Python - 使用HUGE数据集避免内存错误

时间:2017-01-27 09:44:40

标签: python postgresql out-of-memory psycopg2 bigdata

我有一个连接到PostGreSQL数据库的python程序。在这个数据库中,我有很多数据(大约12亿行)。幸运的是,我不必同时分析所有这些行。

这12亿行分布在几张桌子上(大约30张)。目前我正在访问一个名为table_3的表,我想在其中访问具有特定“did”值的所有行(因为该列被调用)。

我使用SQL命令对行进行了计数:

SELECT count(*) FROM table_3 WHERE did='356002062376054';

返回1.57亿行。

我将对所有这些行执行一些“分析”(提取2个特定值)并对这些值进行一些计算,然后将它们写入字典,然后将它们保存在另一个表中的PostGreSQL中。 / p>

问题是我正在创建大量的列表和字典来管理所有这些我最终耗尽内存,即使我使用的是Python 3 64位并且有64 GB的RAM。

一些代码:

CONNECTION = psycopg2.connect('<psycopg2 formatted string>')
CURSOR = CONNECTION.cursor()

DID_LIST = ["357139052424715",
            "353224061929963",
            "356002064810514",
            "356002064810183",
            "358188051768472",
            "358188050598029",
            "356002061925067",
            "358188056470108",
            "356002062376054",
            "357460064130045"]

SENSOR_LIST = [1, 2, 3, 4, 5, 6, 7, 8, 9,
               10, 11, 12, 13, 801, 900, 901,
               902, 903, 904, 905, 906, 907,
               908, 909, 910, 911]

for did in did_list:
    table_name = did
    for sensor_id in sensor_list:
        rows = get_data(did, sensor_id)
        list_object = create_standard_list(sensor_id, rows)  # Happens here
        formatted_list = format_table_dictionary(list_object) # Or here
        pushed_rows = write_to_table(table_name, formatted_list) #write_to_table method is omitted as that is not my problem.

def get_data(did, table_id):
    """Getting data from postgresql."""
    table_name = "table_{0}".format(table_id)
    query = """SELECT * FROM {0} WHERE did='{1}'
               ORDER BY timestamp""".format(table_name, did)

    CURSOR.execute(query)
    CONNECTION.commit()

    return CURSOR

def create_standard_list(sensor_id, data):
    """Formats DB data to dictionary"""
    list_object = []

    print("Create standard list")
    for row in data: # data is the psycopg2 CURSOR
        row_timestamp = row[2]
        row_data = row[3]

        temp_object = {"sensor_id": sensor_id, "timestamp": row_timestamp,
                       "data": row_data}

        list_object.append(temp_object)

    return list_object


def format_table_dictionary(list_dict):
    """Formats dictionary to simple data
       table_name = (dates, data_count, first row)"""
    print("Formatting dict to DB")
    temp_today = 0
    dict_list = []
    first_row = {}
    count = 1

    for elem in list_dict:
        # convert to seconds
        date = datetime.fromtimestamp(elem['timestamp'] / 1000)
        today = int(date.strftime('%d'))
        if temp_today is not today:
            if not first_row:
                first_row = elem['data']
            first_row_str = str(first_row)
            dict_object = {"sensor_id": elem['sensor_id'],
                           "date": date.strftime('%d/%m-%Y'),
                           "reading_count": count,
                           # size in MB of data
                           "approx_data_size": (count*len(first_row_str)/1000),
                           "time": date.strftime('%H:%M:%S'),
                           "first_row": first_row}

            dict_list.append(dict_object)
            first_row = {}
            temp_today = today
            count = 0
        else:
            count += 1

    return dict_list

我的错误发生在创建两个列表中的任何一个,在我的代码中用注释标记。它代表我的电脑停止响应,并最终让我退出。我正在运行Windows 10,如果这有点重要的话。

我知道我使用“create_standard_list”方法创建的第一个列表可以被排除,并且该代码可以在“format_table_dictionary”代码中运行,从而避免在内存中包含157 mio元素的列表,但我认为有些我将遇到的其他表会有类似的问题,可能会更大,所以我想现在就优化它,但我不确定我能做什么?

我想写一个文件并不会真正有用,因为我必须读取该文件,然后将其重新放回内存中?

