class IntegrationDataNode(DataNode):
type = 'integration'
def __init__(self, config, integration_name):
self.config = config
self.integration_name = integration_name
self.default_store = DataStore()
def getType(self):
return self.type
def getTables(self):
return []
def hasTable(self, tableName):
return True
def getTableColumns(self, tableName):
return []
def select(self,
table=None,
columns=None,
where=None,
where_data=None,
order_by=None,
group_by=None,
came_from=None):
if isinstance(where, dict):
where = [where]
if isinstance(where, list):
for el in where:
if isinstance(el, dict):
for key in el:
if isinstance(
el[key],
list) and len(el[key]) > 0 and isinstance(
el[key][0], str) and '.' in el[key][0]:
el[key][0] = el[key][0][el[key][0].find('.') + 1:]
where = {'and': where}
query = format({"from": table, 'select': columns, "where": where})
ds, ds_name = self.default_store.save_datasource(
f'temp_ds_{int(time.time()*100)}', self.integration_name,
{'query': query})
dso = self.default_store.get_datasource_obj(ds_name)
data = dso.df.T.to_dict().values()
self.default_store.delete_datasource(ds_name)
return data
class MindsDBDataNode(DataNode):
type = 'mindsdb'
def __init__(self, config):
self.config = Config()
self.mindsdb_native = NativeInterface()
self.custom_models = CustomModels()
self.ai_table = AITable_store()
self.default_store = DataStore()
def getTables(self):
models = self.mindsdb_native.get_models()
models = [x['name'] for x in models if x['status'] == 'complete']
models += ['predictors', 'commands']
models += [x['name'] for x in self.custom_models.get_models()]
ai_tables = self.ai_table.get_ai_tables()
models += [x['name'] for x in ai_tables]
return models
def hasTable(self, table):
return table in self.getTables()
def _get_ai_table_columns(self, table_name):
aitable_record = self.ai_table.get_ai_table(table_name)
columns = ([x['name'] for x in aitable_record.query_fields] +
[x['name'] for x in aitable_record.predictor_columns])
return columns
def _get_model_columns(self, table_name):
model = self.mindsdb_native.get_model_data(name=table_name)
columns = []
columns += model['columns']
columns += [f'{x}_original' for x in model['predict']]
for col in model['predict']:
if model['data_analysis_v2'][col]['typing'][
'data_type'] == 'Numeric':
columns += [f"{col}_min", f"{col}_max"]
columns += [f"{col}_confidence"]
columns += [f"{col}_explain"]
return columns
def getTableColumns(self, table):
try:
columns = self.custom_models.get_model_data(table)['columns']
columns += [
'external_datasource', 'select_data_query', 'when_data'
]
return columns
except Exception:
pass
if table == 'predictors':
return [
'name', 'status', 'accuracy', 'predict', 'select_data_query',
'external_datasource', 'training_options'
]
if table == 'commands':
return ['command']
columns = []
ai_tables = self.ai_table.get_ai_table(table)
if ai_tables is not None:
columns = self._get_ai_table_columns(table)
elif table in [x['name'] for x in self.mindsdb_native.get_models()]:
columns = self._get_model_columns(table)
columns += [
'when_data', 'select_data_query', 'external_datasource'
]
return columns
def _select_predictors(self):
models = self.mindsdb_native.get_models()
# TODO add custom models
return [
{
'name': x['name'],
'status': x['status'],
'accuracy':
str(x['accuracy']) if x['accuracy'] is not None else None,
'predict': ', '.join(x['predict']),
'select_data_query': '',
'external_datasource': '', # TODO
'training_options': '' # TODO ?
