Skip to content

Edge State Consistency

Problem Statement

Edge agents have local context; cloud agents have remote context. How do we keep them consistent when they diverge? When should state be shared vs local?

Solution Approach

State Classification:

Type Location Sync Strategy
Conversation history Edge (primary) + Cloud (backup) Async sync every 5 turns
User profile Cloud (primary) Fetch on-demand, cache for 1 hour
Intermediate results Edge (temporary) Sync after decision
Decisions Both (immutable log) Sync immediately

Sync Strategies:

  1. Eventual consistency — history eventually syncs, delays acceptable
  2. Use: conversation history

  3. Policy: sync every N turns or on timeout

  4. Causal consistency — one agent sends result, other receives before processing

  5. Use: decision dependencies (B depends on A's result)

  6. Policy: explicit acknowledgment required

  7. Strong consistency — all agents see same state immediately

  8. Use: financial transactions, critical decisions
  9. Policy: two-phase commit or quorum

Implementation (Eventual Consistency Example):

# Edge agent
conversation = [turn1, turn2, ...]
checkpoint_id = "conv:alice:12345:turn:10"
queue.send("sync_checkpoint", {
  "checkpoint_id": checkpoint_id,
  "history": conversation,
  "timestamp": time.time()
})

# Cloud agent (async)
checkpoint = queue.recv("sync_checkpoint", timeout=60s)
if checkpoint:
  # Update cloud history if newer
  if checkpoint.timestamp > cloud_history.timestamp:
    cloud_history = checkpoint.history

When to Use

  • Use eventual consistency for conversation history (delays acceptable)
  • Use causal consistency for decision dependencies
  • Use strong consistency for critical transactions (payments)

Trade-offs

Model Consistency Latency Complexity
Eventual Weak (lags) Low (no wait) Low
Causal Strong (ordered) Medium (wait for ack) Medium
Strong Perfect High (blocking) High

Recommendation: Eventual for history; causal for decisions; strong for payments.

References