def disambiguate_integration_column_identifier(identifier,
integration_name,
table,
initial_name_as_alias=False):
"""Removes integration name from column if it's present, adds table path if it's absent"""
column_table_ref = [table.alias.to_string(
alias=False)] if table.alias else table.parts
parts = list(identifier.parts)
if len(parts) > 1:
if parts[0] == integration_name:
parts = parts[1:]
if len(parts) > 1:
if (len(parts) <= len(column_table_ref)
or parts[:len(column_table_ref)] != column_table_ref):
raise PlanningException(
f'Tried to query column {identifier.to_tree()} from integration {integration_name} table {column_table_ref}, but a different table name has been specified.'
)
elif len(parts) == 1:
# if parts[0] != column_table_ref:
parts = column_table_ref + parts
new_identifier = Identifier(parts=parts)
if identifier.alias:
new_identifier.alias = identifier.alias
elif initial_name_as_alias:
new_identifier.alias = Identifier(parts[-1])
return new_identifier
def test_select_dquote_alias(self, dialect):
sql = """
select
a as "database"
from information_schema.tables "database"
"""
ast = parse_sql(sql, dialect=dialect)
expected_ast = Select(
targets=[Identifier('a', alias=Identifier('database'))],
from_table=Identifier(parts=['information_schema', 'tables'],
alias=Identifier('database')),
)
assert str(ast) == str(expected_ast)
assert ast.to_tree() == expected_ast.to_tree()
def plan_join_two_tables(self, join):
select_left_step = self.plan_integration_select(
Select(targets=[Star()], from_table=join.left))
select_right_step = self.plan_integration_select(
Select(targets=[Star()], from_table=join.right))
left_integration_name, left_table = self.get_integration_path_from_identifier_or_error(
join.left)
right_integration_name, right_table = self.get_integration_path_from_identifier_or_error(
join.right)
left_table_path = left_table.to_string(alias=False)
right_table_path = right_table.to_string(alias=False)
new_condition_args = []
for arg in join.condition.args:
if isinstance(arg, Identifier):
if left_table_path in arg.parts:
new_condition_args.append(
disambiguate_integration_column_identifier(
arg, left_integration_name, left_table))
elif right_table_path in arg.parts:
new_condition_args.append(
disambiguate_integration_column_identifier(
arg, right_integration_name, right_table))
else:
raise PlanningException(
f'Wrong table or no source table in join condition for column: {str(arg)}'
)
else:
new_condition_args.append(arg)
new_join = copy.deepcopy(join)
new_join.condition.args = new_condition_args
new_join.left = Identifier(left_table_path, alias=left_table.alias)
new_join.right = Identifier(right_table_path, alias=right_table.alias)
# FIXME: INFORMATION_SCHEMA with condition
# clear join condition for INFORMATION_SCHEMA
if right_integration_name == 'INFORMATION_SCHEMA':
new_join.condition = None
return self.plan.add_step(
JoinStep(left=select_left_step.result,
right=select_right_step.result,
query=new_join))
def test_show_index(self):
sql = "SHOW INDEX FROM predictors"
ast = parse_sql(sql, dialect='mysql')
expected_ast = Show(category='INDEX',
from_table=Identifier('predictors'))
assert str(ast).lower() == sql.lower()
assert str(ast) == str(expected_ast)
assert ast.to_tree() == expected_ast.to_tree()
def disambiguate_predictor_column_identifier(identifier, predictor):
"""Removes integration name from column if it's present, adds table path if it's absent"""
table_ref = predictor.alias.parts_to_str(
) if predictor.alias else predictor.parts_to_str()
parts = list(identifier.parts)
if parts[0] == table_ref:
parts = parts[1:]
new_identifier = Identifier(parts=parts)
return new_identifier
def plan_fetch_timeseries_partitions(self, query, table,
predictor_group_by_names):
targets = [Identifier(column) for column in predictor_group_by_names]
query = Select(
distinct=True,
targets=targets,
from_table=table,
where=query.where,
)
select_step = self.plan_integration_select(query)
return select_step
def add_order_not_null(condition):
order_field_not_null = BinaryOperation(
