Skip to content

Operations, Monitoring, Backup, and Recovery

A production FL service should have an observability stack covering service health, training progress, security controls, privacy-budget state, audit-log delivery, model performance, and cost. The exact tooling may vary, but the operator should have dashboards and alarms before production data is used.

Signal typeExamples
Service healthCoordinator health, client heartbeats, endpoint latency, CPU/memory/storage.
Training progressRound start/end, selected clients, completed clients, dropout rate, metric trend.
Security controlsmTLS enforcement, certificate expiry, CRL age, failed authentication, image version.
Privacy controlsSecure aggregation quorum, DP budget, privacy-test status.
Audit deliveryLog shipping latency, missing logs, integrity check status.
Cost and energyRuntime, compute type, storage generated, per-run cost estimate.

Operational monitoring should cover both system health and federation-specific behaviour. Where an enterprise SIEM is used, security-relevant coordinator, client, certificate, and audit-log events should be forwarded according to the organisation’s logging standards. Detailed log-forwarding and heartbeat procedures should be maintained in repository runbooks.

DashboardMinimum panels
Operations dashboardService health, client availability, active run, last successful run, failed runs.
Security dashboardmTLS failures, certificate expiry, revocation status, image digest, failed access attempts.
Training dashboardRound metrics, convergence curve, dropout rate, per-participant metrics.
Privacy dashboardDP budget spent, secure aggregation status, privacy-test status.
Cost dashboardPer-run compute, storage, network transfer, monitoring volume, cumulative monthly estimate.
AlarmTriggerAction
Coordinator unavailableHealth check fails or no heartbeat.P1 or P0 depending on production impact; page on-call.
mTLS enforcement regressionUnknown or invalid client unexpectedly accepted, or probe fails.Treat as suspected security regression until verified manually.
Certificate near expiryDays until expiry below threshold, such as 30 days.Trigger rotation runbook.
CRL or revocation staleRevocation source age above threshold or refresh probe missing.Page on-call; fix distribution.
SecAgg abort rateAbort rate above defined threshold over monitoring window.Investigate client health, quorum, timeouts, and network stability.
DP budget exhaustedEpsilon spent at or above configured budget.Halt or pause training and escalate to model owner and DPO.
Validation failedClient reports schema or manifest validation failure.Block run and resolve data integrity issue.
Audit-log delivery lagCross-account or central log delivery exceeds threshold.Verify audit pipeline and retention.

Detailed incident response procedures, severity ladder, dead-man heartbeats, and DR drill guidance are provided in Appendix G.

9.4 Model-performance monitoring and retraining cadence

Section titled “9.4 Model-performance monitoring and retraining cadence”

After release, the model should be monitored for performance degradation, data drift, and operational incidents. Retraining should be scheduled or event-driven, and retraining requires the same evidence discipline as initial training.

TriggerResponse
Scheduled retraining windowRun approved retraining workflow and compare to prior model.
Performance degradationInvestigate per-participant metrics, data drift, and recent schema changes.
Data driftReview feature distributions, label distribution, and missingness.
Participant changeRe-run onboarding, update certificates, update federation agreement where needed.
New model versionComplete model card, privacy tests, security review, and release gate.
Backed-up itemNotes
Infrastructure configurationIaC state, approved configuration templates, deployment parameters.
Certificates and PKI evidenceCertificate inventory, issuance records, revocation records, CA rollover plans.
Model checkpointsApproved model versions and rollback targets.
Run records and manifestsTraining manifests, data-manifest hashes, participant lists, privacy settings.
Audit logsOperational logs, security events, control validation evidence.
RunbooksCurrent operational procedures and escalation contacts.

The default production assumption should be Singapore-resident recovery unless the deploying organisation has an approved data-residency waiver or equivalent governance decision. Disaster recovery plans should define recovery time objective, recovery point objective, backup frequency, restoration steps, and drill cadence.

DR decisionMinimum expectation
Recovery environmentApproved environment in the permitted residency boundary.
BackupsEncrypted, access-controlled, restorable backups for configuration, model artefacts, and evidence.
Restoration procedureDocumented and tested steps for restoring coordinator, client configuration, certificates, and model registry.
Participant coordinationContacts and procedures for client reconnection after recovery.
DR drillPeriodic drill with evidence and lessons learned.