1.问题背景描述

python项目中的时序数据都存放在TD数据库中，数据是秒级存入的，当查询一周数据时将超过50w数据量，这是一次性获取全量数据到python程序很慢，全流程10秒以上，希望进行优化加速

2.排查

首先，分步排查从td取超过7秒，在python程序中处理格式超过3秒；
其次，业务逻辑处理步骤中，时间大量消耗的逻辑是 时间对象转成字符串；
再次，td取数步骤中，时间大量消耗的逻辑是 获取到数据后时间戳转为时间对象
最后，思路确定为，从td获取的ts字段直接按bigint返回，交由业务逻辑处理，直接从bigint转成字符串；

3.查源码尝试修改

TD取的ts字段是用bigint存的（C_TIMESTAMP类型），排查源码发现公共包中会将C_TIMESTAMP类型的字段都转成datetime对象返回，而转化方法_convert_millisecond_to_datetime就是慢的根源；继续查相关源码，若想用bigint返回，发现CONVERT_FUNC_BLOCK这个函数工厂，key是每个字段类型fields[i][“type”]，这个字段类型是在taos_fetch_fields

# TDHelper().db_query(sql) 调用入口
class TDHelper:  # 自定义的TD适配逻辑def db_query(self, sql, return_timestamp=False):with self.cursor() as c:return self._query_handler(c, sql, return_timestamp)def _query_handler(self, cursor, sql, return_timestamp=False):try:cursor.execute(sql)  # execute中会获取_fields属性，由决定后续字段序列化的逻辑# cursor._fields[0]._type = 5  # 测试用result = cursor.fetchall()if not return_timestamp:return resultelse:ret_result = []for one in result:ret_result.append((one[0].timestamp(), *one[1:]))return ret_resultexcept ProgrammingError as e:if e.msg == 'Fail to get table info, error: Table does not exist':# 只输出sql，不需要输出异常信息logger.warning('Table does not exist, sql [{}]'.format(sql))return []raise e# 如下都是TD公共包中的源码def fetchall(self):  # cursor的方法if self._result is None:raise OperationalError("Invalid use of fetchall")fields = self._fields if self._fields is not None else taos_fetch_fields(self._result)buffer = [[] for i in range(len(fields))]self._rowcount = 0while True:block, num_of_rows = taos_fetch_block(self._result, self._fields)   # 关键逻辑errno = taos_errno(self._result)if errno != 0:raise ProgrammingError(taos_errstr(self._result), errno)if num_of_rows == 0:breakself._rowcount += num_of_rowsfor i in range(len(self._fields)):buffer[i].extend(block[i])return list(map(tuple, zip(*buffer)))def taos_fetch_block(result, fields=None, field_count=None):if fields is None:fields = taos_fetch_fields(result)if field_count is None:field_count = taos_field_count(result)pblock = ctypes.c_void_p(0)num_of_rows = _libtaos.taos_fetch_block(result, ctypes.byref(pblock))if num_of_rows == 0:return None, 0precision = taos_result_precision(result)blocks = [None] * field_countfor i in range(len(fields)):data = ctypes.cast(pblock, ctypes.POINTER(ctypes.c_void_p))[i]if fields[i]["type"] not in CONVERT_FUNC_BLOCK_v3 and fields[i]["type"] not in CONVERT_FUNC_BLOCK:raise DatabaseError("Invalid data type returned from database")offsets = []is_null = []if fields[i]["type"] in (FieldType.C_VARCHAR, FieldType.C_NCHAR, FieldType.C_JSON):offsets = taos_get_column_data_offset(result, i, num_of_rows)blocks[i] = CONVERT_FUNC_BLOCK_v3[fields[i]["type"]](data, is_null, num_of_rows, offsets, precision)else:is_null = [taos_is_null(result, j, i) for j in range(num_of_rows)]# 关键逻辑blocks[i] = CONVERT_FUNC_BLOCK[fields[i]["type"]](data, is_null, num_of_rows, offsets, precision)return blocks, abs(num_of_rows)CONVERT_FUNC_BLOCK = {FieldType.C_BOOL: _crow_bool_to_python,FieldType.C_TINYINT: _crow_tinyint_to_python,FieldType.C_SMALLINT: _crow_smallint_to_python,FieldType.C_INT: _crow_int_to_python,FieldType.C_BIGINT: _crow_bigint_to_python,FieldType.C_FLOAT: _crow_float_to_python,FieldType.C_DOUBLE: _crow_double_to_python,FieldType.C_BINARY: _crow_binary_to_python_block,FieldType.C_TIMESTAMP: _crow_timestamp_to_python,   # 关键逻辑FieldType.C_NCHAR: _crow_nchar_to_python_block,FieldType.C_TINYINT_UNSIGNED: _crow_tinyint_unsigned_to_python,FieldType.C_SMALLINT_UNSIGNED: _crow_smallint_unsigned_to_python,FieldType.C_INT_UNSIGNED: _crow_int_unsigned_to_python,FieldType.C_BIGINT_UNSIGNED: _crow_bigint_unsigned_to_python,FieldType.C_JSON: _crow_nchar_to_python_block,
}def _crow_timestamp_to_python(data, is_null, num_of_rows, nbytes=None, precision=FieldType.C_TIMESTAMP_UNKNOWN):"""Function to convert C bool row to python row."""_timestamp_converter = _convert_millisecond_to_datetime   # 关键逻辑if precision == FieldType.C_TIMESTAMP_MILLI:_timestamp_converter = _convert_millisecond_to_datetimeelif precision == FieldType.C_TIMESTAMP_MICRO:_timestamp_converter = _convert_microsecond_to_datetimeelif precision == FieldType.C_TIMESTAMP_NANO:_timestamp_converter = _convert_nanosecond_to_datetimeelse:raise DatabaseError("Unknown precision returned from database")return [None if is_null[i] else _timestamp_converter(ele)for i, ele in enumerate(ctypes.cast(data, ctypes.POINTER(ctypes.c_int64))[: abs(num_of_rows)])]def _convert_millisecond_to_datetime(milli):try:if _priv_tz is None:return _datetime_epoch + timedelta(seconds=milli / 1000.0)return (_utc_datetime_epoch + timedelta(seconds=milli / 1000.0)).astimezone(_priv_tz)  # 万恶之源except OverflowError:# catch OverflowError and passprint("WARN: datetime overflow!")pass