op='is not',
args=[
Identifier(parts=[predictor_time_column_name]),
NullConstant()
])
if condition is not None:
condition = BinaryOperation(
op='and', args=[condition, order_field_not_null])
else:
condition = order_field_not_null
return condition
def test_select_varialbe_complex(self):
sql = f"""SELECT * FROM tab1 WHERE column1 in (SELECT column2 + @variable FROM t2)"""
ast = parse_sql(sql, dialect='mysql')
expected_ast = Select(targets=[Star()],
from_table=Identifier('tab1'),
where=BinaryOperation(
op='in',
args=(Identifier('column1'),
Select(targets=[
BinaryOperation(
op='+',
args=[
Identifier('column2'),
Variable('variable')
])
],
from_table=Identifier('t2'),
parentheses=True))))
assert ast.to_tree() == expected_ast.to_tree()
assert str(ast).lower() == sql.lower()
assert str(ast) == str(expected_ast)
def plan_project(self, query, dataframe, ignore_doubles=False):
out_identifiers = []
for target in query.targets:
if isinstance(target, Identifier) \
or isinstance(target, Star) \
or isinstance(target, Function) \
or isinstance(target, Constant):
out_identifiers.append(target)
else:
new_identifier = Identifier(str(target.to_string(alias=False)),
alias=target.alias)
out_identifiers.append(new_identifier)
return self.plan.add_step(
ProjectStep(dataframe=dataframe,
columns=out_identifiers,
ignore_doubles=ignore_doubles))
def test_between_with_and(self, dialect):
sql = "SELECT col1 FROM t1 WHERE col2 > 1 AND col1 BETWEEN a AND b"
ast = parse_sql(sql, dialect=dialect)
expected_ast = Select(
targets=[Identifier.from_path_str("col1")],
from_table=Identifier.from_path_str('t1'),
where=BinaryOperation(
'and',
args=[
BinaryOperation('>',
args=[
Identifier('col2'),
Constant(1),
]),
BetweenOperation(args=(Identifier.from_path_str('col1'),
Identifier.from_path_str('a'),
Identifier.from_path_str('b'))),
]))
assert ast.to_tree() == expected_ast.to_tree()
assert str(ast).lower() == sql.lower()
assert str(ast) == str(expected_ast)
def plan_join(self, query, integration=None):
join = query.from_table
join_left = join.left
join_right = join.right
if isinstance(join_left, Select):
# dbt query.
# TODO support complex query. Only one table is supported at the moment.
if not isinstance(join_left.from_table, Identifier):
raise PlanningException(
f'Statement not supported: {query.to_string()}')
# move properties to upper query
query = join_left
if query.from_table.alias is not None:
table_alias = [query.from_table.alias.parts[0]]
else:
table_alias = query.from_table.parts
def add_aliases(node, is_table, **kwargs):
if not is_table and isinstance(node, Identifier):
if len(node.parts) == 1:
# add table alias to field
node.parts = table_alias + node.parts
query_traversal(query.where, add_aliases)
if isinstance(query.from_table, Identifier):
# DBT workaround: allow use tables without integration.
# if table.part[0] not in integration - take integration name from create table command
if (integration is not None and query.from_table.parts[0]
not in self.integrations):
# add integration name to table
query.from_table.parts.insert(0, integration)
join_left = join_left.from_table
aliased_fields = self.get_aliased_fields(query.targets)
recursively_check_join_identifiers_for_ambiguity(query.where)
recursively_check_join_identifiers_for_ambiguity(
query.group_by, aliased_fields=aliased_fields)
recursively_check_join_identifiers_for_ambiguity(query.having)
recursively_check_join_identifiers_for_ambiguity(
query.order_by, aliased_fields=aliased_fields)
if isinstance(join_left, Identifier) and isinstance(
join_right, Identifier):
if self.is_predictor(join_left) and self.is_predictor(join_right):
raise PlanningException(
f'Can\'t join two predictors {str(join_left.parts[0])} and {str(join_left.parts[1])}'
)
predictor_namespace = None
predictor = None
table = None
predictor_is_left = False
if self.is_predictor(join_left):
predictor_namespace, predictor = get_predictor_namespace_and_name_from_identifier(
join_left, self.default_namespace)
predictor_is_left = True
else:
table = join_left
if self.is_predictor(join_right):
