def test_scan_empty():
"""Executes a scan where no records are returned. We tst this as it's somewhat peculiar with s3 select, in so much
as s3 does not return column names when selecting data, meaning, unlike a traditional DBMS, no field names tuple
should be present in the results.
:return: None
"""
query_plan = QueryPlan()
# Query plan
ts = query_plan.add_operator(
SQLTableScan('nation.csv', "select * from s3object limit 0", False,
'ts', query_plan, False))
c = query_plan.add_operator(Collate('c', query_plan, False))
ts.connect(c)
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../tests-output"),
gen_test_id())
# Start the query
query_plan.execute()
# Assert the results
# num_rows = 0
# for t in c.tuples():
# num_rows += 1
# print("{}:{}".format(num_rows, t))
assert len(c.tuples()) == 0
# Write the metrics
query_plan.print_metrics()
def test_scan_simple():
"""Executes a scan. The results are then collated.
:return: None
"""
query_plan = QueryPlan()
# Query plan
ts = query_plan.add_operator(
TableScan('nation.csv', 'ts', query_plan, False))
c = query_plan.add_operator(Collate('c', query_plan, False))
ts.connect(c)
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../tests-output"),
gen_test_id())
# Start the query
query_plan.execute()
# Assert the results
# num_rows = 0
# for t in c.tuples():
# num_rows += 1
# print("{}:{}".format(num_rows, t))
assert len(c.tuples()) == 25 + 1
assert c.tuples()[0] == ['_0', '_1', '_2', '_3']
assert c.tuples()[1] == [
'0', 'ALGERIA', '0',
' haggle. carefully final deposits detect slyly agai'
]
assert c.tuples()[2] == [
'1', 'ARGENTINA', '1',
'al foxes promise slyly according to the regular accounts. bold requests alon'
]
assert c.tuples()[3] == [
'2', 'BRAZIL', '1',
'y alongside of the pending deposits. carefully special packages are about '
'the ironic forges. slyly special '
]
# Write the metrics
query_plan.print_metrics()
def test_random_scan_simple():
"""Executes a random scan. The results are then collated.
:return: None
"""
query_plan = QueryPlan()
# Query plan
random_col_defs = [
RandomIntColumnDef(0, 9),
RandomStringColumnDef(10, 20),
RandomDateColumnDef(datetime.strptime('2017-01-01', '%Y-%m-%d'),
datetime.strptime('2018-01-01', '%Y-%m-%d'))
]
random_table_scan = query_plan.add_operator(
RandomTableScan(10,
random_col_defs,
'random_table_scan', query_plan,
False))
collate = query_plan.add_operator(
Collate('collate', query_plan, False))
random_table_scan.connect(collate)
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../tests-output"), gen_test_id())
# Start the query
query_plan.execute()
collate.print_tuples()
# Write the metrics
query_plan.print_metrics()
# Assert the results
assert len(collate.tuples()) == 10 + 1
assert collate.tuples()[0] == ['_0', '_1', '_2']
def test_group_sum():
"""Tests a group by query with a sum aggregate
:return: None
"""
query_plan = QueryPlan()
# Query plan
# select s_nationkey, sum(float(s_acctbal)) from supplier.csv group by s_nationkey
ts = query_plan.add_operator(
SQLTableScan('supplier.csv', 'select * from S3Object;', False, 'ts',
query_plan, False))
g = query_plan.add_operator(
Group(['_3'], [
AggregateExpression(AggregateExpression.SUM,
lambda t_: float(t_['_5']))
], 'g', query_plan, False))
c = query_plan.add_operator(Collate('c', query_plan, False))
ts.connect(g)
g.connect(c)
# Start the query
query_plan.execute()
# Assert the results
# num_rows = 0
# for t in c.tuples():
# num_rows += 1
# print("{}:{}".format(num_rows, t_))
field_names = ['_0', '_1']
assert c.tuples()[0] == field_names
assert len(c.tuples()) == 25 + 1
nation_24 = filter(
lambda t_: IndexedTuple.build(t_, field_names)['_0'] == '24',
c.tuples())[0]
assert round(nation_24[1], 2) == 1833872.56
# Write the metrics
query_plan.print_metrics()
def test_filter_1():
"""
:return:
"""
query_plan = QueryPlan(buffer_size=64, is_async=True, use_shared_mem=False)
# Query plan
ts = query_plan.add_operator(
SQLTableScan(
'tpch-sf1/lineitem_sharded/lineitem.csv.0',
'select * from S3Object where cast(l_extendedprice as float) >= 0 and cast(l_extendedprice as float)<= 910;',
Format.CSV, True, False, False, 'ts', query_plan, False))
# using a 'use_native=True' argument will result in a None object being returned
'''
ts = query_plan.add_operator(
SQLTableScan('tpch-sf1/lineitem_sharded/lineitem.csv.0', 'select l_partkey from S3Object limit 5;',Format.CSV , True, False,False, 'ts', query_plan, False))
#random_strings_2.csv
'''
c = query_plan.add_operator(Collate('c', query_plan, False))
ts.connect(c)
#f.connect(c)
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../tests-output"),
gen_test_id())
# Start the query
query_plan.execute()
#assert 2 + 1 == len(c.tuples())
print("Tuples:")
print(c.tuples())
# Write the metrics
query_plan.print_metrics()
print(ROOT_DIR)
def test_group_empty():
"""Executes a group where no records are returned. We tst this as it's somewhat peculiar with s3 select, in so much
as s3 does not return column names when selecting data, meaning, unlike a traditional DBMS, no field names tuple
should be present in the results.
:return: None
"""
query_plan = QueryPlan()
# Query plan
# select s_nationkey, sum(float(s_acctbal)) from supplier.csv group by s_nationkey
ts = query_plan.add_operator(
SQLTableScan('supplier.csv', 'select * from S3Object limit 0;', False,
'ts', query_plan, False))
g = query_plan.add_operator(
Group(['_3'], [
AggregateExpression(AggregateExpression.SUM,
lambda t_: float(t_['_5']))
], 'g', query_plan, False))
c = query_plan.add_operator(Collate('c', query_plan, False))
ts.connect(g)
g.connect(c)
# Start the query
query_plan.execute()
# Assert the results
# num_rows = 0
# for t in c.tuples():
# num_rows += 1
# print("{}:{}".format(num_rows, t))
field_names = ['_0', '_1']
assert c.tuples()[0] == field_names
assert len(c.tuples()) == 0 + 1
# Write the metrics
query_plan.print_metrics()
def test_filter_arg(object_, sql_query):
# Let's forget about the local filter for now. The pd._expr field of the PredicateExpression class is not well documented and it is needed for the Filter class (on line 102).
"""
:return:
"""
query_plan = QueryPlan(buffer_size=64, is_async=True, use_shared_mem=False)
# Query plan
'''
ts = query_plan.add_operator(
SQLTableScan('lineitem.csv', 'select * from S3Object limit 3;' , False, 'ts', query_plan, False))
'''
# using a 'use_native=True' argument will result in a None object being returned
ts = query_plan.add_operator(
SQLTableScan(object_, sql_query, Format.CSV, True, False, False, 'ts',
query_plan, False))
c = query_plan.add_operator(Collate('c', query_plan, False))
ts.connect(c)
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../tests-output"),
gen_test_id())
# Start the query
query_plan.execute()
#assert 2 + 1 == len(c.tuples())
print("Tuples:")
print(c.tuples())
# Write the metrics
query_plan.print_metrics()
print(ROOT_DIR)
def sample_table(s3key, k, sort_exp):
"""
Given a table name, return a random sample of records. Currently, the returned records are the first k tuples
:param s3key: the s3 object name
:param k: the number of tuples to return (the number added to the SQL Limit clause
:param sort_exp: the sort expression in which the topk is chosen upon
:return: the list of selected keys from the first k tuples in the table
"""
projection = "CAST({} as {})".format(sort_exp.col_name,
sort_exp.col_type.__name__)
sql = "SELECT {} FROM S3Object LIMIT {}".format(projection, k)
q_plan = QueryPlan(None, is_async=False)
select_op = q_plan.add_operator(
SQLTableScan(s3key, sql, True, True, False,
"sample_{}_scan".format(s3key), q_plan, False))
from copy import deepcopy
sample_topk_sort_exp = deepcopy(sort_exp)
sample_topk_sort_exp.col_index = '_0'
topk = q_plan.add_operator(
Top(max_tuples=k,
sort_expression=sample_topk_sort_exp,
use_pandas=True,
name="sampling_topk",
query_plan=q_plan,
log_enabled=False))
collate = q_plan.add_operator(
Collate("sample_{}_collate".format(s3key), q_plan, False))
select_op.connect(topk)
topk.connect(collate)
q_plan.execute()
q_plan.print_metrics()
return collate.tuples(), select_op, q_plan
def main():
parts = 32
query_plan = QueryPlan(is_async=True, buffer_size=0)
# Query plan
lineitem_scan = map(lambda p:
query_plan.add_operator(
SQLTableScan(get_file_key('lineitem', True, p),
"select * from S3Object;", Format.CSV,
use_pandas=True, secure=False, use_native=False,
name='scan_' + str(p), query_plan=query_plan,
log_enabled=False)),
range(0, parts))
collate = query_plan.add_operator(
Collate('collate', query_plan, False))
map(lambda (p, o): o.connect(collate), enumerate(lineitem_scan))
query_plan.execute()
def run_baseline_topk(stats, sort_field_index, sort_field, k, parallel,
use_pandas, sort_order, buffer_size, table_parts_start,
table_parts_end, tbl_s3key, format_, shards_path):
secure = False
use_native = False
print('')
print("Top K Benchmark, Baseline. Sort Field: {}, Order: {}, k: {}".format(
sort_field, sort_order, k))
print("----------------------")
stats += ['baseline', shards_path, sort_field, sort_order, k, 0, 0]
# Query plan
query_plan = QueryPlan(is_async=parallel, buffer_size=buffer_size)
# Sampling
table_parts = table_parts_end - table_parts_start + 1
per_part_samples = int(sample_size / table_parts)
table_name = os.path.basename(tbl_s3key)
# Scan
scan = map(
lambda p: query_plan.add_operator(
SQLTableScan("{}.{}".format(shards_path, p
), "select * from S3Object;", format_,
use_pandas, secure, use_native, 'scan_{}'.format(
p), query_plan, False)),
range(table_parts_start, table_parts_end + 1))
# Project
def project_fn(df):
df.columns = [
sort_field if x == sort_field_index else x for x in df.columns
]
df[[sort_field]] = df[[sort_field]].astype(np.float)
return df
project_exprs = [ProjectExpression(lambda t_: t_['_0'], sort_field)]
project = map(
lambda p: query_plan.add_operator(
Project(project_exprs, 'project_{}'.format(p), query_plan,
False, project_fn)),
range(table_parts_start, table_parts_end + 1))
# TopK
sort_expr = [SortExpression(sort_field, float, sort_order)]
topk = map(
lambda p: query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'topk_{}'.format(p), query_plan,
False)), range(table_parts_start, table_parts_end + 1))
# TopK reduce
topk_reduce = query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'topk_reduce', query_plan, False))
collate = query_plan.add_operator(Collate('collate', query_plan, False))
map(lambda (p, o): o.connect(project[p]), enumerate(scan))
map(lambda (p, o): o.connect(topk[p]), enumerate(project))
map(lambda (p, o): o.connect(topk_reduce), enumerate(topk))
topk_reduce.connect(collate)
# Start the query
query_plan.execute()
print('Done')
# Write the metrics
query_plan.print_metrics()
# Shut everything down
query_plan.stop()
query_time = query_plan.total_elapsed_time
cost, bytes_scanned, bytes_returned, rows = query_plan.cost()
computation_cost = query_plan.computation_cost()
data_cost = query_plan.data_cost()[0]
stats += [
0, 0, 0, query_time, rows, bytes_scanned, bytes_returned, data_cost,
computation_cost, cost
]
def run_head_table_sampling(stats,
sort_field_index,
sort_field,
k,
sample_size,
parallel,
use_pandas,
sort_order,
buffer_size,
table_parts_start,
table_parts_end,
tbl_s3key,
shards_path,
format_,
sampling_only=True):
secure = False
use_native = False
print('')
print(
"Top K Benchmark, Head Table Sampling. Sort Field: {}, Order: {}, k: {}, Sample Size:{}"
.format(sort_field, sort_order, k, sample_size))
print("----------------------")
stats += [
'sampling_{}_{}'.format('head_table', 'non-filtered'), shards_path,
sort_field, sort_order, k, sample_size, 1
]
# Query plan
query_plan = QueryPlan(is_async=parallel, buffer_size=buffer_size)
# Sampling
table_parts = table_parts_end - table_parts_start + 1
per_part_samples = int(sample_size / table_parts)
table_name = os.path.basename(tbl_s3key)
sample_scan = map(
lambda p: query_plan.add_operator(
SQLTableScan(
"{}.{}".format(shards_path, p),
'select {} from S3Object limit {};'.format(
sort_field, per_part_samples), format_, use_pandas, secure,
use_native, 'sample_scan_{}'.format(p), query_plan, False)),
range(table_parts_start, table_parts_end + 1))
# Sampling project
def project_fn(df):
df.columns = [sort_field]
df[[sort_field]] = df[[sort_field]].astype(np.float)
return df
project_exprs = [ProjectExpression(lambda t_: t_['_0'], sort_field)]
sample_project = map(
lambda p: query_plan.add_operator(
Project(project_exprs, 'sample_project_{}'.format(p), query_plan,
False, project_fn)),
range(table_parts_start, table_parts_end + 1))
# TopK samples
sort_expr = [SortExpression(sort_field, float, sort_order)]
sample_topk = map(
lambda p: query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'sample_topk_{}'.format(p),
query_plan, False)),
range(table_parts_start, table_parts_end + 1))
sample_topk_reduce = query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'sample_topk_reduce', query_plan, False))
# Generate SQL command for second scan
sql_gen = query_plan.add_operator(
TopKFilterBuild(sort_order, 'float', 'select * from S3object ',
' CAST({} as float) '.format(sort_field), 'sql_gen',
query_plan, False))
if not sampling_only:
# Scan
scan = map(
lambda p: query_plan.add_operator(
SQLTableScan("{}.{}".format(shards_path, p), "", format_,
use_pandas, secure, use_native, 'scan_{}'.format(
p), query_plan, False)),