} for x in models
]
def delete_predictor(self, name):
self.mindsdb_native.delete_model(name)
def _select_from_ai_table(self, table, columns, where):
aitable_record = self.ai_table.get_ai_table(table)
integration = aitable_record.integration_name
query = aitable_record.integration_query
predictor_name = aitable_record.predictor_name
ds, ds_name = self.default_store.save_datasource(
'temp_ds', integration, {'query': query})
dso = self.default_store.get_datasource_obj(ds_name)
res = self.mindsdb_native.predict(name=predictor_name, when_data=dso)
self.default_store.delete_datasource(ds_name)
keys_map = {}
for f in aitable_record.predictor_columns:
keys_map[f['value']] = f['name']
for f in aitable_record.query_fields:
keys_map[f['name']] = f['name']
keys = list(keys_map.keys())
data = []
for i, el in enumerate(res):
data.append({keys_map[key]: el[key] for key in keys})
return data
def select(self,
table,
columns=None,
where=None,
where_data=None,
order_by=None,
group_by=None,
came_from=None):
''' NOTE WHERE statements can be just $eq joined with 'and'
'''
if table == 'predictors':
return self._select_predictors()
if table == 'commands':
return []
if self.ai_table.get_ai_table(table):
return self._select_from_ai_table(table, columns, where)
original_when_data = None
if 'when_data' in where:
if len(where) > 1:
raise ValueError(
"Should not be used any other keys in 'where', if 'when_data' used"
)
try:
original_when_data = where['when_data']['$eq']
where_data = json.loads(where['when_data']['$eq'])
if isinstance(where_data, list) is False:
where_data = [where_data]
except Exception:
raise ValueError(
f'''Error while parse 'when_data'="{where_data}"''')
external_datasource = None
if 'external_datasource' in where:
external_datasource = where['external_datasource']['$eq']
del where['external_datasource']
select_data_query = None
if came_from is not None and 'select_data_query' in where:
select_data_query = where['select_data_query']['$eq']
del where['select_data_query']
dbtype = self.config['integrations'][came_from]['type']
if dbtype == 'clickhouse':
ch = Clickhouse(self.config, came_from)
res = ch._query(
select_data_query.strip(' ;\n') + ' FORMAT JSON')
data = res.json()['data']
elif dbtype == 'mariadb':
maria = Mariadb(self.config, came_from)
data = maria._query(select_data_query)
elif dbtype == 'mysql':
mysql = MySQL(self.config, came_from)
data = mysql._query(select_data_query)
elif dbtype == 'postgres':
mysql = PostgreSQL(self.config, came_from)
data = mysql._query(select_data_query)
elif dbtype == 'mssql':
mssql = MSSQL(self.config, came_from)
data = mssql._query(select_data_query, fetch=True)
else:
raise Exception(f'Unknown database type: {dbtype}')
if where_data is None:
where_data = data
else:
where_data += data
new_where = {}
if where_data is not None:
where_data = pandas.DataFrame(where_data)
else:
for key, value in where.items():
if isinstance(value, dict) is False or len(
value.keys()) != 1 or list(value.keys())[0] != '$eq':
# TODO value should be just string or number
raise Exception()
new_where[key] = value['$eq']
if len(new_where) == 0:
return []
where_data = [new_where]
try:
model = self.custom_models.get_model_data(name=table)
except Exception:
model = self.mindsdb_native.get_model_data(name=table)
predicted_columns = model['predict']
original_target_values = {}
for col in predicted_columns:
if where_data is not None:
if col in where_data:
original_target_values[col + '_original'] = list(
where_data[col])
else:
original_target_values[col +
'_original'] = [None
] * len(where_data)
else:
original_target_values[col + '_original'] = [None]
if table in [x['name'] for x in self.custom_models.get_models()]:
res = self.custom_models.predict(name=table, when_data=where_data)
data = []
fields = model['columns']
for i, ele in enumerate(res):