predictor_namespace, predictor = get_predictor_namespace_and_name_from_identifier(
join_right, self.default_namespace)
else:
table = join_right
last_step = None
if predictor:
# One argument is a table, another is a predictor
# Apply mindsdb model to result of last dataframe fetch
# Then join results of applying mindsdb with table
predictor_name = self.predictor_names[predictor.to_string(
alias=False).lower()]
if self.predictor_metadata[predictor_name].get('timeseries'):
predictor_steps = self.plan_timeseries_predictor(
query, table, predictor_namespace, predictor)
else:
predictor_steps = self.plan_predictor(
query, table, predictor_namespace, predictor)
# add join
# Update reference
_, table = self.get_integration_path_from_identifier_or_error(
table)
table_alias = table.alias or Identifier(
table.to_string(alias=False).replace('.', '_'))
left = Identifier(
predictor_steps['predictor'].result.ref_name,
alias=predictor.alias
or Identifier(predictor.to_string(alias=False)))
right = Identifier(predictor_steps['data'].result.ref_name,
alias=table_alias)
if not predictor_is_left:
# swap join
left, right = right, left
new_join = Join(left=left,
right=right,
join_type=join.join_type)
left = predictor_steps['predictor'].result
right = predictor_steps['data'].result
if not predictor_is_left:
# swap join
left, right = right, left
last_step = self.plan.add_step(
JoinStep(left=left, right=right, query=new_join))
# limit from timeseries
if predictor_steps.get('saved_limit'):
last_step = self.plan.add_step(
LimitOffsetStep(dataframe=last_step.result,
limit=predictor_steps['saved_limit']))
else:
# Both arguments are tables, join results of 2 dataframe fetches
join_step = self.plan_join_two_tables(join)
last_step = join_step
if query.where:
# FIXME: INFORMATION_SCHEMA with Where
right_integration_name, _ = self.get_integration_path_from_identifier_or_error(
join.right)
if right_integration_name == 'INFORMATION_SCHEMA':
...
else:
last_step = self.plan.add_step(
FilterStep(dataframe=last_step.result,
query=query.where))
if query.group_by:
group_by_targets = []
for t in query.targets:
target_copy = copy.deepcopy(t)
target_copy.alias = None
group_by_targets.append(target_copy)
last_step = self.plan.add_step(
GroupByStep(dataframe=last_step.result,
columns=query.group_by,
targets=group_by_targets))
if query.having:
last_step = self.plan.add_step(
FilterStep(dataframe=last_step.result,
query=query.having))
if query.order_by:
last_step = self.plan.add_step(
OrderByStep(dataframe=last_step.result,
order_by=query.order_by))
if query.limit is not None or query.offset is not None:
limit = query.limit.value if query.limit is not None else None
offset = query.offset.value if query.offset is not None else None
last_step = self.plan.add_step(
LimitOffsetStep(dataframe=last_step.result,
limit=limit,
offset=offset))
else:
raise PlanningException(
f'Join of unsupported objects, currently only tables and predictors can be joined.'
)
return self.plan_project(query, last_step.result)
def plan_timeseries_predictor(self, query, table, predictor_namespace,
predictor):
predictor_name = predictor.to_string(alias=False).lower()
# to original case
predictor_name = self.predictor_names[predictor_name]
predictor_time_column_name = self.predictor_metadata[predictor_name][
'order_by_column']
predictor_group_by_names = self.predictor_metadata[predictor_name][
'group_by_columns']
if predictor_group_by_names is None:
predictor_group_by_names = []
predictor_window = self.predictor_metadata[predictor_name]['window']
if query.order_by:
raise PlanningException(
f'Can\'t provide ORDER BY to time series predictor, it will be taken from predictor settings. Found: {query.order_by}'
)
saved_limit = query.limit
if query.group_by or query.having or query.offset:
raise PlanningException(
f'Unsupported query to timeseries predictor: {str(query)}')
allowed_columns = [predictor_time_column_name.lower()]
if len(predictor_group_by_names) > 0:
allowed_columns += [i.lower() for i in predictor_group_by_names]
