Design and Performance Evaluation of a CRDT-Based Framework for Conflict-Free Real-Time Collaborative Web Applications

Shared documents, whiteboards, design canvases and planning boards all rest on the same promise: several people can change the same thing at the same time, over flaky connections, and nobody ever ends up looking at a different version from everyone else. Most production tools keep that promise with operational transformation (OT), where a central server rewrites every concurrent edit before it is applied. That works, but it makes the server the bottleneck and turns offline work into a special case. Conflict-free replicated data types (CRDTs) take the server out of the correctness story, since every replica converges on its own, yet the textbook versions carry a reputation for bloated metadata, tombstones that never go away and full-state exchanges on every reconnect. In this paper we describe a browser-oriented CRDT framework built to shed those costs, and we measure how far it succeeds. The framework pairs a run-length encoded replicated growable array for text with an observed-remove map, a last-writer-wins register and a counter for structured state, all sharing one identifier scheme. Replicas synchronise by exchanging version-vector digests and then only the deltas the other side lacks, over WebSocket or WebRTC, with a causal buffer for out-of-order arrivals. A persistence layer snapshots the log and collects tombstones once a stable version vector shows that no replica can still need them. On an emulated testbed with up to 64 clients we compared the framework with a Jupiter-style OT server, a server-authoritative last-writer-wins baseline and a stripped-down per-character CRDT. With 16 clients on an 80 ms link, remote edits converged in 94 ms at the median and 146 ms at the 99th percentile, compared with 121 ms and 243 ms for OT, and the relay carried 41,000 operations per second. After 500,000 edits the stored document was 1.9 times the size of its visible text, where the naive CRDT reached 38 times and retained OT history 27 times. A client returning from a ten-minute outage was fully reconciled in 171 ms after exchanging 104 KB. No divergence appeared across 1.2 million randomised test interleavings. We close with practical guidance for teams choosing a collaboration engine.