range(table_parts_start, table_parts_end + 1))
# Project
def project_fn(df):
df.columns = [
sort_field if x == sort_field_index else x for x in df.columns
]
df[[sort_field]] = df[[sort_field]].astype(np.float)
return df
project_exprs = [ProjectExpression(lambda t_: t_['_0'], sort_field)]
project = map(
lambda p: query_plan.add_operator(
Project(project_exprs, 'project_{}'.format(p), query_plan,
False, project_fn)),
range(table_parts_start, table_parts_end + 1))
# TopK
topk = map(
lambda p: query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'topk_{}'.format(p), query_plan,
False)), range(table_parts_start, table_parts_end + 1))
# TopK reduce
topk_reduce = query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'topk_reduce', query_plan, False))
collate = query_plan.add_operator(Collate('collate', query_plan, False))
map(lambda (p, o): o.connect(sample_project[p]), enumerate(sample_scan))
map(lambda (p, o): o.connect(sample_topk[p]), enumerate(sample_project))
map(lambda op: op.connect(sample_topk_reduce), sample_topk)
sample_topk_reduce.connect(sql_gen)
if not sampling_only:
map(lambda (p, o): sql_gen.connect(o), enumerate(scan))
map(lambda (p, o): o.connect(project[p]), enumerate(scan))
map(lambda (p, o): o.connect(topk[p]), enumerate(project))
map(lambda (p, o): o.connect(topk_reduce), enumerate(topk))
topk_reduce.connect(collate)
else:
sql_gen.connect(collate)
# Start the query
query_plan.execute()
print('Done')
# Write the metrics
query_plan.print_metrics()
# Shut everything down
query_plan.stop()
sampling_time = query_plan.total_elapsed_time
cost, bytes_scanned, bytes_returned, rows = query_plan.cost()
computation_cost = query_plan.computation_cost()
data_cost = query_plan.data_cost()[0]
stats += [
sql_gen.threshold, sampling_time, 0, sampling_time, rows,
bytes_scanned, bytes_returned, data_cost, computation_cost, cost
]
def run_local_indexed_sampling(stats,
sort_field_index,
sort_field,
k,
sample_size,
batch_size,
parallel,
use_pandas,
sort_order,
buffer_size,
table_parts_start,
table_parts_end,
tbl_s3key,
shards_path,
format_,
sampling_only=True):
"""
Executes the randomly sampled topk query by firstly building a random sample, then extracting the filtering threshold
Finally scanning the table to retrieve only the records beyond the threshold
:return:
"""
secure = False
use_native = False
n_threads = multiprocessing.cpu_count()
print('')
print("Top K Benchmark, Sampling. Sort Field: {}, Order: {}".format(
sort_field, sort_order))
print("----------------------")
stats += [
'sampling_{}_{}'.format('indexed', 'non-filtered'), shards_path,
sort_field, sort_order, k, sample_size, batch_size
]
# Query plan
query_plan = QueryPlan(is_async=parallel, buffer_size=buffer_size)
# Sampling
tbl_smpler = query_plan.add_operator(
TableRandomSampleGenerator(tbl_s3key, sample_size, batch_size,
"table_sampler", query_plan, False))
sample_scan = map(
lambda p: query_plan.add_operator(
TableRangeAccess(tbl_s3key, use_pandas, secure, use_native,
"sample_scan_{}".format(p), query_plan, False)),
range(table_parts_start, table_parts_end + 1))
map(lambda (i, op): sample_scan[i].set_nthreads(n_threads),
enumerate(sample_scan))
# sample_scan = query_plan.add_operator(
# TableRangeAccess(tbl_s3key, use_pandas, secure, use_native, "sample_scan_{}".format(p),
# query_plan, False))
# sample_scan.set_nthreads(n_threads)
# Sampling project
def project_fn(df):
df.columns = [
sort_field if x == sort_field_index else x for x in df.columns
]
df = df[[sort_field]].astype(np.float, errors='ignore')
return df
project_exprs = [ProjectExpression(lambda t_: t_['_0'], sort_field)]
sample_project = map(
lambda p: query_plan.add_operator(
Project(project_exprs, 'sample_project_{}'.format(p), query_plan,
False, project_fn)),
range(table_parts_start, table_parts_end + 1))
# sample_project = query_plan.add_operator(
# Project(project_exprs, 'sample_project_{}'.format(p), query_plan, False, project_fn))
# TopK samples
sort_expr = [SortExpression(sort_field, float, sort_order)]
sample_topk = map(
lambda p: query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'sample_topk_{}'.format(p),
query_plan, False)),
range(table_parts_start, table_parts_end + 1))
# sample_topk = query_plan.add_operator(
# Top(k, sort_expr, use_pandas, 'sample_topk_{}'.format(p), query_plan, False))
sample_topk_reduce = query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'sample_topk_reduce', query_plan, False))
# Generate SQL command for second scan
sql_gen = query_plan.add_operator(
TopKFilterBuild(sort_order, 'float', 'select * from S3object ',
' CAST({} as float) '.format(sort_field), 'sql_gen',
query_plan, False))
if not sampling_only:
# Scan
scan = map(
lambda p: query_plan.add_operator(
SQLTableScan("{}.{}".format(shards_path, p), "", format_,
use_pandas, secure, use_native, 'scan_{}'.format(
p), query_plan, False)),
range(table_parts_start, table_parts_end + 1))
# Project
def project_fn(df):
df.columns = [
sort_field if x == sort_field_index else x for x in df.columns
]
df[[sort_field]] = df[[sort_field]].astype(np.float)
return df
project_exprs = [ProjectExpression(lambda t_: t_['_0'], sort_field)]
project = map(
lambda p: query_plan.add_operator(
Project(project_exprs, 'project_{}'.format(p), query_plan,
False, project_fn)),
range(table_parts_start, table_parts_end + 1))
# TopK
topk = map(
lambda p: query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'topk_{}'.format(p), query_plan,
False)), range(table_parts_start, table_parts_end + 1))
# TopK reduce
topk_reduce = query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'topk_reduce', query_plan, False))
collate = query_plan.add_operator(Collate('collate', query_plan, False))
map(lambda o: tbl_smpler.connect(o), sample_scan)
map(lambda (p, o): o.connect(sample_project[p]), enumerate(sample_scan))
map(lambda (p, o): o.connect(sample_topk[p]), enumerate(sample_project))
map(lambda o: o.connect(sample_topk_reduce), sample_topk)
sample_topk_reduce.connect(sql_gen)
if not sampling_only:
map(lambda (p, o): sql_gen.connect(o), enumerate(scan))
map(lambda (p, o): o.connect(project[p]), enumerate(scan))
map(lambda (p, o): o.connect(topk[p]), enumerate(project))
map(lambda (p, o): o.connect(topk_reduce), enumerate(topk))
topk_reduce.connect(collate)
else:
sql_gen.connect(collate)
# Plan settings
print('')
print("Settings")
print("--------")
print('')
print('use_pandas: {}'.format(use_pandas))
print("table parts: {}".format(table_parts_end - table_parts_start))
print('')
# Write the plan graph
# query_plan.write_graph(os.path.join(ROOT_DIR, "../benchmark-output"), gen_test_id() + "-" + str(table_parts))
# Start the query
query_plan.execute()
print('Done')
# tuples = collate.tuples()
# collate.print_tuples(tuples)
# Write the metrics
query_plan.print_metrics()
# Shut everything down
query_plan.stop()
sampling_time = query_plan.total_elapsed_time
cost, bytes_scanned, bytes_returned, rows = query_plan.cost()
computation_cost = query_plan.computation_cost()
data_cost = query_plan.data_cost()[0]
stats += [
sql_gen.threshold, sampling_time, 0, sampling_time, rows,
bytes_scanned, bytes_returned, data_cost, computation_cost, cost
]
def run_memory_indexed_sampling(stats,
sort_field_index,
sort_field,
k,
sample_size,
batch_size,
parallel,
use_pandas,
sort_order,
buffer_size,
table_parts_start,
table_parts_end,
tbl_s3key,
shards_path,
format_,
sampling_only=True):
"""
Executes the randomly sampled topk query by firstly building a random sample, then extracting the filtering threshold
Finally scanning the table to retrieve only the records beyond the threshold
:return:
"""
secure = False
use_native = False
n_threads = multiprocessing.cpu_count()
print('')
print(
"Top K Benchmark, Memory Indexed Sampling. Sort Field: {}, Order: {}, K: {}, Sample Size: {}, Batch Size: {}"
.format(sort_field, sort_order, k, sample_size, batch_size))
print("----------------------")
stats += [
'sampling_{}_{}'.format('memory_indexed', 'non-filtered'), shards_path,
sort_field, sort_order, k, sample_size, batch_size
]
# Query plan
query_plan = QueryPlan(is_async=parallel, buffer_size=buffer_size)
# Sampling
tbl_smpler = query_plan.add_operator(
TableRandomSampleGenerator(tbl_s3key, sample_size, batch_size,
"table_sampler", query_plan, False))
sample_scanners = map(
lambda p: query_plan.add_operator(
TableRangeAccess(tbl_s3key, use_pandas, secure, use_native,
"sample_scan_{}".format(p), query_plan, False)),
range(table_parts_start, table_parts_end + 1))
map(lambda op: op.set_nthreads(n_threads), sample_scanners)
# sample_scan = query_plan.add_operator(
# TableRangeAccess(tbl_s3key, use_pandas, secure, use_native, "sample_scan_{}".format(p),
# query_plan, False))
# sample_scan.set_nthreads(n_threads)
# Sampling project
def project_fn(df):
df.columns = [
sort_field if x == sort_field_index else x for x in df.columns
]
df = df[[sort_field]].astype(np.float, errors='ignore')
return df
project_exprs = [ProjectExpression(lambda t_: t_['_0'], sort_field)]
sample_project = map(
lambda p: query_plan.add_operator(
Project(project_exprs, 'sample_project_{}'.format(p), query_plan,
False, project_fn)),
range(table_parts_start, table_parts_end + 1))
# sample_project = query_plan.add_operator(
# Project(project_exprs, 'sample_project_{}'.format(p), query_plan, False, project_fn))
# TopK samples
sort_expr = [SortExpression(sort_field, float, sort_order)]
sample_topk = map(
lambda p: query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'sample_topk_{}'.format(p),
query_plan, False)),
range(table_parts_start, table_parts_end + 1))
# sample_topk = query_plan.add_operator(
# Top(k, sort_expr, use_pandas, 'sample_topk_{}'.format(p), query_plan, False))
sample_topk_reduce = query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'sample_topk_reduce', query_plan, False))
# Generate SQL command for second scan
sql_gen = query_plan.add_operator(
TopKFilterBuild(sort_order, 'float', 'select * from S3object ',
' CAST({} as float) '.format(sort_field),
'sample_sql_gen', query_plan, False))
if not sampling_only:
# Scan
scan = map(
lambda p: query_plan.add_operator(
SQLTableScan("{}.{}".format(shards_path, p), "", format_,
use_pandas, secure, use_native, 'scan_{}'.format(
p), query_plan, False)),
range(table_parts_start, table_parts_end + 1))
# Project
def project_fn(df):
df.columns = [
sort_field if x == sort_field_index else x for x in df.columns
]
df[[sort_field]] = df[[sort_field]].astype(np.float)
return df
project_exprs = [ProjectExpression(lambda t_: t_['_0'], sort_field)]
project = map(
lambda p: query_plan.add_operator(
Project(project_exprs, 'project_{}'.format(p), query_plan,
False, project_fn)),
range(table_parts_start, table_parts_end + 1))
# TopK
topk = map(
lambda p: query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'topk_{}'.format(p), query_plan,
False)), range(table_parts_start, table_parts_end + 1))
# TopK reduce
topk_reduce = query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'topk_reduce', query_plan, False))
collate = query_plan.add_operator(Collate('collate', query_plan, False))
map(lambda op: tbl_smpler.connect(op), sample_scanners)
map(lambda (p, o): o.connect(sample_project[p]),
enumerate(sample_scanners))
map(lambda (p, o): o.connect(sample_topk[p]), enumerate(sample_project))
map(lambda o: o.connect(sample_topk_reduce), sample_topk)
sample_topk_reduce.connect(sql_gen)
if not sampling_only:
map(lambda (p, o): sql_gen.connect(o), enumerate(scan))
map(lambda (p, o): o.connect(project[p]), enumerate(scan))
map(lambda (p, o): o.connect(topk[p]), enumerate(project))
map(lambda (p, o): o.connect(topk_reduce), enumerate(topk))
topk_reduce.connect(collate)
else:
sql_gen.connect(collate)
# Start the query
query_plan.execute()
print('Done')
# tuples = collate.tuples()
# collate.print_tuples(tuples)
# Write the metrics
query_plan.print_metrics()
# Shut everything down
# query_plan.stop()
sampling_threshold = query_plan.retrieve_sampling_threshold()
sampling_runtime = query_plan.get_phase_runtime('sampl')
sampling_num_http_requests, sampling_requests_cost = query_plan.requests_cost(
'sampl')
sampling_returned_bytes, sampling_returned_rows, sampling_transfer_cost = query_plan.data_transfer_cost(
phase_keyword='sampl')
sampling_scanned_bytes, sampling_scan_cost = query_plan.data_scanning_cost(
'sampl')
total_runtime = query_plan.total_elapsed_time
total_http_requests, total_requests_cost = query_plan.requests_cost()
total_returned_bytes, total_returned_rows, total_transfer_cost = query_plan.data_transfer_cost(
)
total_scanned_bytes, total_scan_cost = query_plan.data_scanning_cost()
total_data_cost = query_plan.data_cost()[0]
total_computation_cost = query_plan.computation_cost()
total_cost = query_plan.cost()[0]
stats += [
sampling_threshold, sampling_runtime, total_runtime - sampling_runtime,
total_runtime, sampling_returned_rows,
sampling_scanned_bytes * BYTE_TO_MB,
sampling_returned_bytes * BYTE_TO_MB, sampling_num_http_requests,
sampling_requests_cost, sampling_transfer_cost, sampling_scan_cost,
total_returned_rows, total_scanned_bytes * BYTE_TO_MB,
total_returned_bytes * BYTE_TO_MB, total_http_requests,
total_requests_cost, total_transfer_cost, total_scan_cost,
total_data_cost, total_computation_cost, total_cost
]
def run(buffer_size):
"""The baseline tst uses hash joins with no projection and no filtering pushed down to s3.