validate_ts_where_condition(query.where,
allowed_columns=allowed_columns)
time_filter = find_time_filter(
query.where, time_column_name=predictor_time_column_name)
order_by = [
OrderBy(Identifier(parts=[predictor_time_column_name]),
direction='DESC')
]
preparation_where = copy.deepcopy(query.where)
# add {order_by_field} is not null
def add_order_not_null(condition):
order_field_not_null = BinaryOperation(
op='is not',
args=[
Identifier(parts=[predictor_time_column_name]),
NullConstant()
])
if condition is not None:
condition = BinaryOperation(
op='and', args=[condition, order_field_not_null])
else:
condition = order_field_not_null
return condition
preparation_where2 = copy.deepcopy(preparation_where)
preparation_where = add_order_not_null(preparation_where)
# Obtain integration selects
if isinstance(time_filter, BetweenOperation):
between_from = time_filter.args[1]
preparation_time_filter = BinaryOperation(
'<',
args=[Identifier(predictor_time_column_name), between_from])
preparation_where2 = replace_time_filter(preparation_where2,
time_filter,
preparation_time_filter)
integration_select_1 = Select(
targets=[Star()],
from_table=table,
where=add_order_not_null(preparation_where2),
order_by=order_by,
limit=Constant(predictor_window))
integration_select_2 = Select(targets=[Star()],
from_table=table,
where=preparation_where,
order_by=order_by)
integration_selects = [integration_select_1, integration_select_2]
elif isinstance(
time_filter, BinaryOperation
) and time_filter.op == '>' and time_filter.args[1] == Latest():
integration_select = Select(
targets=[Star()],
from_table=table,
where=preparation_where,
order_by=order_by,
limit=Constant(predictor_window),
)
integration_select.where = find_and_remove_time_filter(
integration_select.where, time_filter)
integration_selects = [integration_select]
elif isinstance(time_filter,
BinaryOperation) and time_filter.op in ('>', '>='):
time_filter_date = time_filter.args[1]
preparation_time_filter_op = {'>': '<=', '>=': '<'}[time_filter.op]
preparation_time_filter = BinaryOperation(
preparation_time_filter_op,
args=[
Identifier(predictor_time_column_name), time_filter_date
])
preparation_where2 = replace_time_filter(preparation_where2,
time_filter,
preparation_time_filter)
integration_select_1 = Select(
targets=[Star()],
from_table=table,
where=add_order_not_null(preparation_where2),
order_by=order_by,
limit=Constant(predictor_window))
integration_select_2 = Select(targets=[Star()],
from_table=table,
where=preparation_where,
order_by=order_by)
integration_selects = [integration_select_1, integration_select_2]
else:
integration_select = Select(
targets=[Star()],
from_table=table,
where=preparation_where,
order_by=order_by,
)
integration_selects = [integration_select]
if len(predictor_group_by_names) == 0:
# ts query without grouping
# one or multistep
if len(integration_selects) == 1:
select_partition_step = self.get_integration_select_step(
integration_selects[0])
else:
select_partition_step = MultipleSteps(steps=[
self.get_integration_select_step(s)
for s in integration_selects
],
reduce='union')
# fetch data step
data_step = self.plan.add_step(select_partition_step)
else:
# inject $var to queries
for integration_select in integration_selects:
condition = integration_select.where
for num, column in enumerate(predictor_group_by_names):
cond = BinaryOperation(
'=',
args=[Identifier(column),
Constant(f'$var[{column}]')])
# join to main condition
if condition is None:
condition = cond
else:
condition = BinaryOperation('and',
args=[condition, cond])
integration_select.where = condition
# one or multistep
if len(integration_selects) == 1:
select_partition_step = self.get_integration_select_step(
integration_selects[0])
else:
select_partition_step = MultipleSteps(steps=[
self.get_integration_select_step(s)
for s in integration_selects
],
reduce='union')
# get groping values
no_time_filter_query = copy.deepcopy(query)
no_time_filter_query.where = find_and_remove_time_filter(
no_time_filter_query.where, time_filter)
