Sorry it has taken me so long to respond, but I spent a long time thinking
about this and started a response several times.
I see no reason not to go to long bit indexes. This would be reflected in
Shape.getNumberOfBits returning a long as well as other places.
I would like to revisit some definitions and use cases before delving into
the suggestions.
Definitions
-
Bloom Filter
-
A set of bits and functions to query bit patterns.
-
Responsible for
-
Management of storage of the set of bits.
-
Producing a standard format external representation of the set of
bits.
-
Importing a standard format external representation of the set of
bits.
-
Performing specific bit queries on the set of bits:
-
Cardinality (number of bits enabled) of the filter.
-
Cardinality of a logical “and” with another filter.
-
Cardinality of a logical “or” with another filter.
-
Cardinality of a logical “xor” with another filter.
-
Determining if this filter “contains” another filter (this
“and” that == that )
-
Merging another filter into this one. This is not technically
required but was a design decision. The alternative for a filter to be
immutable and to create a new filter from the merging of two filters.
-
Shape
-
The definition of a Bloom filter. It includes:
-
The maximum number of bits in the filter. (aka: m)
-
The expected number of items encoded in the filter. (aka: n)
-
The number of hashes generated for each item in the filter. (aka:
t)
-
Responsible for providing the information necessary to identify when
two filters are built with the same strategy.
-
Hasher
-
Converts one or more items into a Bloom filter based upon a Shape.
-
Is repeatable. (i.e. when called multiple times with the same shape
produces the same output).
-
Responsible for:
-
Tracking the number of items being hashed.
-
Generating the number of hash values specified by Shape for each
item.
Types of usage Representation of a single simple object
This is probably the most common usage pattern if measured by number of
constructor calls. In this usage case a single object is converted to a
single item in a Bloom filter. A common use case would be to determine if a
URL has been seen. The filter is constructed by converting the hashing the
URL “t” times, each hash value converted to the range [0,m) and the
associated bits enabled in the Bloom filter.
The single object filter is commonly used to test the presence of the
object in a collection of filters or to register the presence of the object
within the collection.
Representation of a single complex object
In this usage the Bloom filter represents a single object with more than
one property as contrasted to the simple object were a single property is
represented. In this the number of items in the Shape actually refers to
the number of properties for a single complex object time the number of
complex objects. So a collection of 1000 complex objects with 3 properties
each yields 3000 items in the Shape.
There are several hashing strategies used to build complex objects. The
simplest is to hash each property as though it were the property in a
single object. This works well if the individual properties being indexed
do not have any collisions across the representations. A more complex
solution is to seed the hash function with different values based upon the
property being indexed.
Representation of a single complete complex object
This use case is similar to the single complex object in that multiple
properties are represented as individual items in the resulting Bloom
filter. The difference is that the number of expected items, as defined by
the shape, is the number of properties in the object. Thus each object
produces a fully populated Bloom filter. This strategy is used when a
compact searchable representation of the object is required. This type of
filter is generally added to a Multidimensional filter (see below).
Representation of a partial complex object
In this case the representation is built using the same method as the
single complex object but not all properties are specified. This is
commonly used to locate objects in a collection of complex object Bloom
filters.
Representation of multiple objects
In this case multiple single-object Bloom filters are merged together to
create one Bloom filter representing all the objects. This is probably the
most common usage for filters that are persisted across application runs.
This implementation simply the merged bit patters of all the Bloom filters
being added. It is commonly used in applications like Hadoop to determine
which nodes might contain a specific file.
Multidimensional representation Bloom filters
This is a use case that extends the ability to store excess items (>”n”)
into a Bloom filter without significantly modifying the probability of
false positives. To do this efficiently the Multidimensional filter has to
know the Shape of the filters being inserted. In the most efficient cases
it requires the hasher so that it can build filters of different but
associated shapes. In addition, this type of filter can return ID’s
associated with the various filters. So that it in effect becomes an index
of objects associated with Bloom filters.
This type of filter is used in systems that have many (>1000) locations to
check for a Bloom filter match as they can perform the match much faster
than the linear search alternative.
Suggestions Move the hashing into the Bloom filter
Originally the Hasher was called a ProtoBloomFilter. There was a request
for a name change early on. Once a Hasher is created it can be used to
create any concrete Bloom filter. Currently this is done with the
getBits(Shape) method. Perhaps what is needed is a rework in this section.
