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
- 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
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);
ByteArrayOutputStream byteArrayOutputStream = new ByteArrayOutputStream();
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>
|Filter Type. Should always be ||yes|
|The dimension to filter over.||yes|
|Base64 encoded Binary representation of ||yes|
|Extraction function to apply to the dimension values||no|
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
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
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 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
|Aggregator Type. Should always be ||yes|
|Output field name||yes|
|DimensionSpec to add to ||yes|
|Maximum number of distinct values supported by ||no|
"intervals": [ "2015-09-12T00:00:00.000/2015-09-13T00:00:00.000" ],
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
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.