select_partitions_step = self.plan_fetch_timeseries_partitions(
no_time_filter_query, table, predictor_group_by_names)
# sub-query by every grouping value
map_reduce_step = self.plan.add_step(
MapReduceStep(values=select_partitions_step.result,
reduce='union',
step=select_partition_step))
data_step = map_reduce_step
predictor_step = self.plan.add_step(
ApplyTimeseriesPredictorStep(
output_time_filter=time_filter,
namespace=predictor_namespace,
dataframe=data_step.result,
predictor=predictor,
))
return {
'predictor': predictor_step,
'data': data_step,
'saved_limit': saved_limit,
}
def execute_step(self, step, steps_data):
if type(step) == GetPredictorColumns:
predictor_name = step.predictor.parts[-1]
dn = self.datahub.get(self.mindsdb_database_name)
columns = dn.get_table_columns(predictor_name)
columns = [(column_name, column_name) for column_name in columns]
data = {
'values': [],
'columns': {
(self.mindsdb_database_name, predictor_name, predictor_name):
columns
},
'tables':
[(self.mindsdb_database_name, predictor_name, predictor_name)]
}
elif type(step) == GetTableColumns:
table = step.table
dn = self.datahub.get(step.namespace)
ds_query = Select(from_table=Identifier(table), targets=[Star()])
dso, _ = dn.data_store.create_datasource(
dn.integration_name, {'query': ds_query.to_string()})
columns = dso.get_columns()
cols = []
for col in columns:
if not isinstance(col, dict):
col = {'name': col, 'type': 'str'}
cols.append(col)
table_alias = (self.database, table, table)
data = {
'values': [],
'columns': {
table_alias: cols
},
'tables': [table_alias]
}
elif type(step) == FetchDataframeStep:
data = self._fetch_dataframe_step(step)
elif type(step) == UnionStep:
raise ErNotSupportedYet('Union step is not implemented')
# TODO add union support
# left_data = steps_data[step.left.step_num]
# right_data = steps_data[step.right.step_num]
# data = left_data + right_data
elif type(step) == MapReduceStep:
try:
if step.reduce != 'union':
raise Exception(
f'Unknown MapReduceStep type: {step.reduce}')
step_data = steps_data[step.values.step_num]
vars = []
step_data_values = step_data['values']
for row in step_data_values:
var_group = {}
vars.append(var_group)
for row_data in row.values():
for name, value in row_data.items():
if name[0] != '__mindsdb_row_id':
var_group[name[1] or name[0]] = value
data = {'values': [], 'columns': {}, 'tables': []}
substep = step.step
if type(substep) == FetchDataframeStep:
query = substep.query
for var_group in vars:
markQueryVar(query.where)
for name, value in var_group.items():
replaceQueryVar(query.where, value, name)
sub_data = self._fetch_dataframe_step(substep)
if len(data['columns']) == 0:
data['columns'] = sub_data['columns']
if len(data['tables']) == 0:
data['tables'] = sub_data['tables']
data['values'].extend(sub_data['values'])
unmarkQueryVar(query.where)
elif type(substep) == MultipleSteps:
data = self._multiple_steps_reduce(substep, vars)
else:
raise Exception(f'Unknown step type: {step.step}')
except Exception as e:
raise SqlApiException(f'error in map reduce step: {e}') from e
elif type(step) == MultipleSteps:
if step.reduce != 'union':
raise Exception(
f"Only MultipleSteps with type = 'union' is supported. Got '{step.type}'"
)
data = None
for substep in step.steps:
subdata = self.execute_step(substep, steps_data)
if data is None:
data = subdata
else:
data['values'].extend(subdata['values'])
elif type(step) == ApplyPredictorRowStep:
try:
predictor = '.'.join(step.predictor.parts)
dn = self.datahub.get(self.mindsdb_database_name)
where_data = step.row_dict
data = dn.select(
table=predictor,
columns=None,
where_data=where_data,
integration_name=self.session.integration,
integration_type=self.session.integration_type)
data = [{(key, key): value
for key, value in row.items()} for row in data]
table_name = get_preditor_alias(step, self.database)
values = [{table_name: x} for x in data]
columns = {table_name: []}
if len(data) > 0:
row = data[0]
columns[table_name] = list(row.keys())
# TODO else
data = {
'values': values,
'columns': columns,
'tables': [table_name]
