Cuckoo Filter For Java
This library offers a similar interface to Guava's Bloom filters. In most cases it can be used interchangeably and has additional advantages including thread-safety, concurrent operations, deletions/counting and a configurable hashing algorithm.
This Forked version from MGunlogson/CuckooFilter4J has replaced the MurMur hash functions with xxHash which is much quicker. If you are only using the filter with primitive numbers (long,int,short,byte) you should use the filter from the primitive branch (https://github.com/MPdaedalus/CuckooFilter4J/tree/Primitive-Filter) as it is much quicker due to mightContain() calls creating no Garbage compared to this branch that supports Objects.
About Cuckoo Filters
Cuckoo filter is a Bloom filter replacement for approximated set-membership queries. While Bloom filters are well-known space-efficient data structures to serve queries like "if item x is in a set?", they do not support deletion. Their variances to enable deletion (like counting Bloom filters) usually require much more space.
Cuckoo filters provide the flexibility to add and remove items dynamically. A cuckoo filter is based on cuckoo hashing (and therefore named as cuckoo filter). It is essentially a cuckoo hash table storing each key's fingerprint. Cuckoo hash tables can be highly compact, thus a cuckoo filter could use less space than conventional Bloom filters, for applications that require low false positive rates (< 3%).
For details about the algorithm and citations please use this article for now
"Cuckoo Filter: Better Than Bloom" by Bin Fan, Dave Andersen and Michael Kaminsky
If you're looking for an implementation in a different language check out these GitHub links.
My personal thanks goes out to the authors of those libaries, their code helped immensely while writing my own implementation.
Installation
Maven artifact:
<dependency>
<groupId>com.github.mgunlogson</groupId>
<artifactId>cuckoofilter4j</artifactId>
<version>1.0.1</version>
</dependency>
Is performance important to your application?
The regular filter is fairly quick, but MPdaedalus has made significant performance improvement by adding a brand new hash algorithm and reducing GC pressure. These changes speed up most operations by ~2x! His changes can be found here. These improvements will be merged to version 2.0 as soon as the public API is finalized.
Usage
Below is a full example of creating and using the filter. Many more examples can be found in the test folders within the project.
import com.github.mgunlogson.cuckoofilter4j;
import com.google.common.hash.Funnels;
public class Example {
public static void main(String[] args) {
// create
CuckooFilter<Integer> filter = new CuckooFilter.Builder<>(Funnels.integerFunnel(), 2000000).build();
// insert
if (filter.put(42)) {
System.out.println("Insert Success!");
}
// contains
if (filter.mightContain(42)) {
System.out.println("Found 42!");
}
// count
System.out.println("Filter has " + filter.getCount() + " items");
// count
System.out.println("42 has been inserted approximately " + filter.approximateCount(42) + " times");
// % loaded
System.out.println("Filter is " + String.format("%.0f%%", filter.getLoadFactor() * 100) + " loaded");
// delete
if (filter.delete(42)) {
System.out.println("Delete Success!");
}
}
}
Documentation
###False Positives The false positive rate of the filter is the probability that mightContain()
will erroneously return true
for an object that was not added to the filter. Unlike Bloom filters, a Cuckoo filter will fail to insert when it reaches capacity. If an insert fails put()
will return false
Duplicates
Cuckoo filters allow deletion like counting Bloom filters. While counting Bloom filters invariably use more space to allow deletions, Cuckoo filters achieve this with no space or time cost. Like counting variations of Bloom filters, Cuckoo filters have a limit to the number of times you can insert duplicate items. This limit is 8-9 in the current design, depending on internal state. Reaching this limit can cause further inserts to fail and degrades the performance of the filter. Occasional duplicates will not degrade the performance of the filter but will slightly reduce capacity. Existing items can be deleted without affecting the false positive rate or causing false negatives. However, deleting items that were not previously added to the filter can cause false negatives.
Counting
Cuckoo filters support counting items, like counting Bloom filters. The maximum count is still limited by max-duplicates to 7 so this should only be used to count small numbers. The measured count may be higher than actual count due to false positives, but will never be lower since Cuckoo filters have no false negatives.
Capacity
Once the filter reaches capacity (put()
returns false). It's best to either rebuild the existing filter or create a larger one. Deleting items in the current filter is also an option, but you should delete at least ~2% of the items in the filter before inserting again.
Speed
CuckooFilter4J is roughly the same speed as Guava's Bloom filters when running single-threaded. Guava's Bloom is usually faster with small tables, but the trend is reversed with tables too large to fit in the CPU cache. Overall the single-threaded speed of the two libraries is comparable. This library supports concurrent access through multithreading (Guava's Bloom does not). In my tests this scales fairly well, making CuckooFilter4J faster than Bloom filters for multi-threaded applications. On my 4 core machine, running inserts on all cores is roughly 3x faster than single-threaded operation. Cpu architecture will affect this, so your mileage may vary. See the benchmark folder for some tests to run on your own system.
Hashing Algorithms
Hash collision attacks are theoretically possible against Cuckoo filters (as with any hash table based structure). If this is an issue for your application, use one of the cryptographically secure (but slower) hash functions. The default hash function, Murmer3 is not secure. Secure functions include SHA and SipHash. All hashes,including non-secure, are internally seeded and salted. Practical attacks against any of them are unlikely. Also note that the maximum supported size of the filter depends on the hash funciton. Especially in the case of 32 bit Murmur3, the hash will limit table size. Even with a 32 bit hash, the maximum table size is around 270 megabytes. With 64 bit hashes the maximum table size is extremely large, and practically unlimited using 128+bit hash functions. In any case, the library will refuse to create the table using an invalid configuration.
Multi-Threading
All operations are thread-safe. Most also run concurrently for increased performance. Notable exceptions include copy, serialization, and hashcode which nessecarily lock the entire table - running on a single thread until complete. Thread safety should be considered BETA at the moment. Multithreading is notoriously hard to test, and despite my best effort to avoid bugs and deadlocks it is likely that some remain. If you are using multithreading in production I will do my best to provide prompt support and give you my thanks :).
Serializing
Cuckoo filters are serializable.