Edge & Redis locality

How cache GETs stay close to the request — and why that matters once agent and CI traffic start to spike.

Product narrative: Locality & edge. Architecture context: Architecture.

Exact-replay only pays when the GET is near the work. If every identical retry re-crosses a cold path for inventory that already exists, you have rebuilt the waste the pipe was meant to remove. The edge exists so HIT traffic can stay short: Redis first, meters and receipts through the control-plane gate, lab silent.

Ambient fact: locality is not a marketing region claim by itself. Public traffic today runs on a documented single-region deck, with a playbook to re-enable multi-region edges when paid volume justifies it. Honesty about mesh posture lives with the rest of the architecture — not in a slogan.

• Python control plane — full governance; Redis GET/SET for inventory and meters.
• gateway-rs edge — Redis GET on the HIT path; HIT metering gated by the control plane so the edge never holds the receipt signing key.

On a HIT, the Ephemeral Side serves the blob; the Pipeline records the crossing. On a MISS, the Pipeline owns the upstream trip. Streaming skips the edge Redis GET for stream=true and follows the streaming contract instead — Streaming & failover.

• Default tree main: at:{tenant}:cache:v2:{digest}
• Named tree (X-Ohm-Cache-Tree): at:{tenant}:tree:{tree_id}:cache:v3:{digest}

Digest = SHA-256 of the canonical request. Named trees keep preview and agent inventory partitioned under load so main is not everyone’s scratch pad. See Cache trees.

Public deck is single-region today with a documented playbook to re-enable multi-region edges when volume justifies it. Streaming failover scope: Streaming. If you are designing for regional failover of the mesh itself, treat that as an ops roadmap item — not a silent promise of the current public cut.

Status of public limits: Status. Verify plane health with /health and /ready on api.withohm.dev.