}
except Exception as e:
raise SqlApiException(
f'error in apply predictor row step: {e}') from e
elif type(step) in (ApplyPredictorStep, ApplyTimeseriesPredictorStep):
try:
dn = self.datahub.get(self.mindsdb_database_name)
predictor = '.'.join(step.predictor.parts)
where_data = []
for row in steps_data[step.dataframe.step_num]['values']:
new_row = {}
for table_name in row:
keys_intersection = set(new_row) & set(row[table_name])
if len(keys_intersection) > 0:
raise Exception(
f'The predictor got two identical keys from different datasources: {keys_intersection}'
)
new_row.update(row[table_name])
where_data.append(new_row)
where_data = [{key[1]: value
for key, value in row.items()}
for row in where_data]
is_timeseries = self.planner.predictor_metadata[predictor][
'timeseries']
_mdb_make_predictions = None
if is_timeseries:
if 'LATEST' in self.query_str:
_mdb_make_predictions = False
else:
_mdb_make_predictions = True
for row in where_data:
if '__mdb_make_predictions' not in row:
row['__mdb_make_predictions'] = _mdb_make_predictions
for row in where_data:
for key in row:
if isinstance(row[key], datetime.date):
row[key] = str(row[key])
data = dn.select(
table=predictor,
columns=None,
where_data=where_data,
integration_name=self.session.integration,
integration_type=self.session.integration_type)
# if is_timeseries:
# if 'LATEST' not in self.raw:
# # remove additional records from predictor results:
# # first 'window_size' and last 'horizon' records
# # otherwise there are many unxpected rows in prediciton result:
# # ----------------------------------------------------------------------------------------
# # mysql> SELECT tb.time, tb.state, tb.pnew_case, tb.new_case from
# # MYSQL_LOCAL.test_data.covid AS
# # ta JOIN mindsdb.covid_hor3 AS tb
# # WHERE ta.state = "CA" AND ta.time BETWEEN "2020-10-19" AND "2020-10-20";
# # ----------------------------------------------------------------------------------------
# # +------------+-------+-----------+----------+
# # | time | state | pnew_case | new_case |
# # +------------+-------+-----------+----------+
# # | 2020-10-09 | CA | 0 | 2862 |
# # | 2020-10-10 | CA | 0 | 2979 |
# # | 2020-10-11 | CA | 0 | 3075 |
# # | 2020-10-12 | CA | 0 | 3329 |
# # | 2020-10-13 | CA | 0 | 2666 |
# # | 2020-10-14 | CA | 0 | 2378 |
# # | 2020-10-15 | CA | 0 | 3449 |
# # | 2020-10-16 | CA | 0 | 3803 |
# # | 2020-10-17 | CA | 0 | 4170 |
# # | 2020-10-18 | CA | 0 | 3806 |
# # | 2020-10-19 | CA | 0 | 3286 |
# # | 2020-10-20 | CA | 0 | 3474 |
# # | 2020-10-21 | CA | 0 | 3474 |
# # | 2020-10-22 | CA | 0 | 3474 |
# # +------------+-------+-----------+----------+
# # 14 rows in set (2.52 sec)
# window_size = predictor_metadata[predictor]['window']
# horizon = predictor_metadata[predictor]['horizon']
# if len(data) >= (window_size + horizon):
# data = data[window_size:]
# if len(data) > horizon and horizon > 1:
# data = data[:-horizon + 1]
data = [{(key, key): value
for key, value in row.items()} for row in data]
table_name = get_preditor_alias(step, self.database)
values = [{table_name: x} for x in data]
columns = {table_name: []}
if len(data) > 0:
row = data[0]
columns[table_name] = list(row.keys())
# TODO else
data = {
'values': values,
'columns': columns,
'tables': [table_name]
}
except Exception as e:
raise SqlApiException(
f'error in apply predictor step: {e}') from e
elif type(step) == JoinStep:
try:
left_data = steps_data[step.left.step_num]
right_data = steps_data[step.right.step_num]
# FIXME https://github.com/mindsdb/mindsdb_sql/issues/136
# is_timeseries = False
# if True in [type(step) == ApplyTimeseriesPredictorStep for step in plan.steps]:
# right_data = steps_data[step.left.step_num]
# left_data = steps_data[step.right.step_num]
# is_timeseries = True
if step.query.condition is not None:
raise Exception(
'At this moment supported only JOIN without condition')
if step.query.join_type.upper() not in ('LEFT JOIN', 'JOIN'):
raise Exception(
'At this moment supported only JOIN and LEFT JOIN')
if (len(left_data['tables']) != 1