:return: None
"""
print('')
print("TPCH Q14 Baseline Join")
print("----------------------")
query_plan = QueryPlan(is_async=False, buffer_size=buffer_size)
# Query plan
date = '1993-01-01'
min_shipped_date = datetime.strptime(date, '%Y-%m-%d')
max_shipped_date = datetime.strptime(date, '%Y-%m-%d') + timedelta(days=30)
lineitem_scan = query_plan.add_operator(tpch_q14.sql_scan_lineitem_operator_def('lineitem_scan', query_plan))
lineitem_project = query_plan.add_operator(
tpch_q14.project_partkey_extendedprice_discount_shipdate_operator_def('lineitem_project', query_plan))
part_scan = query_plan.add_operator(tpch_q14.sql_scan_part_operator_def('part_scan', query_plan))
part_project = query_plan.add_operator(tpch_q14.project_partkey_brand_type_operator_def('part_project', query_plan))
lineitem_filter = query_plan.add_operator(
tpch_q14.filter_shipdate_operator_def(min_shipped_date, max_shipped_date, 'lineitem_filter', query_plan))
part_filter = query_plan.add_operator(tpch_q14.filter_brand12_operator_def('part_filter', query_plan))
join = query_plan.add_operator(tpch_q14.join_lineitem_part_operator_def('join', query_plan))
aggregate = query_plan.add_operator(tpch_q14.aggregate_promo_revenue_operator_def('aggregate', query_plan))
aggregate_project = query_plan.add_operator(
tpch_q14.project_promo_revenue_operator_def('aggregate_project', query_plan))
collate = query_plan.add_operator(tpch_q14.collate_operator_def('collate', query_plan))
lineitem_scan.connect(lineitem_project)
lineitem_project.connect(lineitem_filter)
join.connect_left_producer(lineitem_filter)
part_scan.connect(part_project)
part_project.connect(part_filter)
join.connect_right_producer(part_filter)
join.connect(aggregate)
aggregate.connect(aggregate_project)
aggregate_project.connect(collate)
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../benchmark-output"), gen_test_id())
# Start the query
query_plan.execute()
# Assert the results
# num_rows = 0
# for t in collate.tuples():
# num_rows += 1
# print("{}:{}".format(num_rows, t))
collate.print_tuples()
# Write the metrics
query_plan.print_metrics()
field_names = ['promo_revenue']
assert len(collate.tuples()) == 1 + 1
assert collate.tuples()[0] == field_names
# NOTE: This result has been verified with the equivalent data and query on PostgreSQL
assert collate.tuples()[1] == [15.090116526324298]
def run(parallel, use_pandas, secure, use_native, buffer_size, format_, customer_parts, order_parts, lineitem_parts,
customer_sharded,
order_sharded, lineitem_sharded, other_parts, fp_rate, sf, expected_result, customer_filter_sql=None,
order_filter_sql=None, lineitem_filter_sql=None):
"""
:return: None
"""
print('')
print("TPCH Q3 Bloom Join")
print("------------------")
query_plan = QueryPlan(is_async=parallel, buffer_size=buffer_size)
customer_scan = map(lambda p:
query_plan.add_operator(
SQLTableScan(get_file_key('customer', customer_sharded, p, sf, format_),
"select "
" c_custkey "
"from "
" S3Object "
"where "
" c_mktsegment = 'BUILDING' "
" {} "
" {} "
.format(
' and ' + customer_filter_sql if customer_filter_sql is not None else '',
get_sql_suffix('customer', customer_parts, p, customer_sharded,
add_where=False)),
format_,
use_pandas, secure, use_native,
'customer_scan' + '_{}'.format(p),
query_plan,
False)),
range(0, customer_parts))
def customer_project_fn(df):
df = df.filter(items=['_0'], axis=1)
df.rename(columns={'_0': 'c_custkey'},
inplace=True)
return df
customer_project = map(lambda p:
query_plan.add_operator(
Project([],
'customer_project' + '_{}'.format(p),
query_plan,
False, customer_project_fn)),
range(0, customer_parts))
customer_bloom_create = query_plan.add_operator(
BloomCreate('c_custkey', 'customer_bloom_create', query_plan, False,
fp_rate))
customer_map = map(lambda p:
query_plan.add_operator(Map('c_custkey', 'customer_map' + '_' + str(p), query_plan, False)),
range(0, customer_parts))
order_scan = map(lambda p:
query_plan.add_operator(
SQLTableScanBloomUse(get_file_key('orders', order_sharded, p, sf, format_),
"select "
" o_custkey, o_orderkey, o_orderdate, o_shippriority "
"from "
" S3Object "
"where "
" cast(o_orderdate as timestamp) < cast('1995-03-01' as timestamp) "
" {} "
" {} "
.format(
' and ' + order_filter_sql if order_filter_sql is not None else '',
get_sql_suffix('orders', order_parts, p, order_sharded,
add_where=False)),
'o_custkey',
format_,
use_pandas, secure, use_native,
'order_scan' + '_{}'.format(p),
query_plan,
False)),
range(0, order_parts))
def order_project_fn(df):
df = df.filter(items=['_0', '_1', '_2', '_3'], axis=1)
df.rename(columns={'_0': 'o_custkey', '_1': 'o_orderkey', '_2': 'o_orderdate', '_3': 'o_shippriority'},
inplace=True)
return df
order_project = map(lambda p:
query_plan.add_operator(
Project([],
'order_project' + '_{}'.format(p),
query_plan,
False, order_project_fn)),
range(0, customer_parts))
order_map_1 = map(lambda p:
query_plan.add_operator(Map('o_custkey', 'order_map_1' + '_' + str(p), query_plan, False)),
range(0, order_parts))
customer_order_join_build = map(lambda p:
query_plan.add_operator(
HashJoinBuild('c_custkey',
'customer_order_join_build' + '_' + str(p), query_plan,
False)),
range(0, other_parts))
customer_order_join_probe = map(lambda p:
query_plan.add_operator(
HashJoinProbe(JoinExpression('c_custkey', 'o_custkey'),
'customer_order_join_probe' + '_' + str(p),
query_plan, False)),
range(0, other_parts))
order_bloom_create = query_plan.add_operator(
BloomCreate('o_orderkey', 'order_bloom_create', query_plan, False,
fp_rate))
lineitem_scan = map(lambda p:
query_plan.add_operator(
SQLTableScanBloomUse(get_file_key('lineitem', lineitem_sharded, p, sf, format_),
"select "
" l_orderkey, l_extendedprice, l_discount "
"from "
" S3Object "
"where "
" cast(l_shipdate as timestamp) > cast('1995-03-01' as timestamp) "
" {} "
" {} "
.format(
' and ' + lineitem_filter_sql if lineitem_filter_sql is not None else '',
get_sql_suffix('lineitem', lineitem_parts, p, lineitem_sharded,
add_where=False)),
'l_orderkey', format_,
use_pandas, secure, use_native,
'lineitem_scan' + '_{}'.format(p),
query_plan,
False)),
range(0, lineitem_parts))
def lineitem_project_fn(df):
df = df.filter(items=['_0', '_1', '_2'], axis=1)
df.rename(columns={'_0': 'l_orderkey', '_1': 'l_extendedprice', '_2': 'l_discount'},
inplace=True)
return df
lineitem_project = map(lambda p:
query_plan.add_operator(
Project([],
'lineitem_project' + '_{}'.format(p),
query_plan,
False, lineitem_project_fn)),
range(0, lineitem_parts))
lineitem_map = map(lambda p:
query_plan.add_operator(Map('l_orderkey', 'lineitem_map' + '_' + str(p), query_plan, False)),
range(0, lineitem_parts))
order_map_2 = map(lambda p:
query_plan.add_operator(Map('o_orderkey', 'order_map_2' + '_' + str(p), query_plan, False)),
range(0, other_parts))
customer_order_lineitem_join_build = map(lambda p:
query_plan.add_operator(
HashJoinBuild('o_orderkey',
'customer_order_lineitem_join_build' + '_' + str(p),
query_plan,
False)),
range(0, other_parts))
customer_order_lineitem_join_probe = map(lambda p:
query_plan.add_operator(
HashJoinProbe(JoinExpression('o_orderkey', 'l_orderkey'),
'customer_order_lineitem_join_probe' + '_' + str(p),
query_plan, False)),
range(0, other_parts))
def groupby_fn(df):
df['l_extendedprice'] = df['l_extendedprice'].astype(np.float)
df['l_discount'] = df['l_discount'].astype(np.float)
df['revenue'] = df['l_extendedprice'] * (1 - df['l_discount'])
grouped = df.groupby(['l_orderkey', 'o_orderdate', 'o_shippriority'])
agg_df = grouped['revenue'].sum()
return agg_df.reset_index()
group = map(lambda p:
query_plan.add_operator(
Group(
['l_orderkey', 'o_orderdate', 'o_shippriority'], # l_partkey
[
AggregateExpression(AggregateExpression.SUM,
lambda t_: float(t_['l_extendedprice'] * (1 - t_['l_discount'])))
],
'group' + '_{}'.format(p), query_plan,
False, groupby_fn)),
range(0, other_parts))
def group_reduce_fn(df):
grouped = df.groupby(['l_orderkey', 'o_orderdate', 'o_shippriority'])
agg_df = grouped['revenue'].sum()
return agg_df.reset_index()
group_reduce = query_plan.add_operator(
Group(
['l_orderkey', 'o_orderdate', 'o_shippriority'], # l_partkey
[
AggregateExpression(AggregateExpression.SUM,
lambda t_: float(t_['l_extendedprice'] * (1 - t_['l_discount'])))
],
'group_reduce', query_plan,
False, group_reduce_fn))
top = query_plan.add_operator(
Top(10, [SortExpression('revenue', float, 'DESC'), SortExpression('o_orderdate', date, 'ASC')], use_pandas,
'top', query_plan,
False))
collate = query_plan.add_operator(tpch_q19.collate_op('collate', query_plan))
# Inline what we can
map(lambda o: o.set_async(False), lineitem_project)
map(lambda o: o.set_async(False), customer_project)
map(lambda o: o.set_async(False), order_project)
map(lambda o: o.set_async(False), lineitem_map)
map(lambda o: o.set_async(False), customer_map)
map(lambda o: o.set_async(False), order_map_1)
map(lambda o: o.set_async(False), order_map_2)
# Connect the operators
connect_many_to_many(customer_scan, customer_project)
connect_many_to_many(customer_project, customer_map)
connect_many_to_one(customer_project, customer_bloom_create)
connect_one_to_many(customer_bloom_create, order_scan)
connect_many_to_many(order_scan, order_project)
connect_many_to_many(order_project, order_map_1)
connect_all_to_all(customer_map, customer_order_join_build)
connect_many_to_many(customer_order_join_build, customer_order_join_probe)
connect_all_to_all(order_map_1, customer_order_join_probe)
# connect_many_to_one(customer_order_join_probe, collate)
connect_many_to_one(order_project, order_bloom_create)
connect_one_to_many(order_bloom_create, lineitem_scan)
connect_many_to_many(lineitem_scan, lineitem_project)
connect_many_to_many(lineitem_project, lineitem_map)
connect_many_to_many(customer_order_join_probe, order_map_2)
connect_all_to_all(order_map_2, customer_order_lineitem_join_build)
connect_many_to_many(customer_order_lineitem_join_build, customer_order_lineitem_join_probe)
connect_all_to_all(lineitem_map, customer_order_lineitem_join_probe)
# connect_many_to_one(customer_order_lineitem_join_probe, collate)
connect_many_to_many(customer_order_lineitem_join_probe, group)
# connect_many_to_one(group, collate)
connect_many_to_one(group, group_reduce)
connect_one_to_one(group_reduce, top)
connect_one_to_one(top, collate)
# Plan settings
print('')
print("Settings")
print("--------")
print('')
print('use_pandas: {}'.format(use_pandas))
print('secure: {}'.format(secure))
print('use_native: {}'.format(use_native))
print("customer_parts: {}".format(customer_parts))
print("order_parts: {}".format(order_parts))
print("lineitem_parts: {}".format(lineitem_parts))
print("customer_sharded: {}".format(customer_sharded))
print("order_sharded: {}".format(order_sharded))
print("lineitem_sharded: {}".format(lineitem_sharded))
print("other_parts: {}".format(other_parts))
print("fp_rate: {}".format(fp_rate))
print("format: {}".format(format_))
print('')
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../benchmark-output"), gen_test_id())
# Start the query
query_plan.execute()
tuples = collate.tuples()
collate.print_tuples(tuples)
# Write the metrics
query_plan.print_metrics()
# Shut everything down
query_plan.stop()
field_names = ['l_orderkey', 'o_orderdate', 'o_shippriority', 'revenue']
assert len(tuples) == 10 + 1
assert tuples[0] == field_names
# NOTE: This result has been verified with the equivalent data and query on PostgreSQL
test_util.assert_tuples(expected_result, tuples)
def run(parallel, use_pandas, buffer_size, table_parts, perc, path, nthreads=16, format_=Format.CSV):
secure = False
use_native = False
print('')
print("Indexing Benchmark")
print("------------------")
# Query plan
query_plan = QueryPlan(is_async=parallel, buffer_size=buffer_size)
# Scan Index Files
upper = perc * 100
index_scan = map(lambda p:
query_plan.add_operator(
SQLTableScan('{}/index/index_f0_{}.csv'.format(path, p),
"select first_byte, last_byte "
" from S3Object "
" where cast(name as float) < {};".format(upper), format_,
use_pandas, secure, use_native,
'index_scan_{}'.format(p), query_plan,
False)),
range(0, table_parts))
# Range accesses
range_access = map(lambda p:
query_plan.add_operator(
TableRangeAccess('{}/data_{}.csv'.format(path, p),
use_pandas, secure, use_native,
'range_access_{}'.format(p), query_plan,
False)),
range(0, table_parts))
map(lambda o: o.set_nthreads(nthreads), range_access)
collate = query_plan.add_operator(
Collate('collate', query_plan, False))
map(lambda (p, o): o.connect(range_access[p]), enumerate(index_scan))
map(lambda (p, o): o.connect(collate), enumerate(range_access))
# Plan settings
print('')
print("Settings")
print("--------")
print('')
print('use_pandas: {}'.format(use_pandas))
print("table parts: {}".format(table_parts))
print('')
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../benchmark-output"), gen_test_id() + "-" + str(table_parts))
# Start the query
query_plan.execute()
print('Done')
tuples = collate.tuples()
# collate.print_tuples(tuples)
# Write the metrics
query_plan.print_metrics()
# Shut everything down
query_plan.stop()
def run(parallel, buffer_size):
"""
:return: None
"""
print('')
print("TPCH Q14 Filtered Join")
print("----------------------")
query_plan = QueryPlan(is_async=parallel, buffer_size=buffer_size)
# Query plan
# DATE is the first day of a month randomly selected from a random year within [1993 .. 1997].
date = '1993-01-01'
min_shipped_date = datetime.strptime(date, '%Y-%m-%d')
max_shipped_date = datetime.strptime(date, '%Y-%m-%d') + timedelta(days=30)
lineitem_scan = query_plan.add_operator(
tpch_q14.
