ContextQMD
Back to Ruview

ADR-083: Per-Cluster Pi Compute Hop

docs/adr/ADR-083-per-cluster-pi-compute-hop.md

1.99.0-pip12.8 KB
Original Source

ADR-083: Per-Cluster Pi Compute Hop

FieldValue
StatusProposed — pending field evidence on three-tier proposal scope
Date2026-04-26
Authorsruv
Supersedes—
RefinesADR-028 (capability audit), ADR-081 (5-layer kernel), ADR-066 (swarm bridge)
Companiondocs/research/architecture/three-tier-rust-node.md, docs/research/architecture/decision-tree.md, docs/research/sota/2026-Q2-rf-sensing-and-edge-rust.md

Context

ADR-028 established the per-node BOM at ~$9 (ESP32-S3 8MB) — ~$15 with a mmWave sensor — and ADR-081 framed the firmware as a 5-layer adaptive kernel running entirely on a single ESP32-S3 die. Both decisions are correct for the per-node dimension; deployments that fit the "sensor talks UDP to a server somewhere" shape work fine on this stack.

The three-tier-node research exploration (docs/research/architecture/three-tier-rust-node.md) raised a separate question: what changes when a deployment scales past one or two rooms, and where should the heavy compute live? The exploration's answer ("dual ESP32-S3 + Pi Zero 2W per node") is one shape, but the companion decision-tree (decision-tree.md §1, §3 L3, §5) identifies a materially cheaper path: keep today's single-S3 sensor node unchanged and add one Pi per cluster of 3–6 sensor nodes. The 2026-Q2 SOTA survey (sota/2026-Q2-rf-sensing-and-edge-rust.md) confirms that the load this path needs to carry — model inference, QUIC backhaul, and a real secure-boot story — fits comfortably on a Pi-class SoC, while the load it doesn't need to carry — CSI capture, ISR-precise wake control — is exactly what the ESP32-S3 already does well.

The three things this ADR is about, all of which the current single-S3 deployment shape pushes onto the cloud or onto every individual node:

  1. Per-deployment ML inference. WiFlow / DT-Pose / GraphPose-Fi class models (4–10M params, 0.5–1.5 GFLOPs) want a Cortex-A53-class target. The ESP32-S3 cannot host these; the cloud can but only at the cost of round-trip latency. A per-cluster Pi inference hop is the natural home.
  2. QUIC backhaul. quinn + rustls is mature on Linux but does not run on ESP32-class hardware in any production-grade form (SOTA §5). A Pi terminating QUIC for a cluster gives every sensor node QUIC's loss/handoff/multiplex properties without porting QUIC to the MCU.
  3. Secure-boot anchor for OTA. ESP-IDF Secure Boot V2 covers each sensor node, but cluster-wide policy (which model is current, which sensor MCU image is canary, what is the rollout ring) needs a higher-trust local store. A Pi running buildroot + dm-verity + signed FIT is a defensible anchor without the BOM hit of CM4 / Pi 5 (the latter is its own decision; see ADR-085 sketch below and decision-tree.md L6).

The cluster-Pi shape does not require any change to ADR-028 or ADR-081. The sensor node continues to be a single-MCU ESP32-S3 running the 5-layer kernel. Everything new lives at the cluster boundary.

Decision

Adopt a per-cluster Pi hop as the canonical RuView mid-scale deployment shape. A "cluster" is 3–6 ESP32-S3 sensor nodes within WiFi mesh range of one Pi.

Specifically:

  1. Sensor nodes are unchanged. They continue to run the ADR-081 5-layer kernel on a single ESP32-S3, emit rv_feature_state_t packets (60 byte, ~5 Hz, ~300 B/s) over UDP, and connect via ESP-WIFI-MESH or direct WiFi to the cluster Pi.
  2. Each cluster has exactly one Pi acting as:
    • Sensor aggregator: ingests UDP from all cluster sensor nodes, runs feature-level fusion (multistatic + viewpoint attention from the existing wifi-densepose-ruvector crate).
    • ML inference target: hosts the WiFi-pose model and runs inference at the cluster boundary, not on each sensor MCU.
    • QUIC client to the cloud / gateway: terminates QUIC mTLS, batches cluster-level events.
    • OTA + secure-boot anchor for its sensor nodes: holds signed manifests, stages canary rollouts, owns provisioning state.
  3. Cluster Pi SoC choice is deferred to a future ADR (sketched below as ADR-085). The acceptable candidates are Pi Zero 2W, Pi 4, Pi 5, and CM4. The decision tree's L6 distinguishes these by secure-boot threat model; this ADR does not pre-commit.
  4. The single-node deployment shape is not deprecated. A home-lab / single-room / development deployment can still run a single ESP32-S3 talking UDP directly to the existing wifi-densepose-sensing-server, no Pi required. The cluster Pi becomes the recommended shape for fleets ≥ 3 sensor nodes.

