Skip to content

A convenient data structure supporting efficient in-memory indexing and querying, including range queries and fuzzy string matching.

License

Notifications You must be signed in to change notification settings

khemichew/Akade.IndexedSet

 
 

Repository files navigation

Akade.IndexedSet

.Net Version CI Build NuGet version (Akade.IndexedSet) MIT Static Badge

A convenient data structure supporting efficient in-memory indexing and querying, including range queries and fuzzy string matching. In a nutshell, it allows you to write LINQ-like queries without enumerating through the entire list. If you are currently completely enumerating through your data, expect huge speedups and much better scalability!

Overview

A sample showing different queries as you might want do for a report:

// typically, you would query this from the db
var data = new Purchase[] {
        new(Id: 1, ProductId: 1, Amount: 1, UnitPrice: 5),
        new(Id: 2, ProductId: 1, Amount: 2, UnitPrice: 5),
        new(Id: 6, ProductId: 4, Amount: 3, UnitPrice: 12),
        new(Id: 7, ProductId: 4, Amount: 8, UnitPrice: 10) // discounted price
        };

IndexedSet<int, Purchase> set = data.ToIndexedSet(x => x.Id)
                                    .WithIndex(x => x.ProductId)
                                    .WithRangeIndex(x => x.Amount)
                                    .WithRangeIndex(x => x.UnitPrice)
                                    .WithRangeIndex(x => x.Amount * x.UnitPrice)
                                    .WithIndex(x => (x.ProductId, x.UnitPrice))
                                    .Build();

// efficient queries on configured indices
// in contrast to standard LINQ, they do not enumerate the entire list!
_ = set.Where(x => x.ProductId, 4);
_ = set.Range(x => x.Amount, 1, 3, inclusiveStart: true, inclusiveEnd: true); 
_ = set.GreaterThanOrEqual(x => x.UnitPrice, 10);
_ = set.MaxBy(x => x.Amount * x.UnitPrice);
_ = set.Where(x => (x.ProductId, x.UnitPrice), (4, 10));

Design Goals

  • Much faster solution than (naive) LINQ-based full-enumeration
  • Syntax close to LINQ-Queries
  • Easy to use with a fluent builder API
  • Reflection & Expression-free to be AOT & Trimming friendly (for example for Blazor/WebASM)
  • It's not a db - in-memory only

Performance and Operation-Support of the different indices:

Below, you find runtime complexities. Benchmarks can be found here

General queries

  • n: total number of elements
  • m: number of elements in the return set
  • ✔: Supported
  • ⚠: Supported but throws if not exactly 1 item was found
  • ❌: Not-supported
Query Unique-Index NonUnique-Index Range-Index
Single ⚠ O(1) ⚠ O(1) ⚠ O(log n)
Where ✔ O(1) ✔ O(m) ✔ O(log n + m)
Range ✔ O(log n + m)
< / <= ✔ O(log n + m)
> / >= ✔ O(log n + m)
OrderBy ✔ O(m)
Max/Min ✔ O(1)

String queries

  • w: length of query word
  • D: maximum distance in fuzzy query
  • r: number of items in result set
Query Prefix-Index FullText-Index
StartWith ✔ O(w+r) ✔ O(w+r)
Contains ✔ O(w+r)
Fuzzy StartWith ✔ O(w+D+r) ✔ O(w+D+r)
Fuzzy Contains ✔ O(w+D+r)

ℹ FullText indices use a lot more memory than prefix indices and are more expensive to construct. Only use FullText indices if you really require it.

Features

Unique index

Dictionary-based, O(1), access on keys:

IndexedSet<int, Data> set = IndexedSetBuilder<Data>.Create(a => a.PrimaryKey)
                                                   .WithUniqueIndex(x => x.SecondaryKey)
                                                   .Build();

_ = set.Add(new(PrimaryKey: 1, SecondaryKey: 5));

// fast access via primary key
Data data = set[1];

// fast access via secondary key
data = set.Single(x => x.SecondaryKey, 5);

ℹ Entities do not require a primary key. IndexedSet<TPrimaryKey, TData> inherits from IndexedSet<TData> but provides convenient access to the automatically added unique index: set[primaryKey] instead of set.Single(x => x.PrimaryKey, primaryKey).

Non-unique index

Dictionary-based, O(1), access on keys (single value) with multiple values (multiple keys):

IndexedSet<int, Data> set = new Data[] { new(PrimaryKey: 1, SecondaryKey: 5), new(PrimaryKey: 2, SecondaryKey: 5) }
        .ToIndexedSet(x => x.PrimaryKey)
        .WithIndex(x => x.SecondaryKey)
        .Build();

// fast access via secondary key
IEnumerable<Data> data = set.Where(x => x.SecondaryKey, 5);

Range index

Binary-heap based O(log(n)) access for range based, smaller than (or equals) or bigger than (or equals) and orderby queries. Also useful to do paging sorted on exactly one index.