sql_scan_lineitem_partkey_extendedprice_discount_where_shipdate_operator_def(
min_shipped_date, max_shipped_date, 'lineitem_scan', query_plan))
lineitem_project = query_plan.add_operator(
tpch_q14.project_partkey_extendedprice_discount_operator_def(
'lineitem_project', query_plan))
part_scan = query_plan.add_operator(
tpch_q14.sql_scan_part_partkey_type_part_where_brand12_operator_def(
'part_scan', query_plan))
part_project = query_plan.add_operator(
tpch_q14.project_partkey_type_operator_def('part_project', query_plan))
join = query_plan.add_operator(
tpch_q14.join_lineitem_part_operator_def('join', query_plan))
aggregate = query_plan.add_operator(
tpch_q14.aggregate_promo_revenue_operator_def('aggregate', query_plan))
aggregate_project = query_plan.add_operator(
tpch_q14.project_promo_revenue_operator_def('aggregate_project',
query_plan))
collate = query_plan.add_operator(
tpch_q14.collate_operator_def('collate', query_plan))
lineitem_scan.connect(lineitem_project)
part_scan.connect(part_project)
join.connect_left_producer(lineitem_project)
join.connect_right_producer(part_project)
join.connect(aggregate)
aggregate.connect(aggregate_project)
aggregate_project.connect(collate)
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../benchmark-output"),
gen_test_id())
# Start the query
query_plan.execute()
tuples = collate.tuples()
collate.print_tuples(tuples)
# Write the metrics
query_plan.print_metrics()
# Shut everything down
query_plan.stop()
field_names = ['promo_revenue']
assert len(tuples) == 1 + 1
assert tuples[0] == field_names
# NOTE: This result has been verified with the equivalent data and query on PostgreSQL
assert tuples[1] == [15.090116526324298]
def run(sf, parallel, use_pandas, secure, use_native, buffer_sizje,
lineitem_parts, sharded, format_):
"""
:return: None
"""
print('')
print("TPCH Q1 Baseline Group By")
print("----------------------")
query_plan = QueryPlan(is_async=parallel, buffer_size=buffer_size)
# Query plan
lineitem_scan = map(
lambda p: query_plan.add_operator(
tpch_q1.sql_scan_lineitem_operator_def(
sharded, p, sf, use_pandas, secure, use_native,
'lineitem_scan' + '_' + str(p), query_plan, format_)),
range(0, lineitem_parts))
'''
lineitem_project = map(lambda p:
query_plan.add_operator(
tpch_q1.project_lineitem_operator_def(
'lineitem_project' + '_' + str(p),
query_plan)),
range(0, lineitem_parts))
date = '1998-12-01'
max_shipped_date = datetime.strptime(date, '%Y-%m-%d') - timedelta(days=117)
lineitem_filter = map(lambda p:
query_plan.add_operator(
tpch_q1.filter_shipdate_operator_def(
max_shipped_date,
'lineitem_filter' + '_' + str(p),
query_plan)),
range(0, lineitem_parts))
profile_file_name = os.path.join(ROOT_DIR, "../tests-output/" + gen_test_id() + ".prof")
groupby = map(lambda p:
query_plan.add_operator(
tpch_q1.groupby_returnflag_linestatus_operator_def(
'groupby' + '_' + str(p),
query_plan)),
range(0, lineitem_parts))
# groupby[0].set_profiled(True, profile_file_name)
groupby_reduce = query_plan.add_operator(
tpch_q1.groupby_reduce_returnflag_linestatus_operator_def(
'groupby_reduce',
query_plan, use_pandas))
# lineitem_scan[0].set_profiled(True,
# os.path.join(ROOT_DIR, "../benchmark-output/", gen_test_id() + "_scan_0" + ".prof"))
# lineitem_project[0].set_profiled(True, os.path.join(ROOT_DIR, "../benchmark-output/",
# gen_test_id() + "_project_0" + ".prof"))
# groupby[0].set_profiled(True,
# os.path.join(ROOT_DIR, "../benchmark-output/", gen_test_id() + "_groupby_0" + ".prof"))
# groupby_reduce.set_profiled(True, os.path.join(ROOT_DIR, "../benchmark-output/",
# gen_test_id() + "_groupby_reduce" + ".prof"))
collate = query_plan.add_operator(
Collate('collate', query_plan, False))
map(lambda o: o.set_async(False), lineitem_project)
map(lambda o: o.set_async(False), lineitem_filter)
# map(lambda o: o.set_async(False), groupby)
map(lambda (p, o): o.connect(lineitem_project[p]), enumerate(lineitem_scan))
map(lambda (p, o): o.connect(lineitem_filter[p]), enumerate(lineitem_project))
map(lambda (p, o): o.connect(groupby[p]), enumerate(lineitem_filter))
map(lambda (p, o): o.connect(groupby_reduce), enumerate(groupby))
groupby_reduce.connect(collate)
'''
groupby = map(
lambda p: query_plan.add_operator(
tpch_q1.groupby_returnflag_linestatus_operator_def(
'groupby' + '_' + str(p), query_plan)),
range(0, lineitem_parts))
collate = query_plan.add_operator(Collate('collate', query_plan, False))
map(lambda (o): o.connect(collate), lineitem_scan)
# Plan settings
print('')
print("Settings")
print("--------")
print('')
print('use_pandas: {}'.format(use_pandas))
print("lineitem parts: {}".format(lineitem_parts))
print('')
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../benchmark-output"),
gen_test_id() + "-" + str(lineitem_parts))
# Start the query
query_plan.execute()
print('Done')
tuples = collate.tuples()
collate.print_tuples(tuples)
# Write the metrics
query_plan.print_metrics()
# Shut everything down
query_plan.stop()
def run(sort_field, k, parallel, use_pandas, sort_order, buffer_size,
table_first_part, table_parts, queried_columns, select_columns, path,
format_):
"""
Executes the baseline topk query by scanning a table and keeping track of the max/min records in a heap
:return:
"""
secure = False
use_native = False
print('')
print("Top K Benchmark, ColumnScan. Sort Field: {}, Order: {}".format(
sort_field, sort_order))
print("----------------------")
# Query plan
query_plan = QueryPlan(is_async=parallel, buffer_size=buffer_size)
# Sampling
sample_scan = map(
lambda p: query_plan.add_operator(
#SQLTableScan("{}/lineitem.snappy.parquet.{}".format(path, p),
SQLTableScan(
"{}/lineitem.typed.1RowGroup.parquet.{}".format(
path, p), 'select {} from S3Object;'.format(
sort_field), format_, use_pandas, secure, use_native,
'column_scan_{}'.format(p), query_plan, False)),
range(table_first_part, table_first_part + table_parts))
# Sampling project
def project_fn1(df):
df.columns = [sort_field]
df[[sort_field]] = df[[sort_field]].astype(np.float)
return df
project_exprs = [ProjectExpression(lambda t_: t_['_0'], sort_field)]
sample_project = map(
lambda p: query_plan.add_operator(
Project(project_exprs, 'sample_project_{}'.format(p), query_plan,
False, project_fn1)),
range(table_first_part, table_first_part + table_parts))
# TopK samples
sort_expr = SortExpression(sort_field, float, sort_order)
sample_topk = query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'sample_topk', query_plan, False))
# Generate SQL command for second scan
sql_gen = query_plan.add_operator(
TopKFilterBuild(
sort_order,
'float',
'select {} from S3object '.format(select_columns),
#' CAST({} as float) '.format(sort_field), 'sql_gen', query_plan, False ))
' {} '.format(sort_field),
'sql_gen',
query_plan,
False))
# Scan
scan = map(
lambda p: query_plan.add_operator(
#SQLTableScan("{}/lineitem.snappy.parquet.{}".format(path, p),
SQLTableScan(
"{}/lineitem.typed.1RowGroup.parquet.{}".format(path, p), "",
format_, use_pandas, secure, use_native, 'scan_{}'.format(
p), query_plan, False)),
range(table_first_part, table_first_part + table_parts))
# Project
def project_fn2(df):
df.columns = queried_columns
df[[sort_field]] = df[[sort_field]].astype(np.float)
return df
project_exprs = [ProjectExpression(lambda t_: t_['_0'], sort_field)]
project = map(
lambda p: query_plan.add_operator(
Project(project_exprs, 'project_{}'.format(p), query_plan, False,
project_fn2)),
range(table_first_part, table_first_part + table_parts))
# TopK
topk = map(
lambda p: query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'topk_{}'.format(p),
query_plan, False)),
range(table_first_part, table_first_part + table_parts))
# TopK reduce
topk_reduce = query_plan.add_operator(
Top(k, sort_expr, use_pandas, 'topk_reduce', query_plan, False))
collate = query_plan.add_operator(Collate('collate', query_plan, False))
#profile_path = '../benchmark-output/groupby/'
#scan[0].set_profiled(True, os.path.join(ROOT_DIR, profile_path, gen_test_id() + "_scan_0" + ".prof"))
#project[0].set_profiled(True, os.path.join(ROOT_DIR, profile_path, gen_test_id() + "_project_0" + ".prof"))
#groupby[0].set_profiled(True, os.path.join(ROOT_DIR, profile_path, gen_test_id() + "_groupby_0" + ".prof"))
#groupby_reduce.set_profiled(True, os.path.join(ROOT_DIR, profile_path, gen_test_id() + "_groupby_reduce" + ".prof"))
#collate.set_profiled(True, os.path.join(ROOT_DIR, profile_path, gen_test_id() + "_collate" + ".prof"))
map(lambda (p, o): o.connect(sample_project[p]), enumerate(sample_scan))
map(lambda (p, o): o.connect(sample_topk), enumerate(sample_project))
sample_topk.connect(sql_gen)
map(lambda (p, o): sql_gen.connect(o), enumerate(scan))
map(lambda (p, o): o.connect(project[p]), enumerate(scan))
map(lambda (p, o): o.connect(topk[p]), enumerate(project))
map(lambda (p, o): o.connect(topk_reduce), enumerate(topk))
topk_reduce.connect(collate)
# Plan settings
print('')
print("Settings")
print("--------")
print('')
print('use_pandas: {}'.format(use_pandas))
print("table parts: {}".format(table_parts))
print('')
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../benchmark-output"),
gen_test_id() + "-" + str(table_parts))
# Start the query
query_plan.execute()
print('Done')
tuples = collate.tuples()
collate.print_tuples(tuples)
# Write the metrics
query_plan.print_metrics()
# Shut everything down
query_plan.stop()
def test_filter_baseline():
"""
:return:
"""
query_plan = QueryPlan(buffer_size=64, is_async=True, use_shared_mem=False)
# Query plan
'''
ts = query_plan.add_operator(
SQLTableScan('lineitem.csv', 'select * from S3Object limit 3;' , False, 'ts', query_plan, False))
'''
# using a 'use_native=True' argument will result in a None object being returned
ts = query_plan.add_operator(
SQLTableScan('random_strings_2.csv', 'select * from S3Object limit 3;',
Format.CSV, True, False, False, 'ts', query_plan, False))
f = query_plan.add_operator(
Filter(
PredicateExpression(lambda t_: cast(t_['_10'], timestamp) >= cast(
'1996-03-01', timestamp)), 'f', query_plan, False))
c = query_plan.add_operator(Collate('c', query_plan, False))
ts.connect(f)
f.connect(c)
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../tests-output"),
gen_test_id())
# Start the query
query_plan.execute()
# Assert the results
# num_rows = 0
# for t in c.tuples():
# num_rows += 1
# print("{}:{}".format(num_rows, t))
print(len(c.tuples()))
assert 2 + 1 == len(c.tuples())
field_names = [
'_0', '_1', '_2', '_3', '_4', '_5', '_6', '_7', '_8', '_9', '_10',
'_11', '_12', '_13', '_14', '_15'
]
assert c.tuples()[0] == field_names
assert c.tuples()[1] == [
'1', '155190', '7706', '1', '17', '21168.23', '0.04', '0.02', 'N', 'O',
'1996-03-13', '1996-02-12', '1996-03-22', 'DELIVER IN PERSON', 'TRUCK',
'egular courts above the'
]
assert c.tuples()[2] == [
'1', '67310', '7311', '2', '36', '45983.16', '0.09', '0.06', 'N', 'O',
'1996-04-12', '1996-02-28', '1996-04-20', 'TAKE BACK RETURN', 'MAIL',
'ly final dependencies: slyly bold '
]
# Write the metrics
query_plan.print_metrics()
def run(parallel,
use_pandas,
buffer_size,
table_parts,
lower,
upper,
sf,
format_=Format.CSV):
secure = False
use_native = False
print('')
print("Indexing Benchmark")
print("------------------")
# Query plan
query_plan = QueryPlan(is_async=parallel, buffer_size=buffer_size)
# SQL scan the file
scan = map(
lambda p: query_plan.add_operator(
SQLTableScan(
get_file_key('lineitem', True, p, sf=sf, format_=format_
), "select * from S3Object "
"where cast(l_extendedprice as float) >= {} and cast(l_extendedprice as float) <= {};"
.format(lower, upper), format_, use_pandas, secure,
use_native, 'scan_{}'.format(p), query_plan, False)),
range(0, table_parts))
'''
scan = map(lambda p:
query_plan.add_operator(
SQLTableScan(get_file_key('lineitem', True, p, sf=sf, format_=format_),
"select * from S3Object "
"where l_extendedprice >= {} and l_extendedprice <= {};".format(
lower, upper), format_,
use_pandas, secure, use_native,
'scan_{}'.format(p), query_plan,
False)),
range(0, table_parts))
'''
# project
def fn(df):
df.columns = [
'l_orderkey', 'l_partkey', 'l_suppkey', 'l_linenumber',
'l_quantity', 'l_extendedprice', 'l_discount', 'l_tax',
'l_returnflag', 'l_linestatus', 'l_shipdate', 'l_commitdate',
'l_receiptdate', 'l_shipinstruct', 'l_shipmode', 'l_comment'
]
df[['l_extendedprice']] = df[['l_extendedprice']].astype(np.float)