or len(right_data['tables']) != 1
or left_data['tables'][0] == right_data['tables'][0]):
raise Exception(
'At this moment supported only JOIN of two different tables'
)
data = {
'values': [],
'columns': {},
'tables':
list(set(left_data['tables'] + right_data['tables']))
}
for data_part in [left_data, right_data]:
for table_name in data_part['columns']:
if table_name not in data['columns']:
data['columns'][table_name] = data_part['columns'][
table_name]
else:
data['columns'][table_name].extend(
data_part['columns'][table_name])
for table_name in data['columns']:
data['columns'][table_name] = list(
set(data['columns'][table_name]))
left_key = left_data['tables'][0]
right_key = right_data['tables'][0]
left_columns_map = {}
left_columns_map_reverse = {}
for i, column_name in enumerate(
left_data['columns'][left_key]):
left_columns_map[f'a{i}'] = column_name
left_columns_map_reverse[column_name] = f'a{i}'
right_columns_map = {}
right_columns_map_reverse = {}
for i, column_name in enumerate(
right_data['columns'][right_key]):
right_columns_map[f'b{i}'] = column_name
right_columns_map_reverse[column_name] = f'b{i}'
left_df_data = []
for row in left_data['values']:
row = row[left_key]
left_df_data.append({
left_columns_map_reverse[key]: value
for key, value in row.items()
})
right_df_data = []
for row in right_data['values']:
row = row[right_key]
right_df_data.append({
right_columns_map_reverse[key]: value
for key, value in row.items()
})
df_a = pd.DataFrame(left_df_data)
df_b = pd.DataFrame(right_df_data)
a_name = f'a{round(time.time() * 1000)}'
b_name = f'b{round(time.time() * 1000)}'
con = duckdb.connect(database=':memory:')
con.register(a_name, df_a)
con.register(b_name, df_b)
resp_df = con.execute(f"""
SELECT * FROM {a_name} as ta full join {b_name} as tb
ON ta.{left_columns_map_reverse[('__mindsdb_row_id', '__mindsdb_row_id')]}
= tb.{right_columns_map_reverse[('__mindsdb_row_id', '__mindsdb_row_id')]}
""").fetchdf()
con.unregister(a_name)
con.unregister(b_name)
con.close()
resp_df = resp_df.where(pd.notnull(resp_df), None)
resp_dict = resp_df.to_dict(orient='records')
for row in resp_dict:
new_row = {left_key: {}, right_key: {}}
for key, value in row.items():
if key.startswith('a'):
new_row[left_key][left_columns_map[key]] = value
else:
new_row[right_key][right_columns_map[key]] = value
data['values'].append(new_row)
# remove all records with empty data from predictor from join result
# otherwise there are emtpy records in the final result:
# +------------+------------+-------+-----------+----------+
# | time | time | state | pnew_case | new_case |
# +------------+------------+-------+-----------+----------+
# | 2020-10-21 | 2020-10-24 | CA | 0.0 | 5945.0 |
# | 2020-10-22 | 2020-10-23 | CA | 0.0 | 6141.0 |
# | 2020-10-23 | 2020-10-22 | CA | 0.0 | 2940.0 |
# | 2020-10-24 | 2020-10-21 | CA | 0.0 | 3707.0 |
# | NULL | 2020-10-20 | NULL | nan | nan |
# | NULL | 2020-10-19 | NULL | nan | nan |
# | NULL | 2020-10-18 | NULL | nan | nan |
# | NULL | 2020-10-17 | NULL | nan | nan |
# | NULL | 2020-10-16 | NULL | nan | nan |
# +------------+------------+-------+-----------+----------+
# 9 rows in set (2.07 sec)
# if is_timeseries:
# data_values = []
# for row in data['values']:
# for key in row:
# if 'mindsdb' in key:
# if not is_empty_prediction_row(row[key]):
# data_values.append(row)
# break
# data['values'] = data_values
except Exception as e:
raise SqlApiException(f'error in join step: {e}') from e
elif type(step) == FilterStep:
raise ErNotSupportedYet('FilterStep is not implemented')
# elif type(step) == ApplyTimeseriesPredictorStep:
# raise Exception('ApplyTimeseriesPredictorStep is not implemented')
elif type(step) == LimitOffsetStep:
try:
step_data = steps_data[step.dataframe.step_num]
data = {
'values': step_data['values'].copy(),
'columns': step_data['columns'].copy(),
'tables': step_data['tables'].copy()
}
if isinstance(step.offset, Constant) and isinstance(
step.offset.value, int):
data['values'] = data['values'][step.offset.value:]
if isinstance(step.limit, Constant) and isinstance(
