Roadmap: metagraphed-core hybrid infra (Cloudflare edge · Railway core · Postgres)

[JSONbored]

Tracking epic for the hybrid deployment topology in ADR 0013 (docs/adr/0013-hybrid-deployment-topology.md), which lands ADR 0012's continuous indexer. Foundation (ADR + portable Timescale schema deploy/postgres/schema.sql + operator runbook deploy/README.md) shipped in #2107.

Shape: one Railway project metagraphed-core (private-networked microservices) · D1 chain sink → Timescale Postgres via Cloudflare Hyperdrive · self-hosted bare-metal core (2-box: archive node + indexer node, per #2518) · Hetzner escape hatch.

These are maintainer-only — each needs Railway/Cloudflare credentials and a human at the verification gates (per ADR 0013 §Sequencing). Not contributor-pickable.

Phases (ordered, each independently shippable + gated)

• infra(railway): provision metagraphed-core (Postgres + Redis), apply schema, add Hyperdrive binding, dual-write #2109 — provision metagraphed-core (Postgres + Redis) + schema + Hyperdrive + dual-write ✅
• infra(indexer): continuous chain indexer service (replaces poller + streamer + */3 drain) + one-time backfill #2110 — continuous chain indexer + bring-up stack (infra(indexer): continuous chain indexer + bare-metal bring-up stack #2117) ✅ · hardening: infra(indexer): harden capture-completeness (timestamp-miss drop, lookback gap) #2118 (closed 2026-07-04, superseded by the Rust rewrite's architecture)
• infra(archive-node): bare-metal buildout — provision → sync → first-party RPC origin #2111 — bare-metal archive node + first-party RPC origin (in progress: full-archive subtensor syncing, ~21% as of 2026-07-04)
• infra(serving): cut serving over D1 → Postgres via Hyperdrive, tier-by-tier #2112 — serving cutover D1 → Postgres via Hyperdrive (tier-by-tier, D1 fallback) ⛔ gate per tier
• infra(railway): migrate health-prober + rollups off CF isolates #2113 — migrate health-prober + rollups off CF isolates
• infra(realtime): Durable Object firehose → SSE/WS + GraphQL subscriptions + MCP, plus alerter #2114 — Durable Object firehose (SSE/WS/GraphQL-subs/MCP) + alerter
• infra(data): exporter (Parquet/CSV → R2) + reconciler (folded-state drift) #2115 — exporter (Parquet/CSV → R2) + reconciler (folded-state drift)
• infra(decommission): retire GitHub poller + streamer + */3 drain; demote D1 to hot cache #2116 — decommission GitHub poller + streamer + */3 drain; D1 → hot cache
• infra(core): cut Postgres/Redis/indexer over from Railway (interim) to the self-hosted box (final) #2956 — cut Postgres/Redis/indexer over from Railway (interim) to the self-hosted box (final), per this epic's own topology table
• infra(cleanup): delete orphaned Railway backfill-volume #2957 — delete orphaned Railway backfill-volume (cleanup, zero risk)
• infra(wss-lb): evaluate relocating the WSS load balancer off Railway #2958 — evaluate relocating wss-lb off Railway (low priority, optional)

Shipped alongside (deploy-agnostic)

• WSS load balancer (deploy/wss-lb, feat(infra): health-aware WSS load balancer for Bittensor RPC (ADR 0013) #2119) ✅ — cosmos.directory-style, health-aware WS reverse proxy over the registry's subtensor-wss pool (reads /api/v1/rpc/pools, connect-time failover). Fills the gap the Cloudflare HTTP proxy explicitly punts (it can't proxy wss). A standalone Railway service, deployable now (needs nothing from the bare-metal node); complements infra(realtime): Durable Object firehose → SSE/WS + GraphQL subscriptions + MCP, plus alerter #2114.

⛔ = verification gate that must not be run unsupervised.

2026-07-04 status note: the self-hosted archive + indexer nodes now exist and are live (2-box split per #2518, resolving #2111's previously-flagged "network path... currently undocumented" gap via a Tailscale ACL grant, tag:edge-indexer → tag:edge-archive:9944). The self-hosted indexer-rs is already live-following chain head into a self-hosted Postgres/TimescaleDB, running in parallel with Railway's copy. #2956 tracks the actual cutover from interim (Railway) to final (owned box), gated on the archive node finishing its historical sync (~8 more days as of 2026-07-04) before full capture-parity verification is possible.