return df
project = map(
lambda p: query_plan.add_operator(
Project([], 'project_{}'.format(p), query_plan, False, fn)),
range(0, table_parts))
# aggregation
def agg_fun(df):
return pd.DataFrame({'count': [len(df)]})
aggregate = query_plan.add_operator(
Aggregate([], True, 'agg', query_plan, False, agg_fun))
collate = query_plan.add_operator(Collate('collate', query_plan, False))
map(lambda (p, o): o.connect(project[p]), enumerate(scan))
map(lambda (p, o): o.connect(aggregate), enumerate(project))
aggregate.connect(collate)
# Plan settings
print('')
print("Settings")
print("--------")
print('')
print('use_pandas: {}'.format(use_pandas))
print("table parts: {}".format(table_parts))
print('')
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../benchmark-output"),
gen_test_id() + "-" + str(table_parts))
# Start the query
query_plan.execute()
print('Done')
tuples = collate.tuples()
collate.print_tuples(tuples)
# Write the metrics
query_plan.print_metrics()
# Shut everything down
query_plan.stop()
def test_pandas_filter_baseline():
"""
:return:
"""
query_plan = QueryPlan(buffer_size=8192)
# Query plan
ts = query_plan.add_operator(
SQLTableScan('lineitem.csv', 'select * from S3Object limit 3;', True,
'ts', query_plan, False))
# f = query_plan.add_operator(
# Filter(PredicateExpression(lambda t_: cast(t_['_10'], timestamp) >= cast('1996-03-01', timestamp)),
# 'f', query_plan,
# False))
def pd_expr(df):
# df['_10'] = pd.to_datetime(df['_10'])
return pd.to_datetime(df['_10']) >= '1996-03-01'
f = query_plan.add_operator(
Filter(PredicateExpression(None, pd_expr), 'f', query_plan, True))
c = query_plan.add_operator(Collate('c', query_plan, False))
ts.connect(f)
f.connect(c)
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../tests-output"),
gen_test_id())
# Start the query
query_plan.execute()
# Assert the results
# num_rows = 0
# for t in c.tuples():
# num_rows += 1
# print("{}:{}".format(num_rows, t))
assert 2 + 1 == len(c.tuples())
field_names = [
'_0', '_1', '_2', '_3', '_4', '_5', '_6', '_7', '_8', '_9', '_10',
'_11', '_12', '_13', '_14', '_15'
]
assert c.tuples()[0] == field_names
assert c.tuples()[1] == [
'1', '155190', '7706', '1', '17', '21168.23', '0.04', '0.02', 'N', 'O',
'1996-03-13', '1996-02-12', '1996-03-22', 'DELIVER IN PERSON', 'TRUCK',
'egular courts above the'
]
assert c.tuples()[2] == [
'1', '67310', '7311', '2', '36', '45983.16', '0.09', '0.06', 'N', 'O',
'1996-04-12', '1996-02-28', '1996-04-20', 'TAKE BACK RETURN', 'MAIL',
'ly final dependencies: slyly bold '
]
# Write the metrics
query_plan.print_metrics()
def run(parallel,
use_pandas,
buffer_size,
table_parts,
lower,
upper,
sf,
nthreads=16,
format_=Format.CSV):
secure = False
use_native = False
print('')
print("Indexing Benchmark")
print("------------------")
# Query plan
query_plan = QueryPlan(is_async=parallel, buffer_size=buffer_size)
assert sf == 1 or sf == 10
# Scan Index Files
index_scan = map(
lambda p: query_plan.add_operator(
SQLTableScan(
'tpch-sf{}/lineitem_sharded/index/index_l_extendedprice.csv.{}'
.format(sf, p
if sf == 1 else p + 1), "select first_byte, last_byte "
" from S3Object "
" where col_values >= {} and col_values <= {};".format(
lower, upper), format_, use_pandas, secure, use_native,
'index_scan_{}'.format(p), query_plan, False)),
range(0, table_parts))
# Range accesses
range_access = map(
lambda p: query_plan.add_operator(
TableRangeAccess(
get_file_key('lineitem', True, p, sf=sf, format_=Format.CSV),
use_pandas, secure, use_native, 'range_access_{}'.format(
p), query_plan, False)), range(0, table_parts))
map(lambda o: o.set_nthreads(nthreads), range_access)
# Aggregation
def agg_fun(df):
return pd.DataFrame({'count': [len(df)]})
aggregate = query_plan.add_operator(
Aggregate([], True, 'agg', query_plan, False, agg_fun))
collate = query_plan.add_operator(Collate('collate', query_plan, False))
map(lambda (p, o): o.connect(range_access[p]), enumerate(index_scan))
map(lambda (p, o): o.connect(aggregate), enumerate(range_access))
aggregate.connect(collate)
# Plan settings
print('')
print("Settings")
print("--------")
print('')
print('use_pandas: {}'.format(use_pandas))
print("table parts: {}".format(table_parts))
print('')
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../benchmark-output"),
gen_test_id() + "-" + str(table_parts))
# Start the query
query_plan.execute()
print('Done')
tuples = collate.tuples()
collate.print_tuples(tuples)
# Write the metrics
query_plan.print_metrics()
# Shut everything down
query_plan.stop()
def query_plan(settings):
# type: (SyntheticSemiJoinSettings) -> QueryPlan
"""
:return: None
"""
if settings.use_shared_mem:
system = WorkerSystem(settings.shared_memory_size)
else:
system = None
query_plan = QueryPlan(system,
is_async=settings.parallel,
buffer_size=settings.buffer_size,
use_shared_mem=settings.use_shared_mem)
# Define the operators
scan_a = \
map(lambda p:
query_plan.add_operator(
SQLTableScan(get_file_key(settings.table_A_key, settings.table_A_sharded, p, settings.sf),
"select "
" {} "
"from "
" S3Object "
"where "
" {} "
" {} "
.format(settings.table_A_AB_join_key,
settings.table_A_filter_sql,
get_sql_suffix(settings.table_A_key, settings.table_A_parts, p,
settings.table_A_sharded)), settings.format_,
settings.use_pandas,
settings.secure,
settings.use_native,
'scan_a' + '_{}'.format(p),
query_plan,
False)),
range(0, settings.table_A_parts))
field_names_map_a = OrderedDict(
zip([
'_{}'.format(i)
for i, name in enumerate(settings.table_A_field_names)
], settings.table_A_field_names))
def project_fn_a(df):
df = df.rename(columns=field_names_map_a, copy=False)
return df
project_a = map(
lambda p: query_plan.add_operator(
Project([
ProjectExpression(k, v)
for k, v in field_names_map_a.iteritems()
], 'project_a' + '_{}'.format(p), query_plan, False, project_fn_a)
), range(0, settings.table_A_parts))
bloom_create_ab_join_key = map(
lambda p: query_plan.add_operator(
BloomCreate(settings.table_A_AB_join_key,
'bloom_create_ab_join_key' + '_{}'.format(p),
query_plan,
False,
fp_rate=settings.fp_rate)),
range(0, settings.table_A_parts))
scan_b_on_ab_join_key = \
map(lambda p:
query_plan.add_operator(
SQLTableScanBloomUse(get_file_key(settings.table_B_key, settings.table_B_sharded, p, settings.sf),
"select "
" {},{} "
"from "
" S3Object "
"where "
" {} "
" {} "
.format(settings.table_B_BC_join_key,
settings.table_B_AB_join_key,
settings.table_B_filter_sql,
get_sql_suffix(settings.table_B_key, settings.table_B_parts, p,
settings.table_B_sharded, add_where=False)), settings.format_,
settings.table_B_AB_join_key,
settings.use_pandas,
settings.secure,
settings.use_native,
'scan_b_on_ab_join_key' + '_{}'.format(p),
query_plan,
False)),
range(0, settings.table_B_parts))
if settings.table_C_key is None:
scan_b_detail_on_b_pk = \
map(lambda p:
query_plan.add_operator(
SQLTableScanBloomUse(get_file_key(settings.table_B_key, settings.table_B_sharded, p, settings.sf),
"select "
" {},{} "
"from "
" S3Object "
"where "
" {} "
" {} "
.format(settings.table_B_primary_key,
settings.table_B_detail_field_name,
settings.table_B_filter_sql,
get_sql_suffix(settings.table_B_key, settings.table_B_parts, p,
settings.table_B_sharded, add_where=False)), settings.format_,
settings.table_B_primary_key,
settings.use_pandas,
settings.secure,
settings.use_native,
'scan_c_detail_on_b_pk' + '_{}'.format(p),
query_plan,
False)),
range(0, settings.table_B_parts))
field_names_map_b_detail = OrderedDict([
('_0', settings.table_B_primary_key),
('_1', settings.table_B_detail_field_name)
])
def project_fn_b_detail(df):
df.rename(columns=field_names_map_b_detail, inplace=True)
return df
project_b_detail = map(
lambda p: query_plan.add_operator(
Project([
ProjectExpression(k, v)
for k, v in field_names_map_b_detail.iteritems()
], 'project_b_detail' + '_{}'.format(p), query_plan,
False, project_fn_b_detail)),
range(0, settings.table_B_parts))
map_b_pk_1 = map(
lambda p: query_plan.add_operator(
Map(settings.table_B_primary_key, 'map_b_pk_1' + '_{}'.format(
p), query_plan, False)), range(0, settings.table_B_parts))
map_b_pk_2 = map(
lambda p: query_plan.add_operator(
Map(settings.table_B_primary_key, 'map_b_pk_2' + '_{}'.format(
p), query_plan, False)), range(0, settings.table_B_parts))
bloom_create_b_pk = map(
lambda p: query_plan.add_operator(
BloomCreate(settings.table_B_primary_key,
'bloom_create_b_pk' + '_{}'.format(p),
query_plan,
False,
fp_rate=settings.fp_rate)),
range(0, settings.table_B_parts))
join_probe_ab_and_b_on_b_pk = map(
lambda p: query_plan.add_operator(
HashJoinProbe(
JoinExpression(settings.table_B_primary_key, settings.
table_B_primary_key),
'join_probe_ab_and_b_on_b_pk' + '_{}'.format(
p), query_plan, False)),
range(0, settings.table_B_parts))
join_build_ab_and_b_on_b_pk = map(
lambda p: query_plan.add_operator(
HashJoinBuild(settings.table_B_primary_key,
'join_build_ab_and_b_on_b_pk' + '_{}'.format(
p), query_plan, False)),
range(0, settings.table_B_parts))
else:
scan_c_on_bc_join_key = \
map(lambda p:
query_plan.add_operator(
SQLTableScanBloomUse(get_file_key(settings.table_C_key, settings.table_C_sharded, p, settings.sf),
"select "
" {}, {} "
"from "
" S3Object "
"where "
" {} "
" {} "
.format(settings.table_C_primary_key,
settings.table_C_BC_join_key,
settings.table_C_filter_sql,
get_sql_suffix(settings.table_C_key, settings.table_C_parts, p,
settings.table_C_sharded, add_where=False)),
settings.table_C_BC_join_key, settings.format_,
settings.use_pandas,
settings.secure,
settings.use_native,
'scan_c_on_bc_join_key' + '_{}'.format(p),
query_plan,
False)),
range(0, settings.table_C_parts))
field_names_map_c = OrderedDict(
zip([
'_{}'.format(i)
for i, name in enumerate(settings.table_C_field_names)
], settings.table_C_field_names))
def project_fn_c(df):
df.rename(columns=field_names_map_c, inplace=True)
return df
project_c = map(
lambda p: query_plan.add_operator(
Project([
ProjectExpression(k, v)
for k, v in field_names_map_c.iteritems()
], 'project_c' + '_{}'.format(p), query_plan, False,
project_fn_c)), range(0, settings.table_C_parts))
scan_c_detail_on_c_pk = \
map(lambda p:
query_plan.add_operator(
SQLTableScanBloomUse(get_file_key(settings.table_C_key, settings.table_C_sharded, p, settings.sf),
"select "
" {},{} "
"from "
" S3Object "
"where "
" {} "
" {} "
.format(settings.table_C_primary_key,
settings.table_C_detail_field_name,
settings.table_C_filter_sql,
get_sql_suffix(settings.table_C_key, settings.table_C_parts, p,
settings.table_C_sharded, add_where=False)),
settings.table_C_primary_key, settings.format_,
settings.use_pandas,
settings.secure,
settings.use_native,
'scan_c_detail_on_c_pk' + '_{}'.format(p),
query_plan,
False)),
range(0, settings.table_C_parts))
field_names_map_c_detail = OrderedDict([
('_0', settings.table_C_primary_key),
('_1', settings.table_C_detail_field_name)
])
def project_fn_c_detail(df):
df.rename(columns=field_names_map_c_detail, inplace=True)
return df
project_c_detail = map(
lambda p: query_plan.add_operator(
Project([
ProjectExpression(k, v)
for k, v in field_names_map_c_detail.iteritems()
], 'project_c_detail' + '_{}'.format(p), query_plan,
False, project_fn_c_detail)),
range(0, settings.table_C_parts))
map_bc_b_join_key = map(
lambda p: query_plan.add_operator(
Map(settings.table_B_BC_join_key, 'map_bc_b_join_key' + '_{}'.
format(p), query_plan, False)),
range(0, settings.table_C_parts))
map_c_pk_1 = map(
lambda p: query_plan.add_operator(
Map(settings.table_C_primary_key, 'map_c_pk_1' + '_{}'.format(
p), query_plan, False)), range(0, settings.table_C_parts))
map_c_pk_2 = map(
lambda p: query_plan.add_operator(
Map(settings.table_C_primary_key, 'map_c_pk_2' + '_{}'.format(
p), query_plan, False)), range(0, settings.table_C_parts))
bloom_create_c_pk = map(
lambda p: query_plan.add_operator(
BloomCreate(settings.table_C_primary_key,
'bloom_create_bc_b_to_c_join_key_{}'.format(p),
query_plan,
False,
fp_rate=settings.fp_rate)),
range(0, settings.table_C_parts))
join_build_ab_and_c_on_bc_join_key = map(
lambda p: query_plan.add_operator(
HashJoinBuild(
settings.table_B_BC_join_key,
'join_build_ab_and_c_on_bc_join_key' + '_{}'.format(
p), query_plan, False)),
range(0, settings.table_C_parts))
join_probe_ab_and_c_on_bc_join_key = map(
lambda p: query_plan.add_operator(
HashJoinProbe(
JoinExpression(settings.table_B_BC_join_key, settings.