If the Hasher had a method to return a LongSupplier for each item then the
Bloom filter constructor could be in control of the number of hash
functions that were called for each item.
So assume Hasher has 2 new methods:
int items(); // the number of items
LongSupplier supplier( int itemNumber )
Then BloomFilter(Hasher, Shape) constructor could call:
for (int i =0; i<hasher.items();i++) {
LongSupplier sup = hasher.supplier(i);
for (int t=0;t<shape.getNumberOfHashFunctions();t++)
{
enableBit( Math.floorMod( sup.getAsLong(), shape.getNumberOfBits() );
}
}
Remove the requirement for a full length byte[]
The original Hash.Factory was a simple implementation to allow multiple
items in the Hash.
We could create a DataSink interface along the lines you are talking about.
Then have Hash.Factory extend that so that all those base data types can be
used as items in their own right and add a method with(DataSink) that would
take all the items in the DataSink and combine them into a single item for
the Hasher.
Create an add(T)/contains(T) method pair or Make a simple BloomFilter<T>
First I think this would be merge(T)/contains(T) pair. But I think it is a
bad idea to confuse the general BloomFilter responsibility with the Hasher
responsibility. I suggest leaving the general structure as defined above in
place and create a TypedBloomFilter<T> class that works much as you
suggest. We have to make a number of assumptions. For example we assume
that BitSetBloomFilter is the appropriate implementation and the
DynamicHasher the appropriate Hasher.
TypedBloomFilter<T> extends BitSetBloomFilter {
private Function<T,DataSink> fn;
TypedBloomFilter( Function<T,DataSink> fn, Shape shape ) {
super( shape );
this.fn = fn;
}
public void merge( T t ) {
merge( new DynamicHasher.Builder().with( fn.apply( t ) ) );
}
public boolean contains( T t ) {
return contains( new DynamicHasher.Builder().with( fn.apply( t ) ) );
}
}
Thread Safety
I think we note that Bloom filters are not generally safe for multithreaded
use without external synchronization.
Rework the HashFunction and HashFunctionIdentifier issues
We recognize that there are issues with HashFunctions and the identifier as
preserved in the Shape. I think that it is overly complex and prone to
error. Perhaps the proper solution is to remove the HasherIdentity from
Shape and create a method on Hasher String getID(). In most cases the
string would return the fully qualified class name, in cases where
difference hashing algorithms can be defined the string should return the
className prepended to the algorithm name or something similar. We can then
remove the HashFunction interface and all the stuff that derives from it.
Bloom filter equivalence will then depend upon Shape and Hasher equivalence.
On Fri, Mar 20, 2020 at 10:51 AM Alex Herbert <alex.d.herb...@gmail.com>
wrote:
> Sorry for an absence. I have been thinking on ways to move the BloomFilter
> API forward that consolidates the current functionality but makes it
> simpler to use for the common case.
>
> > On 18 Mar 2020, at 17:12, Claude Warren <cla...@xenei.com> wrote:
> >
> > bf.getBits() * Long.BYTES may be as long as Math.Ceil(
> > Shape.getNumberOfBits() / 8.0 ) or it may be shorter.
>
> I presume you mean:
>
> bf.getBits().length * Long.BYTES may be equal to Math.Ceil(
> Shape.getNumberOfBits() / 8.0 ) or it may be longer.
>
> >
> > I am building byte buffers of fixed length that is the maximum size that
> > any valid bf.getBits() * Long.BYTES I need to know
> > Shape.getNumberOfBytes().
> > The conversion is required for some Bloom filter indexing techniques.
>
> OK. I gave you an example of how to do this. In my example the
> Shape.getNumberOfBytes() was 1 line of code in 8.
>
> This method has a practical use if you are creating a byte[]
> representation of the bits in a filter. So you already have to be writing
> code to do that. In the context of this process it is not going to save you
> a lot of code. It seems like this function is very specific to your need
> and not something generic required within the API.
>
> >
> > And while serialization is outside the scope of the library, it is only
> > reasonable that we provide enough information to allow developers to
> > serialize/deserialse the data. For example BloomFilter allows you to get
> > either the long[] representation or the list of bit indexes (via OfInt)
> and
> > there are ways to reconstruct a BloomFilter if you were to write that out
> > and read it back.
>
> Yes and you are free to do data transfer how you please with that
> information.
