Skip to content

MLOps Platform SRE and Reliability by Scale

Last updated: 2026-05-24

MLOps platform SRE is the operating discipline for the services that make ML change safe: experiment trackers, data catalogs, model registries, pipeline orchestrators, GPU schedulers, evaluation/replay services, serving platforms, feature/embedding stores, attestation systems, monitoring pipelines, audit logs, and release-policy engines. Fleet SRE keeps vehicles and operations recoverable; MLOps platform SRE keeps the ML evidence and release-control plane recoverable.

Use this page with mlops-scale-research-scope.md for maturity, mlops-reference-architectures-by-scale.md for platform boundaries, pipeline-orchestration-release-workflows-by-scale.md for workflow lanes, model-registry-artifact-lifecycle-by-scale.md for registry authority, evaluation-platform-replay-gates-by-scale.md for eval service SLOs, serving-inference-operations-by-scale.md for endpoint and batch serving SLOs, gpu-queueing-finops-by-scale.md for capacity and incident lanes, model-monitoring-drift-response-by-scale.md for alert actionability, and ../operations/fleet-sre-incident-response.md for fleet incident command.

The core rule is: the MLOps platform is part of the safety and release system once it can approve, block, serve, monitor, or reconstruct a model change. Its reliability cannot be measured only as cloud uptime; it must also preserve evidence correctness, artifact identity, rollback readiness, and auditability.


What Platform SRE Owns

SurfaceWhat must be reliableFailure it prevents
Artifact identityRegistry versions, aliases, digests, artifact-set compatibilityA release loads an ambiguous or wrong artifact
Evidence generationEvaluation, replay, runtime smoke, shadow/canary report, release packetGreen dashboards replace reproducible evidence
Evidence storageDataset manifests, eval reports, approvals, audit logs, incident recordsIncident review cannot reconstruct why a release was approved
Workflow executionPipelines, retries, queues, artifact handoffs, policy hooksFailed or partial DAGs produce silent side effects
Serving and endpoint controlEndpoint manifests, traffic routes, autoscaling, readiness, rollbackApproved artifacts are served outside approved scope
Monitoring and alertingSLIs, alert routing, suppression expiry, delayed-label joinsDrift and failures become dashboard noise
CapacityGPU queues, eval workers, replay clusters, registry API, metadata DBsRelease/evidence/incident jobs starve behind research workloads
Security and tenancyIAM, namespaces, secrets, signing keys, tenant partitions, data residencyCross-tenant leakage or unauthorized artifact promotion
Backup and restoreMetadata DBs, object stores, registry aliases, pipeline states, signing materialPlatform outage destroys release history or rollback path

The MLOps platform has two reliability dimensions: service reliability and evidence reliability. An API can be available while returning stale lineage, accepting unsupported alias changes, or losing approval history. That is not reliable MLOps.


Scale Ladder

MLOps scalePlatform SRE postureMinimum reliability controlAnti-pattern to block
S0 notebook researchLocal hygieneGit, local backup, run note, data pointerIrreplaceable notebook or checkpoint on one workstation
S1 repeatable prototypeShared but low-criticality servicesTracker/catalog backup, owner, basic availability note, restore smokeShared tracker becomes source of truth with no export path
S2 single-product productionProduct release control planeService inventory, SLOs for registry/eval/serving, backup/restore test, release freeze modePlatform outage forces unsafe manual release
S3 fleet and multi-siteSite-aware release platformIncident lane, ODD-cell blast-radius query, tenant/site partitions, capacity SLOsOne platform outage blocks rollback, evaluation, and fleet containment
S4 regulated safety-criticalEvidence-critical platformImmutable audit log, retention hold, RPO/RTO, dual-control recovery, restore drill, reportability pathEvidence exists only in mutable tickets or dashboards
S5 platform scaleMulti-tenant platform SREError budgets, platform SLOs, DR plan, policy-as-code, audit API, tenant isolation, support modelPlatform becomes slower or less trusted than bypass paths

At S2 and above, platform reliability is release readiness. A candidate should not move to shadow, site_canary, or champion if registry, evaluation, serving, attestation, monitoring, or rollback evidence is degraded beyond its SLO.


