Low-resource mode: suspend copy trading, schedule ML training, resource view

[yonixnetwork]

Keeps the trading core reliable on the 2 GB / 2 vCPU droplet. The database is about 16 GB; measured load was 8–12, with 345 MB free and 1.3 GB of swap in use.

This PR deletes no data, keeps the copy-trading code, and does not change Solana execution, risk or position management. Full report: docs/LOW_RESOURCE_MODE.md.

Changes

Modes and status

• SYSTEM_RESOURCE_MODE: NORMAL / LOW_RESOURCE / EMERGENCY. The default is LOW_RESOURCE.
• COPY_TRADING_STATUS: ACTIVE / THROTTLED / SUSPENDED. The default is SUSPENDED.
• Both are defaults from .env. The dashboard can change them at runtime; the value is stored in platform_settings.operating_mode.
• Code: yonixalpha_core.operating_mode.

Resource level

• NORMAL / WARNING / CRITICAL, read from /proc: available memory, load per CPU, and Linux memory pressure (PSI).
• Thresholds are the RESOURCE_* settings.
• CRITICAL pauses priority-3 work only (copy trading, ML training, automatic review refreshes).
• Code: yonixalpha_core.resources.

Copy engine

• SUSPENDED stops:
  • target watching on all chains;
  • copy entries and mirrored sells;
  • outcome evaluation, wallet profiles and enrichment.
• SUSPENDED keeps stop loss, trailing and exits for copy positions that are already open.
• THROTTLED watches less often: EVM every 10 s, outcomes every 10 minutes.
• Step timers now start due on the first pass.

ml service

• In LOW_RESOURCE, training runs once a day and never while the level is CRITICAL. A skipped run is recorded as SKIPPED with the reason.
• Wallet ML pauses together with copy trading.
• Inference is unchanged.

ML Review

• New review_status values: CURRENT / STALE / RUNNING / PENDING / FAILED.
• Returns HTTP 202 while there is no result yet.
• Automatic refreshes wait while the level is CRITICAL. The Refresh button still computes, except in EMERGENCY.
• The page shows last calculated, age and status.

API

• GET /system/resources
• PUT /system/resource-mode
• GET /copy/trading-status and POST /copy/trading-status. Resuming returns 409 unless all four COPY_* thresholds hold. Auto-resume is off. Changes are audited.

UI

• System Health gets a "Server resources" panel: mode badge, level, host / Postgres / Redis figures, memory and CPU per service, and the mode selector.
• The Copy Trading page gets the status banner ("SUSPENDED — LOW SERVER RESOURCES") and a "Resume Copy Trading" check with a WAIT / OK recommendation.

Infrastructure

• Heartbeats now include CPU seconds.
• Worker database pools are 5 + 10 (were 10 + 10). The API keeps 10 + 10.
• Postgres in prod compose: explicit defaults, effective_cache_size=512MB, max_parallel_workers_per_gather=0, jit=off. A deploy restarts Postgres for a few seconds.

Measurement

• db_health now prints the mode, per-service memory and CPU, Redis, and Solana pipeline latency.

Tests

• New tests:
  • core: resources and operating mode;
  • copy-engine: policy against a real DB; SUSPENDED and THROTTLED loops;
  • ml: training schedule and wallet pause;
  • api: resources, resume refusal and audit, review 202/PENDING, review cache statuses.
• Full suite: core 887, api 175, data-solana 10, data-evm 14, copy-engine 17, discovery 18, migration 11, momentum 11, decision-engine 58, ml 46, paper-trading 81.
• ruff clean; web tsc / lint / build OK.

Rollout

Staged, as docs/LOW_RESOURCE_MODE.md section 4 describes:

1. Deploy PR Unblock the copy-engine loop and trim the remaining long EVM queries #56 alone and measure.
2. Then deploy this PR and measure.

Not yet measured on the server: per-process memory/CPU (docker stats), and the after-deploy figures.

🤖 Generated with Claude Code

https://claude.ai/code/session_015vpuZ1UX4MXz7NCF7wrtDv

---

Generated by Claude Code