IndexedSet<Data> set = IndexedSetBuilder.Create(new Data[] { new(1, SecondaryKey: 3), new(2, SecondaryKey: 4) })
                                        .WithRangeIndex(x => x.SecondaryKey)
                                        .Build();

// fast access via range query
IEnumerable<Data> data = set.Range(x => x.SecondaryKey, 1, 5);

// fast max & min key value or elements
int maxKey = set.Max(x => x.SecondaryKey);
data = set.MaxBy(x => x.SecondaryKey);

// fast larger or smaller than
data = set.LessThan(x => x.SecondaryKey, 4);

// fast ordering & paging
data = set.OrderBy(x => x.SecondaryKey, skip: 10).Take(10); // second page of 10 elements

String indices and fuzzy matching

Prefix- & Suffix-Trie based indices for efficient StartWith & String-Contains queries including support for fuzzy matching.

IndexedSet<Type> data = typeof(object).Assembly.GetTypes()
                                               .ToIndexedSet()
                                               .WithPrefixIndex(x => x.Name)
                                               .WithFullTextIndex(x => x.FullName)
                                               .Build();

// fast prefix or contains queries via indices
_ = data.StartsWith(x => x.Name, "Int");
_ = data.Contains(x => x.FullName, "Int");

// fuzzy searching is supported by prefix and full text indices
// the following will also match "String"
_ = data.FuzzyStartsWith(x => x.Name, "Strang", 1);
_ = data.FuzzyContains(x => x.FullName, "Strang", 1);

Multi-key indices: All indices can be used with multiple keys

There are overloads for all indices that allow to use multiple keys.

You can have a unique index where each element can have multiple keys:

IndexedSet<int, Data> set = IndexedSetBuilder<Data>.Create(a => a.PrimaryKey)
												   .WithUniqueIndex(x => x.AlternativeKeys) // Where AlternativeKeys returns an IEnumerable<int>
												   .Build();

_ = set.Add(new(PrimaryKey: 1, AlternativeKeys: new[] { 3, 4 }));
set.Single(x => x.AlternativeKeys, 3); // returns above element

The same applies for all other index types, for example for non-unique indices:

IndexedSet<int, GraphNode> set = IndexedSetBuilder<GraphNode>.Create(a => a.Id)
                                                             .WithIndex(x => x.ConnectsTo) // Where ConnectsTo returns an IEnumerable<int>
                                                             .Build();

//   1   2
//   |\ /
//   | 3
//    \|
//     4

_ = set.Add(new(Id: 1, ConnectsTo: new[] { 3, 4 }));
_ = set.Add(new(Id: 2, ConnectsTo: new[] { 3 }));
_ = set.Add(new(Id: 3, ConnectsTo: new[] { 1, 2, 3 }));
_ = set.Add(new(Id: 4, ConnectsTo: new[] { 1, 3 }));

// For readability, it is recommended to write the name for the parameter contains
IEnumerable<GraphNode> nodesThatConnectTo1 = set.Where(x => x.ConnectsTo, contains: 1); // returns nodes 3 & 4
IEnumerable<GraphNode> nodesThatConnectTo3 = set.Where(x => x.ConnectsTo, contains: 1); // returns nodes 1 & 2 & 3

// Non-optimized Where(x => x.Contains(...)) query:
nodesThatConnectTo1 = set.FullScan().Where(x => x.ConnectsTo.Contains(1)); // returns nodes 3 & 4, but enumerates through the entire set

ℹ️ For range queries, this introduces a small overhead as the results are filtered to be distinct: i.e. O(log n + m log m) instead of O(log n + m).

ℹ️ Multi-key string indices are marked experimental. Read more at Experimental Features

Computed or compound key

The data structure also allows to use computed or compound keys:

var data = new RangeData[] { new(Start: 2, End: 10) };
IndexedSet<RangeData> set = data.ToIndexedSet()
                                .WithIndex(x => (x.Start, x.End))
                                .WithIndex(x => x.End - x.Start)
                                .WithIndex(ComputedKey.SomeStaticMethod)
                                .Build();
// fast access via indices
IEnumerable<RangeData> result = set.Where(x => (x.Start, x.End), (2, 10));
result = set.Where(x => x.End - x.Start, 8);
result = set.Where(ComputedKey.SomeStaticMethod, 42);

ℹ For more samples, take a look at the unit tests.