table_C_BC_join_key),
'join_probe_ab_and_c_on_bc_join_key' + '_{}'.format(
p), query_plan, False)),
range(0, settings.table_C_parts))
join_build_abc_and_c_on_c_pk = map(
lambda p: query_plan.add_operator(
HashJoinBuild(settings.table_C_primary_key,
'join_build_abc_and_c_on_c_pk' + '_{}'.format(
p), query_plan, False)),
range(0, settings.table_C_parts))
join_probe_abc_and_c_on_c_pk = map(
lambda p: query_plan.add_operator(
HashJoinProbe(
JoinExpression(settings.table_C_primary_key, settings.
table_C_primary_key),
'join_probe_abc_and_c_on_c_pk' + '_{}'.format(
p), query_plan, False)),
range(0, settings.table_C_parts))
bloom_create_bc_join_key = map(
lambda p: query_plan.add_operator(
BloomCreate(settings.table_B_BC_join_key,
'bloom_create_bc_join_key' + '_{}'.format(
p), query_plan, False)),
range(0, settings.table_B_parts))
map_bc_c_join_key = map(
lambda p: query_plan.add_operator(
Map(settings.table_C_BC_join_key, 'map_bc_c_join_key' + '_{}'.
format(p), query_plan, False)),
range(0, settings.table_B_parts))
field_names_map_b = OrderedDict(
zip([
'_{}'.format(i)
for i, name in enumerate(settings.table_B_field_names)
], settings.table_B_field_names))
def project_fn_b(df):
df.rename(columns=field_names_map_b, inplace=True)
return df
project_b = map(
lambda p: query_plan.add_operator(
Project([
ProjectExpression(k, v)
for k, v in field_names_map_b.iteritems()
], 'project_b' + '_{}'.format(p), query_plan, False, project_fn_b)
), range(0, settings.table_B_parts))
map_ab_a_join_key = map(
lambda p: query_plan.add_operator(
Map(settings.table_A_AB_join_key, 'map_ab_a_join_key' + '_{}'
.format(p), query_plan, False)),
range(0, settings.table_A_parts))
map_ab_b_join_key = map(
lambda p: query_plan.add_operator(
Map(settings.table_B_AB_join_key, 'map_ab_b_join_key' + '_{}'
.format(p), query_plan, False)),
range(0, settings.table_B_parts))
join_build_a_and_b_on_ab_join_key = map(
lambda p: query_plan.add_operator(
HashJoinBuild(
settings.table_A_AB_join_key,
'join_build_a_and_b_on_ab_join_key' + '_{}'.format(
p), query_plan, False)), range(0, settings.table_B_parts))
join_probe_a_and_b_on_ab_join_key = map(
lambda p: query_plan.add_operator(
HashJoinProbe(
JoinExpression(settings.table_A_AB_join_key, settings.
table_B_AB_join_key),
'join_probe_a_and_b_on_ab_join_key' + '_{}'.format(
p), query_plan, False)), range(0, settings.table_B_parts))
if settings.table_C_key is None:
def part_aggregate_fn(df):
sum_ = df[settings.table_B_detail_field_name].astype(
np.float).sum()
return pd.DataFrame({'_0': [sum_]})
part_aggregate = map(
lambda p: query_plan.add_operator(
Aggregate([
AggregateExpression(
AggregateExpression.SUM, lambda t: float(t[
settings.table_B_detail_field_name]))
], settings.use_pandas, 'part_aggregate_{}'.format(p),
query_plan, False, part_aggregate_fn)),
range(0, settings.table_B_parts))
else:
def part_aggregate_fn(df):
sum_ = df[settings.table_C_detail_field_name].astype(
np.float).sum()
return pd.DataFrame({'_0': [sum_]})
part_aggregate = map(
lambda p: query_plan.add_operator(
Aggregate([
AggregateExpression(
AggregateExpression.SUM, lambda t: float(t[
settings.table_C_detail_field_name]))
], settings.use_pandas, 'part_aggregate_{}'.format(p),
query_plan, False, part_aggregate_fn)),
range(0, settings.table_C_parts))
def aggregate_reduce_fn(df):
sum_ = df['_0'].astype(np.float).sum()
return pd.DataFrame({'_0': [sum_]})
aggregate_reduce = query_plan.add_operator(
Aggregate([
AggregateExpression(AggregateExpression.SUM,
lambda t: float(t['_0']))
], settings.use_pandas, 'aggregate_reduce', query_plan, False,
aggregate_reduce_fn))
aggregate_project = query_plan.add_operator(
Project([ProjectExpression(lambda t: t['_0'], 'total_balance')],
'aggregate_project', query_plan, False))
collate = query_plan.add_operator(Collate('collate', query_plan, False))
# Inline some of the operators
map(lambda o: o.set_async(False), project_a)
map(lambda o: o.set_async(False), project_b)
map(lambda o: o.set_async(False), map_ab_a_join_key)
map(lambda o: o.set_async(False), map_ab_b_join_key)
if settings.table_C_key is None:
map(lambda o: o.set_async(False), map_b_pk_1)
map(lambda o: o.set_async(False), map_b_pk_2)
map(lambda o: o.set_async(False), project_b_detail)
else:
map(lambda o: o.set_async(False), map_bc_b_join_key)
map(lambda o: o.set_async(False), map_bc_c_join_key)
map(lambda o: o.set_async(False), map_c_pk_1)
map(lambda o: o.set_async(False), map_c_pk_2)
map(lambda o: o.set_async(False), project_c)
map(lambda o: o.set_async(False), project_c_detail)
aggregate_project.set_async(False)
# Connect the operators
connect_many_to_many(scan_a, project_a)
connect_many_to_many(project_a, map_ab_a_join_key)
connect_all_to_all(map_ab_a_join_key, join_build_a_and_b_on_ab_join_key)
connect_all_to_all(project_a, bloom_create_ab_join_key)
# connect_all_to_all(map_A_to_B, join_build_a_and_b_on_ab_join_key)
connect_many_to_many(join_build_a_and_b_on_ab_join_key,
join_probe_a_and_b_on_ab_join_key)
# connect_all_to_all(map_bloom_A_to_B, bloom_create_ab_join_key)
connect_many_to_many(bloom_create_ab_join_key, scan_b_on_ab_join_key)
connect_many_to_many(scan_b_on_ab_join_key, project_b)
# connect_many_to_many(project_b, join_probe_a_and_b_on_ab_join_key)
# connect_all_to_all(map_B_to_B, join_probe_a_and_b_on_ab_join_key)
connect_many_to_many(project_b, map_ab_b_join_key)
connect_all_to_all(map_ab_b_join_key, join_probe_a_and_b_on_ab_join_key)
# connect_many_to_many(join_probe_a_and_b_on_ab_join_key, map_bloom_B_to_B)
if settings.table_C_key is None:
# connect_all_to_all(join_probe_a_and_b_on_ab_join_key, part_aggregate)
connect_many_to_many(scan_b_detail_on_b_pk, project_b_detail)
connect_many_to_many(project_b_detail, map_b_pk_2)
connect_many_to_many(bloom_create_b_pk, scan_b_detail_on_b_pk)
connect_all_to_all(join_probe_a_and_b_on_ab_join_key,
bloom_create_b_pk)
connect_all_to_all(map_b_pk_2, join_probe_ab_and_b_on_b_pk)
connect_many_to_many(join_probe_ab_and_b_on_b_pk, part_aggregate)
connect_many_to_many(join_build_ab_and_b_on_b_pk,
join_probe_ab_and_b_on_b_pk)
connect_many_to_many(join_probe_a_and_b_on_ab_join_key, map_b_pk_1)
connect_all_to_all(map_b_pk_1, join_build_ab_and_b_on_b_pk)
else:
connect_all_to_all(join_probe_a_and_b_on_ab_join_key,
bloom_create_bc_join_key)
connect_many_to_many(bloom_create_bc_join_key, scan_c_on_bc_join_key)
connect_many_to_many(scan_c_on_bc_join_key, project_c)
# connect_many_to_many(project_c, join_probe_ab_and_c_on_bc_join_key)
connect_all_to_all(map_bc_c_join_key,
join_probe_ab_and_c_on_bc_join_key)
# connect_many_to_many(join_probe_a_and_b_on_ab_join_key, join_build_ab_and_c_on_bc_join_key)
connect_many_to_many(join_probe_a_and_b_on_ab_join_key,
map_bc_b_join_key)
connect_all_to_all(map_bc_b_join_key,
join_build_ab_and_c_on_bc_join_key)
connect_all_to_all(join_probe_ab_and_c_on_bc_join_key,
bloom_create_c_pk)
# connect_many_to_many(join_probe_ab_and_c_on_bc_join_key, join_build_abc_and_c_on_c_pk)
connect_many_to_many(join_probe_ab_and_c_on_bc_join_key, map_c_pk_1)
connect_all_to_all(map_c_pk_1, join_build_abc_and_c_on_c_pk)
connect_many_to_many(bloom_create_c_pk, scan_c_detail_on_c_pk)
# connect_all_to_all(bloom_create_bc_join_key, scan_c_detail_on_c_pk)
connect_many_to_many(join_build_abc_and_c_on_c_pk,
join_probe_abc_and_c_on_c_pk)
# connect_many_to_many(join_probe_a_and_b_on_ab_join_key, map_B_to_C)
# connect_all_to_all(join_probe_a_and_b_on_ab_join_key, join_build_abc_and_c_on_c_pk)
connect_many_to_many(scan_c_detail_on_c_pk, project_c_detail)
# connect_many_to_many(project_c_detail, map_C_to_C)
# connect_all_to_all(project_c_detail, join_probe_abc_and_c_on_c_pk)
connect_many_to_many(project_c_detail, map_c_pk_2)
connect_many_to_many(project_c, map_bc_c_join_key)
connect_many_to_many(join_build_ab_and_c_on_bc_join_key,
join_probe_ab_and_c_on_bc_join_key)
connect_all_to_all(map_c_pk_2, join_probe_abc_and_c_on_c_pk)
connect_many_to_many(join_probe_abc_and_c_on_c_pk, part_aggregate)
connect_many_to_one(part_aggregate, aggregate_reduce)
connect_one_to_one(aggregate_reduce, aggregate_project)
connect_one_to_one(aggregate_project, collate)
return query_plan
def run(parallel, use_pandas, secure, use_native, buffer_size, lineitem_parts,
part_parts, lineitem_sharded, part_sharded, other_parts, sf, fp_rate,
expected_result, format_):
"""
:return: None
"""
print('')
print("TPCH Q17 Bloom Join")
print("-------------------")
query_plan = QueryPlan(is_async=parallel, buffer_size=buffer_size)
# Query plan
part_scan = map(
lambda p: query_plan.add_operator(
tpch_q17.sql_scan_select_partkey_where_brand_container_op(
part_sharded, p, part_parts, use_pandas, secure, use_native,
'part_scan' + '_' + str(p), query_plan, sf, format_)),
range(0, part_parts))
part_project = map(
lambda p: query_plan.add_operator(
tpch_q17.project_partkey_op('part_project' + '_' + str(p),
query_plan)), range(0, part_parts))
part_lineitem_join_build_map = map(
lambda p: query_plan.add_operator(
Map('p_partkey', 'part_lineitem_join_build_map' + '_' + str(p),
query_plan, False)), range(0, part_parts))
part_bloom_create = query_plan.add_operator(
tpch_q17.bloom_create_partkey_op(fp_rate, 'part_bloom_create',
query_plan))
lineitem_bloom_use = \
map(lambda p:
query_plan.add_operator(
tpch_q17.bloom_scan_lineitem_select_orderkey_partkey_quantity_extendedprice_bloom_partkey_op(
lineitem_sharded,
p,
part_parts,
use_pandas,
secure,
use_native,
'lineitem_bloom_use' + '_' + str(p),
query_plan,
sf, format_)),
range(0, lineitem_parts))
lineitem_project = map(
lambda p: query_plan.add_operator(
tpch_q17.