Boundary contract

The cluster Pi exposes two interfaces:

InterfaceDirectionSchema
UDP rv_feature_state_t ingestsensor → PiExisting 60-byte packed struct from ADR-081 (magic 0xC5110006)
QUIC mTLS uplinkPi → gateway/cloudNew: cluster-level event envelope (CBOR), batched, ~10 KB/min upper bound

Sensor → Pi is the same wire as today's sensor → server. Cluster Pi uplink is new and is what the existing wifi-densepose-sensing-server becomes — relocated from the user's laptop / container to the cluster node. Concretely: the sensing server already exists in crates/wifi-densepose-sensing-server; it cross-compiles to ARMv7 / AArch64 today via cargo build --target aarch64-unknown-linux-gnu. The relocation is a deployment change, not a re-implementation.

Three-tier vs cluster hop

This ADR's cluster-Pi shape is the L3-hybrid path in decision-tree.md §2 — not the full three-tier (dual-MCU + per-node Pi) shape. It captures most of the value (ML, QUIC, secure-boot anchor) at minimal BOM impact. The full three-tier shape remains the long-term exploration target, blocked behind L4 (no_std CSI maturity) and L2 (per-node ISR-jitter evidence).

Consequences

Positive

  • Pose-grade ML on edge becomes deployable, not just possible. A Pi (any of the eligible SoCs) hosts WiFlow-class models with ≤ 100 ms latency per cluster, vs ≥ 1 s round-trip if pose runs in the cloud (SOTA §1, §3).
  • QUIC arrives without an MCU port. quinn + rustls runs on the Pi as it does on a server (SOTA §5). The sensor MCU keeps UDP — the cheapest, highest-tested wire it already speaks.
  • Cluster-level secure boot becomes coherent. Per-sensor Secure Boot V2 + flash encryption (ADR-028 baseline) is unchanged. The Pi buildroot + dm-verity image is the cluster trust anchor and signs the OTA manifests for its sensors. The cluster-level threat model is expressible without per-sensor BOM regression.
  • No PCB respin. Sensor nodes are bit-for-bit identical to today's ADR-028 baseline. The cluster Pi is a separate device on the cluster WiFi (and / or Ethernet, if available).
  • Deployment cost scales sub-linearly with sensor count. One $25–$60 Pi per 3–6 sensor nodes adds ~$5–$20 per sensor amortized, vs ~$25–$50 per sensor for the per-node-Pi shape.

Negative

  • The cluster Pi is a new piece of infrastructure to provision, monitor, and update. It is the right place for cluster-level responsibilities, but it is not free; it adds a Linux box to every multi-room deployment. Mitigated by buildroot images and the existing OTA tooling story (see Implementation §4).
  • Cluster Pi failure takes the cluster offline (sensor nodes cannot uplink without a working aggregator on the WiFi LAN). For high-availability deployments, this ADR is the floor; an HA-pair cluster Pi would be a follow-up.
  • One more network hop on the sensing path. Sensor → Pi → cloud adds ~5–20 ms over Sensor → cloud (depending on link quality). Pose latency budgets are 100s of ms, so this is well inside spec.

Neutral

  • ADR-028 (capability audit), ADR-081 (5-layer kernel), and ADR-066 (swarm bridge) are unchanged. This ADR adds a new device class above the sensor; it does not modify the sensor itself.
  • The home-lab single-node shape continues to work; this ADR adds a recommended path for fleets, it does not deprecate the existing one.

Implementation

The implementation is intentionally light because most of the pieces already exist; the ADR is largely about formalizing where they live.

