mirror
/
mongo
mirror of https://github.com/mongodb/mongo
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
mongo/buildscripts/cost_model/end_to_end.py

466 lines
20 KiB
Python
Raw Permalink Blame History

# Copyright (C) 2022-present MongoDB, Inc.
#
# This program is free software: you can redistribute it and/or modify
# it under the terms of the Server Side Public License, version 1,
# as published by MongoDB, Inc.
#
# This program is distributed in the hope that it will be useful,
# but WITHOUT ANY WARRANTY; without even the implied warranty of
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
# Server Side Public License for more details.
#
# You should have received a copy of the Server Side Public License
# along with this program. If not, see
# <http://www.mongodb.com/licensing/server-side-public-license>.
#
# As a special exception, the copyright holders give permission to link the
# code of portions of this program with the OpenSSL library under certain
# conditions as described in each individual source file and distribute
# linked combinations including the program with the OpenSSL library. You
# must comply with the Server Side Public License in all respects for
# all of the code used other than as permitted herein. If you modify file(s)
# with this exception, you may extend this exception to your version of the
# file(s), but you are not obligated to do so. If you do not wish to do so,
# delete this exception statement from your version. If you delete this
# exception statement from all source files in the program, then also delete
# it in the license file.
"""
End2End testing.
The test executes the given query pipelines with the given Cost Model Coefficients and compares
the predicted cost of every ABT node with the actual running time of the nodes.
It produces descriptive statistics (mean, stddev, min, max) and calculates R2 to
estimate quality of the tested Cost Model.
"""
from typing import Callable, Sequence, Tuple
import os
import asyncio
import dataclasses
import pandas as pd
import numpy as np
from sklearn.metrics import r2_score
from calibration_settings import main_config, HIDDEN_STRING_VALUE, distributions
from database_instance import DatabaseInstance, get_database_parameter
from random_generator import RandomDistribution
from data_generator import CollectionInfo, DataGenerator
from benchmark import CostModelCoefficients
from workload_execution import Query
import workload_execution
import config
import experiment as exp
import physical_tree as pt
import execution_tree as et
from parameters_extractor import get_excution_stats
from cost_estimator import ExecutionStats
class CostEstimator:
"""Estimates execution cost of ABT nodes."""
def __init__(self, cost_model: CostModelCoefficients):
"""Initialize cost estimator."""
self.cost_model = cost_model
self.estimators = {
'PhysicalScan': self.physical_scan,
'IndexScan': self.index_scan,
'Seek': self.seek,
'Filter': self.filter,
'Evaluation': self.evaluation,
'GroupBy': self.group_by,
'Unwind': self.unwind,
'NestedLoopJoin': self.nested_loop_join,
'HashJoin': self.hash_join,
'MergeJoin': self.merge_join,
'Unique': self.unique,
'Union': self.union,
'LimitSkip': self.limit_skip,
'Root': self.root,
}
def estimate(self, abt_node_name: str, cardinality: int) -> float:
"""Estimate ABT node cost."""
estimator = self.estimators.get(abt_node_name, self.default_estimator)
return estimator(cardinality)
def physical_scan(self, cardinality: int) -> float:
"""Estinamate PhysicalScan ABT node."""
return self.cost_model.scan_startup_cost + cardinality * self.cost_model.scan_incremental_cost
def index_scan(self, cardinality: int) -> float:
"""Estinamate IndexScan ABT node."""
return self.cost_model.index_scan_startup_cost + cardinality * self.cost_model.index_scan_incremental_cost
def seek(self, cardinality: int) -> float:
"""Estinamate Seek ABT node."""
return self.cost_model.seek_startup_cost + cardinality * self.cost_model.seek_cost
def filter(self, cardinality: int) -> float:
"""Estinamate Filter ABT node."""
return self.cost_model.filter_startup_cost + cardinality * self.cost_model.filter_incremental_cost
def evaluation(self, cardinality: int) -> float:
"""Estinamate Evaluation ABT node."""
return self.cost_model.eval_startup_cost + cardinality * self.cost_model.eval_incremental_cost
def group_by(self, cardinality: int) -> float:
"""Estinamate GroupBy ABT node."""
return self.cost_model.group_by_startup_cost + cardinality * self.cost_model.group_by_incremental_cost
def unwind(self, cardinality: int) -> float:
"""Estinamate Unwind ABT node."""
return self.cost_model.unwind_startup_cost + cardinality * self.cost_model.unwind_incremental_cost
def nested_loop_join(self, cardinality: int) -> float:
"""Estinamate NestedLoopJoin ABT node."""
return self.cost_model.nested_loop_join_startup_cost + cardinality * self.cost_model.nested_loop_join_incremental_cost
def hash_join(self, cardinality: int) -> float:
"""Estinamate HashJoin ABT node."""
