PhysicalOperator, source_op_estimates: OperatorCostEstimates | None = None) -> PlanCost:
-        # NOTE: some physical operators may not have any sample execution data in this cost model;
-        #       these physical operators are filtered out of the Optimizer, thus we can assume that
-        #       we will have execution data for each operator passed into __call__; nevertheless, we
-        #       still perform a sanity check
-        # look up physical and logical op ids associated with this physical operator
-        full_op_id = operator.get_full_op_id()
-        logical_op_id = operator.logical_op_id
-        physical_op_to_stats = self.operator_to_stats.get(logical_op_id)
-        assert physical_op_to_stats is not None, f"No execution data for logical operator: {str(operator)}"
-        assert physical_op_to_stats.get(full_op_id) is not None, f"No execution data for physical operator: {str(operator)}"
-        logger.debug(f"Calling __call__ for {str(operator)}")
-
-        # look up stats for this operation
-        est_cost_per_record = self.operator_to_stats[logical_op_id][full_op_id]["cost"]
-        est_time_per_record = self.operator_to_stats[logical_op_id][full_op_id]["time"]
-        est_quality = self.operator_to_stats[logical_op_id][full_op_id]["quality"]
-        est_selectivity = self.operator_to_stats[logical_op_id][full_op_id]["selectivity"]
-
-        # create source_op_estimates for scan operators if they are not provided
-        if isinstance(operator, ScanPhysicalOp):
-            # get handle to scan operator and pre-compute its size (number of records)
-            datareader_len = len(operator.datareader)
-
-            source_op_estimates = OperatorCostEstimates(
-                cardinality=datareader_len,
-                time_per_record=0.0,
-                cost_per_record=0.0,
-                quality=1.0,
-            )
-
-        # generate new set of OperatorCostEstimates
-        op_estimates = OperatorCostEstimates(
-            cardinality=est_selectivity * source_op_estimates.cardinality,
-            time_per_record=est_time_per_record,
-            cost_per_record=est_cost_per_record,
-            quality=est_quality,
-        )
-
-        # compute estimates for this operator
-        op_time = op_estimates.time_per_record * source_op_estimates.cardinality
-        op_cost = op_estimates.cost_per_record * source_op_estimates.cardinality
-        op_quality = op_estimates.quality
-
-        # construct and return op estimates
-        plan_cost = PlanCost(cost=op_cost, time=op_time, quality=op_quality, op_estimates=op_estimates)
-        logger.debug(f"Done calling __call__ for {str(operator)}")
-        logger.debug(f"Plan cost: {plan_cost}")
-        return plan_cost
-
-
-class CostModel(BaseCostModel):
-    """
-    This class takes in a list of RecordOpStats and performs cost estimation on a given operator
-    by taking the average of any sample execution that the CostModel has for that operator. If no
-    such data exists, it returns a naive estimate.
-    """
-    def __init__(
-        self,
-        sample_execution_data: list[RecordOpStats] | None = None,
-        available_models: list[Model] | None = None,
-    ) -> None:
-        if sample_execution_data is None:
-            sample_execution_data = []
-        if available_models is None:
-            available_models = []
-
-        # construct full dataset of samples
-        self.sample_execution_data_df = (
-            pd.DataFrame(sample_execution_data)
-            if len(sample_execution_data) > 0
-            else None
-        )
-        # df contains a column called record_state, that sometimes contain a dict
-        # we want to extract the keys from the dict and create a new column for each key
-
-        # set available models
-        self.available_models = available_models
-
-        # compute per-operator estimates
-        self.operator_estimates = self._compute_operator_estimates()
-
-        # compute set of costed full op ids from operator_to_stats
-        self.costed_full_op_ids = None if self.operator_estimates is None else set(self.operator_estimates.keys())
-        logger.info("Initialized CostModel.")
-        logger.debug(f"Initialized CostModel with params: {self.__dict__}")
-
-    def get_costed_full_op_ids(self):
-        return self.costed_full_op_ids
-
-    def _compute_mean(self, df: pd.DataFrame, col: str, model_name: str | None = None) -> float:
-        """
-        Compute the mean for the given column and dataframe. If the model_name is provided, filter
-        for the subset of rows belonging to the model.