  1. Cluster-Pi cross-compile target. Add to rust-port/wifi-densepose-rs/.cargo/config.toml (or the equivalent per-crate target spec) an aarch64-unknown-linux-gnu target so wifi-densepose-sensing-server builds for Pi 4 / 5 / CM4 by default. Also retain armv7-unknown-linux-gnueabihf for Pi Zero 2W compatibility while the Pi-SoC decision (ADR-085 sketch) is open.
  2. Cluster-Pi service unit. Add a systemd unit file under firmware/cluster-pi/ (new directory) that runs wifi-densepose-sensing-server with the cluster's UDP/QUIC ports and drops privileges. Buildroot integration is a separate ADR if the SoC choice goes to Pi Zero 2W (where there's no RPi-OS path).
  3. QUIC uplink module. Add wifi-densepose-sensing-server a feature-gated quic-uplink module using quinn + rustls. The feature is off by default in the home-lab shape and on for the cluster Pi.
  4. OTA + signed-manifest flow. Out of scope for this ADR; tracked as I4 in decision-tree.md §4. The cluster Pi's role is to hold the manifest store, not to define the manifest format. Use the existing ADR-066 swarm bridge channel for OTA staging.
  5. Documentation update. README's hardware-table gains a "Cluster compute" row. CLAUDE.md gets a one-paragraph cluster-Pi section under Architecture. User-guide gets a cluster-deployment section.
  6. Validation. A 3-sensor cluster + 1 Pi fixture in the lab. Pass criteria: end-to-end CSI → cluster fusion → cloud ingest; measured latency under 100 ms per cluster; cluster Pi reboot without sensor data loss > 5 s; OTA staging round-trip across all sensors in the cluster.

Validation

This ADR is proposed, not accepted. Acceptance requires:

  1. The cluster-Pi wifi-densepose-sensing-server cross-compiles cleanly on aarch64-unknown-linux-gnu and armv7-unknown-linux-gnueabihf targets with the existing workspace tests passing.
  2. A 3-sensor + 1-Pi field test demonstrates ≥ 4 hours stable end-to-end CSI → fusion → cloud round-trip with latency ≤ 100 ms per cluster and zero phantom-skeleton regressions (ADR-082 holds across the new uplink).
  3. The cluster-Pi ↔ sensor secure-boot story is approved alongside ADR-085's SoC choice.

When the above pass, this ADR moves from Proposed → Accepted and the README + CLAUDE.md are updated to reflect cluster-Pi as the recommended fleet-shape.

Related ADRs (current and proposed)

  • ADR-028 (Accepted) — ESP32 capability audit. Single-node BOM baseline. Unchanged by this ADR.
  • ADR-029 (Proposed) — RuvSense multistatic sensing mode. Pairs naturally with cluster-Pi: cluster Pi is the natural home for multi-sensor fusion.
  • ADR-066 — Swarm bridge to coordinator. The cluster-Pi is the per-cluster swarm coordinator endpoint.
  • ADR-081 (Accepted) — 5-layer adaptive CSI mesh firmware kernel. Unchanged by this ADR.
  • ADR-082 (Accepted) — Pose tracker confirmed-track output filter. Holds across UDP and QUIC uplinks identically.
  • Future ADR (sketched in decision-tree.md L4) — no_std CSI capture maturity benchmark. Gates the dual-MCU shape; not required for the cluster-Pi shape proposed here.
  • Future ADR (sketched in decision-tree.md L6) — Cluster-Pi SoC choice (Pi Zero 2W vs CM4 vs Pi 5). Pure secure-boot decision.

Open questions

  • Cluster size sweet spot. "3–6 nodes" is a planning estimate. The 3-sensor lab fixture in §Implementation will inform whether the upper bound is closer to 4, 6, or 8 in practice.
  • Cluster-Pi failure semantics. Default behavior: sensor MCUs hold the last 60 s of feature packets in RAM and replay on reconnect. HA-pair cluster Pi is a separate ADR if needed.
  • Mesh control-plane interaction. If the deployment moves to Thread (decision-tree.md L5), the cluster Pi may need a Thread Border Router role. This ADR doesn't pre-commit; it's compatible with both ESP-WIFI-MESH and Thread futures.