return self.cost_model.hash_join_startup_cost + cardinality * self.cost_model.hash_join_incremental_cost
def merge_join(self, cardinality: int) -> float:
"""Estinamate MergeJoin ABT node."""
return self.cost_model.merge_join_startup_cost + cardinality * self.cost_model.merge_join_incremental_cost
def unique(self, cardinality: int) -> float:
"""Estinamate Unique ABT node."""
return self.cost_model.unique_startup_cost + cardinality * self.cost_model.unique_incremental_cost
def union(self, cardinality: int) -> float:
"""Estinamate Union ABT node."""
return self.cost_model.union_startup_cost + cardinality * self.cost_model.union_incremental_cost
def limit_skip(self, cardinality: int) -> float:
"""Estinamate LimitSkip ABT node."""
return self.cost_model.limit_skip_startup_cost + cardinality * self.cost_model.limit_skip_incremental_cost
def root(self, _: int) -> float:
"""Root ABT node is always 0."""
return 0.0
def default_estimator(self, _: int) -> float:
"""Used if no ABT nodes matched."""
return -1e10
class AbtCostEstimator:
"""Calculates a cost for the given ABT tree."""
def __init__(self, estimate_node: Callable[[str, int], float]):
self.estimate_node = estimate_node
def estimate(self, abt: pt.Node, sbe: et.Node,
estimations: Sequence[Tuple[str, ExecutionStats, float]], level=0):
stats = get_excution_stats(sbe, abt.plan_node_id)
local_cost = self.estimate_node(abt.node_type, stats.n_processed)
estimations.append((abt.node_type, stats, local_cost))
child_cost = sum((self.estimate(child, sbe, estimations, level + 1)
for child in abt.children), start=0.0)
return local_cost + child_cost
@dataclasses.dataclass(init=False)
class EndToEndStatisticsRow:
"""Represents a row with descriptive statistics of one query execution."""
def __init__(self, pipeline: str = None, abt_type: str = None, abt_type_id: int = 0,
execution_time: float = 0.0, estimated_cost: float = 0.0, n_documents: int = 0):
self.pipeline = pipeline if pipeline is not None else ''
self.abt_type = abt_type if abt_type is not None else ''
self.abt_type_id = abt_type_id
self.execution_time = execution_time
self.estimated_cost = estimated_cost
self.estimation_error = execution_time - estimated_cost
self.estimation_error_per_doc = self.estimation_error / n_documents if n_documents != 0 else 0
self.relative_error = self.estimation_error / self.execution_time if self.execution_time != 0 else 0
pipeline: str
abt_type: str
abt_type_id: int
execution_time: float
estimated_cost: float
estimation_error: float
estimation_error_per_doc: float
relative_error: float
def make_config():
def create_end2end_collection_template(name: str,
cardinality: int) -> config.CollectionTemplate:
values = [
'iqtbr5b5is', 'vt5s3tf8o6', 'b0rgm58qsn', '9m59if353m', 'biw2l9ok17', 'b9ct0ue14d',
'oxj0vxjsti', 'f3k8w9vb49', 'ec7v82k6nk', 'f49ufwaqx7'
]
start_weight = 30
step_weight = 250
finish_weight = start_weight + len(values) * step_weight
weights = list(range(start_weight, finish_weight, step_weight))
fill_up_weight = cardinality - sum(weights)
if fill_up_weight > 0:
values.append(HIDDEN_STRING_VALUE)
weights.append(fill_up_weight)
distr = RandomDistribution.choice(values, weights)
return config.CollectionTemplate(
name=name, fields=[
config.FieldTemplate(name="indexed_choice", data_type=config.DataType.STRING,
distribution=distr, indexed=True),
config.FieldTemplate(name="int1", data_type=config.DataType.INTEGER,
distribution=distributions["int_normal"], indexed=True),
config.FieldTemplate(name="non_indexed_choice", data_type=config.DataType.STRING,
distribution=distributions['string_choice'], indexed=False),
config.FieldTemplate(name="uniform1", data_type=config.DataType.STRING,
distribution=distributions["string_uniform"], indexed=False),
config.FieldTemplate(name="int2", data_type=config.DataType.INTEGER,
distribution=distributions["int_normal"], indexed=True),
config.FieldTemplate(name="choice2", data_type=config.DataType.STRING,
distribution=distributions["string_choice"], indexed=False),
config.FieldTemplate(name="mixed2", data_type=config.DataType.STRING,
distribution=distributions["string_mixed"], indexed=False),
], compound_indexes=[], cardinalities=[cardinality])
col_end2end = create_end2end_collection_template('end2end', 2000000)
data_generator_config = config.DataGeneratorConfig(
enabled=True, create_indexes=True, batch_size=10000, collection_templates=[col_end2end],
write_mode=config.WriteMode.REPLACE, collection_name_with_card=True)
workload_execution_config = config.WorkloadExecutionConfig(
enabled=True, output_collection_name='end2endData', write_mode=config.WriteMode.APPEND,
warmup_runs=3, runs=30)
