HINT Protocol Bidirectional Hints for Predictive Caching
Tell your cache what you need. It learns what works.
A transport-agnostic protocol for closed-loop predictive caching. Clients declare intent, servers prefetch proactively, feedback drives continuous adaptation.
The Cache Feedback Gap
HTTP/2 Server Push promised proactive delivery. CDNs prefetch assets they think you'll need. GraphQL dataloaders batch requests. But none of them learn.
When a server pushes data you never use, it doesn't know. When a prefetcher wastes bandwidth, there's no feedback. The cache is blind to outcomes.
HINT closes the loop. Clients declare what they anticipate needing. Servers tag proactive payloads with prediction identifiers. Clients report what actually happened. The system adapts.
How It Works
Three messages. One closed loop. Continuous learning.
Hint
Client declares anticipated fields, scope, and TTL
Proactive Payload
Server prefetches and tags with prediction_id
Feedback
Client reports outcome: HIT, MISS, EVICTED_UNUSED
The Closed Loop
The prediction_id correlates every proactive payload with its eventual outcome. This enables selective reinforcement (prefetch more of what works) and suppression (stop prefetching what doesn't).
Protocol Specification
Transport-agnostic message formats for HTTP, GraphQL, and gRPC.
Hint Message
{ "version": "1.0", "request_id": "req-abc-123", "scope": { "tenant": "app-123", "namespace": "users" }, "ttl": 30000, "anticipated_fields": [ "user.profile", "user.preferences" ] }
Feedback Signal
{ "version": "1.0", "client_id": "client-xyz", "outcomes": [{ "prediction_id": "pred-456", "outcome": "HIT", "timestamp": 1704153600, "details": { "latency_saved_ms": 45 } }] }
Proactive Payload
{ "version": "1.0", "prediction_id": "pred-456", "request_id": "req-abc-123", "prefetched": [ { "field": "user.profile", "value": { ... }, "ttl": 30000, "confidence": 0.87 } ], "budget_remaining": { "bytes": 102400, "items": 8 } }
Transport Bindings
Headers
Request:
PCE-Hint: {base64}
PCE-Feedback: {base64}
Response:
PCE-Proactive: {base64}
Extensions field
Request:
extensions.pce.hint
extensions.pce.feedback
Response:
extensions.pce.proactive
Metadata keys
Request:
pce-hint: {binary}
pce-feedback: {binary}
Response:
pce-proactive: {binary}
Key Concepts
Scope Isolation
Hints include a scope definition (tenant, namespace, resources) that bounds where proactive effects apply. Multi-tenant safe by design.
Budget Gating
Proactive plans are subject to budgets: concurrency limits, byte-read quotas, CPU time, energy, and cost constraints. Prefetch with guardrails.
Negative Prefetch
Hints can explicitly exclude fields from prefetching. "Don't fetch user.password" is as important as "do fetch user.profile."
Idempotency
Each hint includes a request_id for deduplication. Replays with the same ID are safely ignored. Network retries don't cause duplicate prefetches.
Use Cases
GraphQL N+1 Mitigation
Client hints anticipated nested fields before query execution. Server prefetches related entities in parallel. Feedback teaches which patterns recur.
Kubernetes Control Plane
API server hints to etcd about anticipated dependent resources during RAFT replication. Scheduler and controllers benefit from pre-warmed caches.
Edge/CDN Prefetching
Multi-tier hint propagation from client to edge to origin. Each tier refines hints based on local context. Feedback flows back for global optimization.
ML Inference Caching
Pre-warm GPU memory with anticipated embeddings or model inputs. Eliminate PCIe round-trips for hot paths. Feedback identifies which prefetches reduce latency.
HINT-Native in Voxell Products
HINT is an open standard. Voxell products implement it natively.
GPU-Native Key-Value Cache
HINT-driven prefetch into HBM for ML inference pipelines.
Cache-Resident Rate Limiting
HINT-driven token bucket pre-computation for burst prediction.
HINT is transport-agnostic and implementation-independent. Build your own PCE or use Voxell's.
Get Involved
HINT is an open standard. Read the spec, join the discussion, or implement it yourself.
HINT Protocol is released under the Apache License 2.0
GitHub repository coming soon: voxell-ai/hint-spec