row = {}
row['select_data_query'] = select_data_query
row['external_datasource'] = external_datasource
row['when_data'] = original_when_data
for key in ele:
row[key] = ele[key]['predicted_value']
# FIXME prefer get int from mindsdb_native in this case
if model['data_analysis_v2'][key]['typing'][
'data_subtype'] == 'Int':
row[key] = int(row[key])
for k in fields:
if k not in ele:
if isinstance(where_data, list):
if k in where_data[i]:
row[k] = where_data[i][k]
else:
row[k] = None
elif k in where_data.columns:
row[k] = where_data[k].iloc[i]
else:
row[k] = None
for k in original_target_values:
row[k] = original_target_values[k][i]
data.append(row)
field_types = {
f: model['data_analysis_v2'][f]['typing']['data_subtype']
for f in fields if 'typing' in model['data_analysis_v2'][f]
}
for row in data:
cast_row_types(row, field_types)
return data
else:
res = self.mindsdb_native.predict(name=table, when_data=where_data)
keys = [x for x in list(res._data.keys()) if x in columns]
min_max_keys = []
for col in predicted_columns:
if model['data_analysis_v2'][col]['typing'][
'data_type'] == 'Numeric':
min_max_keys.append(col)
data = []
explains = []
for i, el in enumerate(res):
data.append({key: el[key] for key in keys})
explains.append(el.explain())
field_types = {
f: model['data_analysis_v2'][f]['typing']['data_subtype']
for f in model['columns']
if 'typing' in model['data_analysis_v2'][f]
}
for i, row in enumerate(data):
cast_row_types(row, field_types)
row['select_data_query'] = select_data_query
row['external_datasource'] = external_datasource
row['when_data'] = original_when_data
for k in original_target_values:
row[k] = original_target_values[k][i]
explanation = explains[i]
for key in predicted_columns:
row[key + '_confidence'] = explanation[key]['confidence']
row[key + '_explain'] = json.dumps(explanation[key],
cls=NumpyJSONEncoder,
ensure_ascii=False)
for key in min_max_keys:
row[key + '_min'] = min(
explanation[key]['confidence_interval'])
row[key + '_max'] = max(
explanation[key]['confidence_interval'])
return data
class IntegrationDataNode(DataNode):
type = 'integration'
def __init__(self, config, integration_name):
self.config = config
self.integration_name = integration_name
self.default_store = DataStore()
def getType(self):
return self.type
def getTables(self):
return []
def hasTable(self, tableName):
return True
def getTableColumns(self, tableName):
return []
def select(self,
table=None,
columns=None,
where=None,
where_data=None,
order_by=None,
group_by=None,
came_from=None):
has_where = isinstance(where, (dict, list)) and len(where) > 0
if isinstance(where, dict):
where = [where]
if isinstance(where, list):
for el in where:
if isinstance(el, dict):
for key in el:
if isinstance(
el[key],
list) and len(el[key]) > 0 and isinstance(
el[key][0], str) and '.' in el[key][0]:
el[key][0] = el[key][0][el[key][0].find('.') + 1:]
where = {'and': where}
format_data = {'from': table, 'select': columns}
if has_where:
format_data['where'] = where
query = format(format_data)
ds, ds_name = self.default_store.save_datasource(
f'temp_ds_{int(time.time()*100)}', self.integration_name,
{'query': query})
dso = self.default_store.get_datasource_obj(ds_name)
data = dso.df.to_dict(orient='records')
for column_name in dso.df.columns:
if pd.core.dtypes.common.is_datetime_or_timedelta_dtype(
dso.df[column_name]):
pass_data = dso.df[column_name].dt.to_pydatetime()
for i, rec in enumerate(data):
rec[column_name] = pass_data[i].timestamp()
self.default_store.delete_datasource(ds_name)
return data
class StreamController:
def __init__(self,
name,
predictor,
stream_in,
stream_out,
anomaly_stream=None,
learning_stream=None,
learning_threshold=100):
self.name = name
self.predictor = predictor
self.stream_in = stream_in
self.stream_out = stream_out
self.anomaly_stream = anomaly_stream
