crdt

crdt

English | 简体中文

crdt is a small Go library for composable state-based CRDTs. It provides deterministic binary state and delta frames so replicas can converge despite duplicate delivery, reordering, and temporary partitions.

Status: stable releases are published from main; APIs follow semantic versioning.

Architecture at a glance

This is a conceptual comparison of integration boundaries, not a wire-compatibility, feature-parity, or performance claim.

Features

• State-based G-Counter with joinable, type-isolated deltas.
• Grow-only G-Set with a caller-defined element codec and joinable deltas.
• Add-wins observed-remove OR-Set with a caller-defined element codec.
• Experimental delta-replicated LWW-Set with a caller-defined element codec and deterministic HLC conflict resolution.
• Experimental delta-replicated LWW-Map with opaque byte values and deterministic HLC conflict resolution.
• Experimental attachment references for images, audio, video, and data: bounded metadata only, backed by an authenticated application object store.
• Causally replicated MV-Register that preserves concurrent opaque-byte writes instead of resolving them by wall clock.
• Hybrid logical clock (HLC) tags and a persistable clock state for replica restarts.
• Canonical, checksummed binary frames with bounded decoding and deterministic encoding.
• Delta batching/coalescing, versioned snapshots, and Merkle digests for anti-entropy workflows.
• Optional exact-acknowledgement tombstone collection with membership epochs.
• Safe concurrent access for the provided CRDT implementations.
• Opt-in manifest-bound WebSocket and HTTP/SSE live-relay reference surfaces; disabled unless an application explicitly enables them.
• A single-writer durable WebSocket relay reference with a bbolt operation log, exact-dot binding, bounded replay, and reconnect support.
• Experimental framed LWW-Set, LWW-Map, legacy scalar RGA v1, and OR-Tree collections, enabled only by an explicit per-replication-group protocol policy. New Go RGA groups use compact run-v2 frames; the RGA tombstone lifecycle still requires careful retention and exact-acknowledgement handling.

Scope

The core library provides CRDT data types and wire primitives. It deliberately does not choose a membership protocol, authentication scheme, storage backend, or retry policy. The optional extensions package adds explicitly enabled WebSocket and HTTP/SSE live-relay reference surfaces, but it does not start a listener or provide durability, replay, reconnect, TLS, anti-entropy, or identity/session management. Those remain application-owned. The optional WebSocket provider reference is likewise a bounded, manifest-bound integration adapter; it does not provide durable delivery, recovery, TLS, membership, authorization policy, or production operations. The separate durable reference adds one persistent operation log and reconnect/replay flow for a single process and one protected storage volume. It still does not provide clustered storage, application CRDT checkpoint transactions, TLS, identity/session lifecycle, membership, or tombstone GC. tombstonegc.Coordinator performs safe automatic collection only after the application supplies an authoritative, authenticated active-membership view. It does not discover, authenticate, or persist that view. A checksum detects accidental frame corruption; it is not an authenticity or encryption mechanism.

Experimental LWW-Set, LWW-Map, legacy RGA v1, and OR-Tree protocols

LWW-Set (lww.Set, TypeIDs 7/8) encodes generic elements through an application-supplied canonical lww.ElementCodec. It retains remove metadata, so persist SnapshotCurrentState(codec) (or Snapshot(codec, frontier)) and restore a same-ID replica only with NewSetFromSnapshot. Its new wire format is experimental and must be explicitly negotiated.

RGA text v1 (text, TypeIDs 11/12) accepts out-of-order deltas through a bounded delayed-integration queue, rejects incomplete snapshots, and uses an incremental indexed sequence rather than rebuilding the full visible projection after each edit. It remains experimental while its full tombstone lifecycle is validated. Persist its HLC-backed snapshot atomically.

New Go RGA groups select compact run-v2 frames (TypeIDs 19/20) through crdt.DefaultRGAFrameType(). A run-v2 group must encode with Delta.MarshalRunBinary, use RGA.MarshalRunBinary and RGA.SnapshotRunCurrentState for complete state, and bind those same IDs in its manifest. The run encoding preserves scalar RGA position semantics, but a manifest still represents exactly one wire protocol: do not mix it with a legacy v1 client or frame stream.