project_lineitem_filtered_orderkey_partkey_quantity_extendedprice_op(
'lineitem_project' + '_' + str(p), query_plan)),
range(0, lineitem_parts))
part_lineitem_join_probe_map = map(
lambda p: query_plan.add_operator(
Map('l_partkey', 'part_lineitem_join_probe_map' + '_' + str(p),
query_plan, False)), range(0, lineitem_parts))
part_lineitem_join_build = map(
lambda p: query_plan.add_operator(
HashJoinBuild('p_partkey', 'part_lineitem_join_build' + '_' + str(
p), query_plan, False)), range(0, other_parts))
part_lineitem_join_probe = map(
lambda p: query_plan.add_operator(
HashJoinProbe(JoinExpression('p_partkey', 'l_partkey'),
'part_lineitem_join_probe' + '_' + str(
p), query_plan, False)), range(0, other_parts))
lineitem_part_avg_group = map(
lambda p: query_plan.add_operator(
tpch_q17.group_partkey_avg_quantity_op(
'lineitem_part_avg_group' + '_' + str(p), query_plan)),
range(0, other_parts))
lineitem_part_avg_group_project = map(
lambda p: query_plan.add_operator(
tpch_q17.project_partkey_avg_quantity_op(
'lineitem_part_avg_group_project' + '_' + str(p), query_plan)),
range(0, other_parts))
part_lineitem_join_avg_group_join_build = \
map(lambda p:
query_plan.add_operator(
HashJoinBuild('l_partkey',
'part_lineitem_join_avg_group_join_build' + '_' + str(p),
query_plan,
False)),
range(0, other_parts))
part_lineitem_join_avg_group_join_probe = \
map(lambda p:
query_plan.add_operator(
HashJoinProbe(JoinExpression('l_partkey', 'l_partkey'),
'part_lineitem_join_avg_group_join_probe' + '_' + str(p),
query_plan,
False)),
range(0, other_parts))
lineitem_filter = map(
lambda p: query_plan.add_operator(
tpch_q17.filter_lineitem_quantity_op(
'lineitem_filter' + '_' + str(p), query_plan)),
range(0, other_parts))
extendedprice_sum_aggregate = map(
lambda p: query_plan.add_operator(
tpch_q17.aggregate_sum_extendedprice_op(
use_pandas, 'extendedprice_sum_aggregate' + '_' + str(p),
query_plan)), range(0, other_parts))
def aggregate_reduce_fn(df):
sum1_ = df['_0'].astype(np.float).sum()
return pd.DataFrame({'_0': [sum1_]})
aggregate_reduce = query_plan.add_operator(
Aggregate([
AggregateExpression(AggregateExpression.SUM,
lambda t: float(t['_0']))
], use_pandas, 'aggregate_reduce', query_plan, False,
aggregate_reduce_fn))
extendedprice_sum_aggregate_project = query_plan.add_operator(
tpch_q17.project_avg_yearly_op('extendedprice_sum_aggregate_project',
query_plan))
collate = query_plan.add_operator(
tpch_q17.collate_op('collate', query_plan))
# Inline what we can
map(lambda o: o.set_async(False), lineitem_project)
map(lambda o: o.set_async(False), part_project)
map(lambda o: o.set_async(False), lineitem_filter)
map(lambda o: o.set_async(False), part_lineitem_join_probe_map)
map(lambda o: o.set_async(False), part_lineitem_join_build_map)
map(lambda o: o.set_async(False), lineitem_part_avg_group)
map(lambda o: o.set_async(False), lineitem_part_avg_group_project)
map(lambda o: o.set_async(False), extendedprice_sum_aggregate)
extendedprice_sum_aggregate_project.set_async(False)
# Connect the operators
# part_scan.connect(part_project)
map(lambda (p, o): o.connect(part_project[p]), enumerate(part_scan))
# map(lambda (p, o): o.connect(part_bloom_create_map[p]), enumerate(part_project))
map(lambda (p, o): o.connect(part_lineitem_join_build_map[p]),
enumerate(part_project))
connect_many_to_one(part_project, part_bloom_create)
connect_one_to_many(part_bloom_create, lineitem_bloom_use)
# part_project.connect(part_bloom_create)
# map(lambda (p1, o1): map(lambda (p2, o2): o1.connect(o2), enumerate(part_bloom_create)),
# enumerate(part_bloom_create_map))
# part_bloom_create.connect(lineitem_bloom_use)
# map(lambda (p1, o1): map(lambda (p2, o2): o1.connect(o2), enumerate(lineitem_bloom_use)),
# enumerate(part_bloom_create))
# lineitem_bloom_use.connect(lineitem_project)
map(lambda (p, o): o.connect(lineitem_project[p]),
enumerate(lineitem_bloom_use))
# part_lineitem_join.connect_left_producer(part_project)
map(
lambda (p1, o1): map(lambda (p2, o2): o1.connect(o2),
enumerate(part_lineitem_join_build)),
enumerate(part_lineitem_join_build_map))
map(lambda (p, o): part_lineitem_join_probe[p].connect_build_producer(o),
enumerate(part_lineitem_join_build))
# part_lineitem_join.connect_right_producer(lineitem_project)
# map(lambda (p, o): o.connect(part_lineitem_join_probe_map[p]), enumerate(lineitem_project))
connect_many_to_many(lineitem_project, part_lineitem_join_probe_map)
map(
lambda (p1, o1): map(lambda (p2, o2): o2.connect_tuple_producer(o1),
enumerate(part_lineitem_join_probe)),
enumerate(part_lineitem_join_probe_map))
# part_lineitem_join.connect(lineitem_part_avg_group)
map(lambda (p, o): o.connect(lineitem_part_avg_group[p]),
enumerate(part_lineitem_join_probe))
# lineitem_part_avg_group.connect(lineitem_part_avg_group_project)
map(lambda (p, o): o.connect(lineitem_part_avg_group_project[p]),
enumerate(lineitem_part_avg_group))
# part_lineitem_join_avg_group_join.connect_left_producer(lineitem_part_avg_group_project)
map(lambda (p, o): o.connect(part_lineitem_join_avg_group_join_build[p]),
enumerate(lineitem_part_avg_group_project))
# part_lineitem_join_avg_group_join.connect_right_producer(part_lineitem_join)
map(
lambda (p, o): part_lineitem_join_avg_group_join_probe[p].
connect_build_producer(o),
enumerate(part_lineitem_join_avg_group_join_build))
map(
lambda (p, o): part_lineitem_join_avg_group_join_probe[p].
connect_tuple_producer(o), enumerate(part_lineitem_join_probe))
# part_lineitem_join_avg_group_join.connect(lineitem_filter)
map(lambda (p, o): o.connect(lineitem_filter[p]),
enumerate(part_lineitem_join_avg_group_join_probe))
# lineitem_filter.connect(extendedprice_sum_aggregate)
map(lambda (p, o): o.connect(extendedprice_sum_aggregate[p]),
enumerate(lineitem_filter))
# extendedprice_sum_aggregate.connect(extendedprice_sum_aggregate_project)
map(lambda (p, o): o.connect(aggregate_reduce),
enumerate(extendedprice_sum_aggregate))
aggregate_reduce.connect(extendedprice_sum_aggregate_project)
# extendedprice_sum_aggregate_project.connect(collate)
extendedprice_sum_aggregate_project.connect(collate)
# Plan settings
print('')
print("Settings")
print("--------")
print('')
print('use_pandas: {}'.format(use_pandas))
print('secure: {}'.format(secure))
print('use_native: {}'.format(use_native))
print("lineitem parts: {}".format(lineitem_parts))
print("part_parts: {}".format(part_parts))
print("lineitem_sharded: {}".format(lineitem_sharded))
print("part_sharded: {}".format(part_sharded))
print("other_parts: {}".format(other_parts))
print("fp_rate: {}".format(fp_rate))
print("format: {}".format(format_))
print('')
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../benchmark-output"),
gen_test_id())
# Start the query
query_plan.execute()
tuples = collate.tuples()
collate.print_tuples(tuples)
# Write the metrics
query_plan.print_metrics()
# Shut everything down
query_plan.stop()
field_names = ['avg_yearly']
assert len(tuples) == 1 + 1
assert tuples[0] == field_names
# NOTE: This result has been verified with the equivalent data and query on PostgreSQL
if s3filter.util.constants.TPCH_SF == 10:
assert round(
float(tuples[1][0]), 10
) == 372414.2899999995 # TODO: This isn't correct but haven't checked tpch17 on 10 sf yet
elif s3filter.util.constants.TPCH_SF == 1:
numpy.testing.assert_approx_equal(float(tuples[1][0]), expected_result)
def test_sort_desc():
"""Executes a sorted query. The results are collated.
:return: None
"""
query_plan = QueryPlan()
# Query plan
ts = query_plan.add_operator(
SQLTableScan('supplier.csv', 'select * from S3Object '
'limit 3;', False, 'ts', query_plan, False))
s = query_plan.add_operator(
Sort([SortExpression('_5', float, 'DESC')], 's', query_plan, False))
c = query_plan.add_operator(Collate('c', query_plan, False))
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../tests-output"),
gen_test_id())
ts.connect(s)
s.connect(c)
# Start the query
query_plan.execute()
# Assert the results
# num_rows = 0
# for t in c.tuples():
# num_rows += 1
# print("{}:{}".format(num_rows, t))
assert len(c.tuples()) == 3 + 1
assert c.tuples()[0] == ['_0', '_1', '_2', '_3', '_4', '_5', '_6']
assert c.tuples()[1] == [
'1', 'Supplier#000000001', ' N kD4on9OM Ipw3,gf0JBoQDd7tgrzrddZ', '17',
'27-918-335-1736', '5755.94', 'each slyly above the careful'
]
assert c.tuples()[2] == [
'3', 'Supplier#000000003', 'q1,G3Pj6OjIuUYfUoH18BFTKP5aU9bEV3', '1',
'11-383-516-1199', '4192.40',
'blithely silent requests after the express dependencies are sl'
]
assert c.tuples()[3] == [
'2', 'Supplier#000000002', '89eJ5ksX3ImxJQBvxObC,', '5',
'15-679-861-2259', '4032.68', ' slyly bold instructions. idle dependen'
]
# Write the metrics
query_plan.print_metrics()
def query_plan(settings):
# type: (SyntheticBaselineJoinSettings) -> QueryPlan
"""
:type settings:
:return: None
"""
query_plan = QueryPlan(is_async=settings.parallel,
buffer_size=settings.buffer_size)
def scan_A_fun(df):
df.columns = settings.table_A_field_names
criterion = settings.table_A_filter_fn(df)
df = df[criterion]
return df
# Define the operators
scan_A = \
map(lambda p:
query_plan.add_operator(
SQLTableScan(get_file_key(settings.table_A_key, settings.table_A_sharded, p),
"select "
" * "
"from "
" S3Object "
"{}"
.format(
get_sql_suffix(settings.table_A_key, settings.table_A_parts, p,
settings.table_A_sharded, add_where=True)), settings.format_,
settings.use_pandas,
settings.secure,
settings.use_native,
'scan_A_{}'.format(p),
query_plan,
False, fn=scan_A_fun)),
range(0, settings.table_A_parts))
def scan_B_fun(df):
df.columns = settings.table_B_field_names
criterion = settings.table_B_filter_fn(df)
return df[criterion]
scan_B = \
map(lambda p:
query_plan.add_operator(
SQLTableScan(get_file_key(settings.table_B_key, settings.table_B_sharded, p),
"select "
" * "
"from "
" S3Object "
"{}"
.format(
get_sql_suffix(settings.table_B_key, settings.table_B_parts, p,
settings.table_B_sharded, add_where=True)), settings.format_,
settings.use_pandas,
settings.secure,
settings.use_native,
'scan_B_{}'.format(p),
query_plan,
False, fn=scan_B_fun)),
range(0, settings.table_B_parts))
"""
field_names_map_B = OrderedDict(
zip(['_{}'.format(i) for i, name in enumerate(settings.table_B_field_names)], settings.table_B_field_names))
def project_fn_B(df):
df.rename(columns=field_names_map_B, inplace=True)
return df
project_B = map(lambda p:
query_plan.add_operator(Project(
[ProjectExpression(k, v) for k, v in field_names_map_B.iteritems()],
'project_B_{}'.format(p),
query_plan,
True,
project_fn_B)),
range(0, settings.table_B_parts))
filter_b = map(lambda p:
query_plan.add_operator(Filter(
PredicateExpression(None, pd_expr=settings.table_B_filter_fn), 'filter_b' + '_{}'.format(p), query_plan,
False)),
range(0, settings.table_B_parts))
"""
def scan_C_fun(df):
df.columns = settings.table_C_field_names
criterion = settings.table_C_filter_fn(df)
return df[criterion]
scan_C = \
map(lambda p:
query_plan.add_operator(
SQLTableScan(get_file_key(settings.table_C_key, settings.table_C_sharded, p),
"select "
" * "
"from "
" S3Object "
"{}"
.format(
get_sql_suffix(settings.table_C_key, settings.table_C_parts, p,
settings.table_C_sharded, add_where=True)), settings.format_,
settings.use_pandas,
settings.secure,
settings.use_native,
'scan_C_{}'.format(p),
query_plan,
False, fn=scan_C_fun)),
range(0, settings.table_C_parts))
"""
field_names_map_C = OrderedDict(
zip(['_{}'.format(i) for i, name in enumerate(settings.table_C_field_names)], settings.table_C_field_names))
def project_fn_C(df):
df.rename(columns=field_names_map_C, inplace=True)
return df
project_C = map(lambda p:
query_plan.add_operator(Project(
[ProjectExpression(k, v) for k, v in field_names_map_C.iteritems()],
'project_C_{}'.format(p),
query_plan,
True,
project_fn_C)),
range(0, settings.table_C_parts))
filter_c = map(lambda p:
query_plan.add_operator(Filter(
PredicateExpression(None, pd_expr=settings.table_C_filter_fn), 'filter_c' + '_{}'.format(p), query_plan,
False)),
range(0, settings.table_C_parts))
"""
map_A_to_B = map(
lambda p: query_plan.add_operator(
Map(settings.table_A_AB_join_key, 'map_A_to_B_{}'.format(p),
query_plan, False)), range(0, settings.table_A_parts))
map_B_to_B = map(
lambda p: query_plan.add_operator(
Map(settings.table_B_AB_join_key, 'map_B_to_B_{}'.format(p),
query_plan, False)), range(0, settings.table_B_parts))
map_B_to_C = map(
lambda p: query_plan.add_operator(
Map(settings.table_B_BC_join_key, 'map_B_to_C_{}'.format(p),
query_plan, False)), range(0, settings.table_B_parts))
map_C_to_C = map(
lambda p: query_plan.add_operator(
Map(settings.table_C_BC_join_key, 'map_C_to_C_{}'.format(p),
query_plan, False)), range(0, settings.table_C_parts))
join_build_A_B = map(
lambda p: query_plan.add_operator(
HashJoinBuild(settings.table_A_AB_join_key, 'join_build_A_B_{}'.
format(p), query_plan, False)),
range(0, settings.table_B_parts))
join_probe_A_B = map(
lambda p: query_plan.add_operator(
HashJoinProbe(
JoinExpression(settings.table_A_AB_join_key, settings.
table_B_AB_join_key), 'join_probe_A_B_{}'
.format(p), query_plan, False)),
range(0, settings.table_B_parts))
join_build_AB_C = map(
lambda p: query_plan.add_operator(
HashJoinBuild(settings.table_B_BC_join_key, 'join_build_AB_C_{}'.
format(p), query_plan, False)),
range(0, settings.table_C_parts))
join_probe_AB_C = map(
lambda p: query_plan.add_operator(
HashJoinProbe(
JoinExpression(settings.table_B_BC_join_key, settings.