# The cost model to test.
cost_model = CostModelCoefficients(
scan_incremental_cost=422.31145989, scan_startup_cost=6175.527218993269,
index_scan_incremental_cost=403.68075869, index_scan_startup_cost=14054.983953111061,
seek_cost=1223.35513997, seek_startup_cost=7488.662376624863,
filter_incremental_cost=83.7274685, filter_startup_cost=1461.3148783443378,
eval_incremental_cost=430.6176946, eval_startup_cost=1103.4048573163343,
group_by_incremental_cost=413.07932374, group_by_startup_cost=1199.8878012735659,
unwind_incremental_cost=586.57200195, unwind_startup_cost=1.0,
nested_loop_join_incremental_cost=161.62301944,
nested_loop_join_startup_cost=402.8455479458652, hash_join_incremental_cost=250.61365634,
hash_join_startup_cost=1.0, merge_join_incremental_cost=111.23423304,
merge_join_startup_cost=1517.7970800404169, unique_incremental_cost=269.71368614,
unique_startup_cost=1.0, union_incremental_cost=111.94945268,
union_startup_cost=69.88096657391543, limit_skip_incremental_cost=62.42111111,
limit_skip_startup_cost=655.1342592592522)
cost_estimator = CostEstimator(cost_model)
processor_config = config.End2EndProcessorConfig(
enabled=True, estimator=cost_estimator.estimate,
input_collection_name=workload_execution_config.output_collection_name)
return config.EntToEndTestingConfig(
database=main_config.database, data_generator=data_generator_config,
workload_execution=workload_execution_config, processor=processor_config,
result_csv_filepath="end2end.csv")
async def execute_queries(database: DatabaseInstance, we_config: config.WorkloadExecutionConfig,
collections: Sequence[CollectionInfo]):
collection = [ci for ci in collections if ci.name.startswith('end2end')][0]
requests = []
limits = [5, 10, 15, 20, 25, 50]
skips = [15, 10, 5]
for field in [f for f in collection.fields if f.name == 'indexed_choice']:
for val in field.distribution.get_values():
if val.startswith('_'):
continue
limit = limits[len(requests) % len(limits)]
skip = skips[len(requests) % len(skips)]
requests.append(
Query(pipeline=[{'$match': {field.name: val}}, {"$skip": skip}, {"$limit": limit},
{"$project": {"int1": 1}}]))
for field in [f for f in collection.fields if f.name == 'non_indexed_choice']:
for val in ['chisquare', 'hi']:
limit = limits[len(requests) % len(limits)]
skip = skips[len(requests) % len(skips)]
requests.append(
Query(pipeline=[{'$match': {field.name: val}}, {"$skip": skip}, {"$limit": limit},
{"$project": {"int1": 1}}]))
for i in range(100, 1000, 250):
limit = limits[len(requests) % len(limits)]
skip = skips[len(requests) % len(skips)]
requests.append(
Query(pipeline=[{'$match': {'in1': i, 'in2': 1000 -
i}}, {"$skip": skip}, {"$limit": limit}]))
requests.append(
Query(pipeline=[{'$match': {'in1': {'$lte': i}, 'in2': 1000 - i}}, {"$skip": skip},
{"$limit": limit}]))
await workload_execution.execute(database, we_config, [collection], requests)
async def execute_index_intersect_queries(database: DatabaseInstance,
we_config: config.WorkloadExecutionConfig,
collections: Sequence[CollectionInfo]):
collection = [ci for ci in collections if ci.name.startswith('end2end')][0]
requests = []
limits = [5, 10, 15, 20, 25, 50]
skips = [15, 10, 5]
for i in range(100, 1000, 250):
limit = limits[len(requests) % len(limits)]
skip = skips[len(requests) % len(skips)]
requests.append(
Query(pipeline=[{'$match': {'in1': i, 'in2': 1000 -
i}}, {"$skip": skip}, {"$limit": limit}]))
requests.append(
Query(pipeline=[{'$match': {'in1': {'$lte': i}, 'in2': 1000 - i}}, {"$skip": skip},
{"$limit": limit}]))
async with get_database_parameter(
database, 'internalCostModelCoefficients') as cost_model_param, get_database_parameter(
database, 'internalCascadesOptimizerDisableMergeJoinRIDIntersect'
) as merge_join_param, get_database_parameter(
database,
'internalCascadesOptimizerDisableHashJoinRIDIntersect') as hash_join_param:
await cost_model_param.set('{"filterIncrementalCost": 10000.0}')
await merge_join_param.set(False)
await hash_join_param.set(False)
await workload_execution.execute(database, we_config, [collection], requests)
await merge_join_param.set(True)
await hash_join_param.set(True)
await workload_execution.execute(database, we_config, [collection], requests)
def extract_abt_nodes(df: pd.DataFrame, estimate_cost) -> pd.DataFrame:
"""Extract ABT Nodes and execution statistics from calibration DataFrame."""
