Clickhouse Dictionaries 在内存中的存储方式

Clickhouse Dictionaries -Storing Dictionaries in Memory

Clickhouse支持多种方式将字典存储在内存中

一般推荐flathashedcomplex_key_hashed,这些提供了最佳的处理速度,但是不推荐使用cache,因为可能会出现性能差且难以选择最佳参数的问题。

有以下几种方式提升字典的使用性能:

  • 在使用Group By之后再调用函数处理字典
  • 将属性标记为单射(injective).如果不同的属性值对应不同的键,则属性被称为单射。因此,当group by 中使用通过key获取字典value的函数时,此函数将自动从group by中取出。

出现以下情况,会产生异常:

  • 无法加载正在访问的字典
  • 查询cache字典时出错

可以通过查询system.dictionaries表来获取已导入的外部字典列表以及字典的状态

存储方式的配置节点在字典配置文件的dictionary - layout - layout_type节点:

1
2
3
4
5
6
7
8
9
10
11
<yandex>
<dictionary>
...
<layout>
<layout_type>
<!-- layout settings -->
</layout_type>
</layout>
...
</dictionary>
</yandex>

clickhouse 目前支持以下存储方式:

  • flat
  • hashed
  • cache
  • range_hashed
  • complex_key_hashed
  • complex_key_cache
  • ip_trie

flat

字典以平面阵列(flat arrays)的形式完全存储在内存中。具体使用了多少内存?占用内存量和最大key的size成正比

字典的键(key)必须是UInt64类型,而且键值被限制在500,000,超过最大值,则不会创建字典并且抛出异常

支持所有的数据源,更新时,数据(from a file or a table)被整体加载

在所有的存储方式中,flat具有最佳性能

配置

1
2
3
<layout>
<flat />
</layout>

hashed

字典以hash table 的形式完全存储在内存中。该种形式存储的字典可以包含任意数量的带有任何标识符的元素。实际上,键的数量可以达到数千万(tens of millions)。

支持所有的数据源,更新时,数据(from a file or a table)被整体加载

配置

1
2
3
<layout>
<hashed />
</layout>

complex_key_hashed

从《Dictionary Key and Fields》章节可以知道,clickhouse 字典是支持Numeric keyComposite key。如果使用Composite key,存储方式必须为complex_key_hashed 或者complex_key_cache

其他同 hashed

配置

1
2
3
<layout>
<complex_key_hashed />
</layout>

range_hashed

字典以 具有有序的范围数组和对应值的hash table 的形式完全存储在内存中

该存储方式和 hashed 工作方式相同,除了普通的key外,还可以使用日期/时间范围的键

Example: 该表包含以下格式的每个广告客户的折扣::

1
2
3
4
5
6
7
8
9
+---------------+---------------------+-------------------+--------+
| advertiser id | discount start date | discount end date | amount |
+===============+=====================+===================+========+
| 123 | 2015-01-01 | 2015-01-15 | 0.15 |
+---------------+---------------------+-------------------+--------+
| 123 | 2015-01-16 | 2015-01-31 | 0.25 |
+---------------+---------------------+-------------------+--------+
| 456 | 2015-01-01 | 2015-01-15 | 0.05 |
+---------------+---------------------+-------------------+--------+

使用时间range的样例, 在配置中定义 range\_min and range\_max

1
2
3
4
5
6
7
8
9
10
11
<structure>
<id>
<name>Id</name>
</id>
<range_min>
<name>first</name>
</range_min>
<range_max>
<name>last</name>
</range_max>
....

要使用这些词典,需要将一个额外的日期参数传递给dictGetT函数:

1
dictGetT('dict_name', 'attr_name', id, date)

此函数返回指定的id并且date在符合的日期范围内的值(id = x and date between first and last)。

1
2
-- 举例广告客户123在2015-01-18的折扣
dictGetT('dict_name', 'attr_name', 123, 2015-01-18) = 0.25 --id是uint64,date 是日期型,只是举例

Details of the algorithm:

  • 如果未找到id或未找到符合传递date的范围,则返回字典的默认值。
  • 如果存在重叠范围,则可以使用任何范围。
  • 如果范围分隔符为NULL或无效日期(例如1900-01-01或2039-01-01),则the range is left open。The range can be open on both sides.。

配置

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
<yandex>
<dictionary>

...

