Operate remote workers and platform services

Identify the service you need#

AMM's application API coordinates research and trading. A remote worker is another prepared AMM API executor on Windows, macOS or Linux. Platform Server supplies hosted services such as the economic calendar. Broker relay connections are another responsibility. A healthy calendar service does not prove broker connectivity or research capacity.

One backtest/optimisation runs entirely on one executor. Batch distribution places whole jobs on workers; it does not split one chronological engine across machines. Each worker has its own data, plugins and result storage. Completed remote evidence is downloaded to the manager so it can appear in ordinary history.

Prepare a worker connection#

Remote workers require authenticated HTTPS with a provisioned certificate, trusted host identity and owner-only token file. Non-loopback plaintext listeners and ad hoc listener overrides are rejected. Arrange service provisioning with your deployment administrator before adding an address in the UI.

The worker's native engine must match its OS and architecture. A Windows library cannot serve as a macOS/Linux native runtime. Plugin payload and historical data also need to be compatible with that executor. Do not interpret an HTTP health response as proof of native readiness.

The current automated Deploy action fails closed because it cannot atomically provision the full TLS/token/service boundary. Use a prepared secure service. The presence of a Deploy button is not a promise of one-click installation.

Add and test the worker#

1. Add Remote Workers to your workspace. Its Fleet panel lists health, platform/lifecycle, version, CPU/memory, job counts and last-seen information.
2. Choose Add worker. Enter a stable name, host/port, API token-file reference and optional description. Names use letters, numbers, underscores and hyphens.
3. Expand SSH and lifecycle only when remote start/stop/log access is configured. SSH authority is distinct from API authentication.
4. Use Test connection and inspect the exact failure if it cannot authenticate or validate the endpoint.
5. Save/enable the worker and verify fresh health/telemetry. Some configuration changes require API restart before routing changes take effect; preserve active jobs first.

Disabled workers are not polled or offered for ordinary routing. Local remains enabled. A failed refresh keeps the last fleet list visibly stale; check its timestamp before treating it as current.

Submit work and synchronize inputs#

Choose the worker in a research form's Run On field. The history badge identifies where the job runs. Algorithm sync compares manifest-backed plugin payload hashes and transfers changed runtime packages before submission unless separately managed synchronization is configured.

Historical sync is separate. Batch optimisation offers Distributed, a worker pool, Local, Sync data, Preflight and Apply sync. Read per-worker coverage/results for the actual symbol/feed/account and warmup range. Do not assume a similarly named local archive exists remotely.

A manager configured as an execution gateway can redirect nominally Local work to a designated worker. Inspect effective running-on identity; the UI machine need not be the executor.

Monitor and control safely#

The jobs board contains Running, Recent remote and Queued views, with type/worker/text filters. Progress differs by workload; a live session is not a percent-complete research job. Assign… appears where implemented but is disabled when the server reports unsupported reassignment. Live sessions remain on their starting executor.

Select a worker for Overview, Jobs, Logs and Connection. Logs can load 200/500/1,000 lines, filter severity and copy the fetched evidence. Read the fetch time and worker identity.

Start, Stop and Restart depend on enabled state and configured lifecycle access. Disruptive operations require verifiable idle telemetry and are refused when jobs are active or unverifiable. An unhealthy process may still offer a recovery action, but the server guard remains authoritative. Do not bypass it by killing a host with unknown broker exposure.

Removing a worker changes manager configuration; it does not delete the remote machine or its files. Retained Backtest/Optimization history remains separate. Read the confirmation and restart requirement.

Recover distributed research#

The manager schedules batch slots and handles retryable transport failures. Coordinator states such as draining, quarantined and faulted explain why a healthy-looking machine may receive no new work. Drain finishes current work without accepting more; cancel/requeue is a different intentional action where offered through the supported API.

For repeated failures, check HTTPS/token trust, native readiness, plugin payload, exact data coverage, available memory and manager events. Increasing concurrency can worsen an input or memory problem. Reopen durable batch/history after reconnecting and distinguish completed, interrupted, failed and requeued attempts.

Remote live sessions also need broker credentials, account approval and required history on their executor. A manager UI disconnect does not stop those sessions or close their positions. Verify the worker and broker state through automated operations.

Diagnose Platform Server calendar service#

The application accesses hosted calendar data through its normal API relay; the hosted client key stays server-side. An administrator configures the approved HTTPS origin and optional client credential. The relay can be disabled, so an empty calendar is not necessarily a failed worker.

Platform Server exposes service health, provider/status/event queries and revision pushes. Manual provider refresh and forecast recalculation are administrator operations with separate authorization. In the application, inspect calendar provider status, selected filters, date range and last refresh before concluding events are missing.

A new revision causes the relay to fetch an authoritative snapshot and update the UI. If the platform is unavailable, resolve the hosted origin, credential and provider refresh error; restarting a trading session is unrelated. Economic forecasts, source coverage and stale events should be read using the calendar guide, including the distinction between published values, estimates and unavailable provider data.