-        """
-        # use model-specific estimate if possible
-        if model_name is not None:
-            model_df = df[df.model_name == model_name]
-            if not model_df.empty:
-                return model_df[col].mean()
-
-        # compute aggregate mean across all models
-        return df[col].mean()
-
-    def _est_time_per_record(self, op_df: pd.DataFrame, model_name: str | None = None) -> tuple[float, float, float]:
-        """
-        Given sample cost data observations for a specific operation, compute the mean and CI
-        for the time per record.
-        """
-        return self._compute_mean(df=op_df, col="time_per_record", model_name=model_name)
-
-    def _est_cost_per_record(self, op_df: pd.DataFrame, model_name: str | None = None) -> tuple[float, float, float]:
-        """
-        Given sample cost data observations for a specific operation, compute the mean and CI
-        for the cost per record.
-        """
-        return self._compute_mean(df=op_df, col="cost_per_record", model_name=model_name)
-
-    def _est_tokens_per_record(self, op_df: pd.DataFrame, model_name: str | None = None) -> tuple[float, float]:
-        """
-        Given sample cost data observations for a specific operation, compute the mean and CI
-        for the total input tokens and total output tokens.
-        """
-        total_input_tokens = self._compute_mean(df=op_df, col="total_input_tokens", model_name=model_name)
-        total_output_tokens = self._compute_mean(df=op_df, col="total_output_tokens", model_name=model_name)
-
-        return total_input_tokens, total_output_tokens
-
-    def _est_cardinality(self, op_df: pd.DataFrame, model_name: str | None = None) -> float:
-        """
-        Given sample cost data observations for a specific operation, compute the number of
-        rows output by the operation.
-
-        NOTE: right now, this should only be used by the ApplyGroupByOp as a way to gauge the
-        number of output groups. Using this to estimate, e.g. the cardinality of a filter,
-        convert, or base scan will lead to wildly inaccurate results because the absolute value
-        of these cardinalities will simply be a reflection of the sample size.
-
-        For those operations, we use the `_est_selectivity` function to estimate the operator's
-        selectivity, which we can apply to an est. of the operator's input cardinality.
-        """
-        return op_df.shape[0] / len(op_df.plan_id.unique())
-
-    def _est_selectivity(self, df: pd.DataFrame, op_df: pd.DataFrame, model_name: str | None = None) -> float:
-        """
-        Given sample cost data observations for the plan and a specific operation, compute
-        the ratio of records between this operator and its source operator.
-        """
-        # compute whether or not this operation is a filter
-        is_filter_op = "filter" in str(op_df.op_name.iloc[0]).lower()
-
-        # use model-specific estimate if possible
-        if model_name is not None:
-            model_op_df = op_df[op_df.model_name == model_name]
-            if not model_op_df.empty:
-                num_input_records = model_op_df.shape[0]
-
-                # get subset of records that were the source to this operator
-                num_output_records = None
-                if is_filter_op:
-                    num_output_records = model_op_df.passed_operator.sum()
-                else:
-                    full_op_ids = model_op_df.full_op_id.unique().tolist()
-                    plan_ids = model_op_df.plan_id.unique().tolist()
-                    num_output_records = df[df.source_full_op_id.isin(full_op_ids) & df.plan_id.isin(plan_ids)].shape[0]
-
-                # estimate the selectivity / fan-out
-                return num_output_records / num_input_records
-
-        # otherwise average selectivity across all ops
-        num_input_records = op_df.shape[0]
-
-        # get subset of records that were the source to this operator
-        num_output_records = None
-        if is_filter_op:
-            num_output_records = op_df.passed_operator.sum()
-        else:
-            full_op_ids = op_df.full_op_id.unique().tolist()
-            num_output_records = df[df.source_full_op_id.isin(full_op_ids)].shape[0]
-
-        # estimate the selectivity / fan-out
-        return num_output_records / num_input_records
-
-    def _compute_quality(self, row):
-        # compute accuracy for filter
-        if "filter" in row["op_name"].lower():
-            row["correct"] = int(row["answer"] == row["accepted_answer"])
-            row["num_answers"] = 1
-            return row
-
-        # otherwise, compute recall on a per-key basis
-        try:
-            # we'll measure recall on accepted_answer, as extraneous info is often not an issue
-            answer = row["answer"]
-            accepted_answer = row["accepted_answer"]
-            correct = 0
-            for key, value in accepted_answer.items():
-                if key in answer and answer[key] == value:
-                    correct += 1
-            
-            row["correct"] = correct
-            row["num_answers"] = len(accepted_answer.keys())
-            return row
-
-        except Exception as e:
-            print(f"WARNING: error decoding answer or accepted_answer: {str(e)}")
-            row["correct"] = 0
-            row["num_answers"] = 1
-            return row
-
-    def _est_quality(self, op_df: pd.DataFrame, model_name: str | None = None) -> float:
-        """
-        Given sample cost data observations for a specific operation, compute the an estimate
-        of the quality of its outputs by using GPT-4 as a champion model.
