Skip to content
Research2SVII Internal Work

Offline Synchronisation for Intermittently Connected Applications

A conflict resolution strategy for mobile clients that spend long periods offline, evaluated against three months of field traces.

Authors
Sidney Okine2SVII, co-founderSamuel Okine2SVII, co-founder
Published
Updated
Reading time
3 min

Research summary

Problem
Applications used on intermittent mobile networks accumulate large divergent local histories. Last-write-wins discards work that the user believes was saved, and full operational transformation is heavier than most product teams can justify maintaining.
Objective
Establish whether a bounded, per-field merge strategy can resolve the majority of real divergence without server-side operation replay, and quantify what it fails to resolve.
Methodology
We instrumented a production field-data application across 41 devices for 94 days, recording every local mutation and its eventual server reconciliation. Divergence events were replayed offline against three strategies: last-write-wins, per-field last-write-wins with vector clocks, and a state-based CRDT baseline.
Findings
  1. Per-field resolution with vector clocks resolved 96.2% of observed divergence events without user intervention, against 71.4% for whole-record last-write-wins.
  2. The residual 3.8% clustered almost entirely in list reordering, which per-field resolution cannot express.
  3. Metadata cost was 180 bytes per record at the observed replica count, roughly 11% of mean record size.
Limitations
  1. A single application domain with 41 devices. Divergence patterns are workload-specific and these results should not be assumed to transfer.
  2. Median offline period was 4.2 hours. Behaviour beyond roughly 72 hours offline is not characterised.
  3. The CRDT baseline was implemented for this evaluation and is not a tuned production implementation, so its overhead figures are an upper bound.

Applications that assume connectivity behave badly without it. On networks where a session may drop for hours, the client accumulates a local history that the server has never seen, and the reconciliation strategy chosen at design time determines whether that work survives.

The prevailing default is last-write-wins on the whole record. It is simple to implement and it silently destroys data: two users editing unrelated fields of the same record will see one edit vanish. The alternative usually proposed is a CRDT [shapiro-2011], which is correct but carries a maintenance and storage cost that small teams frequently underestimate.

Method

We instrumented an existing field-data collection application already deployed on intermittent networks. Every local mutation was recorded with a vector clock [fidge-1988] and a monotonic device counter, then shipped alongside the mutation when the device next reconnected.

type FieldVersion<T> = {
  value: T;
  clock: VectorClock;
  deviceId: string;
};

function mergeField<T>(local: FieldVersion<T>, remote: FieldVersion<T>): FieldVersion<T> {
  const ordering = compare(local.clock, remote.clock);

  if (ordering === "descends") return local;
  if (ordering === "ascends") return remote;

  // Concurrent. Neither clock dominates, so there is no causal answer and any
  // choice here is arbitrary. Ordering by device id at least makes it the same
  // arbitrary choice on every replica, which is what convergence requires.
  return local.deviceId > remote.deviceId ? local : remote;
}
typescriptThe per-field merge, reduced to its essential case.

Results

Divergence events resolved without user intervention, by strategy (n = 2,914).
StrategyResolvedMetadata / recordEscalated to user
Whole-record last-write-wins71.4%24 B0%
Per-field + vector clocks96.2%180 B3.8%
State-based CRDT baseline99.1%1.4 kB0.9%

The CRDT baseline resolves more, as expected. The question is whether the remaining 2.9 percentage points justify roughly eight times the metadata and a substantially larger implementation surface. For the workload measured here we judged that they did not, and the residual list-reordering cases were escalated to an explicit user prompt instead.

Reproducing this

The anonymised trace set and the replay harness are published alongside this record. Device identifiers are salted per-device and the field payloads are omitted; the harness replays divergence structure only.

References

  1. [shapiro-2011]Shapiro, M., Preguiça, N., Baquero, C., Zawirski, M. (2011). Conflict-free Replicated Data Types. Stabilization, Safety, and Security of Distributed Systems, pp. 386–400.
  2. [fidge-1988]Fidge, C. J. (1988). Timestamps in message-passing systems that preserve the partial ordering. Proceedings of the 11th Australian Computer Science Conference, pp. 56–66.

Related