CompactTombstones is intentionally conservative: it can collect only deleted leaves after an authenticated exact-acknowledgement epoch has durably saved a post-compaction snapshot and retired old deltas. Nodes with descendants remain structural anchors. LWW-Set, LWW-Map, legacy scalar RGA v1 (TypeIDs 11/12), and OR-Tree remain experimental and require explicit opt-in:

policy := crdt.ProtocolPolicy{AllowExperimental: true}
for _, kind := range policy.FrameTypes() {
	// This is only a local capability allowlist, not the full handshake.
	_ = kind
}

Before an experimental frame is accepted, bind it to a replica.Manifest in the authenticated handshake. The manifest includes the group, schema, epoch, codec, and semantics version. To keep one explicit opt-in while constructing a group's local replica objects, create a replica.NewSessionBuilder(..., policy) and use its NewChange, NewInbox, NewCheckpoint, and NewSession methods. The builder remains a local helper, not a handshake: the zero policy still rejects experimental manifests, and frame type IDs alone do not establish wire-semantic compatibility. The individual *WithPolicy constructors remain available for isolated use.

The zero-value policy advertises G-Counter, G-Set, OR-Set, MV-Register, PN-Counter, and default RGA run-v2 protocols. The policy is neither a global switch nor a plugin registry: unknown and reserved frame types remain unsupported. Experimental LWW-Set, LWW-Map, legacy RGA v1, and OR-Tree replicas must persist HLC state with snapshots and retain their tombstones.

Browser and JavaScript mobile clients

The repository includes a bounded TypeScript frame-envelope decoder and a Go/ Wasm RGA client runtime under clients/typescript. The default artifact uses run-v2 TypeIDs 19/20 and semantics version 2, matching crdt.DefaultRGAFrameType(). The Wasm layer reuses the Go RGA merge, out-of-order, tombstone, and HLC implementation so a browser/WebView can merge locally instead of waiting for server arbitration.

Build and exercise the client boundary with:

make wasm
make typescript-test
make wasm-test

Authenticate the exact manifest (including protocol IDs and semantic version) before passing a received frame to the runtime; CRC-32C does not authenticate a peer. Persist the client snapshot's state, clock, and frontier atomically. Native mobile apps without a compatible WebAssembly runtime must follow the RGA run-v2 wire protocol and its vectors before joining a run-v2 group. Split a local editor transaction above 64 KiB or 16,384 runes before insertion.

Experimental attachment references

attachment.Register represents a document's images, audio, video, or other binary data as a bounded LWW-Map of immutable references. A reference contains an opaque object ID, canonical MIME type, declared byte length, and SHA-256 digest; it never carries media bytes in a CRDT delta, snapshot, log, or diagnostic. Text that users edit remains text.RGA; ordinary structured data remains lww.Map, OR-Set, or OR-Tree according to its conflict semantics.

Attachment references use the experimental LWW-Map frame IDs (9/10). Bind each replication group to a replica.Manifest with schema ID github.com/DarkInno/crdt/attachment-reference/v1, an empty codec ID, and attachment.SemanticsVersion; enable AllowExperimental on every boundary. Persist SnapshotCurrentState() with its HLC state, and retain delete metadata until the LWW tombstone lifecycle is complete.

The application owns authorization, object-store lifecycle, content scanning, rate limits, and download policy. After a fetch, call Reference.Verify before decoding or rendering: it streams the object without buffering it and rejects a short, oversized, or digest-mismatched response. Do not put signed URLs, credentials, personal data, or raw media content in Reference.ObjectID.

Requirements

• Go 1.21 or later

Install

Install the latest stable release:

go get github.com/DarkInno/crdt@latest

For development, use a local checkout:

git clone https://github.com/DarkInno/crdt.git
cd crdt
go test ./...

