Skip to content

Consistency Model

Eventual Consistency Design

Constraint: Vectorize operates under an eventually consistent model. Solution: Queue-based writes on retreival ( for temporal data as it depends on access frequency ) maintain consistency and minimize search latency by updating temporal features on access. We provide PMQ, a free and open-source queue implementation for Cloudflare Workers.

Note: We do not have access to Cloudflare’s low-level APIs, nor are we affiliated with them. However, this solution works well within the limits of Cloudflare’s free plan.

The design ensures that you can experiment with MeridianDB without barriers no need to pay for Cloudflare Workers to get started.

Our Enterprise Edition uses Cloudflare Queues by default for maximum scalibilty.

Write Pattern

Store → Queue → [Process] → Vectorize + D1
                    ↓
           [Retrieve] ← Consistent State

Redundant Storage

  • Vectorize: Embeddings + agentId metadata
  • D1: Full multidimensional features with memory content
  • KV: Counters, session state.
  • We currently only support text.

Queue Implementation

Temporal Features Queue

  • Updates recency/frequency on access on memories we implemented our queue using cloudflare workers, with Write-ahead log and buffer-flush.
  • Uses PMQ (Poor Man's Queue) for free tier.
  • Configurable processing frequency.

Behavioral Endpoint

  • Processes user feedback (success/failure).
  • Updates success rate metrics on all involved memory based on user feedback.
  • Independent of retrieval process simple webhook ( think like dislike generation on ChatGPT).

Trade-offs

  • ✅ Scalable, cost-effective
  • ✅ Automatic retries/failover
  • ❌ Reads may lag writes
  • ❌ May Requires cleanup jobs ( the algorithm only filter out decayed, bad behavioral score data ).