Skip to main content

Bloom Filter

To use this Apache Druid extension, include druid-bloom-filter in the extensions load list.

This extension adds the ability to both construct bloom filters from query results, and filter query results by testing against a bloom filter. A Bloom filter is a probabilistic data structure for performing a set membership check. A bloom filter is a good candidate to use with Druid for cases where an explicit filter is impossible, e.g. filtering a query against a set of millions of values.

Following are some characteristics of Bloom filters:

  • Bloom filters are highly space efficient when compared to using a HashSet.
  • Because of the probabilistic nature of bloom filters, false positive results are possible (element was not actually inserted into a bloom filter during construction, but test() says true)
  • False negatives are not possible (if element is present then test() will never say false).
  • The false positive probability of this implementation is currently fixed at 5%, but increasing the number of entries that the filter can hold can decrease this false positive rate in exchange for overall size.
  • Bloom filters are sensitive to number of elements that will be inserted in the bloom filter. During the creation of bloom filter expected number of entries must be specified. If the number of insertions exceed the specified initial number of entries then false positive probability will increase accordingly.

This extension is currently based on org.apache.hive.common.util.BloomKFilter from hive-storage-api. Internally, this implementation uses Murmur3 as the hash algorithm.

To construct a BloomKFilter externally with Java to use as a filter in a Druid query:

BloomKFilter bloomFilter = new BloomKFilter(1500);
bloomFilter.addString("value 1");
bloomFilter.addString("value 2");
bloomFilter.addString("value 3");
ByteArrayOutputStream byteArrayOutputStream = new ByteArrayOutputStream();
BloomKFilter.serialize(byteArrayOutputStream, bloomFilter);
String base64Serialized = Base64.encodeBase64String(byteArrayOutputStream.toByteArray());

This string can then be used in the native or SQL Druid query.

Filtering queries with a Bloom Filter

JSON Specification of Bloom Filter

{
"type" : "bloom",
"dimension" : <dimension_name>,
"bloomKFilter" : <serialized_bytes_for_BloomKFilter>,
"extractionFn" : <extraction_fn>
}
PropertyDescriptionrequired?
typeFilter Type. Should always be bloomyes
dimensionThe dimension to filter over.yes
bloomKFilterBase64 encoded Binary representation of org.apache.hive.common.util.BloomKFilteryes
extractionFnExtraction function to apply to the dimension valuesno

Serialized Format for BloomKFilter

Serialized BloomKFilter format:

  • 1 byte for the number of hash functions.
  • 1 big endian int(That is how OutputStream works) for the number of longs in the bitset
  • big endian longs in the BloomKFilter bitset

Note: org.apache.hive.common.util.BloomKFilter provides a serialize method which can be used to serialize bloom filters to outputStream.

Filtering SQL Queries

Bloom filters can be used in SQL WHERE clauses via the bloom_filter_test operator:

SELECT COUNT(*) FROM druid.foo WHERE bloom_filter_test(<expr>, '<serialized_bytes_for_BloomKFilter>')

Expression and Virtual Column Support

The bloom filter extension also adds a bloom filter Druid expression which shares syntax with the SQL operator.

bloom_filter_test(<expr>, '<serialized_bytes_for_BloomKFilter>')

Bloom Filter Query Aggregator

Input for a bloomKFilter can also be created from a druid query with the bloom aggregator. Note that it is very important to set a reasonable value for the maxNumEntries parameter, which is the maximum number of distinct entries that the bloom filter can represent without increasing the false positive rate. It may be worth performing a query using one of the unique count sketches to calculate the value for this parameter in order to build a bloom filter appropriate for the query.

JSON Specification of Bloom Filter Aggregator

{
"type": "bloom",
"name": <output_field_name>,
"maxNumEntries": <maximum_number_of_elements_for_BloomKFilter>
"field": <dimension_spec>
}
PropertyDescriptionrequired?
typeAggregator Type. Should always be bloomyes
nameOutput field nameyes
fieldDimensionSpec to add to org.apache.hive.common.util.BloomKFilteryes
maxNumEntriesMaximum number of distinct values supported by org.apache.hive.common.util.BloomKFilter, default 1500no

Example

{
"queryType": "timeseries",
"dataSource": "wikiticker",
"intervals": [ "2015-09-12T00:00:00.000/2015-09-13T00:00:00.000" ],
"granularity": "day",
"aggregations": [
{
"type": "bloom",
"name": "userBloom",
"maxNumEntries": 100000,
"field": {
"type":"default",
"dimension":"user",
"outputType": "STRING"
}
}
]
}

response

[{"timestamp":"2015-09-12T00:00:00.000Z","result":{"userBloom":"BAAAJhAAAA..."}}]

These values can then be set in the filter specification described above.

Ordering results by a bloom filter aggregator, for example in a TopN query, will perform a comparatively expensive linear scan of the filter itself to count the number of set bits as a means of approximating how many items have been added to the set. As such, ordering by an alternate aggregation is recommended if possible.

SQL Bloom Filter Aggregator

Bloom filters can be computed in SQL expressions with the bloom_filter aggregator:

SELECT BLOOM_FILTER(<expression>, <max number of entries>) FROM druid.foo WHERE dim2 = 'abc'

but requires the setting druid.sql.planner.serializeComplexValues to be set to true. Bloom filter results in a SQL response are serialized into a base64 string, which can then be used in subsequent queries as a filter.