step.limit.value, int):
data['values'] = data['values'][:step.limit.value]
except Exception as e:
raise SqlApiException(
f'error in limit offset step: {e}') from e
elif type(step) == ProjectStep:
try:
step_data = steps_data[step.dataframe.step_num]
columns_list = []
for column_identifier in step.columns:
table_name = None
if type(column_identifier) == Star:
for table_name, table_columns_list in step_data[
'columns'].items():
for column in table_columns_list:
columns_list.append(
Column(database=table_name[0],
table_name=table_name[1],
table_alias=table_name[2],
name=column[0],
alias=column[1]))
elif type(column_identifier) == Identifier:
column_name_parts = column_identifier.parts
column_alias = None if column_identifier.alias is None else '.'.join(
column_identifier.alias.parts)
if len(column_name_parts) > 2:
raise Exception(
f'Column name must contain no more than 2 parts. Got name: {column_identifier}'
)
elif len(column_name_parts) == 1:
column_name = column_name_parts[0]
appropriate_table = None
if len(step_data['tables']) == 1:
appropriate_table = step_data['tables'][0]
else:
for table_name, table_columns in step_data[
'columns'].items():
table_column_names_list = [
x[1] or x[0] for x in table_columns
]
column_exists = get_column_in_case(
table_column_names_list, column_name)
if column_exists:
if appropriate_table is not None:
raise Exception(
'Found multiple appropriate tables for column {column_name}'
)
else:
appropriate_table = table_name
if appropriate_table is None:
# it is probably constaint
# FIXME https://github.com/mindsdb/mindsdb_sql/issues/133
# column_name = column_name.strip("'")
# name_or_alias = column_alias or column_name
# column_alias = name_or_alias
# for row in step_data['values']:
# for table in row:
# row[table][(column_name, name_or_alias)] = row[table][(column_name, column_name)]
# appropriate_table = step_data['tables'][0]
# FIXME: must be exception
columns_list.append(
Column(database=appropriate_table[0],
table_name=appropriate_table[1],
table_alias=appropriate_table[2],
name=column_alias))
else:
columns_list.append(
Column(database=appropriate_table[0],
table_name=appropriate_table[1],
table_alias=appropriate_table[2],
name=column_name,
alias=column_alias)) # column_name
elif len(column_name_parts) == 2:
table_name_or_alias = column_name_parts[0]
column_name = column_name_parts[1]
appropriate_table = None
for table_name, table_columns in step_data[
'columns'].items():
table_column_names_list = [
x[1] or x[0] for x in table_columns
]
checkig_table_name_or_alias = table_name[
2] or table_name[1]
if table_name_or_alias.lower(
) == checkig_table_name_or_alias.lower():
column_exists = get_column_in_case(
table_column_names_list, column_name)
if column_exists:
appropriate_table = table_name
break
else:
raise Exception(
f'Can not find column "{column_name}" in table "{table_name}"'
)
if appropriate_table is None:
raise Exception(
f'Can not find approproate table for column {column_name}'
)
columns_to_copy = None
table_column_names_list = [
x[1] or x[0] for x in table_columns
]
checking_name = get_column_in_case(
table_column_names_list, column_name)
for column in step_data['columns'][
appropriate_table]:
if column[0] == checking_name and (
column[1] is None
or column[1] == checking_name):
columns_to_copy = column
break
else:
raise Exception(
f'Can not find approproate column in data: {(column_name, column_alias)}'
)
for row in step_data['values']:
row[appropriate_table][(
column_name, column_alias
)] = row[appropriate_table][columns_to_copy]
columns_list.append(
Column(database=appropriate_table[0],
table_name=appropriate_table[1],
table_alias=appropriate_table[2],
name=column_name,
alias=column_alias))
else:
raise Exception('Undefined column name')
else:
raise Exception(
f'Unexpected column name type: {column_identifier}'
)
self.columns_list = columns_list
data = step_data
except Exception as e:
raise SqlApiException(f'error on project step:{e} ') from e
else:
raise SqlApiException(F'Unknown planner step: {step}')
return data