CarbonData BloomFilter DataMap

DataMap Management

Creating BloomFilter DataMap

CREATE DATAMAP [IF NOT EXISTS] datamap_name
ON TABLE main_table
USING 'bloomfilter'
DMPROPERTIES ('index_columns'='city, name', 'BLOOM_SIZE'='640000', 'BLOOM_FPP'='0.00001')

Dropping specified datamap

DROP DATAMAP [IF EXISTS] datamap_name
ON TABLE main_table

Showing all DataMaps on this table

SHOW DATAMAP
ON TABLE main_table

Disable Datamap

The datamap by default is enabled. To support tuning on query, we can disable a specific datamap during query to observe whether we can gain performance enhancement from it. This is effective only for current session.

// disable the datamap
SET carbon.datamap.visible.dbName.tableName.dataMapName = false
// enable the datamap
SET carbon.datamap.visible.dbName.tableName.dataMapName = true

BloomFilter DataMap Introduction

A Bloom filter is a space-efficient probabilistic data structure that is used to test whether an element is a member of a set. Carbondata introduced BloomFilter as an index datamap to enhance the performance of querying with precise value. It is well suitable for queries that do precise match on high cardinality columns(such as Name/ID). Internally, CarbonData maintains a BloomFilter per blocklet for each index column to indicate that whether a value of the column is in this blocklet. Just like the other datamaps, BloomFilter datamap is managed along with main tables by CarbonData. User can create BloomFilter datamap on specified columns with specified BloomFilter configurations such as size and probability.

For instance, main table called datamap_test which is defined as:

CREATE TABLE datamap_test (
  id string,
  name string,
  age int,
  city string,
  country string)
STORED AS carbondata
TBLPROPERTIES('SORT_COLUMNS'='id')

In the above example, id and name are high cardinality columns and we always query on id and name with precise value. since id is in the sort_columns and it is orderd, query on it will be fast because CarbonData can skip all the irrelative blocklets. But queries on name may be bad since the blocklet minmax may not help, because in each blocklet the range of the value of name may be the same -- all from A* to z*. In this case, user can create a BloomFilter datamap on column name. Moreover, user can also create a BloomFilter datamap on the sort_columns. This is useful if user has too many segments and the range of the value of sort_columns are almost the same.

User can create BloomFilter datamap using the Create DataMap DDL:

CREATE DATAMAP dm
ON TABLE datamap_test
USING 'bloomfilter'
DMPROPERTIES ('INDEX_COLUMNS' = 'name,id', 'BLOOM_SIZE'='640000', 'BLOOM_FPP'='0.00001', 'BLOOM_COMPRESS'='true')

Properties for BloomFilter DataMap

Property Is Required Default Value Description
INDEX_COLUMNS YES Carbondata will generate BloomFilter index on these columns. Queries on these columns are usually like 'COL = VAL'.
BLOOM_SIZE NO 640000 This value is internally used by BloomFilter as the number of expected insertions, it will affect the size of BloomFilter index. Since each blocklet has a BloomFilter here, so the default value is the approximate distinct index values in a blocklet assuming that each blocklet contains 20 pages and each page contains 32000 records. The value should be an integer.
BLOOM_FPP NO 0.00001 This value is internally used by BloomFilter as the False-Positive Probability, it will affect the size of bloomfilter index as well as the number of hash functions for the BloomFilter. The value should be in the range (0, 1). In one test scenario, a 96GB TPCH customer table with bloom_size=320000 and bloom_fpp=0.00001 will result in 18 false positive samples.
BLOOM_COMPRESS NO true Whether to compress the BloomFilter index files.

Loading Data

When loading data to main table, BloomFilter files will be generated for all the index_columns given in DMProperties which contains the blockletId and a BloomFilter for each index column. These index files will be written inside a folder named with datamap name inside each segment folders.

Querying Data

User can verify whether a query can leverage BloomFilter datamap by executing EXPLAIN command, which will show the transformed logical plan, and thus user can check whether the BloomFilter datamap can skip blocklets during the scan. If the datamap does not prune blocklets well, you can try to increase the value of property BLOOM_SIZE and decrease the value of property BLOOM_FPP.

Data Management With BloomFilter DataMap

Data management with BloomFilter datamap has no difference with that on Lucene datamap. You can refer to the corresponding section in CarbonData Lucene DataMap.

Useful Tips

  • BloomFilter DataMap is suggested to be created on the high cardinality columns. Query conditions on these columns are always simple equal or in, such as 'col1=XX', 'col1 in (XX, YY)'.
  • We can create multiple BloomFilter datamaps on one table, but we do recommend you to create one BloomFilter datamap that contains multiple index columns, because the data loading and query performance will be better.
  • BLOOM_FPP is only the expected number from user, the actually FPP may be worse. If the BloomFilter datamap does not work well, you can try to increase BLOOM_SIZE and decrease BLOOM_FPP at the same time. Notice that bigger BLOOM_SIZE will increase the size of index file and smaller BLOOM_FPP will increase runtime calculation while performing query.
  • '0' skipped blocklets of BloomFilter datamap in explain output indicates that BloomFilter datamap does not prune better than Main datamap. (For example since the data is not ordered, a specific value may be contained in many blocklets. In this case, bloom may not work better than Main DataMap.) If this occurs very often, it means that current BloomFilter is useless. You can disable or drop it. Sometimes we cannot see any pruning result about BloomFilter datamap in the explain output, this indicates that the previous datamap has pruned all the blocklets and there is no need to continue pruning.
  • In some scenarios, the BloomFilter datamap may not enhance the query performance significantly but if it can reduce the number of spark task, there is still a chance that BloomFilter datamap can enhance the performance for concurrent query.
  • Note that BloomFilter datamap will decrease the data loading performance and may cause slightly storage expansion (for datamap index file).