Quick start

G-Counter

Each replica increments only its own component. Merge takes the per-replica maximum, so it is commutative, associative, and idempotent.

package main

import (
	"fmt"
	"log"

	"github.com/DarkInno/crdt/counter"
)

func main() {
	left, err := counter.NewGCounter("left")
	if err != nil {
		log.Fatal(err)
	}
	right, err := counter.NewGCounter("right")
	if err != nil {
		log.Fatal(err)
	}

	if _, err := left.Increment(2); err != nil {
		log.Fatal(err)
	}
	if _, err := right.Increment(3); err != nil {
		log.Fatal(err)
	}
	if err := left.Merge(right); err != nil { // Delivery order does not matter.
		log.Fatal(err)
	}

	value, err := left.Value()
	if err != nil {
		log.Fatal(err)
	}
	fmt.Println(value)
	// Output: 5
}

PN-Counter

A PN-Counter supports independent increments and decrements. It stores positive and negative per-replica G-Counter components, so merges remain commutative, associative, and idempotent. Value returns an exact *big.Int; use ValueInt64 when the application requires a bounded machine integer.

counter, err := counter.NewPNCounter("cart")
if err != nil {
	log.Fatal(err)
}
if _, err := counter.Increment(7); err != nil {
	log.Fatal(err)
}
if _, err := counter.Decrement(2); err != nil {
	log.Fatal(err)
}
value, err := counter.Value()
if err != nil {
	log.Fatal(err)
}
fmt.Println(value)
// Output: 5

OR-Set delta replication

An OR-Set uses a stable codec ID and stable encoded element bytes to identify a set's element type across replicas.

package main

import (
	"fmt"
	"log"

	"github.com/DarkInno/crdt/set"
)

type stringCodec struct{}

func (stringCodec) ID() string                            { return "example.com/string/v1" }
func (stringCodec) Marshal(value string) ([]byte, error)  { return []byte(value), nil }
func (stringCodec) Unmarshal(data []byte) (string, error) { return string(data), nil }

func main() {
	codec := stringCodec{}
	left, err := set.NewORSet("left", codec)
	if err != nil {
		log.Fatal(err)
	}
	right, err := set.NewORSet("right", codec)
	if err != nil {
		log.Fatal(err)
	}

	delta, err := left.Add("item")
	if err != nil {
		log.Fatal(err)
	}
	if err := right.ApplyDelta(delta); err != nil {
		log.Fatal(err)
	}

	fmt.Println(right.Contains("item"))
	// Output: true
}

For a remove to be observed by other replicas, send the returned remove delta or merge the state. An add concurrent with a remove that did not observe its tag remains present (add-wins semantics).

End-to-end integration

For a reproducible local HTTP delivery exercise, a production-integration checklist, snapshot/restart guidance, and the expected convergence evidence, see the end-to-end integration tutorial. The runnable collaborative-workboard example models duplicate delivery, a partitioned add/remove conflict, and recovery from an OR-Set snapshot:

go run ./examples/collaborative-board

The warehouse replication example shows framed G-Set and MV-Register deltas, duplicate delivery, concurrent register values, and safe MV-Register recovery before reusing a replica ID:

go run ./examples/warehouse-replication

The experimental collaboration example uses low, explicit receive and RGA retention limits for LWW-Map, RGA, and OR-Tree. Run it only after the replication group has completed the authenticated experimental-protocol handshake described above:

go run ./examples/experimental-collaboration

The attachment collaboration example uses separate manifest-bound groups for RGA text and attachment references, persists both receiver states through snapshots, and streams an authorized download through Reference.Verify before accepting it:

go run ./examples/attachment-collaboration

The opt-in transport extension guide and its runnable provider example show one application-owned mux exposing both WebSocket and HTTP/SSE surfaces. The example demonstrates a WebSocket-to-HTTP delivery followed by an HTTP-to- WebSocket delivery; it is a bounded live relay, not a durable replication service:

go run ./examples/extensions-provider

For persistent operation replay and reconnect behavior, use the separate durable WebSocket relay reference. It is a single-writer bbolt deployment shape with bounded replay rather than a replacement for a replicated database or application checkpoint transaction.

See attachment reference integration for the manifest fields, limits, storage boundary, deletion retention, and verification requirements.

