Production-grade streaming data sketching algorithms for Elixir.
ExDataSketch provides probabilistic data structures for approximate counting,
frequency estimation, quantile computation, heavy-hitter detection, membership
testing with deletion, and set reconciliation on streaming data. Stream-native
integration with Elixir's Collectable, GenStage, Broadway, and Flow,
plus :telemetry/OpenTelemetry instrumentation, persistence backends (ETS,
DETS, CubDB, Mnesia, Ecto), and 23 production-oriented Livebooks (7
cross-cutting integration guides plus a per-family tutorial for every
sketch).
| Algorithm | Purpose | Status |
|---|---|---|
| HyperLogLog (HLL) | Cardinality estimation | Implemented (Pure + Rust) |
| Count-Min Sketch (CMS) | Frequency estimation | Implemented (Pure + Rust) |
| Theta Sketch | Set operations on cardinalities | Implemented (Pure + Rust) |
| KLL Quantiles | Rank and quantile estimation | Implemented (Pure + Rust) |
| DDSketch | Relative-error quantile estimation | Implemented (Pure + Rust) |
| FrequentItems (SpaceSaving) | Heavy-hitter / top-k detection | Implemented (Pure + Rust) |
| Bloom Filter | Probabilistic membership testing | Implemented (Pure + Rust) |
| Cuckoo Filter | Membership testing with deletion | Implemented (Pure + Rust) |
| Quotient Filter | Membership with deletion and merge | Implemented (Pure + Rust) |
| CQF (Counting Quotient) | Multiset membership with counting | Implemented (Pure + Rust) |
| XorFilter | Static immutable membership testing | Implemented (Pure + Rust) |
| IBLT | Set reconciliation | Implemented (Pure + Rust) |
| REQ Sketch | Relative-error quantile estimation | Implemented (Pure) |
| Misra-Gries | Deterministic heavy-hitter detection | Implemented (Pure) |
| UltraLogLog (ULL) | Improved cardinality estimation | Implemented (Pure + Rust) |
| Structure | insert | delete | merge | count | serialize | static | reconciliation |
|---|---|---|---|---|---|---|---|
| Bloom | yes | -- | yes | yes* | yes | -- | -- |
| Cuckoo | yes | yes | -- | yes | yes | -- | -- |
| Quotient | yes | yes | yes | yes | yes | -- | -- |
| CQF | yes | yes | yes | yes | yes | -- | -- |
| XorFilter | -- | -- | -- | yes | yes | yes | -- |
| IBLT | yes | yes | yes | yes | yes | -- | yes |
*Bloom count is a popcount-based cardinality estimate.
- Bloom -- Default choice for membership testing. No deletion needed, mergeable.
- Cuckoo -- Need to delete items. Better space efficiency than Bloom at low FPR.
- Quotient -- Need deletion and merge. Good when filter resizing may be needed.
- CQF -- Need to count how many times each item was inserted (multiset).
- XorFilter -- Static set known at build time. Smallest footprint, fastest lookups.
- IBLT -- Need to find differences between two sets (reconciliation).
Add ex_data_sketch to your list of dependencies in mix.exs:
def deps do
[
{:ex_data_sketch, "~> 0.10.1"}
]
end# HLL: count distinct elements
hll = ExDataSketch.HLL.new() |> ExDataSketch.HLL.update_many(1..100_000)
ExDataSketch.HLL.estimate(hll) # ~100_000
# Stream API: build from lazy enumerables
sketch = 1..100_000 |> Stream.map(&to_string/1) |> ExDataSketch.Stream.hll(p: 14)
# Collectable: Enum.into works with any mergeable sketch
sketch = Enum.into(1..50_000, ExDataSketch.ULL.new(p: 14))
# Partitioned parallel reduction
sketch = ExDataSketch.Stream.reduce_partitioned(1..1_000_000, ExDataSketch.HLL, partitions: 8, p: 14)
# Persistence: save and load from ETS
:ets.new(:sketches, [:set, :public, :named_table])
ExDataSketch.Storage.ETS.save(sketch, :sketches, "daily:2024-01-15")
{:ok, loaded} = ExDataSketch.Storage.ETS.load(ExDataSketch.HLL, :sketches, "daily:2024-01-15")
# KLL: quantile estimation
kll = ExDataSketch.KLL.new() |> ExDataSketch.KLL.update_many(1..100_000)
ExDataSketch.KLL.quantile(kll, 0.5) # approximate median (~50_000)
ExDataSketch.KLL.quantile(kll, 0.99) # 99th percentile (~99_000)
# Bloom: membership testing
bloom = ExDataSketch.Bloom.new(capacity: 100_000)
bloom = ExDataSketch.Bloom.put_many(bloom, 1..50_000)
ExDataSketch.Bloom.member?(bloom, 42) # true
ExDataSketch.Bloom.member?(bloom, 99_999) # false (probably)See the Quick Start Guide for more examples.
23 production-oriented Livebooks demonstrate real-world patterns: a per-family tutorial for all 16 sketches (each generating and caching its own sample data), plus 7 cross-cutting guides covering stream consumption, distributed merge semantics, framework integration, and a 1-billion-row-style case study. See Livebooks Guide for the recommended reading order and what each Livebook teaches.
Full documentation is available at HexDocs.
- Binary state: All sketch state is canonical Elixir binaries. No opaque NIF resources.
- Backend system: Pure Elixir reference implementation with optional Rust NIF acceleration. The Rust backend falls back to Pure automatically when unavailable.