>
> There is a lot of this API that is changing. I don’t want to keep adding
> methods before the basic functionality is fixed. So leave
> Shape.getNumberOfBytes() for now.
>
> On to the main topic for discussion ...
>
>
> IMO the basic functionality is not yet fixed. It is hard to use due to the
> separation of the BloomFilter from the hashing of objects you want to put
> into it. Thus at all times to use a BloomFilter you need to also have
> something that can create Hashers. Currently they are easy to create if you
> have a byte[]. So your primary use case of hashing Strings is easy. But
> what about other things? Here are my current issues with the design:
>
> 1. The decoupling of the hashing of an item from the BloomFilter, thus the
> filter does not control the number of indexes generated
> 2. The use of 32-bit integers for bit indexes
> 3. Hashing currently requires a full length byte[] to pass to a hash
> function
> 4. Thread safety
>
> In order:
>
> 1. The decoupling of the hashing of an item from the BloomFilter, thus the
> filter does not control the number of indexes generated
>
> I still think the API needs to have some simple way to use a BloomFilter
> that allows adding objects. Currently you cannot use a BloomFilter on its
> own. You can only use them if you have another BloomFilter or a Hasher. So
> you need to have the ability to create one of those. It is not very in the
> spirit of a collection. A collection should be able to interact directly
> with what it is a collection of. Currently the BloomFilter cannot. We have
> this situation:
>
> BloomFilter bf = BloomFilters.create(shape); // Construct somehow
> T t;
> // lots of code (maybe with ByteBuffer): T => byte[] bytes
> bf.merge(new DynamicHasher.Builder(bf.getShape()).with(bytes).build());
>
> The final merge operation also relies on the fact that the Hasher will
> create the correct number of indexes for the shape. A contract that it can
> easily violate and break the BloomFilter functionality.
>
> I am thinking about a better way to do this. I like the typed
> BloomFilter<T> approach in Guava. To me it makes more sense for a
> BloomFilter in commons-collections to be typed like all other collections.
> Having an untyped raw interface is strange for a generic collection
> framework. I think we need an abstraction layer between the raw processing
> of indexes done in the BloomFilter and the objects T that you put in.
>
> If you wish to maintain flexibility of defining Hashing then perhaps we
> extend the functionality in the Guava approach (which uses a fixed internal
> index generator) and allow the generator of indexes to be specified:
>
> interface IndexGenerator<T> {
> // infinite stream of 64-bit longs. Not yet mapped to a BloomFilter
> shape.
> LongSupplier generate(T item);
> }
>
> IndexGenerator<T> generator = ...;
> BloomFilter<T> bf = BloomFilter.create(generator, shape);
> T t;
> bf.add(t);
> bf.contains(t);
>
> This moves the hashing of the items inside the BloomFilter but allows the
> method to hash them to be configured as it is inside the IndexGenerator.
> This encapsulation prevents the filter state from being abused by adding
> hashers, which as discussed before should adhere to a contract to generate
> the correct number of indexes in the correct range for the number of bits
> but do not have to. If the BloomFilter is controlling the mapping of a
> stream of indexes to the appropriate range and number then this prevents
> state corruption.
>
> You previously mentioned separation of concerns as a reason to have a
> BloomFilter and the process for creation of a Hasher separate. This was to
> allow you to send a specialised Hasher over a network. Under this system
> you can create a BloomFilter<long[]> that stores a representation of a
> cyclic hash. The IndexGenerator<T> would be a IndexGenerator<long[]> that
> simply cycles the pair of longs to create indexes. Sending the pre-hashed
> long[2] over a network is simple. You thus set up a remote hash algorithm
> to convert your items to long[2] and send them for remote query.
>
> I’ll show an example later of how this approach could be incorporated on
> top of out current code base.
>
>
> 2. The use of 32-bit integers for bit indexes
>
> The lack of 64-bit indexes for the bit positions in a filter prevents
> construction of very large filters. This is a limitation as shown here:
>
> P = 1e-6
> M = 2^31, N = 74,681,641
> M= 2^31 * 64, N = 4,779,624,982
>
> So by using a full length long[] representation that could be addressed in
> memory from a long index divided by 64 allows you to increase capacity of
> an array base filter by 64-fold. This pushes it into a range where the
> number of items that can be stored is much larger than that fully supported
> by the JDK collections (which use 32-bit integers for size). It future
> proofs the API for expansion when RAM is cheap. The actual change to the
> code underneath is minimal. It would eliminate some data structures that we
> currently have e.g. for the CountingBloomFilter, but these could be marked
> to throw exceptions if the size is too big.