Criticality Tiers

TierServicesReliability posture
T0 release authorityModel/artifact registry, alias policy, release packet store, compatibility manifest store, attestation/signing, approval audit logStrongest backup, access control, restore test, retention hold, and change review
T1 evidence generationEvaluation/replay service, runtime package smoke, split/leakage checker, scenario catalog, safety evidence workflowsCapacity reservation, flake monitoring, deterministic rerun, evidence immutability
T2 production operationServing platform, monitoring pipeline, feature/embedding store used in production, fleet telemetry joinsAvailability, latency, staleness, rollback, site/tenant scope, incident runbooks
T3 data and training platformData catalog, lakehouse tables, annotation queue, training orchestrator, GPU schedulerQueue SLOs, lineage correctness, quota, cost visibility, restore for metadata
T4 research convenienceExploratory trackers, notebooks, scratch storage, ad hoc dashboardsExport path, owner, TTL, no release authority

Do not apply one uniform SLO to every tool. T0/T1 systems need correctness and recoverability more than low latency. T3/T4 systems need usability, cost, and data hygiene but should not be able to change release truth without promotion.


SLIs and SLOs

Platform areaSLIsSLO examples
RegistryAPI success, alias mutation latency, alias policy pass/fail, stale alias count, audit-log write success99.9% read availability for release hours; zero unlogged alias mutation
Evaluation/replayQueue wait by priority, flake rate, evaluator availability, stale scenario age, evidence completenessRelease eval starts within target window; incident replay preempts research jobs
OrchestrationWorkflow start latency, success rate by class, retry cause, policy hook availability, artifact handoff completenessRelease/evidence workflows have higher SLO than exploratory sweeps
ServingReadiness, p99 latency, model load success, traffic policy match, rollback load successCanary and rollback endpoints meet p99 and load-test SLO before promotion
Data catalogMetadata freshness, lineage completeness, quality report availability, deletion propagationRelease datasets have complete lineage before training/eval
MonitoringEvent ingestion latency, missing artifact IDs, alert precision, suppression expiry, delayed-label join successSafety-relevant alerts route to owner within target time
GPU/computeQueue wait, utilization, preemption correctness, quota violations, reserved incident capacityIncident and release lanes retain reserved capacity
AttestationSigning success, verification success, policy false positive/negative, key availabilityRelease-affecting artifacts verify before alias movement or activation
Backup/restoreBackup age, restore success, RPO/RTO, restore drill frequency, immutable-retention coverageT0 metadata restore tested on schedule and before major platform migration

SLOs should be scoped by scale and artifact authority. A research tracker can tolerate downtime. A release registry with active vehicles cannot silently lose alias history, approval state, or rollback target.


Error Budgets and Change Freezes

Error budgets are useful only when they affect behavior.

Budget typeWhat consumes itPolicy when exhausted
Availability budgetRegistry/eval/serving/platform API failuresFreeze non-urgent platform changes and focus on reliability
Evidence-correctness budgetMissing artifact IDs, incomplete manifests, stale lineage, audit write failureBlock release authority until evidence completeness is restored
Eval lead-time budgetRelease/evidence jobs miss queue or runtime SLOReserve capacity, reduce exploratory load, fix flakes
Alert-quality budgetLow alert precision, stale suppressions, missing ownersTune monitors, audit suppressions, update runbooks
Rollback-readiness budgetRollback load test fails or cache expiresHold promotion and rebuild compatible rollback set
Platform-bypass budgetTeams deploy outside approved pathImprove golden path, narrow policy, escalate unsupported bypasses

For autonomy, error budgets should never justify safety-evidence loss. If the evidence system is unhealthy, the correct response is to hold releases, not to accept undocumented risk.


Backup, Restore, and Disaster Recovery

AssetBackup requirementRestore question
Registry metadataVersioned database backup, alias-history export, immutable audit logCan we prove which artifact was active at a past time?
Object artifactsDigest-pinned object storage, retention class, cross-zone/region policy where justifiedCan we fetch the exact model/map/eval/replay artifact used in release?
Evaluation reportsImmutable report store plus source evaluator/container IDsCan we reproduce or defend the release decision?
Pipeline stateWorkflow definitions, run history, logs, parameters, output artifact IDsCan we resume or replay failed evidence generation without hidden side effects?
Feature/vector storesSnapshot metadata, index build IDs, corpus versions, deletion stateCan we avoid stale retrieval or point-in-time leakage after restore?
Secrets and signing keysKey management, break-glass policy, rotation history, dual controlCan we recover without allowing unauthorized signing?
Monitoring dataActive artifact IDs, incident windows, alert state, suppression historyCan we reconstruct a field anomaly after platform outage?
Policy rulesVersioned policy-as-code and exception recordsCan we tell whether a release passed the policy active at the time?