-        """
-        # get unique set of records
-        record_ids = op_df.record_id.unique()
-
-        # get champion model name
-        vision = ("image_operation" in op_df.columns and op_df.image_operation.any())
-        champion_model_name = get_champion_model_name(self.available_models, vision)
-
-        # compute champion's answer (per-record) across all models; fall-back to most common answer if champion is not present
-        record_id_to_answer = {}
-        for record_id in record_ids:
-            record_df = op_df[op_df.record_id == record_id]
-            champion_most_common_answer = record_df[
-                record_df.model_name == champion_model_name
-            ].answer.mode()
-            all_models_most_common_answer = record_df.answer.mode()
-
-            if not champion_most_common_answer.empty:
-                record_id_to_answer[record_id] = champion_most_common_answer.iloc[0]
-            elif not all_models_most_common_answer.empty:
-                record_id_to_answer[record_id] = all_models_most_common_answer.iloc[0]
-            else:
-                record_id_to_answer[record_id] = None
-
-        # compute accepted answers and clean all answers
-        pd.options.mode.chained_assignment = None  # turn off copy warnings
-        op_df.loc[:, "accepted_answer"] = op_df.record_id.apply(lambda id: record_id_to_answer[id])
-        op_df = op_df.apply(lambda row: self._compute_quality(row), axis=1)
-
-        # get subset of observations for model_name and estimate quality w/fraction of answers that match accepted answer
-        model_df = (
-            op_df[op_df.model_name == model_name]
-            if model_name is not None
-            else op_df[op_df.model_name.isna()]
-        )
-
-        # compute quality as the fraction of answers which are correct (recall on expected output)
-        num_correct = model_df.correct.sum() if not model_df.empty else op_df.correct.sum()
-        total_answers = model_df.num_answers.sum() if not model_df.empty else op_df.num_answers.sum()
-        est_quality = num_correct / total_answers
-
-        return est_quality
-
-    def _compute_operator_estimates(self) -> dict[str, Any] | None:
-        """
-        Compute per-operator estimates of runtime, cost, and quality.
