SYSTEM Cited by 1 source

Atlassian ML Studio

Definition

Atlassian ML Studio is Atlassian's unified enterprise-scale ML development platform that standardizes modular development, centralizes workflow orchestration, and embeds data governance directly into the execution layer. It is the mission-critical backbone for AI systems including Rovo Search and Chat, the Teamwork Graph, and Confluence AI, enabling ~120k monthly workflow runs across 100+ ML teams and serving models to 5M+ monthly active Rovo users.

Architecture

Three architectural pillars:

Pillar 1 — Composable, Reusable ML Modules

Each module is a self-contained, executable unit (data processing, model training, packaging, evaluation, deployment). Modules are versioned, shareable, and composable into end-to-end pipelines via the module-as-versioned-artifact pattern:

• Every code push produces a module artifact used in workflows
• Artifact tags (latestAlpha, latest) enable rollback without redeploying infrastructure
• Dynamic CI/CD pipelines provide independent automated tests per module
• Teams own module namespaces autonomously; shared libraries form a reusable catalog
• 2,000+ reusable modules with 200k+ monthly iterations

The local dev loop builds Python modules in <30 seconds (down from minutes), with RDEs mirroring production. Repository is reserved for peer review and promotion, not iteration.

Pillar 2 — Workflow Orchestrator

Central service that manages, schedules, and automates ML workflows via CLI, portal, and API:

• Cloning/rerun — reproduce or adapt previous runs trivially
• Flexible triggering — portal, CLI, or programmatic API
• Composable workflows — nested and joined sub-workflows
• Hot clusters — pre-provisioned clusters stay active between runs (patterns/hot-cluster-for-iterative-ml)
• CRON scheduling — automated recurring tasks (retraining, data refreshes, monitoring)
• Automatic caching — ~80% of workflows use daily, saving 1,000+ hours/month (patterns/deterministic-task-caching)

Orchestrates jobs across platforms including Databricks, with access to relevant data sources.

Pillar 3 — Embedded Compliance Controls

Multi-layer governance framework:

1. User identity-based access control — only authorized users can initiate/manage workflows
2. Domain-level access control — clear data boundaries per workflow type (experimentation vs. production vs. ad-hoc)
3. Column-level classification — data classification tags at column granularity with automatic tag propagation through pipeline stages; unclassified columns treated conservatively by default

Integration Layer

ML Studio connects to Atlassian's broader ML ecosystem:

• Experiment tracking (compare runs and iterate)
• Central feature store (reuse trusted features)
• Model registry (version, approve, reuse models)
• Monitoring — ML Lens (multi-dimensional metrics for training/production regressions)
• Deployment & serving (internal serving platform)
• Service integrations (APIs, microservice triggers)

Workloads Served

• GenAI — LLM fine-tuning, automated prompt optimization, batch inference with cost/priority-aware workload management
• Predictive AI — end-to-end pipelines from labeling through feature engineering with data-residency-compatible storage
• Training & evaluation — complex evaluation pipelines, automated retraining triggers, safe rollout via versioning/tagging

Powers: Rovo Search (ranking, Q&A, chat), Dev-AI (code review), AIOps (alert grouping), Knowledge Discovery (topic extraction), Confluence AI (summaries, snippets, labeling).

Operational Scale

• 5M+ monthly active Rovo users served by models built on ML Studio
• 900k+ datasets generated with built-in access control
• ~120k monthly workflow runs by 100+ ML teams
• ~20k monthly model iterations
• 2,000+ reusable modules, 200k+ monthly iterations

Seen In

• sources/2026-06-10-atlassian-architecting-scalable-ml-platforms

(Source: sources/2026-06-10-atlassian-architecting-scalable-ml-platforms)