table_C_BC_join_key), 'join_probe_AB_C_{}'
.format(p), query_plan, False)),
range(0, settings.table_C_parts))
def agg_fun(df):
return pd.DataFrame({
'sum':
[df[settings.table_C_detail_field_name].astype(float).sum()]
})
#return pd.DataFrame( { 'sum' : [ len(df) ] } )
part_aggregate = map(
lambda p: query_plan.add_operator(
Aggregate([
AggregateExpression(
AggregateExpression.SUM, lambda t: float(t[
settings.table_C_detail_field_name]))
], settings.use_pandas, 'part_aggregate_{}'.format(p), query_plan,
False, agg_fun)), range(0, settings.table_C_parts))
def agg_reduce_fun(df):
return pd.DataFrame({'sum': [df['sum'].sum()]})
aggregate_reduce = query_plan.add_operator(
Aggregate([
AggregateExpression(AggregateExpression.SUM,
lambda t: float(t['_0']))
], settings.use_pandas, 'aggregate_reduce', query_plan, False,
agg_reduce_fun))
"""
aggregate_project = query_plan.add_operator(Project(
[
ProjectExpression(lambda t: t['_0'], 'total_balance')
],
'aggregate_project', query_plan,
False))
"""
collate = query_plan.add_operator(Collate('collate', query_plan, False))
# Connect the operators
connect_many_to_many(scan_A, map_A_to_B)
#connect_many_to_many(project_A, filter_A)
#connect_many_to_many(filter_A, map_A_to_B)
connect_all_to_all(map_A_to_B, join_build_A_B)
connect_many_to_many(join_build_A_B, join_probe_A_B)
connect_many_to_many(scan_B, map_B_to_B)
#connect_many_to_many(project_B, filter_b)
#connect_many_to_many(filter_b, map_B_to_B)
connect_all_to_all(map_B_to_B, join_probe_A_B)
#connect_many_to_many(join_probe_A_B, part_aggregate)
#connect_many_to_one(part_aggregate, aggregate_reduce)
#connect_one_to_one(aggregate_reduce, collate)
connect_many_to_many(join_build_AB_C, join_probe_AB_C)
connect_many_to_many(join_probe_A_B, map_B_to_C)
connect_all_to_all(map_B_to_C, join_build_AB_C)
connect_many_to_many(scan_C, map_C_to_C)
#connect_many_to_many(project_C, filter_c)
#connect_many_to_many(filter_c, map_C_to_C)
connect_all_to_all(map_C_to_C, join_probe_AB_C)
connect_many_to_many(join_probe_AB_C, part_aggregate)
connect_many_to_one(part_aggregate, aggregate_reduce)
connect_one_to_one(aggregate_reduce, collate)
#connect_one_to_one(aggregate_project, collate)
return query_plan
def query_plan(settings):
# type: (SyntheticFilteredJoinSettings) -> QueryPlan
"""
:return: None
"""
if settings.use_shared_mem:
system = WorkerSystem(settings.shared_memory_size)
else:
system = None
query_plan = QueryPlan(system,
is_async=settings.parallel,
buffer_size=settings.buffer_size,
use_shared_mem=settings.use_shared_mem)
# Define the operators
scan_A = \
map(lambda p:
query_plan.add_operator(
SQLTableScan(get_file_key(settings.table_A_key, settings.table_A_sharded, p, settings.sf),
"select "
" {} "
"from "
" S3Object "
" {} "
" {} "
.format(','.join(settings.table_A_field_names),
' where {} '.format(
settings.table_A_filter_sql) if settings.table_A_filter_sql is not None else '',
get_sql_suffix(settings.table_A_key, settings.table_A_parts, p,
settings.table_A_sharded,
add_where=settings.table_A_filter_sql is None)), settings.format_,
settings.use_pandas,
settings.secure,
settings.use_native,
'scan_A_{}'.format(p),
query_plan,
False)),
range(0, settings.table_A_parts))
field_names_map_A = OrderedDict(
zip([
'_{}'.format(i)
for i, name in enumerate(settings.table_A_field_names)
], settings.table_A_field_names))
def project_fn_A(df):
df = df.rename(columns=field_names_map_A, copy=False)
return df
project_A = map(
lambda p: query_plan.add_operator(
Project([
ProjectExpression(k, v)
for k, v in field_names_map_A.iteritems()
], 'project_A_{}'.format(p), query_plan, False, project_fn_A)),
range(0, settings.table_A_parts))
scan_B = \
map(lambda p:
query_plan.add_operator(
SQLTableScan(
get_file_key(settings.table_B_key, settings.table_B_sharded, p, settings.sf),
"select "
" {} "
"from "
" S3Object "
" {} "
" {} "
.format(','.join(settings.table_B_field_names),
' where {} '.format(
settings.table_B_filter_sql) if settings.table_B_filter_sql is not None else '',
get_sql_suffix(settings.table_B_key, settings.table_B_parts, p,
settings.table_B_sharded,
add_where=settings.table_B_filter_sql is None)), settings.format_,
settings.use_pandas,
settings.secure,
settings.use_native,
'scan_B_{}'.format(p),
query_plan,
False)),
range(0, settings.table_B_parts))
field_names_map_B = OrderedDict(
zip([
'_{}'.format(i)
for i, name in enumerate(settings.table_B_field_names)
], settings.table_B_field_names))
def project_fn_B(df):
df.rename(columns=field_names_map_B, inplace=True)
return df
project_B = map(
lambda p: query_plan.add_operator(
Project([
ProjectExpression(k, v)
for k, v in field_names_map_B.iteritems()
], 'project_B_{}'.format(p), query_plan, False, project_fn_B)),
range(0, settings.table_B_parts))
if settings.table_C_key is not None:
scan_C = \
map(lambda p:
query_plan.add_operator(
SQLTableScan(
get_file_key(settings.table_C_key, settings.table_C_sharded, p, settings.sf),
"select "
" {} "
"from "
" S3Object "
"where "
" {} "
" {} "
.format(','.join(settings.table_C_field_names),
settings.table_C_filter_sql,
get_sql_suffix(settings.table_C_key, settings.table_C_parts, p,
settings.table_C_sharded, add_where=False)), settings.format_,
settings.use_pandas,
settings.secure,
settings.use_native,
'scan_C_{}'.format(p),
query_plan,
False)),
range(0, settings.table_C_parts))
field_names_map_C = OrderedDict(
zip([
'_{}'.format(i)
for i, name in enumerate(settings.table_C_field_names)
], settings.table_C_field_names))
def project_fn_C(df):
df = df.rename(columns=field_names_map_C, copy=False)
return df
project_C = map(
lambda p: query_plan.add_operator(
Project([
ProjectExpression(k, v)
for k, v in field_names_map_C.iteritems()
], 'project_C_{}'.format(p), query_plan, False, project_fn_C)),
range(0, settings.table_C_parts))
map_B_to_C = map(
lambda p: query_plan.add_operator(
Map(settings.table_B_BC_join_key, 'map_B_to_C_{}'.format(p),
query_plan, False)), range(0, settings.table_B_parts))
map_C_to_C = map(
lambda p: query_plan.add_operator(
Map(settings.table_C_BC_join_key, 'map_C_to_C_{}'.format(p),
query_plan, False)), range(0, settings.table_C_parts))
join_build_AB_C = map(
lambda p: query_plan.add_operator(
HashJoinBuild(settings.table_B_BC_join_key,
'join_build_AB_C_{}'.format(
p), query_plan, False)),
range(0, settings.table_C_parts))
join_probe_AB_C = map(
lambda p: query_plan.add_operator(
HashJoinProbe(
JoinExpression(settings.table_B_BC_join_key, settings.
table_C_BC_join_key), 'join_probe_AB_C_{}'.
format(p), query_plan, False)),
range(0, settings.table_C_parts))
map_A_to_B = map(
lambda p: query_plan.add_operator(
Map(settings.table_A_AB_join_key, 'map_A_to_B_{}'.format(p),
query_plan, False)), range(0, settings.table_A_parts))
map_B_to_B = map(
lambda p: query_plan.add_operator(
Map(settings.table_B_AB_join_key, 'map_B_to_B_{}'.format(p),
query_plan, False)), range(0, settings.table_B_parts))
join_build_A_B = map(
lambda p: query_plan.add_operator(
HashJoinBuild(settings.table_A_AB_join_key, 'join_build_A_B_{}'.
format(p), query_plan, False)),
range(0, settings.other_parts))
join_probe_A_B = map(
lambda p: query_plan.add_operator(
HashJoinProbe(
JoinExpression(settings.table_A_AB_join_key, settings.
table_B_AB_join_key), 'join_probe_A_B_{}'.
format(p), query_plan, False)), range(0, settings.other_parts))
if settings.table_C_key is None:
def part_aggregate_fn(df):
sum_ = df[settings.table_B_detail_field_name].astype(
np.float).sum()
return pd.DataFrame({'_0': [sum_]})
part_aggregate = map(
lambda p: query_plan.add_operator(
Aggregate([
AggregateExpression(
AggregateExpression.SUM, lambda t: float(t[
settings.table_B_detail_field_name]))
], settings.use_pandas, 'part_aggregate_{}'.format(p),
query_plan, False, part_aggregate_fn)),
range(0, settings.other_parts))
else:
def part_aggregate_fn(df):
sum_ = df[settings.table_C_detail_field_name].astype(
np.float).sum()
return pd.DataFrame({'_0': [sum_]})
part_aggregate = map(
lambda p: query_plan.add_operator(
Aggregate([
AggregateExpression(
AggregateExpression.SUM, lambda t: float(t[
settings.table_C_detail_field_name]))
], settings.use_pandas, 'part_aggregate_{}'.format(p),
query_plan, False, part_aggregate_fn)),
range(0, settings.table_C_parts))
def aggregate_reduce_fn(df):
sum_ = df['_0'].astype(np.float).sum()
return pd.DataFrame({'_0': [sum_]})
aggregate_reduce = query_plan.add_operator(
Aggregate([
AggregateExpression(AggregateExpression.SUM,
lambda t: float(t['_0']))
], settings.use_pandas, 'aggregate_reduce', query_plan, False,
aggregate_reduce_fn))
aggregate_project = query_plan.add_operator(
Project([ProjectExpression(lambda t: t['_0'], 'total_balance')],
'aggregate_project', query_plan, False))
collate = query_plan.add_operator(Collate('collate', query_plan, False))
# Inline some of the operators
map(lambda o: o.set_async(False), project_A)
map(lambda o: o.set_async(False), project_B)
map(lambda o: o.set_async(False), map_A_to_B)
map(lambda o: o.set_async(False), map_B_to_B)
if settings.table_C_key is not None:
map(lambda o: o.set_async(False), map_B_to_C)
map(lambda o: o.set_async(False), map_C_to_C)
map(lambda o: o.set_async(False), project_C)
map(lambda o: o.set_async(False), part_aggregate)
aggregate_project.set_async(False)
# Connect the operators
connect_many_to_many(scan_A, project_A)
connect_many_to_many(project_A, map_A_to_B)
connect_all_to_all(map_A_to_B, join_build_A_B)
connect_many_to_many(join_build_A_B, join_probe_A_B)
connect_many_to_many(scan_B, project_B)
connect_many_to_many(project_B, map_B_to_B)
connect_all_to_all(map_B_to_B, join_probe_A_B)
if settings.table_C_key is None:
connect_many_to_many(join_probe_A_B, part_aggregate)
else:
connect_many_to_many(join_probe_A_B, map_B_to_C)
connect_all_to_all(map_B_to_C, join_build_AB_C)
connect_many_to_many(join_build_AB_C, join_probe_AB_C)
connect_many_to_many(scan_C, project_C)
connect_many_to_many(project_C, map_C_to_C)
connect_all_to_all(map_C_to_C, join_probe_AB_C)
connect_many_to_many(join_probe_AB_C, part_aggregate)
connect_many_to_one(part_aggregate, aggregate_reduce)
connect_one_to_one(aggregate_reduce, aggregate_project)
connect_one_to_one(aggregate_project, collate)
return query_plan
def run(parallel,
use_pandas,
buffer_size,
table_first_part,
table_parts,
queried_columns,
queried_aliases,
castable_aliases,
select_str,
aggregate_column,
filter_str,
path,
format_=Format.PARQUET):
secure = False
use_native = False
print('')
print("Parquet Aggregate Benchmark")
print("------------------")
# Query plan
query_plan = QueryPlan(is_async=parallel, buffer_size=buffer_size)
# SQL scan the file
scan = map(
lambda p: query_plan.add_operator(
SQLTableScan(
"{}/lineitem.{}.parquet".format(path, p
), "select {} from S3Object "
"{};".format(select_str, filter_str), format_, use_pandas,
secure, use_native, 'scan_{}'.format(p), query_plan, False)),
range(table_first_part, table_first_part + table_parts))
# project
def fn(df):
df.columns = queried_aliases
df[castable_aliases] = df[castable_aliases].astype(np.double)
return df
project = map(
lambda p: query_plan.add_operator(
Project([], 'project_{}'.format(p), query_plan, False, fn)),
range(table_first_part, table_first_part + table_parts))
# aggregation
def agg_fun(df):
if aggregate_column in df:
return pd.DataFrame({'sum': [sum(df[aggregate_column])]})
else:
return pd.DataFrame({'count': len(df)}, index=[0])
aggregate = query_plan.add_operator(
Aggregate([], True, 'agg', query_plan, False, agg_fun))
collate = query_plan.add_operator(Collate('collate', query_plan, False))
map(lambda (p, o): o.connect(project[p]), enumerate(scan))
map(lambda (p, o): o.connect(aggregate), enumerate(project))
aggregate.connect(collate)
# Plan settings
print('')
print("Settings")
print("--------")
print('')
print('use_pandas: {}'.format(use_pandas))
print("table parts: {}".format(table_parts))
print('')
# Write the plan graph
# query_plan.write_graph(os.path.join(ROOT_DIR, "../benchmark-output"), gen_test_id() + "-" + str(table_parts))
# Start the query
query_plan.execute()
print('Done')
tuples = collate.tuples()
collate.print_tuples(tuples)
# Write the metrics
query_plan.print_metrics()
# Shut everything down
query_plan.stop()
def test_filter_empty():
"""Executes a filter where no records are returned. We tst this as it's somewhat peculiar with s3 select, in so much
as s3 does not return column names when selecting data, meaning, unlike a traditional DBMS, no field names tuple
should be present in the results.
:return: None
"""
query_plan = QueryPlan()
# Query plan
ts = query_plan.add_operator(
SQLTableScan('lineitem.csv', 'select * from S3Object limit 0;', False,
'ts', query_plan, False))
f = query_plan.add_operator(
Filter(
PredicateExpression(lambda t_: cast(t_['_10'], timestamp) >= cast(
'1996-03-01', timestamp)), 'f', query_plan, False))
c = query_plan.add_operator(Collate('c', query_plan, False))
ts.connect(f)
f.connect(c)
# Write the plan graph
query_plan.write_graph(os.path.join(ROOT_DIR, "../tests-output"),
gen_test_id())
# Start the query
query_plan.execute()
# Assert the results
# num_rows = 0
# for t in c.tuples():
# num_rows += 1
# print("{}:{}".format(num_rows, t))
assert len(c.tuples()) == 0
# Write the metrics
query_plan.print_metrics()