def extract(df_seq):
es_dict = exp.extract_execution_stats(df_seq['sbe'], df_seq['abt'], [])
rows = []
for abt_type, es in es_dict.items():
for stat in es:
if stat.n_processed == 0:
continue
estimated_cost = estimate_cost(abt_type, stat.n_processed)
rows.append(
EndToEndStatisticsRow(abt_type=abt_type, execution_time=stat.execution_time,
estimated_cost=estimated_cost,
n_documents=stat.n_processed))
return rows
return pd.DataFrame(list(df.apply(extract, axis=1).explode()))
def build_abt_nodes_report(df: pd.DataFrame, processor_config: config.End2EndProcessorConfig):
return extract_abt_nodes(df, processor_config.estimator)
def build_queries_report(df: pd.DataFrame, processor_config: config.End2EndProcessorConfig):
abt_estimator = AbtCostEstimator(processor_config.estimator)
def calculate_cost(row):
rows = []
estimations = []
total_estimated_cost = abt_estimator.estimate(row['abt'], row['sbe'], estimations)
local_id = 0
rows.append(
EndToEndStatisticsRow(pipeline=row['pipeline'], abt_type_id=local_id,
execution_time=row['total_execution_time'],
estimated_cost=total_estimated_cost))
for (abt_type, stats, local_cost) in estimations:
local_id += 1
rows.append(
EndToEndStatisticsRow(pipeline=row['pipeline'], abt_type=abt_type,
abt_type_id=local_id,
execution_time=row['total_execution_time'],
estimated_cost=local_cost, n_documents=stats.n_processed))
return rows
return pd.DataFrame(list(df.apply(calculate_cost, axis=1).explode()))
async def conduct_end2end(database: DatabaseInstance,
processor_config: config.End2EndProcessorConfig):
if not processor_config.enabled:
return {}
df = await exp.load_calibration_data(database, processor_config.input_collection_name)
noout_df = exp.remove_outliers(df, 0.0, 0.90)
abt_report = build_abt_nodes_report(noout_df, processor_config)
queries_report = build_queries_report(noout_df, processor_config)
report = pd.concat([abt_report, queries_report], axis=0)
group_columns = ['pipeline', 'abt_type', 'abt_type_id']
def calc_r2(group):
return r2_score(group['execution_time'], group['estimated_cost'])
r2_scores = report.groupby(group_columns).apply(calc_r2).reset_index()
r2_scores.columns = [group_columns + ['r2'], [''] * (len(group_columns) + 1)]
agg_stats = report.groupby(group_columns)[[
'execution_time', 'estimated_cost', 'estimation_error', 'estimation_error_per_doc',
'relative_error'
]].agg([np.mean, np.std, np.min, np.max])
report = pd.merge(r2_scores, agg_stats, on=group_columns)
del report['abt_type_id']
return report
async def end2end(e2e_config: config.EntToEndTestingConfig):
script_directory = os.path.abspath(os.path.dirname(__file__))
os.chdir(script_directory)
# 1. Database Instance provides connectivity to a MongoDB instance, it loads data optionally
# from the dump on creating and stores data optionally to the dump on closing.
with DatabaseInstance(e2e_config.database) as database:
# 2. Data generation (optional), generates random data and populates collections with it.
generator = DataGenerator(database, e2e_config.data_generator)
await generator.populate_collections()
# 3. Collecting data for calibration (optional).
# It runs the pipelines and stores explains to the database.
execution_query_functions = [execute_queries, execute_index_intersect_queries]
for execute_query in execution_query_functions:
await execute_query(database, e2e_config.workload_execution, generator.collection_infos)
e2e_config.workload_execution.write_mode = config.WriteMode.APPEND
#4. Process end to end testing. Compare the estimated and actual costs and return results.
report = await conduct_end2end(database, e2e_config.processor)
if e2e_config.result_csv_filepath is not None:
report.to_csv(e2e_config.result_csv_filepath, index=False)
async def main():
e2e_config = make_config()
await end2end(e2e_config)
if __name__ == '__main__':
loop = asyncio.new_event_loop()
asyncio.set_event_loop(loop)
try:
asyncio.run(main())
except KeyboardInterrupt:
pass
Reference in New Issue View Git Blame Copy Permalink