<layout>
<range_hashed />
</layout>

<structure>
<id>
<name>Abcdef</name>
</id>
<range_min>
<name>StartDate</name>
</range_min>
<range_max>
<name>EndDate</name>
</range_max>
<attribute>
<name>XXXType</name>
<type>String</type>
<null_value />
</attribute>
</structure>

</dictionary>
</yandex>

cache

字典存储在具有固定数量单元格的缓存中。 这些单元格包含常用元素。

当进行字典搜索时,将首先搜索cache。对于每个数据块,将会使用SELECT attrs... FROM db.table WHERE id IN (k1, k2, ...)sql 语句从源中请求那些不能被缓存命中或者已经超时的keys。然后将获得的数据写入cache。

对于缓存字典,可以设置缓存中数据的到期生存期(lifetime)。 如果数据从加载到cell开始超过了lifetime,则不使用cell’s value,并在下次需要使用时重新请求该数据。这也是存储词典的所有方法中效率最低的。 缓存字典的速度在很大程度上取决于正确的设置和使用场景。 只有当命中率足够高时(建议99%或更高),缓存类型字典才能很好地执行。 您可以在system.dictionaries表中查看平均命中率。

要提高缓存性能,请使用带有LIMIT的子查询,并在外部使用字典调用该函数。

Supported sources: MySQL, ClickHouse, executable, HTTP

配置

1
2
3
4
5
6
<layout>
<cache>
<!-- The size of the cache, in number of cells. Rounded up to a power of two. -->
<size_in_cells>1000000000</size_in_cells>
</cache>
</layout>

设置足够大的缓存大小。 您需要尝试选择单元格数量:

  1. 设定一些value。
  2. 运行查询,直到缓存完全填满。
  3. 使用system.dictionaries表评估内存消耗。
  4. 增加或减少单元数,直到达到所需的内存消耗。

Warning:

不要用Click house作为数据源,因为clickhouse 处理随机读太慢

complex_key_cache

This type of storage is for use with composite keys. Similar to cache.

ip_trie

这种类型的存储用于将网络前缀(IP地址)映射到诸如ASN的元数据

示例:该表包含网络前缀及其对应的AS编号和国家/地区代码:

1
2
3
4
5
6
7
8
9
10
11
+-----------------+-------+--------+
| prefix | asn | cca2 |
+=================+=======+========+
| 202.79.32.0/20 | 17501 | NP |
+-----------------+-------+--------+
| 2620:0:870::/48 | 3856 | US |
+-----------------+-------+--------+
| 2a02:6b8:1::/48 | 13238 | RU |
+-----------------+-------+--------+
| 2001:db8::/32 | 65536 | ZZ |
+-----------------+-------+--------+

When using this type of layout, the structure must have a composite key

Example:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
<structure>
<key>
<attribute>
<name>prefix</name>
<type>String</type>
</attribute>
</key>
<attribute>
<name>asn</name>
<type>UInt32</type>
<null_value />
</attribute>
<attribute>
<name>cca2</name>
<type>String</type>
<null_value>??</null_value>
</attribute>
...

key必须只有一个包含允许的IP前缀的String类型属性。 其他类型尚不支持。

对于查询,您必须使用与Composite key字典相同的函数(dictGetT with a tuple):

1
dictGetT('dict_name', 'attr_name', tuple(ip))

该函数支持UInt32的IPv4或FixedString(16)的IPv6:

1
dictGetString('prefix', 'asn', tuple(IPv6StringToNum('2001:db8::1')))

其他类型尚不支持。 该函数返回与此IP地址对应的前缀的属性。 如果存在重叠前缀,则返回最具体的一个。

数据以字典树(Trie)形式完全存储在随机存取存储器中(RAM)


Your browser is out-of-date!

Update your browser to view this website correctly. Update my browser now

×