For the Chinese versions, see 集成教程 and 协作任务示例 and 仓库复制示例 and 实验协作示例 and 可选传输扩展. 另见可持久化 WebSocket relay 参考实现。

Correct use in a distributed system

• Give every live logical replica a globally unique, non-blank replica ID.
• Persist an OR-Set snapshot atomically with its HLC state. Use ORSet.SnapshotCurrentState() when the set's own frontier is sufficient, or ORSet.Snapshot(frontier) when the replication layer has a broader acknowledgement frontier; restore with NewORSetFromSnapshot. Do not restore a same-ID OR-Set from bytes alone.
• For automatic tombstone collection, create a coordinator with a stable replication-group ID. Each active member reports its exact ORSet.TombstoneTags() (or experimental ORTree.TombstoneTags()) under that ID and the current tombstonegc.Coordinator membership epoch; pass both values to AcknowledgeAndCompact (or AcknowledgeAndCompactTarget for the tree) for every received report. The tree target additionally refuses to remove a tombstoned node with any known child. Do not derive acknowledgements from Frontier() when delta delivery can be out of order: a maximum tag does not prove that prior tombstones were received. Removing a member requires retiring it from replication; a rejoining member must bootstrap from a post-compaction snapshot. Persist that checkpoint and bind every new frame to the next membership epoch before accepting compaction.
• ORSet.Compact remains available only for transports that independently prove a gap-free causal prefix for every supplied frontier.
• Persist an MV-Register state snapshot before reusing its replica ID. Its version vector, not a wall clock, proves which writes a later Set observes; recover with register.NewMVRegisterFromSnapshot.
• Use ProtocolPolicy.FrameTypes() as an authenticated connection/setup capability advertisement. Send LWW-Set, LWW-Map, RGA, or OR-Tree frames only when both peers opt in. Persist HLC-backed snapshots atomically. RGA tombstone compaction additionally requires an authenticated exact-acknowledgement epoch, durable post-compaction checkpoint, and retirement of old deltas.
• Keep ElementCodec.ID, Marshal, and Unmarshal deterministic and safe for concurrent calls. Encoded values must round-trip canonically.
• Treat received bytes as untrusted. Use UnmarshalBinaryWithLimits and Unmarshal*DeltaWithLimits with limits appropriate to the transport.
• Authenticate, authorize, encrypt, retry, and persist messages in the surrounding application. CRDT convergence does not provide those guarantees.

JSON diagnostics

Concrete CRDT state and delta objects implement json.Marshaler for structured logs and human inspection. The output is a compact, stable summary such as:

{"type":"gcounter","replica_id":"left","element_count":2,"tombstone_count":0}

It deliberately excludes application values, element keys, tags, clock state, and binary frames. JSON diagnostics cannot restore or apply a CRDT state or delta and are not a replication format; use the bounded canonical binary encoders for that.

Packages

Development and verification

go test ./...
go test -race ./...
go vet ./...
make coverage

make verify additionally runs fuzzing, staticcheck, and golangci-lint. Those two tools must be installed on PATH locally; GitHub Actions installs pinned versions. To reproduce the coverage gate in Docker:

make docker-test

Diagnostic and synchronization probes

crdt-analyze verifies one bounded frame before emitting JSON metadata (type, codec, payload size, and SHA-256 fingerprint):

go run ./cmd/crdt-analyze -file ./state.frame

crdt-sync-probe is a short-lived HTTP test utility for exercising duplicate delta delivery across hosts. It is not a production replication service. Its default listener is loopback-only; a non-empty token is required for every endpoint. Prefer -token-file (mode 0600) over -token, and bind a public address only for a controlled test window.

# On each receiver.
go run ./cmd/crdt-sync-probe -mode serve -replica receiver -token-file ./probe.token

# Generate one delta and send that same byte sequence to every target.
go run ./cmd/crdt-sync-probe -mode send \
  -target http://receiver-a:49511,http://receiver-b:49511 \
  -replica sender -token-file ./probe.token -duplicates 3

Use make test-unit to run packages independently and make test-integration for the three-replica, recovery, batching, encoding, and anti-entropy flow.