- Streaming native:
ExDataSketch.Stream,Collectable, andfrom_enumerable/2for lazy stream consumption. Broadway, GenStage, and Flow integration for production pipelines. - Persistence: ETS, DETS, CubDB, Mnesia, and Ecto backends with atomic
save/load/merge/deleteoperations. All storage uses EXSK v2 binary format with CRC32C checksum. - Observability: Structured
:telemetryevents at batch boundaries, optional OpenTelemetry span bridge, category-based enable/disable. - Serialization: ExDataSketch-native format (EXSK) for all sketches, plus Apache DataSketches CompactSketch interop for Theta. v1 format escape hatch for backward compatibility.
- Deterministic hashing: Stable 64-bit hash (
ExDataSketch.Hash) for reproducible results. - Backend parity: Both backends produce byte-identical serialized output for the same inputs.
The following guarantees apply within the v0.x release series:
- EXSK serialization: The ExDataSketch-native binary format is stable. Binaries produced by any v0.x release can be deserialized by any other v0.x release, with one exception: see the ULL v1/v2 state-format note below.
- Pure vs Rust parity: Given identical inputs, both backends produce byte-identical serialized state and identical estimates.
- Deterministic output: The same input sequence always produces the same sketch state and estimate, regardless of backend.
- Backward compatibility: v0.7.x EXSK v1 frames remain decodable.
HLL.serialize(sketch, format: :v1)(and, as of v0.10.0, every other family exceptFilterChain) produces backward-compatible v0.7.x-style output (requires:phash2hash strategy where applicable).
Not guaranteed:
- ULL binary format (v0.10.1+):
ExDataSketch.ULL's register encoding and estimator were rewritten in v0.10.1 to fix a correctness bug (the previous implementation was an HLL-derived approximation, not real UltraLogLog, and overestimated cardinality by orders of magnitude once a sketch's registers filled up). Its internalULL1state format version was bumped 1 -> 2 as part of the fix; sketches serialized by v0.10.0 or earlier fail to decode on v0.10.1+ with a clear error instead of being silently misread. Rebuild any persisted ULL sketches from source data after upgrading. No other family is affected. - ULL estimate stability: beyond the v0.10.1 binary-format break above, ULL estimates also shifted at low cardinalities between v0.8.0 and v0.9.0 due to an earlier accuracy correction.
- Cross-language interop:
ThetaandKLL(as of v0.10.0) support Apache DataSketches compact binary formats. HLL and CMS DataSketches interop is not implemented. - Performance stability: Benchmark results may vary across hardware and OTP versions.
- EXSK format across major versions: The binary format may change in future major releases.
# Get dependencies
mix deps.get
# Run tests with coverage
mix test --cover
# Run lints
mix lint
# Run benchmarks
mix bench
# Generate docs
mix docs| Version | Focus | Status |
|---|---|---|
| v0.1.0 | Core sketches (HLL, CMS, Theta) + Rust NIFs | Released |
| v0.2.0 | KLL quantiles | Released |
| v0.2.1 | DDSketch relative-error quantiles | Released |
| v0.3.0 | FrequentItems (SpaceSaving) | Released |
| v0.4.0 | Bloom filter (membership testing) | Released |
| v0.5.0 | Advanced membership filters (Cuckoo, Quotient, CQF, XorFilter, IBLT, FilterChain) | Released |
| v0.6.0 | REQ sketch, Misra-Gries, XXHash3, Rust NIF parity for all membership filters | Released |
| v0.7.0 | ULL (UltraLogLog) -- improved cardinality estimation with Pure + Rust NIF | Released |
| v0.7.1 | NIF batch hashing, hash customization, quotient filter fix, merge safety | Released |
| v0.8.0 | Deterministic Foundations -- pluggable hash registry (XXHash3 + Murmur3), binary stability and corruption detection, HLL hot-path optimization, precompiled NIFs, property-based validation | Released |
| v0.9.0 | Streaming Integrations -- Stream/Collectable API, Broadway/GenStage/Flow integration, persistence (ETS/DETS/CubDB/Mnesia/Ecto), telemetry + OpenTelemetry, ULL accuracy fix, v1 serialization escape hatch | Released |
| v0.10.0 | Production Ergonomics -- unified sketch contract & facade dispatch, storage behaviour, windowing, supervised sketches (Server/Sketches), Telemetry.Metrics + LiveDashboard, filter NIF raw-hashing, Apache KLL interop, v1 serialization escape hatch for every family | Released |
| v0.10.1 | Correctness and polish -- ULL rewritten to the real UltraLogLog algorithm (was an HLL-derived approximation; binary format bumped v1->v2), KLL compaction weight-invariant fix, HLL precision range widened to p=4..26, 16 new per-family tutorial Livebooks, phoenix_demo/ sample app, plus the original post-review fixes (Server graceful-shutdown snapshotting, storage merge crash-safety, filter :hash_strategy build/round-trip fix, hardened opencode.yml workflow) |
Released |
| v0.11.0 | Apache HLL Interoperability & New Sketch Families -- full cross-language HLL exchange, CPC (Compressed Probabilistic Counting), Tuple Sketch (weighted distinct counting) | Planned |
| v0.12.0 | Similarity & Sampling -- MinHash, Weighted MinHash, VarOpt sampling | Planned |
| v1.0.0 | Stable Binary Contract -- locked EXSK format, full benchmark suite, Nx / Arrow ecosystem integrations | Planned |
MIT License. See LICENSE for details.