-        """
-        # if we don't have sample execution data, we cannot compute per-operator estimates
-        if self.sample_execution_data_df is None:
-            return None
-
-        # get the set of operator ids for which we have sample data
-        full_op_ids = self.sample_execution_data_df.full_op_id.unique()
-
-        # compute estimates of runtime, cost, and quality (and intermediates like cardinality) for every operator
-        operator_estimates = {}
-        for full_op_id in full_op_ids:
-            # filter for subset of sample execution data related to this operation
-            op_df = self.sample_execution_data_df[
-                self.sample_execution_data_df.full_op_id == full_op_id
-            ]
-
-            # skip computing an estimate if we didn't capture any sampling data for this operator
-            # (this can happen if/when upstream filter operation(s) filter out all records) 
-            if op_df.empty:
-                continue
-
-            # initialize estimates
-            estimates = {}
-
-            # get the op_name for this operation
-            model_name = op_df.model_name.iloc[0] if op_df.model_name.iloc[0] is not None else None
-            op_name = str(op_df.op_name.iloc[0])
-            if model_name is not None:
-                # compute estimates per-model, and add None which forces computation of avg. across all models
-                model_names = [m.value for m in get_models(include_vision=True)] + [None]
-                # model_names = op_df.model_name.unique().tolist()
-                estimates = {model_name: None for model_name in model_names}
-                for model_name in model_names:
-                    time_per_record = self._est_time_per_record(op_df, model_name=model_name)
-                    cost_per_record = self._est_cost_per_record(op_df, model_name=model_name)
-                    input_tokens, output_tokens = self._est_tokens_per_record(op_df, model_name=model_name)
-                    selectivity = self._est_selectivity(self.sample_execution_data_df, op_df, model_name=model_name)
-                    quality = self._est_quality(op_df, model_name=model_name)
-
-                    model_estimates = {
-                        "time_per_record": time_per_record,
-                        "cost_per_record": cost_per_record,
-                        "total_input_tokens": input_tokens,
-                        "total_output_tokens": output_tokens,
-                        "selectivity": selectivity,
-                        "quality": quality,
-                    }
-                    estimates[model_name] = model_estimates
-
-            # TODO pre-compute lists of op_names in groups
-            elif op_name in ["NonLLMFilter"]:
-                time_per_record = self._est_time_per_record(op_df)
-                selectivity = self._est_selectivity(self.sample_execution_data_df, op_df)
-                estimates = {"time_per_record": time_per_record, "selectivity": selectivity}
-
-            elif op_name in ["MarshalAndScanDataOp", "CacheScanDataOp", "LimitScanOp", "CountAggregateOp", "AverageAggregateOp"]:
-                time_per_record = self._est_time_per_record(op_df)
-                estimates = {"time_per_record": time_per_record}
-
-            elif op_name in ["ApplyGroupByOp"]:
-                time_per_record = self._est_time_per_record(op_df)
-                cardinality = self._est_cardinality(op_df)
-                estimates = {"time_per_record": time_per_record, "cardinality": cardinality}
-
-            operator_estimates[full_op_id] = estimates
-
-        return operator_estimates
-
-    def __call__(self, operator: PhysicalOperator, source_op_estimates: OperatorCostEstimates | None = None) -> PlanCost:
-        # get identifier for operation which is unique within sentinel plan but consistent across sentinels
+    def _compute_naive_plan_cost(self, operator: PhysicalOperator, source_op_estimates: OperatorCostEstimates | None = None, right_source_op_estimates: OperatorCostEstimates | None = None) -> PlanCost:
+        # get identifier for operator which is unique within sentinel plan but consistent across sentinels
         full_op_id = operator.get_full_op_id()
         logger.debug(f"Calling __call__ for {str(operator)} with full_op_id: {full_op_id}")
 
         # initialize estimates of operator metrics based on naive (but sometimes precise) logic
         if isinstance(operator, MarshalAndScanDataOp):
             # get handle to scan operator and pre-compute its size (number of records)
-            datareader_len = len(operator.datareader)
+            datasource_len = len(operator.datasource)
 
             source_op_estimates = OperatorCostEstimates(
-                cardinality=datareader_len,
+                cardinality=datasource_len,
                 time_per_record=0.0,
                 cost_per_record=0.0,
                 quality=1.0,
@@ -498,88 +164,30 @@ class CostModel(BaseCostModel):
 
             op_estimates = operator.naive_cost_estimates(source_op_estimates, input_record_size_in_bytes=NAIVE_BYTES_PER_RECORD)
 
-        elif isinstance(operator, CacheScanDataOp):
-            datareader_len = len(operator.datareader)
-
+        elif isinstance(operator, ContextScanOp):
             source_op_estimates = OperatorCostEstimates(
-                cardinality=datareader_len,
+                cardinality=1.0,
                 time_per_record=0.0,
                 cost_per_record=0.0,
                 quality=1.0,
             )
 
-            op_estimates = operator.naive_cost_estimates(source_op_estimates, input_record_size_in_bytes=NAIVE_BYTES_PER_RECORD)
+            op_estimates = operator.naive_cost_estimates(source_op_estimates)
+
+        elif isinstance(operator, JoinOp):
+            op_estimates = operator.naive_cost_estimates(source_op_estimates, right_source_op_estimates)
 
         else:
             op_estimates = operator.naive_cost_estimates(source_op_estimates)
 
-        # if we have sample execution data, update naive estimates with more informed ones
-        sample_op_estimates = self.operator_estimates
-        if sample_op_estimates is not None and full_op_id in sample_op_estimates:
-            if isinstance(operator, (MarshalAndScanDataOp, CacheScanDataOp)):
-                op_estimates.time_per_record = sample_op_estimates[full_op_id]["time_per_record"]
-
-            elif isinstance(operator, ApplyGroupByOp):
-                # NOTE: in theory we should also treat this cardinality est. as a random variable, but in practice we will
-                #       have K samples of the number of groups produced by the groupby operator, where K is the number of
-                #       plans we generate sample data with. For now, we will simply use the estimate without bounds.