>
> It does make the API suitable for huge filters that can for example be
> used to check database keys exist before a call to a database.
>
> I suspect the use or 32-bit indexes was to allow the use of Set<Integer>
> and BitSet. We no longer use Set<Integer> for the counting bloom filter. A
> TreeSet<Long> would function just as well where we do use it. We can create
> a custom implementation of BitSet to allow long indexes. The BitSet has
> overhead due to its dynamic sizing that we can remove. A simple BitArray
> implementation would be easy to do.
>
>
> 3. Hashing currently requires a full length byte[] to pass to a hash
> function
>
> Hashing requires a preallocated byte[]. I would prefer to see hashing as a
> dynamic process where bytes are fed into an algorithm. For example
> MurmurHash3 32-bit algorithm only requires 4 bytes at a time. The 128-bit
> algorithm requires 2 longs. Currently the Objects32 hasher uses 1 byte at a
> time so could be totally online.
>
> The Guava code has the concept of a Hasher as a sink that can accepts data
> of any primitive type. In their implementation bytes are always processed
> in the input order as the code is also tied to their implementation of
> various hash algorithms so the output must be correct. However there is no
> reason that the bytes have to be processed in order. This allows for an
> optimisation where a DynamicHasher can process bytes as it chooses such as
> immediately process int/float and long/double input and for smaller input
> put those in a temp storage until 4 bytes have been received. The
> DynamicHasher need only ensure all bytes contribute to the final hash but
> can choose the order in which bytes are processed.
>
> I’ll show an example later of how this could be incorporated.
>
>
> 4. Thread safety
>
> Currently the hash functions are not thread safe. If the function is moved
> inside a BloomFilter then the architecture must exist to create the stream
> of indexes for each input object in a thread safe manner. Thus multiple
> threads would be able to query the same BloomFilter instance. If we create
> the correct design this feature will be naturally included. Obviously
> adding to a BloomFilter may not be concurrent (it would be for the standard
> implementation using a bit array) but if a filter is just a gateway filter
> which never changes then I think concurrent query should be expected to
> work.
>
>
> Here is a modification to the current code using an abstraction layer to
> convert items T to indexes. I was thinking about something like the
> following skeleton:
>
> // Represent a hash
> interface Hash {
> // Size of bits of the hash that is created
> int bits();
> // 64 bits of the hash (little-endian), starting from offset * 64.
> long getLongBits(int offset);
> }
>
> // Generate a Hash from a pre-existing byte[].
> // (This is modified from the current HashFunction.)
> interface StaticHashFunction {
> Hash apply(byte[], int seed);
> }
>
> // Generate indexes for an item.
> // Left to the BloomFilter to call this appropriately for its own shape.
> // The IndexGenerator would be constructed with:
> // - A system to convert T to data for hashing (e.g. Function<T, byte[]>)
> // - A system to convert data to a Hash (e.g. StaticHashFunction)
> // - Other details of how a hash is incremented, e.g. cyclic/iterative
> process type
> interface IndexGenerator<T> {
> // The supplier provides an indefinite stream
> LongSupplier generate(T item);
> }
>
> // Note:
> // Initial implementation of IndexGenerator would be
> // StaticIndexGenerator using StaticHashFunction with variants for
> iterative or cyclic.
> // In either case the user supplies the Function<T, byte[]> to create a
> byte[] from
> // the object. This is then used to create a Hash for each call of the
> LongSupplier
> // or using a cyclic process from a single call to create the base Hash.
>
> // Note: A PrehashedCyclicIndexGenerator<long[]> class can be used to just
> generate
> // the indexes directly using the first 2 longs in the long[] object using
> a cyclic function.
> // This allows constructing a BloomFilter<long[]> and storing pre-hashed
> items in it. The
> // hashed representation is simple to send over a network (it is only 2
> longs) allowing remote
> // hashing and query.