Concurrency and Thread-Safety

The "normal" indexedset is not thread-safe, however, a ReaderWriterLock-based implementation is available. Just call BuildConcurrent() instead of Build():

ConcurrentIndexedSet<RangeData> set = data.ToIndexedSet()
                                          .WithIndex(x => (x.Start, x.End))
                                          .BuildConcurrent();

⚠ The concurrent implementation needs to materialize all query results.
OrderBy and OrderByDescending take an additional count parameter to avoid unnecessary materialization. You can judge the overhead here

No reflection and no expressions - convention-based index naming

We are using the CallerArgumentExpression-Feature of .Net 6/C# 10 to provide convention-based naming of the indices:

  • set.Where(x => (x.Prop1, x.Prop2), (1, 2)) tries to use an index named "x => (x.Prop1, x.Prop2)"
  • set.Where(ComputedKeys.NumberOfDays, 5) tries to use an index named "ComputedKeys.NumberOfDays"
  • **Hence, be careful what you pass in.

ℹ️ The following naming conventions are recommended:

  • Use x as parameter name in any lambdas that determines an index name.
  • Do not use parentheses in any lambda that determines an index name.
  • Do not use block bodied in any lambda that determines an index name.
  • For complex indices, use a static method. C# Analyzers are shipped with the package to spot incorrect index names.

Reasons

  • Simple and yet effective:
    • Allows computed, compound, custom values etc. to be indexed without adding complexity...
  • Performance: No reflection at work and no (runtime) code-gen necessary
  • AOT-friendly including full trimming support

FAQs

How do I use multiple index types for the same property?

Use "named" indices by using static methods:

record Data(int PrimaryKey, int SecondaryKey);

IndexedSet<int, Data> set = IndexedSetBuilder<Data>.Create(x => x.PrimaryKey)
                                                   .WithUniqueIndex(DataIndices.UniqueIndex)
                                                   .WithRangeIndex(x => x.SecondaryKey)
                                                   .Build();
_ = set.Add(new(1, 4));
// querying unique index:
Data data = set.Single(DataIndices.UniqueIndex, 4); // Uses the unique index
Data data2 = set.Single(x => x.SecondaryKey, 4); // Uses the range index
IEnumerable<Data> inRange = set.Range(x => x.SecondaryKey, 1, 10); // Uses the range index

ℹ We recommend using the lambda syntax for "simple" properties and static methods for more complicated ones. It's easy to read, resembles "normal" LINQ-Queries and all the magic strings are compiler generated.

How do I update key values if the elements are already in the set?

The implementation requires any keys of any type to never change the value while the instance is within the set. You can manually remove, update and add an object. However, there are some helper methods for that - which is especially useful for the concurrent variant as it provides thread-safe serialized access.

// updating a mutable property
_ = set.Update(dataElement, e => e.MutableProperty = 7);
// updating an immutable property
_ = set.Update(dataElement, e => e with { SecondaryKey = 12 });
// be careful: the dataElement still refers to the "old" record after the update method
_ = set.Update(dataElement, e => e with { SecondaryKey = 12 });

// updating in an concurrent set
concurrentSet.Update(set =>
{
    // serialized access to the inner IndexedSet, where you can safely use above update methods
    // in an multi-threaded environment
});

How do I do case-insensitve (fuzzy) string matching (Prefix, FullTextIndex)?

Remember that you can index whatever you want, including computed properties. This also applies for fuzzy matching:

IndexedSet<Data> set = IndexedSetBuilder<Data>.Create(x => x.PrimaryKey)
                                              .WithFullTextIndex(x => x.Text.ToLowerInvariant())
                                              .Build();
IEnumerable<Data> matches = set.FuzzyContains(x => x.Text.ToLowerInvariant(), "Search", maxDistance: 2);

Roadmap

Potential features (not ordered):

  • Thread-safe version
  • Easier updating of keys
  • More index types (Trie)
  • Range insertion and corresponding .ToIndexedSet().WithIndex(x => ...).[...].Build()
  • Refactoring to allow a primarykey-less set: this was an artificial restriction that is not necessary
  • Benchmarks
  • Simplification of string indices, i.e. Span/String based overloads to avoid AsMemory()...
  • Analyzers to help with best practices
  • Multi-key everything: All index types can be used with multiple keys per element.
  • Tree-based range index for better insertion performance
  • Aggregates (i.e. sum or average: interface based on state & add/removal state update functions)
  • Custom (equality) comparer for indices
  • Helper functions for search scenarios (Searching in multiple properties, text-reprocessing & result merging)
  • Becnhmark vs elastic search

If you have any suggestion or found a bug / unexpected behavior, open an issue! I will also review PRs and integrate them if they fit the project.

About

A convenient data structure supporting efficient in-memory indexing and querying, including range queries and fuzzy string matching.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C# 99.9%
  • PowerShell 0.1%