极简主义示例

我有一张桌子

---------------------------------------------------------------
|Row 1 | did | timestamp | data | unused value | unused value |
|Row 2 | did | timestamp | data | unused value | unused value |
....
---------------------------------

table = [{ values from above row1 }, { values from above row2},...]

connection = psycopg2.connect(<connection string>)
cursor = connection.cursor()

table = cursor.execute("""SELECT * FROM table_3 WHERE did='356002062376054'
                          ORDER BY timestamp""")

extracted_list = extract(table)
calculated_list = calculate(extracted_list)
... write to db ...

def extract(table):
    """extract all but unused values"""
    new_list = []
    for row in table:
        did = row[0]
        timestamp = row[1]
        data = row[2]

        a_dict = {'did': did, 'timestamp': timestamp, 'data': data}
        new_list.append(a_dict)

    return new_list


def calculate(a_list):
    """perform calculations on values"""
    dict_list = []
    temp_today = 0
    count = 0
    for row in a_list:
        date = datetime.fromtimestamp(row['timestamp'] / 1000) # from ms to sec
        today = int(date.strfime('%d'))
        if temp_today is not today:
            new_dict = {'date': date.strftime('%d/%m-%Y'),
                        'reading_count': count,
                        'time': date.strftime('%H:%M:%S')}
            dict_list.append(new_dict)

    return dict_list

2 个答案:

答案 0 :(得分:6)

create_standard_list()format_table_dictionary()可以构建生成器(yield每个项目而不是return完整列表),这会停止将整个列表保存在内存中,因此应该解决您的问题,例如:

def create_standard_list(sensor_id, data):
    for row in data:
        row_timestamp = row[2]
        row_data = row[3]

        temp_object = {"sensor_id": sensor_id, "timestamp": row_timestamp,
                       "data": row_data}
        yield temp_object
       #^ yield each item instead of appending to a list

有关generatorsyield keyword的更多信息。

答案 1 :(得分:3)

您在此尝试执行的操作IIUC是在Python代码中模拟SQL GROUP BY表达式。这永远不会像直接在数据库中那样快速和有效。 您的示例代码似乎有一些问题,但我理解为:您想要 计算给定did每天发生的每天行数。而且,你是 感兴趣的是每组价值的最小值(或最大值,或中位数,无关紧要),即每天。

让我们设置一个小的示例表(在Oracle上测试):

create table t1 (id number primary key, created timestamp, did number, other_data varchar2(200));  

insert into t1 values (1, to_timestamp('2017-01-31 17:00:00', 'YYYY-MM-DD HH24:MI:SS'), 9001, 'some text');
insert into t1 values (2, to_timestamp('2017-01-31 19:53:00', 'YYYY-MM-DD HH24:MI:SS'), 9001, 'some more text');
insert into t1 values (3, to_timestamp('2017-02-01 08:10:00', 'YYYY-MM-DD HH24:MI:SS'), 9001, 'another day');
insert into t1 values (4, to_timestamp('2017-02-01 15:55:00', 'YYYY-MM-DD HH24:MI:SS'), 9001, 'another day, rainy afternoon');
insert into t1 values (5, to_timestamp('2017-02-01 15:59:00', 'YYYY-MM-DD HH24:MI:SS'), 9002, 'different did');
insert into t1 values (6, to_timestamp('2017-02-03 01:01:00', 'YYYY-MM-DD HH24:MI:SS'), 9001, 'night shift');

对于9001,我们有几天分散了几天。还有9002的价值,我们会这样做 忽视。现在让我们将您要写入第二个表的行作为简单的SELECT .. GROUP BY

select 
    count(*) cnt, 
    to_char(created, 'YYYY-MM-DD') day, 
    min(to_char(created, 'HH24:MI:SS')) min_time 
from t1 
where did = 9001
group by to_char(created, 'YYYY-MM-DD')
;

我们按照created列(时间戳)的日期对所有行进行分组。我们选择了 每组的行数,日期本身,以及 - 只是为了好玩 - 每个的最小时间部分 组。结果:

cnt day         min_time
2   2017-02-01  08:10:00
1   2017-02-03  01:01:00
2   2017-01-31  17:00:00

所以现在你的第二个表格为SELECT。从中创建表格是微不足道的:

create table t2 as
select
    ... as above
;

HTH!