-                #
-                # NOTE: this cardinality is the only cardinality we estimate directly b/c we can observe how many groups are
-                #       produced by the groupby in our sample and assume it may generalize to the full workload. To estimate
-                #       actual cardinalities of operators we estimate their selectivities / fan-outs and multiply those by
-                #       the input cardinality (where the initial input cardinality from the datareader is known).
-                op_estimates.cardinality = sample_op_estimates[full_op_id]["cardinality"]
-                op_estimates.time_per_record = sample_op_estimates[full_op_id]["time_per_record"]
-
-            elif isinstance(operator, (CountAggregateOp, AverageAggregateOp)):  # noqa: SIM114
-                op_estimates.time_per_record = sample_op_estimates[full_op_id]["time_per_record"]
-
-            elif isinstance(operator, LimitScanOp):
-                op_estimates.time_per_record = sample_op_estimates[full_op_id]["time_per_record"]
-
-            elif isinstance(operator, NonLLMFilter):
-                op_estimates.cardinality = source_op_estimates.cardinality * sample_op_estimates[full_op_id]["selectivity"]
-                op_estimates.time_per_record = sample_op_estimates[full_op_id]["time_per_record"]
-
-            elif isinstance(operator, LLMFilter):
-                model_name = operator.model.value
-                op_estimates.cardinality = source_op_estimates.cardinality * sample_op_estimates[full_op_id][model_name]["selectivity"]
-                op_estimates.time_per_record = sample_op_estimates[full_op_id][model_name]["time_per_record"]
-                op_estimates.cost_per_record = sample_op_estimates[full_op_id][model_name]["cost_per_record"]
-                op_estimates.quality = sample_op_estimates[full_op_id][model_name]["quality"]
-
-            elif isinstance(operator, LLMConvert):
-                # NOTE: code synthesis does not have a model attribute
-                model_name = operator.model.value if hasattr(operator, "model") else None
-                op_estimates.cardinality = source_op_estimates.cardinality * sample_op_estimates[full_op_id][model_name]["selectivity"]
-                op_estimates.time_per_record = sample_op_estimates[full_op_id][model_name]["time_per_record"]
-                op_estimates.cost_per_record = sample_op_estimates[full_op_id][model_name]["cost_per_record"]
-                op_estimates.quality = sample_op_estimates[full_op_id][model_name]["quality"]
-
-                # NOTE: if code synth. fails, this will turn into ConventionalQuery calls to GPT-3.5,
-                #       which would wildly mess up estimate of time and cost per-record
-                # do code synthesis adjustment
-                if isinstance(operator, CodeSynthesisConvert):
-                    op_estimates.time_per_record = 1e-5
-                    op_estimates.cost_per_record = 1e-4
-                    op_estimates.quality = op_estimates.quality * (GPT_4o_MODEL_CARD["code"] / 100.0)
-
-                # rag convert adjustment
-                if isinstance(operator, RAGConvert):
-                    total_input_tokens = operator.num_chunks_per_field * operator.chunk_size
-                    total_output_tokens = sample_op_estimates[full_op_id][model_name]["total_output_tokens"]
-                    op_estimates.cost_per_record = (
-                        MODEL_CARDS[model_name]["usd_per_input_token"] * total_input_tokens
-                        + MODEL_CARDS[model_name]["usd_per_output_token"] * total_output_tokens
-                    )
-                    op_estimates.quality = op_estimates.quality * operator.naive_quality_adjustment
-
-            else:
-                raise Exception("Unknown operator")
-
         # compute estimates for this operator
-        op_time = op_estimates.time_per_record * source_op_estimates.cardinality
-        op_cost = op_estimates.cost_per_record * source_op_estimates.cardinality