self.learning_stream = learning_stream
self.learning_threshold = learning_threshold
self.learning_data = []
self.company_id = os.environ.get('MINDSDB_COMPANY_ID', None)
self.stop_event = Event()
self.model_interface = ModelInterfaceWrapper(ModelInterface())
self.data_store = DataStore()
self.config = Config()
p = db.session.query(db.Predictor).filter_by(
company_id=self.company_id, name=self.predictor).first()
if p is None:
raise Exception(f'Predictor {predictor} doesn\'t exist')
self.target = p.to_predict[0]
ts_settings = p.learn_args.get('timeseries_settings', None)
if not ts_settings['is_timeseries']:
ts_settings = None
if ts_settings is None:
self.thread = Thread(target=StreamController._make_predictions,
args=(self, ))
else:
self.ts_settings = ts_settings
self.thread = Thread(target=StreamController._make_ts_predictions,
args=(self, ))
self.thread.start()
def _is_anomaly(self, res):
for k in res:
if k.endswith('_anomaly') and res[k] is not None:
return True
return False
def _consider_learning(self):
if self.learning_stream is not None:
self.learning_data.extend(self.learning_stream.read())
if len(self.learning_data) >= self.learning_threshold:
p = db.session.query(db.Predictor).filter_by(
company_id=self.company_id, name=self.predictor).first()
ds_record = db.session.query(
db.Datasource).filter_by(id=p.datasource_id).first()
df = pd.DataFrame.from_records(self.learning_data)
name = 'name_' + str(time()).replace('.', '_')
path = os.path.join(self.config['paths']['datasources'], name)
df.to_csv(path)
from_data = {
'class': 'FileDS',
'args': [path],
'kwargs': {},
}
self.data_store.save_datasource(name=name,
source_type='file',
source=path,
file_path=path,
company_id=self.company_id)
ds = self.data_store.get_datasource(name, self.company_id)
self.model_interface.adjust(p.name, from_data, ds['id'],
self.company_id)
self.learning_data.clear()
def _make_predictions(self):
while not self.stop_event.wait(0.5):
self._consider_learning()
for when_data in self.stream_in.read():
preds = self.model_interface.predict(self.predictor, when_data,
'dict')
for res in preds:
if self.anomaly_stream is not None and self._is_anomaly(
res):
self.anomaly_stream.write(res)
else:
self.stream_out.write(res)
def _make_ts_predictions(self):
window = self.ts_settings['window']
order_by = self.ts_settings['order_by']
order_by = [order_by] if isinstance(order_by, str) else order_by
group_by = self.ts_settings.get('group_by', None)
group_by = [group_by] if isinstance(group_by, str) else group_by
cache = Cache(self.name)
while not self.stop_event.wait(0.5):
self._consider_learning()
for when_data in self.stream_in.read():
for ob in order_by:
if ob not in when_data:
raise Exception(
f'when_data doesn\'t contain order_by[{ob}]')
for gb in group_by:
if gb not in when_data:
raise Exception(
f'when_data doesn\'t contain group_by[{gb}]')
gb_value = tuple(
when_data[gb]
for gb in group_by) if group_by is not None else ''
# because cache doesn't work for tuples
# (raises Exception: tuple doesn't have "encode" attribute)
gb_value = str(gb_value)
with cache:
if gb_value not in cache:
cache[gb_value] = []
# do this because shelve-cache doesn't support
# in-place changing
records = cache[gb_value]
records.append(when_data)
cache[gb_value] = records
with cache:
for gb_value in cache.keys():
if len(cache[gb_value]) >= window:
cache[gb_value] = [
*sorted(
cache[gb_value],
# WARNING: assuming wd[ob] is numeric
key=lambda wd: tuple(wd[ob]
for ob in order_by))
]
while len(cache[gb_value]) >= window:
res_list = self.model_interface.predict(
self.predictor, cache[gb_value][:window],
'dict')
if self.anomaly_stream is not None and self._is_anomaly(
res_list[-1]):
self.anomaly_stream.write(res_list[-1])
else:
self.stream_out.write(res_list[-1])
cache[gb_value] = cache[gb_value][1:]