The CI workflow enforces formatting, unit tests, race detection, vet, decoder fuzzing, static analysis, per-package coverage of at least 90%, and a Go 1.26 container verification.

Quality and performance snapshot — 2026-07-28

This historical pre-release snapshot was collected on Go 1.26.5. It is recorded evidence for that revision, not a latency or throughput guarantee for every workload; the checks below were not re-executed as part of this documentation update.

• The recorded make verify run passed: formatting, independent-package tests, integration and extreme scenarios, the race detector, vet, four 10-second decoder fuzz campaigns, staticcheck, golangci-lint, and a per-package coverage gate of at least 90%.
• The recorded make docker-test run passed with Go 1.26; govulncheck ./... found no known vulnerabilities.
• A controlled three-host delivery probe confirmed idempotent duplicate delivery and rejected unauthorized, malformed, and oversized requests.
• The supplied benchmarks cover G-Counter, PN-Counter, G-Set, OR-Set, and MV-Register Merge, ApplyDelta, and MarshalBinary. Run make benchmark on your target hardware before choosing capacity limits.

The scenario evaluation at the end of this README records the current local sample for duplicate delivery and both live-state and tombstone-heavy serialization. Exact results depend on the CPU, Go version, element codec, set size, and mutation mix.

Run make test-extreme to repeat the high-cardinality scenario in normal and race-instrumented modes. Internal investigation data and deployment runbooks are intentionally kept outside the public release tree.

Publishing releases

Before merging beta into main, run the verification commands above, review the public API, and ensure the repository is publicly reachable at the module path. The main push workflow creates the next immutable semantic-version tag and a GitHub Release with generated notes. Do not create a competing stable tag by hand.

go mod tidy
make verify
# Merge the reviewed beta -> main pull request, then wait for the main workflow.
GOPROXY=proxy.golang.org go list -m github.com/DarkInno/crdt@latest

Do not move or reuse a published tag. For Go modules, breaking changes after the first stable release require a new major-version module path such as github.com/DarkInno/crdt/v2.

Contributing

Please open an issue before proposing an API expansion. Contributions should include focused tests, preserve deterministic wire encoding, keep untrusted input bounded, and pass the verification commands above.

Scenario performance evaluation — 2026-07-28

Measured locally on an Apple M4 Pro with Go 1.26.5 (darwin/arm64). The fixture contains 128 string elements; each result is the rounded mean of three two-second samples. Allocation values were stable across samples.

Merge, ordinary ApplyDelta, and MarshalBinary use serial benchmark loops; their GOMAXPROCS=4 values are runtime-setting samples, not four-core throughput measurements. Only the parallel duplicate-delivery row uses RunParallel. Compared with an earlier local sample using the same live-state fixture and method, MarshalBinary now uses 29,952 B and 132 allocations per operation, down from 96,312 B and 778 allocations. These are local pre-release measurements, not capacity planning or SLA guarantees; rerun make benchmark on the deployment target before setting limits.

PN-Counter performance evaluation — 2026-07-28

Measured on two independent Debian 13 (linux/amd64) hosts, each with four Intel Xeon Platinum 8272CL vCPUs and 3.8 GiB memory. The benchmark binary was built from this revision with Go 1.26.5 and run three times per setting with -benchtime=2s; values below are rounded means. The fixture has 128 replica components in each positive and negative map. MarshalBinary includes its reported encoded-throughput sample in parentheses; allocation counts were identical in all three runs.

The GOMAXPROCS=4 rows remain serial benchmark measurements, not aggregate four-core throughput. These controlled host samples are public regression evidence for this revision, not capacity limits or SLA guarantees; rerun the same command on the deployment target before setting production limits:

GOMAXPROCS=4 go test -run='^$' \
  -bench='^BenchmarkPNCounter(Merge|ApplyDelta|Value|MarshalBinary)$' \
  -benchmem -benchtime=2s ./counter

License

SPDX-License-Identifier: MIT

Licensed under the MIT License. Copyright (c) 2026 DarkInno.