>
>
> // Typed to items
> interface BloomFilter<T> {
> boolean add(T item);
> }
>
>
> // To use it you require the function to convert objects to data
> Function<T, byte[]> converter = t -> t.getBytes();
>
> IndexGenerator<T> hasher = IndexGeneratores.create(converter,
> staticHashFunction, processType);
> BloomFilter<T> bf = new BloomFilter<>(hasher, shape);
>
> // or common use case with default hash function and process type
> BloomFilter<T> bf = BloomFilters.create(n, p, converter);
>
>
> // common API
> T t;
> bf.add(t);
> bf.contains(t);
>
>
> The following is an extension to allow dynamic hashing similar to the
> Guava API
>
> // Accepts primitive data
> interface DataSink {
> DataSink putByte(byte b);
> DataSink putBytes(byte[] b, int offset, int length);
> DataSink putChars(CharSequence cs, Charset encoding);
> DataSink putUnencodedChars(CharSequence cs);
> DataSink putInt(int i);
> // All other primitives
> }
>
> // Specify how item T is decomposed into primitive data
> interface Pipe<T> {
> void connect(T item, DataSink sink);
> }
>
> // Accepts data and dynamically hash it.
> interface Hasher extends DataSink {
> Hash build();
> }
>
> // Provides an implementation to create a Hash dynamically
> interface DynamicHashFunction {
> Hasher newHasher(int seed);
> // Size of bits of the hash that is created
> int bits();
> }
>
> // To use it you require the pipe implementation to convert objects to data
> Pipe<T> pipe = (t, sink) -> sink.putInt(t.getIntProperty());
>
> // explicit control over hash function and conversion of hashes to indexes
> IndexGenerator<T> hasher = IndexGeneratores.create(pipe,
> dyanmicHashFunction, processType);
> BloomFilter<T> bf = new BloomFilter<>(hasher, shape);
>
> // or common use case with default hash function and process type
> BloomFilter<T> bf = BloomFilters.create(n, p, pipe);
>
>
>
> With various factory methods to use defaults for a hash function and
> process type and a few implementations of Pipe for common things (e.g.
> String, Integer).
>
> There are lots of holes to fill in. But it allows you to directly add
> items T to a BloomFilter<T> and control the hash function and the iterative
> process.
>
> One main issue is how to deal with filter compatibility. We can compare
> filter properties k and m easily. But how to compare the IndexGenerator<T>
> used by the filter? IMO the current HashFunctionIdentity is subject to
> error. No system is perfect. Perhaps we go with the simple option of using
> Object.equals(Object) to test equality of the IndexGenerator. In the common
> use case you will have the same object for your IndexGenerator<T> when
> merging two filters. The equals will be fast. All the implementations we
> provide can override equals to make this efficient using an internal
> signature. The exact method can be private until we have the rest of the
> framework complete.
>
> There are issues. There is a lot of object creation. I’d like to minimise
> this. But I don’t see how you can have a flexible API without object
> creation. For example you could make the BloomFilter convert the Hash to
> indexes and avoid the index generator object. But then you lose the ability
> to use an item of type long[] as a prehashed item for conversion with a
> customised index generator.
>
> If this is not the best way then I still think the API needs to have some
> simple way to use a BloomFilter that allows adding objects. Currently you
> cannot use a BloomFilter on its own. You can only use them if you have
> another BloomFilter or a Hasher. So you need to have the ability to create
> one of those. It is not very in the spirit of a collection. A collection
> should be able to interact directly with what it is a collection of.
> Currently the BloomFilter cannot. So this suggestion is to make the
> conversion of objects to data for hashing part of the API.
>
> Note also that here we drop the ability to query using a Hasher build with
> multiple objects as it is covered by query using a BloomFilter<T> as a
> collection of T.
>
> These are just thoughts. But I’d like you opinion on whether we can make
> this work for all the scenarios you are applying this to. I think it would
> work for your common use cases of storing String and the remote filter
> query. The goal would be to make a BloomFilter<T> like a fuzzy Set<T> (may
> or may not contain T) and just as easy to use. Otherwise I think you risk
> making the API too complex for the common use case. That would be a user
> wants to store items T and knows the fields in T that uniquely define the
> item. They just build a function to create a byte[] representation of the
> object or a pipe to send data from T to a downstream receiver that will
> collate the data. The reset of the process to create indexes from the item
> data is provided by the framework.
>
> Alex
>
>
>
> ---------------------------------------------------------------------
> To unsubscribe, e-mail: dev-unsubscr...@commons.apache.org
> For additional commands, e-mail: dev-h...@commons.apache.org
>
>
--
I like: Like Like - The likeliest place on the web
<http://like-like.xenei.com>
LinkedIn: http://www.linkedin.com/in/claudewarren