Backups are not enough. Restore drills should be scheduled, measured, and recorded as platform evidence. At S4/S5, a restore drill should include one sample release packet, one registry alias history query, one eval report, one serving manifest, and one incident evidence query.


Architecture Options

Platform reliability architectureAdvantagesDisadvantagesBest use
Local files plus GitTransparent, cheap, easy to exportWeak concurrency, weak audit, manual restoreS0-S1 research and baselines
Managed cloud MLOps platformFast reliability baseline, integrated registry/pipelines/endpoints, managed backupsVendor coupling and limited autonomy metadataS2-S4 cloud-native product lanes
Self-hosted Kubernetes platformStrong control, air-gapped/on-prem fit, custom metadata and policyRequires platform SRE, upgrades, DR, tenancy designS3-S5 autonomy platforms and private deployments
Hybrid managed plus autonomy metadata layerBalances managed reliability with custom release semanticsIntegration and ownership complexityS2-S5 teams using managed services but needing map/calibration/ODD evidence
Multi-region active/passive control planeStrong DR and business continuityCost and consistency complexityS4-S5 high-availability release platforms
Air-gapped or site-local platformData sovereignty and offline operationHard updates, constrained capacity, manual evidence synchronizationRegulated sites, customer isolation, low-connectivity managed sites

Do not overbuild multi-region MLOps before artifact contracts exist. But once a platform is the only path to rollback, evidence, or release authority, restore and DR become part of the safety case.


Incident Response for the MLOps Platform

Incident typeFirst containmentEvidence to preserve
Registry alias policy failureFreeze alias mutation and deployment admissionAlias history, policy version, affected artifacts, approver record
Evaluation service outageHold candidate promotion and route incident/release jobs to reserve laneQueue state, failed workflow IDs, missing evidence, workaround decision
Serving platform outageRoll back or route traffic by approved manifestEndpoint manifest, traffic split, model server logs, active artifact IDs
Monitoring ingestion lossRestrict rollout expansion and preserve raw telemetryMissing window, impacted sites/cohorts, monitor config, delayed-label plan
Data catalog lineage corruptionBlock training/eval consuming affected datasetsSnapshot IDs, lineage diff, downstream model/replay consumers
Signing or attestation outageBlock release-affecting artifact activationKey state, policy decision, build provenance, affected subject digests
Tenant isolation breachDisable affected namespace or data pathAccess logs, data products, artifacts, consumers, notification record
Backup restore failureFreeze authority changes until recoverability is provenLast good backup, restore logs, missing artifacts, compensating controls

MLOps platform incidents should update release records. If an evaluation outage forced a waiver, if a registry restore changed alias history, or if monitoring missed a canary window, the affected candidate or champion record must show that fact.


Operating Model

ScaleOwnershipReview cadenceOn-call expectation
S0-S1Model owner or research leadBaseline reviewNo formal on-call; document recovery
S2Product MLOps owner plus runtime/data ownersRelease readiness and weekly platform healthBusiness-hours or release-window coverage
S3Platform owner, product owners, site operationsRelease train, incident review, capacity reviewOn-call for release/eval/serving/monitoring
S4Platform SRE, safety owner, release manager, securityFormal evidence and restore reviewIncident-ready coverage for safety-relevant releases
S5Central platform SRE with tenant product ownersSLO/error-budget review, tenant council, policy exception review24/7 or risk-based support for critical services

The platform team owns reliability of the path. Product teams still own release decisions. A healthy platform makes the approved path faster than bypassing it.


Managed-Site and Autonomy Rules

Airside, port, yard, campus, warehouse, construction, mining, and urban-district mapping deployments need MLOps platform SRE rules that account for local operations:

  • Release/eval/rollback capacity should be reserved for active sites, not consumed by exploratory sweeps.
  • Site and tenant partitions must carry through data catalog, registry, evaluation, serving, monitoring, and audit logs.
  • Local outages should degrade to safe operational states: hold rollout, freeze map publication, use cached rollback artifacts, or restrict ODD cells.
  • Edge and site-local artifacts need a synchronization policy for registry aliases, compatibility manifests, and incident evidence after reconnect.
  • Map, calibration, semantic taxonomy, and serving manifests must remain queryable during incidents even when training systems are unavailable.
  • For offline aggregated-map segmentation, platform SLOs include batch backlog age, map-tile evidence completeness, source-map QA availability, and publication rollback readiness.

