learning_ai_invt_trdg/docs/OPERATIONS.md

236 lines
7.5 KiB
Markdown

# Trading Monorepo Operations
## Purpose
This document is the operator and engineer runbook for `learning_ai_invt_trdg`.
It covers:
- local development setup
- verification and CI expectations
- staged rollout of the new monorepo deployment
- rollback rules
- release go/no-go checks
- post-cutover monitoring
## Local Development
### Prerequisites
- Node.js `>=20`
- `pnpm` `>=10`
- local checkout of:
- `learning_ai_invt_trdg`
- `learning_ai_common_plat`
- access to:
- platform-service
- Azure Cosmos DB
### Workspace bootstrap
```bash
pnpm install
cp .env.example .env
pnpm verify
```
### Core commands
```bash
pnpm verify
pnpm lint
pnpm typecheck
pnpm test
pnpm build
pnpm smoke:release
```
### Surface-specific commands
```bash
pnpm --filter @bytelyst/trading-backend dev
pnpm --filter @bytelyst/trading-web dev
pnpm --filter @bytelyst/trading-mobile dev
```
## Environment Model
### Platform-service
- `PLATFORM_API_URL`
- `PLATFORM_AUTH_ENABLED`
- `PLATFORM_JWT_ISSUER`
- `PLATFORM_JWT_PUBLIC_KEY` or `PLATFORM_JWT_JWKS_URL`
- `JWT_SECRET` only for HS256 compatibility environments
### Cosmos
- `COSMOS_ENDPOINT`
- `COSMOS_KEY`
- `COSMOS_DATABASE`
Rule:
- platform-service and Cosmos are the only supported production systems for this repo
- legacy repos may still be consulted as code references, but they are not runtime dependencies
- trading user profiles, dynamic config, trading controls, snapshots, capital ledgers, and strategy presets already use Cosmos-backed authority paths
- dynamic config runtime refresh and admin updates no longer seed from or mirror to legacy storage in the active backend runtime path
## Verification Standard
Before merge or release, all of the following must pass from repo root:
```bash
pnpm verify
pnpm lint
```
`pnpm verify` currently gates:
- backend, web, and mobile typecheck
- backend and web test suites
- backend and web build plus mobile typecheck
`pnpm lint` currently gates:
- backend contract and safety verification scripts
- web lint
- mobile lint
## Request Tracing
- the main web and mobile API paths, operator actions, and lifecycle fetches now attach `x-request-id`
- backend HTTP responses echo `x-request-id` so browser/app logs can be correlated with backend logs
- during incident review, treat `x-request-id` as the primary request correlation key across client and backend traces
## Feature Flag Ownership
- backend `GET /api/feature-flags` is the authoritative runtime contract for user-facing feature access
- web feature gates must read explicit feature-flag contracts instead of scraping generic config payloads
- dynamic config may still store the underlying values, but the product surfaces should consume the typed feature-flag API
## Staged Cutover
### Order
1. Backend internal validation
2. Web internal adoption
3. Mobile internal beta
4. Controlled operator rollout
5. Broader production cutover
### Backend cutover
- deploy backend with platform JWT support and Cosmos-backed control-plane and execution persistence enabled
- confirm runtime control reads/writes work through backend APIs
- confirm `dynamic_config`, trading-control, order, trade-history, and manual-entry containers are readable and writable
- confirm unauthorized requests are rejected and tenant-scoped reads are enforced
### Web cutover
- move operators to the monorepo web dashboard
- validate sign-in, session restore, kill-switch handling, and admin controls
- validate dynamic config writes through backend APIs
- keep legacy direct-table workflows disabled where backend API replacements exist
### Mobile cutover
- release to internal beta first
- validate sign-in, session restore, live state, degraded-state handling, and safe interventions
- do not enable broader rollout until backend/web contracts stay stable through at least one backend deploy cycle
## Rollback Rules
### Hard rollback triggers
- auth/session failures prevent sign-in or session refresh
- incorrect tenant scoping leaks another user's profile, orders, alerts, or history
- global trade halt or scoped disable controls do not apply correctly
- dynamic config writes fail or partially apply without clear operator visibility
- mobile/web clients cannot recover from degraded platform-service or backend states
### Rollback actions
1. stop rollout to additional users immediately
2. revert the most recent monorepo deployment
3. restore traffic to the previous stable web/backend/mobile release
4. keep backend trade-halt authority available during rollback
5. preserve audit logs and operational events for incident review
### Data rollback rule
- do not rewrite or delete Cosmos control-plane state as part of first-response rollback
- prefer application rollback first, then explicit state repair if needed
## Release Go/No-Go
Release is `go` only if all of the following are true:
- `pnpm verify` passes
- `pnpm lint` passes
- `pnpm smoke:release` passes
- platform-service auth is reachable from web and mobile
- Cosmos control-plane reads and writes succeed
- Cosmos execution-data reads and writes succeed
- kill-switch and maintenance behavior are validated on web and mobile
- backend tenant isolation checks are green
- operator-safe mobile interventions are limited to approved actions only
- no legacy runtime data dependency remains in critical public flows
Release is `no-go` if any of the following are true:
- auth source of truth is ambiguous in production
- admin/runtime-control actions are not fully audited
- rollback owner or rollback commands are unclear
## Release Smoke Checklist
`pnpm smoke:release` currently validates:
- web sign-in flow behavior
- web password reset flow behavior
- web authenticated session bootstrap behavior
- web websocket auth token gating
- web product kill-switch accessibility gating
- mobile auth and product-availability surfaces still compile against the shared platform contracts
Manual mobile release smoke is still required before broad rollout:
1. Sign in on a fresh install.
2. Confirm session restore after app restart.
3. Confirm product-disabled state blocks the app shell.
4. Confirm maintenance/availability messaging is visible.
5. Confirm the app recovers after re-enabling the product.
## Post-Cutover Monitoring
### Watch immediately after rollout
- platform auth failures
- token refresh failures
- backend `401` and `403` spikes
- websocket connection failure rate
- dynamic config update failures
- trading-control update failures
- mobile degraded/offline state frequency
- unexpected operator intervention failures
### Watch for the first 24 hours
- tenant isolation anomalies
- runtime control drift between backend memory and Cosmos control state
- kill-switch misfires
- stale session behavior across web and mobile
- build or chunk-size regressions affecting web load
## Known Remaining Gaps
- Cosmos-only execution persistence is now in place for the main backend runtime paths, but dormant legacy code and one-off reference scripts still need cleanup
- web now uses platform-session handling end to end; the remaining auth cleanup is removing dormant compatibility stubs and aligning profile bootstrap contracts fully with backend-owned product APIs
- root `pnpm verify` is green again after aligning the web Vitest harness with platform-session storage and current API contracts
- mobile does not yet include push notification infrastructure
- broader feature-flag ownership beyond the current shared backtest contract is not fully standardized yet
- exchange/order-level correlation-ID propagation is not fully standardized yet
These are follow-up items, not hidden defects. They should remain tracked in `docs/ROADMAP.md`.