import importlib import gzip import pickle from pprint import pprint from tqdm import tqdm from tqdm.contrib.concurrent import process_map import numpy as np import pandas as pd import dolphindb as ddb import dolphindb.settings as keys import sqlalchemy as sa import ProtoBuffEntitys def make_symbol(series): return series.astype('int32').astype('str')\ .apply(str.zfill, args=(6,))\ .apply(lambda code : \ code + '.SH' if code[:2] == '60' \ else code + '.SZ') def make_date(series): return pd.to_datetime( series.astype(str), format='%Y%m%d') def make_nparray(series): return series.apply(lambda x : np.array(x)) def make_time(series): s_hr = series // 10000000 * 3600000 s_min = series % 1000000 // 100000 * 60000 s_sec = series % 100000 // 1000 s_ms = series % 1000 return pd.to_timedelta(s_hr + s_min + s_sec + s_ms, unit='ms') class DDBLoader(object): """ 0. 从sql-server中读取calendar数据,并创建成员变量df_calendar,df_calendar可以保存在本地pickle作为缓存 |- `def make_calendar_df(self) -> df_calendar` 1. 创建ddb中的数据库,分区性质从calendar数据中获取 |- `def create_ddb_database(self, df_calendar) -> void` |- `def load_ddb_database(self) -> void` 2. 在ddb数据库中创建calendar表 |- `def create_ddb_calendar(self, df_calendar) -> void` 3. 创建ddb的分布式表结构 |- `create_ddb_partition_table(self, hft_type_name)` |- `_make_table_skeleton(self, hft_type_name, capacity) -> memory_table_name` 4. 从sql server的高频数据转录到dolpindb数据库中 |- `dump_hft_to_ddb(self, type_name, stock_id, trade_date=None)` """ hft_type_list = ['KLine', 'Order', 'Tick', 'TickQueue', 'Transe'] protobuff_name_dict = { name : f"{name}Message_pb2" for name in hft_type_list } protobuff_module_dict = { type_name : importlib.import_module(f".{module_name}", package='ProtoBuffEntitys') for type_name, module_name in protobuff_name_dict.items() } protobuff_desc_dict = { type_name : eval(f"ProtoBuffEntitys.{module_name}.{type_name}Array.{type_name}Data.DESCRIPTOR") for type_name, module_name in protobuff_name_dict.items() } mssql_name_dict = { type_name : ( f"{type_name}" if type_name != 'TickQueue' \ else f"TickQue" ) for type_name in hft_type_list } # 数据库路径和数据库名可以不一致 ddb_path = "dfs://hft_ts_stock" ddb_dbname = "db_ts_stock" ddb_memory_table_suffix = "Memroy" ddb_partition_table_suffix = "Partitioned" # calendar表不需要分区,因此需要创建一个新的数据库 # 该数据库可以是一个简单的csv,现在还不清楚两者的差别 #ddb_calendar_path = "dfs://daily_calendar" #ddb_calendar_dbname = "db_calendar" ddb_calendar_table_name = "Calendar" col_type_mapping = { 'code' : 'SYMBOL', 'm_nDate' : 'DATE', 'm_nTime' : 'TIME', 1 : 'FLOAT', 3 : 'INT', 5 : 'INT', 13 : 'INT', } mssql_config = { 'host' : '192.168.1.7', 'username' : 'sa', 'password' : 'passw0rd!' } ddb_config = { 'host' : '192.168.1.167', 'username' : 'admin', 'password' : '123456' } default_table_capacity = 10000 def __init__(self): self.mssql_engine = sa.create_engine( "mssql+pyodbc://{username}:{password}@{host}/master?driver=ODBC+Driver+18+for+SQL+Server".format(**self.mssql_config), connect_args = { "TrustServerCertificate": "yes" }, echo=False ) self.ddb_sess = ddb.session(self.ddb_config['host'], 8848) self.ddb_sess.login(self.ddb_config['username'], self.ddb_config['password']) def init_ddb_database(self, df_calendar): """ 1. 创建ddb_database 2. 创建calendar表 3. 创建数据分区表 """ # df_calendar还是由外部输入比较方便 #df_calendar = self.make_calendar_df() self.create_ddb_database(df_calendar) self.create_ddb_calendar(df_calendar) for hft_type_name in self.hft_type_list: self.create_ddb_partition_table(hft_type_name) def init_ddb_table_data(self, df_calendar): """ 对每个股票进行循环,转录数据到分区表 """ stock_list = df_calendar['code'].unique().astype('str') for hft_type_name in self.hft_type_list: print('Will work on hft type:', hft_type_name) with tqdm(stock_list) as pbar: for stock_id in pbar: pbar.set_description(f"Working on stock {stock_id}") self.dump_hft_to_ddb(hft_type_name, stock_id, pbar=pbar) def _get_stock_date_list(self, cache=False): """ Deprecated: This function is replaced by `create_ddb_calendar()`. """ if cache: with open('tmp.pkl', 'rb') as fin: stock_list, date_list = pickle.load(fin) else: with self.mssql_engine.connect() as conn: # 