The non-road setting often has fewer vehicles than road AVs, but each site has stronger local procedures and higher map dependence. Platform SRE should preserve local authority instead of forcing one global release lane.


Release Blockers

BlockerApplies fromExample
No service inventoryS2Registry, eval, serving, and monitoring dependencies are unknown
Untested restoreS2-S5Registry or eval metadata backup exists but has never been restored
Missing artifact IDs in platform telemetryS3-S5Platform incident cannot identify affected model/map/runtime/prompt artifacts
T0 service degraded during promotionS2-S5Registry, attestation, release packet, or compatibility store is unhealthy
Incident lane starvedS3-S5Release replay or safety evidence waits behind research jobs
Audit log mutable or incompleteS4-S5Alias movement or approval cannot be defended after incident
Tenant/site boundary missingS3-S5Data, models, prompts, or evidence can cross customer/site boundaries unsupported
Platform SLO bypassS5Teams deploy outside platform because supported path is too slow or unreliable

Release blockers should preserve the blocked state. A frozen alias, held candidate, or paused rollout is useful evidence when it records the unhealthy dependency and owner.


KPIs

KPIMeaning
Evidence completenessRelease-affecting workflows with complete manifests, artifact IDs, and audit links
T0 availabilityRegistry, release packet, attestation, compatibility, and policy service health
Restore drill successPercentage of required restore tests passing within RPO/RTO
Release/eval queue SLORelease and incident workflows starting within target windows
Platform incident MTTRTime to contain, diagnose, and restore MLOps platform incidents
Alias mutation audit coveragePercentage of alias changes with policy result, approver, evidence, and scope
Platform bypass rateRelease-affecting artifacts outside approved platform path
Tenant isolation findingsCross-tenant/site access or evidence-boundary violations
Alert actionabilityPlatform alerts that produce owner, action, evidence, or suppression expiry
Rollback readinessRollback artifacts and serving routes that pass scheduled load/restore tests

Failure Modes

Failure modeConsequenceControl
Platform uptime hides evidence lossAPI is up but audit, lineage, or report fields are missingEvidence-correctness SLOs and manifest coverage checks
Registry restored without alias historyActive deployment cannot be reconstructedImmutable alias log and restore drill
Eval service is reliable but flakyResults arrive quickly but cannot support releaseFlake SLO and rerun policy
Serving platform bypassEndpoint ships outside registry/eval/attestation pathAdmission policy and golden-path SLOs
One shared queueIncident and release jobs starvePriority lanes and reserved assurance capacity
S5 governance too earlySmall team stops using the platformScale-gated controls and exportable simple manifests
S5 governance too lateMany teams create incompatible platformsStandard interfaces and policy at artifact boundaries
Backups miss object artifactsMetadata restores but models/maps/eval packs are goneDigest inventory and object-retention audit
Tenant isolation is dashboard-onlyData or evidence crosses site/customer boundariesNamespace/IAM/data-product enforcement and audit
Platform team becomes release authorityProduct accountability is blurredProduct/safety owners remain approvers; platform owns path reliability

  • mlops-scale-research-scope.md - maturity ladder and lifecycle controls.
  • mlops-reference-architectures-by-scale.md - platform placement, service boundaries, and durable interfaces.
  • mlops-migration-checklist-by-scale.md - transition gates before adding shared platform SRE.
  • mlops-scorecards-and-kpis-by-scale.md - KPIs, release blockers, operating cadence, and anti-metrics.
  • pipeline-orchestration-release-workflows-by-scale.md - workflow SLOs, incident lanes, and evidence state machines.
  • model-registry-artifact-lifecycle-by-scale.md - registry authority, aliases, audit, rollback, and retention.
  • evaluation-platform-replay-gates-by-scale.md - evaluation service health and replay/evidence SLOs.
  • serving-inference-operations-by-scale.md - endpoint, batch, edge, traffic, autoscaling, and rollback reliability.
  • model-monitoring-drift-response-by-scale.md - monitoring event contract, alert quality, and response states.
  • gpu-queueing-finops-by-scale.md - capacity, queueing, reserved incident lanes, and unit economics.
  • secure-artifact-attestation-profile.md - trust-chain service reliability and policy verification.
  • ../operations/fleet-sre-incident-response.md - fleet SRE, incident command, and safety evidence.
  • ../data-platform/data-catalog-lineage-quality-ops.md - lineage, quality, and data-product promotion reliability.
  • ../ota/perception-slam-artifact-compatibility-matrix.md - compatibility manifest and activation gates.

Sources

Public research notes collected from public sources.