Storing Dictionaries in Memory

There are a variety of ways to store dictionaries in memory.

We recommend flat, hashed and complex_key_hashed. which provide optimal processing speed.

Caching is not recommended because of potentially poor performance and difficulties in selecting optimal parameters. Read more in the section “cache“.

There are several ways to improve dictionary performance:

  • Call the function for working with the dictionary after GROUP BY.
  • Mark attributes to extract as injective. An attribute is called injective if different attribute values correspond to different keys. So when GROUP BY uses a function that fetches an attribute value by the key, this function is automatically taken out of GROUP BY.

ClickHouse generates an exception for errors with dictionaries. Examples of errors:

  • The dictionary being accessed could not be loaded.
  • Error querying a cached dictionary.

You can view the list of external dictionaries and their statuses in the system.dictionaries table.

The configuration looks like this:

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

Corresponding DDL-query:

  1. CREATE DICTIONARY (...)
  2. ...
  3. LAYOUT(LAYOUT_TYPE(param value)) -- layout settings
  4. ...

Ways to Store Dictionaries in Memory

flat

The dictionary is completely stored in memory in the form of flat arrays. How much memory does the dictionary use? The amount is proportional to the size of the largest key (in space used).

The dictionary key has the UInt64 type and the value is limited to 500,000. If a larger key is discovered when creating the dictionary, ClickHouse throws an exception and does not create the dictionary.

All types of sources are supported. When updating, data (from a file or from a table) is read in its entirety.

This method provides the best performance among all available methods of storing the dictionary.

Configuration example:

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

or

  1. LAYOUT(FLAT())

hashed

The dictionary is completely stored in memory in the form of a hash table. The dictionary can contain any number of elements with any identifiers In practice, the number of keys can reach tens of millions of items.

All types of sources are supported. When updating, data (from a file or from a table) is read in its entirety.

Configuration example:

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

or

  1. LAYOUT(HASHED())

sparse_hashed

Similar to hashed, but uses less memory in favor more CPU usage.

Configuration example:

  1. <layout>
  2. <sparse_hashed />
  3. </layout>
  1. LAYOUT(SPARSE_HASHED())

complex_key_hashed

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

Configuration example:

  1. <layout>
  2. <complex_key_hashed />
  3. </layout>
  1. LAYOUT(COMPLEX_KEY_HASHED())

range_hashed

The dictionary is stored in memory in the form of a hash table with an ordered array of ranges and their corresponding values.

This storage method works the same way as hashed and allows using date/time (arbitrary numeric type) ranges in addition to the key.

Example: The table contains discounts for each advertiser in the format:

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

To use a sample for date ranges, define the range_min and range_max elements in the structure. These elements must contain elements name andtype (if type is not specified, the default type will be used - Date). type can be any numeric type (Date / DateTime / UInt64 / Int32 / others).

Example:

  1. <structure>
  2. <id>
  3. <name>Id</name>
  4. </id>
  5. <range_min>
  6. <name>first</name>
  7. <type>Date</type>
  8. </range_min>
  9. <range_max>
  10. <name>last</name>
  11. <type>Date</type>
  12. </range_max>
  13. ...

or

  1. CREATE DICTIONARY somedict (
  2. id UInt64,
  3. first Date,
  4. last Date
  5. )
  6. PRIMARY KEY id
  7. LAYOUT(RANGE_HASHED())
  8. RANGE(MIN first MAX last)

To work with these dictionaries, you need to pass an additional argument to the dictGetT function, for which a range is selected:

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

This function returns the value for the specified ids and the date range that includes the passed date.

Details of the algorithm:

  • If the id is not found or a range is not found for the id, it returns the default value for the dictionary.
  • If there are overlapping ranges, you can use any.
  • If the range delimiter is NULL or an invalid date (such as 1900-01-01 or 2039-01-01), the range is left open. The range can be open on both sides.

Configuration example:

  1. <yandex>
  2. <dictionary>
  3. ...
  4. <layout>
  5. <range_hashed />
  6. </layout>
  7. <structure>
  8. <id>
  9. <name>Abcdef</name>
  10. </id>
  11. <range_min>
  12. <name>StartTimeStamp</name>
  13. <type>UInt64</type>
  14. </range_min>
  15. <range_max>
  16. <name>EndTimeStamp</name>
  17. <type>UInt64</type>
  18. </range_max>
  19. <attribute>
  20. <name>XXXType</name>
  21. <type>String</type>
  22. <null_value />
  23. </attribute>
  24. </structure>
  25. </dictionary>
  26. </yandex>

or

  1. CREATE DICTIONARY somedict(
  2. Abcdef UInt64,
  3. StartTimeStamp UInt64,
  4. EndTimeStamp UInt64,
  5. XXXType String DEFAULT ''
  6. )
  7. PRIMARY KEY Abcdef
  8. RANGE(MIN StartTimeStamp MAX EndTimeStamp)

cache

The dictionary is stored in a cache that has a fixed number of cells. These cells contain frequently used elements.

When searching for a dictionary, the cache is searched first. For each block of data, all keys that are not found in the cache or are outdated are requested from the source using SELECT attrs... FROM db.table WHERE id IN (k1, k2, ...). The received data is then written to the cache.

For cache dictionaries, the expiration lifetime of data in the cache can be set. If more time than lifetime has passed since loading the data in a cell, the cell’s value is not used, and it is re-requested the next time it needs to be used.
This is the least effective of all the ways to store dictionaries. The speed of the cache depends strongly on correct settings and the usage scenario. A cache type dictionary performs well only when the hit rates are high enough (recommended 99% and higher). You can view the average hit rate in the system.dictionaries table.

To improve cache performance, use a subquery with LIMIT, and call the function with the dictionary externally.

Supported sources: MySQL, ClickHouse, executable, HTTP.

Example of settings:

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

or

  1. LAYOUT(CACHE(SIZE_IN_CELLS 1000000000))

Set a large enough cache size. You need to experiment to select the number of cells:

  1. Set some value.
  2. Run queries until the cache is completely full.
  3. Assess memory consumption using the system.dictionaries table.
  4. Increase or decrease the number of cells until the required memory consumption is reached.

Warning

Do not use ClickHouse as a source, because it is slow to process queries with random reads.

complex_key_cache

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

ip_trie

This type of storage is for mapping network prefixes (IP addresses) to metadata such as ASN.

Example: The table contains network prefixes and their corresponding AS number and country code:

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

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

Example:

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

or

  1. CREATE DICTIONARY somedict (
  2. prefix String,
  3. asn UInt32,
  4. cca2 String DEFAULT '??'
  5. )
  6. PRIMARY KEY prefix

The key must have only one String type attribute that contains an allowed IP prefix. Other types are not supported yet.

For queries, you must use the same functions (dictGetT with a tuple) as for dictionaries with composite keys:

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

The function takes either UInt32 for IPv4, or FixedString(16) for IPv6:

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

Other types are not supported yet. The function returns the attribute for the prefix that corresponds to this IP address. If there are overlapping prefixes, the most specific one is returned.

Data is stored in a trie. It must completely fit into RAM.

Original article