从KLine表查询,主要是因为KLine表最小 stat = "select distinct S_INFO_WINDCODE, TRADE_DT from Level2BytesKline.dbo.KLine" rs = conn.execute(stat) stock_date_list = [(stock_name, date) for stock_name, date in rs.fetchall()] stock_list, date_list = zip(*stock_date_list) # cache #with open('tmp.pkl', 'wb') as fout: # pickle.dump((stock_list, date_list), fout) return pd.Series(stock_list, dtype='str').unique(), \ pd.Series(date_list, dtype='datetime64[D]').unique() def create_ddb_database(self, pd_calendar): # 从`pd_calendar`中创建`stock_list`和`date_list` stock_list = pd_calendar['code'].unique().astype('str') date_list = pd_calendar['m_nDate'].unique().astype('datetime64[D]') # 可以把所有股票高频数据放在一个数据库中不同的表 # 分区策略是跟数据库绑定的,因此需要保证同一个数据库中的表都使用同样的分区额策略 # 对于股票高频数据,我们可以使用COMPO的分区策略,并且两个子db的分区策略都是VALUE类型的code和m_nDate字段 if self.ddb_sess.existsDatabase(self.ddb_path): print('Wiil drop database:', self.ddb_path) self.ddb_sess.dropDatabase(self.ddb_path) # 要创建一个COMPO分区的数据库,需要首先创建两个简单分区的子数据库 # 这里我们使用先按日期,然后按股票分区的子数据库 # Please note that when creating a DFS database with COMPO domain, # the parameter dbPath for each partition level must be either an empty string or unspecified. db_date = self.ddb_sess.database('db_date', partitionType=keys.VALUE, partitions=date_list, dbPath='') # 这里看起来直接使用dolphindb的脚本语句更方便一些 #db_stock = self.ddb_sess.database('db_stock', partitionType=keys.VALUE, partitions=stock_list, dbPath='') self.ddb_sess.run(""" db_stock = database("", 1, symbol({partitions})) """.format( partitions = '`' + '`'.join(stock_list) )) self.ddb_sess.run(""" {dbName} = database( directory = '{dbPath}', partitionType = COMPO, partitionScheme = [db_date, db_stock], engine = "TSDB") """.format( dbName = self.ddb_dbname, dbPath = self.ddb_path )) def load_ddb_database(self): db_date = self.ddb_sess.database('db_date', dbPath='') db_stock = self.ddb_sess.database('db_stock', dbPath='') self.ddb_sess.run("{dbName} = database(directory='{dbPath}')".format( dbName = self.ddb_dbname, dbPath = self.ddb_path )) def create_ddb_calendar(self, df_calendar): mem_table = self.ddb_calendar_table_name + self.ddb_memory_table_suffix per_table = self.ddb_calendar_table_name # 1. 创建临时内存表 # calendar的行数大概是股票数量 * 交易日数量 self.ddb_sess.run(""" {table_name} = table({capacity}:0, {col_names}, [{col_types}]); """.format( table_name = mem_table, capacity = 5000 * 1000, col_names = '`code`m_nDate', col_types = "SYMBOL, DATE" )) print('Did create the memory table') # 2. 向内存表中插入所有(code, date)数据 appender = ddb.tableAppender(tableName=mem_table, ddbSession=self.ddb_sess) num = appender.append(df_calendar) print('Did append calendar data into ddb memory table, return code', num) # 3. 创建持久化表格之前需要先根据路径创建一个database对象 # 但研究了一下,发现好像一个database里面可以同时存在分区表和非分区表, # 所以在这里暂时就不创建新的database了 # 但因为原database设置成了TSDB,所以必须在createTable的时候指定sortKey #self.ddb_sess.run(""" # {db_name} = #""") # 4. 直接从内存表创建一个持久化表格 if self.ddb_sess.existsTable(self.ddb_path, per_table): self.ddb_sess.dropTable(self.ddb_path, per_table) self.ddb_sess.run(""" tableInsert(createTable( dbHandle={ddb_dbname}, table={mem_table}, tableName=`{per_table}, sortCOlumns=`code`m_nDate, compressMethods={{"m_nDate":"delta"}} ), {mem_table}) """.format( ddb_dbname = self.ddb_dbname, mem_table = mem_table, per_table = per_table )) print('Did create the persistent table with the memory table') def make_calendar_df(self): # 从KLine表查询,主要是因为KLine表最小 with self.mssql_engine.connect() as conn: stat = "select distinct S_INFO_WINDCODE, TRADE_DT from Level2BytesKline.dbo.KLine" rs = conn.execute(stat) stock_date_list = [(stock_name, date) for stock_name, date in rs.fetchall()] df_calendar = pd.DataFrame(stock_date_list, columns=['code', 'm_nDate']) df_calendar['m_nDate'] = make_date(df_calendar['m_nDate']) print('Did make the DataFrame for calendar') return df_calendar def _make_table_skeleton(self, hft_type_name, table_capacity=default_table_capacity): def _make_tbl_config(field_list): """ 根据ProtoBuffEntity对象的Descriptor.fields,创建ddb标准的列名列表和列类型列表。 """ col_name_list, col_type_list = [], [] for desc in field_list: col_name_list.append(desc.name) # 如果对列明有特殊设定,目前仅包括`code`m_nDate和`m_nTime三个字段 if desc.name in self.col_type_mapping: col_type_list.append(self.col_type_mapping[desc.name]) # 通过对ProtoBuffEntity的类型编号,映射到ddb的类型编号 # 如果默认值是一个数组,那么ddb类型要额外增加说明是数组 # ProtoBuffEntity的类型编号只针对基本类型,数组需要通过`default_value`来判断 else: col_type = self.col_type_mapping[desc.type] if isinstance(desc.default_value, list): col_type += '[]' col_type_list.append(col_type) return col_name_list, col_type_list desc_obj = self.protobuff_desc_dict[hft_type_name] col_name_list, col_type_list = _make_tbl_config(desc_obj.fields) table_name = hft_type_name + self.ddb_memory_table_suffix print('-' * 80) print('Will create table structure:', table_name) self.ddb_sess.run(""" {table_name} = table({capacity}:0, {col_names}, [{col_types}]); """.format( table_name = table_name, capacity = table_capacity, col_names = '`' + '`'.join(col_name_list), col_types = ', '.join([f"'{type_name}'" for type_name in col_type_list]) )) res = self.ddb_sess.run(f"schema({table_name}).colDefs") pprint(res) print('-' * 80) return table_name def create_ddb_partition_table(self, hft_type_name): memory_table_name = self._make_table_skeleton(hft_type_name, 10) partition_table_name = hft_type_name + self.ddb_partition_table_suffix print('-' * 80) print('Will create partitioned table:', partition_table_name) self.ddb_sess.run(""" db_ts_stock.createPartitionedTable( table = {memory_table_name}, tableName = `{partition_table_name}, partitionColumns = `m_nDate`code, sortColumns = `code`m_nDate`m_nTime, compressMethods = {{m_nDate:"delta", m_nTime:"delta"}} ) """.format( memory_table_name = memory_table_name, partition_table_name = partition_table_name )) res = self.ddb_sess.run(f"schema(loadTable('{self.ddb_path}', '{partition_table_name}')).colDefs") pprint(res) print('-' * 80) def _make_stock_daily_df(self, blob, type_name): blob = gzip.decompress(blob) dataArray = eval(f"ProtoBuffEntitys.{type_name}Message_pb2.{type_name}Array()") dataArray.ParseFromString(blob) data_dict_list = [ {field.name : val for field, val in entry.ListFields()} for entry in dataArray.dataArray ] array_type_list = [ field.name for field, val in dataArray.dataArray[0].ListFields() if isinstance(field.default_value, list) ] #pprint(array_type_list) df = pd.DataFrame(data_dict_list) df['code'] = make_symbol(df['code']) df['m_nDate'] = make_date(df['m_nDate']) df['m_nTime'] = df['m_nDate'] + make_time(df['m_nTime']) for field_name in array_type_list: df[field_name] = make_nparray(df[field_name]) #print(f"Did create ddb table for dataframe of shape {df.shape}") # self.make_table_skeleton(type_name, df.shape[0]) self.ddb_sess.upload({type_name : df}) return type_name def dump_hft_to_ddb(self, type_name, stock_id, trade_date=None, pbar=None, num_workers=4): def _dump_stock_daily(row): df_table_name = self._make_stock_daily_df(row[2], type_name) self.ddb_sess.run("tableInsert(loadTable('{dbPath}', `{partitioned_table_name}), {df_table_name})".format( dbPath = self.ddb_path, partitioned_table_name = type_name + self.ddb_partition_table_suffix, df_table_name = df_table_name )) # 经过尝试,按个股来做batch查询效率还是可以接受的 # mssql中,索引字段是(S_INFO_WINDCODE, TRADE_DT) with self.mssql_engine.connect() as conn: stat = """ select * from [Level2Bytes{mssql_type_name}].dbo.[{mssql_type_name}] where S_INFO_WINDCODE='{stock_id}' """.format( mssql_type_name = self.mssql_name_dict[type_name], stock_id = stock_id ) rs = conn.execute(stat) if pbar: pbar.set_description(f"Did get the result set for stock {stock_id} from mssql") else: print(f"Did get the result set for stock {stock_id} from mssql") # 每一行是当个个股某一日的所有高频交易信息 # 使用多进程来加快速度 #with tqdm(rs.fetchall(), leave=False) as pbar: # for row in pbar: # pbar.set_description(f"Working on {row[0]} {row[1]}") process_map(_dump_stock_daily, rs.fetchall(), max_workers=num_workers) def main(): loader = DDBLoader() df_calendar = loader.make_calendar_df() loader.init_ddb_database(df_calendar) print('Did finish init_ddb_database') loader.init_ddb_table_data(df_calendar) print('Did finish init_table_data') if __name__ == '__main__': main()