+        est_input_cardinality = (
+            source_op_estimates.cardinality * right_source_op_estimates.cardinality
+            if isinstance(operator, JoinOp)
+            else source_op_estimates.cardinality
+        )
+        op_time = op_estimates.time_per_record * est_input_cardinality
+        op_cost = op_estimates.cost_per_record * est_input_cardinality
         op_quality = op_estimates.quality
 
         # create and return PlanCost object for this op's statistics
@@ -593,3 +201,62 @@ class CostModel(BaseCostModel):
         logger.debug(f"Plan cost: {op_plan_cost}")
 
         return op_plan_cost
+
+    def __call__(self, operator: PhysicalOperator, source_op_estimates: OperatorCostEstimates | None = None, right_source_op_estimates: OperatorCostEstimates | None = None) -> PlanCost:
+        # for non-sentinel execution, we use naive estimates
+        full_op_id = operator.get_full_op_id()
+        unique_logical_op_id = operator.unique_logical_op_id
+        if self.operator_to_stats is None or unique_logical_op_id not in self.operator_to_stats:
+            return self._compute_naive_plan_cost(operator, source_op_estimates, right_source_op_estimates)
+
+        # NOTE: some physical operators may not have any sample execution data in this cost model;
+        #       these physical operators are filtered out of the Optimizer, thus we can assume that
+        #       we will have execution data for each operator passed into __call__; nevertheless, we
+        #       still perform a sanity check
+        # look up physical and logical op ids associated with this physical operator
+        physical_op_to_stats = self.operator_to_stats.get(unique_logical_op_id)
+        assert physical_op_to_stats is not None, f"No execution data for logical operator: {str(operator)}"
+        assert physical_op_to_stats.get(full_op_id) is not None, f"No execution data for physical operator: {str(operator)}"
+        logger.debug(f"Calling __call__ for {str(operator)}")
+
+        # look up stats for this operation
+        est_cost_per_record = self.operator_to_stats[unique_logical_op_id][full_op_id]["cost"]
+        est_time_per_record = self.operator_to_stats[unique_logical_op_id][full_op_id]["time"]
+        est_quality = self.operator_to_stats[unique_logical_op_id][full_op_id]["quality"]
+        est_selectivity = self.operator_to_stats[unique_logical_op_id][full_op_id]["selectivity"]
+
+        # create source_op_estimates for scan operators if they are not provided
+        if isinstance(operator, ScanPhysicalOp):
+            # get handle to scan operator and pre-compute its size (number of records)
+            datasource_len = len(operator.datasource)
+
+            source_op_estimates = OperatorCostEstimates(
+                cardinality=datasource_len,
+                time_per_record=0.0,
+                cost_per_record=0.0,
+                quality=1.0,
+            )
+
+        # generate new set of OperatorCostEstimates
+        est_input_cardinality = (
+            source_op_estimates.cardinality * right_source_op_estimates.cardinality
+            if isinstance(operator, JoinOp)
+            else source_op_estimates.cardinality
+        )
+        op_estimates = OperatorCostEstimates(
+            cardinality=est_selectivity * est_input_cardinality,
+            time_per_record=est_time_per_record,
+            cost_per_record=est_cost_per_record,
+            quality=est_quality,
+        )
+
+        # compute estimates for this operator
+        op_time = op_estimates.time_per_record * est_input_cardinality
+        op_cost = op_estimates.cost_per_record * est_input_cardinality
+        op_quality = op_estimates.quality
+
+        # construct and return op estimates
+        plan_cost = PlanCost(cost=op_cost, time=op_time, quality=op_quality, op_estimates=op_estimates)
+        logger.debug(f"Done calling __call__ for {str(operator)}")
+        logger.debug(f"Plan cost: {plan_cost}")
+        return plan_cost