Vai al contenuto principale

Architettura

Ilum is a modular data lakehouse platform built on Kubernetes. It combines Apache Spark, Trino, DuckDBe Apache Flink into a unified multi-engine architecture, managed through a primary control plane (ilum-core) and a dedicated module-management microservice (ilum-api), deployed entirely via Helm charts. Ilum follows an "all-in-one but optional" philosophy where every component beyond the core services can be enabled or disabled independently at runtime, allowing deployments to range from a lightweight development setup to a full enterprise data platform.

The platform is built around several key architectural principles:

  • Decoupled compute and storage: processing engines and data storage scale independently.
  • Stateless services: both ilum-core e ilum-api hold no in-memory state, enabling horizontal scaling and crash recovery.
  • Multi-engine execution: Spark, Trino, DuckDB, and Flink are peer engines, with an automatic router that selects the right engine for each query based on data size, workload type, and locality.
  • Modular extensibility: optional modules install, upgrade, and disable at runtime via ilum-api. Future releases extend ilum-api con Model Context Protocol (MCP) capabilities and open APIs for third-party integration.
  • Open standards: OpenAPI REST, JDBC/ODBC, Apache Iceberg/Delta/Hudi table formats, OpenLineage for lineage.
  • Kubernetes-native: all workloads run as pods with native resource management, scheduling, and observability. Compatible with any CNCF-compliant distribution including Red Hat OpenShift, EKS, AKS, GKE, k3s, Rancher, and bare-metal Kubernetes.
Ilum platform architecture diagram showing Spark jobs running as Kubernetes microservices

Componenti chiave

                        ┌─────────────┐
│ ilum-ui │
│ (React) │
└──────┬──────┘
│ REST API
┌──────────┐ ┌──────▼──────┐ ┌──────────────────┐
│ ilum-cli │───▶│ ilum-core │───▶│ PostgreSQL │
└──────────┘ └──┬───┬───┬──┘ │ (primary) │
│ │ │ │ MongoDB (legacy)│
┌────────────┘ │ └─────┐ └──────────────────┘
│ │ │
│ ┌──────▼──────┐ │
│ │ ilum-api │ │ Module management
│ │ (modules) │ │ Helm install / upgrade
│ │ + MCP │ │ Roadmap: MCP, open APIs
│ │ (roadmap) │ │
│ └─────────────┘ │
│ │
┌────────▼───────┐ ┌──────────┐ ┌──▼─────────────┐
│ Kafka / gRPC │ │ Catalog │ │ Object Storage │
│ (messaging) │ │ Layer │ │ (MinIO / S3) │
└────────┬───────┘ └─────┬────┘ └───────┬────────┘
│ │ │
┌────────▼───────────────▼──────────────▼────────────────┐
│ Engine Layer │
│ Spark · Trino · DuckDB · Flink (Beta) │
│ fronted by Apache Kyuubi │
└────────────────────────────────────────────────────────┘
  1. ilum-core: The main platform service. Manages cluster lifecycle, job scheduling, session management, multi-engine SQL execution, security, lineage capture, and REST API exposure (OpenAPI 3.0). Stateless by design; all persistent state is stored in PostgreSQL (primary, accessed via R2DBC with jOOQ-generated SQL DSL), with MongoDB retained for legacy deployments.

  2. ilum-ui: React-based web console for managing clusters, submitting jobs, browsing tables and lineage, running multi-engine SQL queries through the SQL Editor, configuring security, and toggling modules. Communicates exclusively with ilum-core e ilum-api via REST.

  3. ilum-api: Dedicated module-management microservice. Drives Helm-based install, upgrade, and disable of optional Ilum modules (Trino, JupyterHub, Superset, Airflow, etc.) at runtime via cluster-scoped RBAC. Stateless. Future releases extend ilum-api con Model Context Protocol (MCP) e open APIs to act as the platform-wide extensibility surface.

  4. ilum-cli: Command-line interface for scripting and automation. Supports all ilum-core REST API operations including job submission, cluster management, and configuration. Useful for CI/CD pipelines and headless environments.

  5. Kubernetes Cluster: The primary execution environment. Spark, Trino, Flink, and other engine pods run with configurable executor counts, resource limits, and node affinities. Ilum manages the full pod lifecycle.

  6. Storage di oggetti: S3-compatible storage (MinIO, Ceph, RustFS, AWS S3, GCS, Azure Blob) serves as the persistent data layer. All table data, job artifacts, and Spark event logs are stored here, decoupled from compute.

  7. PostgreSQL (primary) and MongoDB (legacy): Internal metadata store for job history, cluster configuration, user accounts, session state, and operational data. PostgreSQL is the recommended store for new deployments; MongoDB is supported for existing deployments with built-in migration tooling (M001 through M009 scripts).

  8. Messaging Layer (Kafka / gRPC): Handles communication between ilum-core and running Spark jobs. See Tipi di comunicazione for details.

  9. Catalog Layer: Persistent metadata services (Hive Metastore, Nessie, Unity Catalog, DuckLake) that enable SQL access across all engines. See Data Catalog Layer for details.

  10. Engine Layer: Execution engines (Spark, Trino, DuckDB, Flink) fronted by the Apache Kyuubi SQL gateway. See Multi-Engine Query Architecture.

Integrated Modules

All modules below are optional Helm-deployed components that share authentication, networking, and catalog configuration with ilum-core:

CategoryModuli
OrchestrazioneAirflow, Kestra, Mage, n8n, NiFi
NotebookJupyterLab, JupyterHub (Enterprise), Zeppelin
BI & VisualizationSuperset, Streamlit
ML e intelligenza artificialeMLflow, LangFuse, AI Data Analyst
EnginesTrino, Apache Flink
Cataloghi datiHive Metastore, Nessie, Unity Catalog, DuckLake
OsservabilitàGrafana, Prometheus, Loki, Marquez (default-on)
IdentityOry Hydra (Ilum as IdP), OpenLDAP
Controllo della versioneGitea
Analytics storeClickHouse

Each module is enabled via a single Helm flag and automatically inherits cluster networking, LDAP/OAuth2 authentication, and catalog connection settings. See Panoramica for detailed feature descriptions.

Flusso di lavoro

Batch Job Workflow

  1. A user submits a Spark job via the ilum-ui, REST API, or ilum-cli, specifying the job JAR/Python file, Spark configuration, and target cluster.
  2. ilum-core validates the request, schedules the job, and creates a Spark driver pod on the target Kubernetes cluster.
  3. The Spark driver provisions executor pods according to the configured parallelism. Executors scale horizontally across cluster nodes.
  4. The job executes, reading from and writing to object storage through the catalog layer.
  5. Results and execution metadata are returned to ilum-core and stored in PostgreSQL (or MongoDB on legacy deployments). Users view results in the UI or fetch them via the REST API.

Interactive Session Workflow

  1. A user creates an interactive session (via UI, API, or CLI), which launches a long-running Spark application pod.
  2. The session remains alive, ready to accept code execution requests without Spark initialization overhead.
  3. Users submit code snippets (Scala, Python, SQL) via REST API. Each snippet executes within the existing Spark context and returns results immediately.
  4. Multiple users can share a session through Code Groups - shared Spark contexts that enable collaborative analysis while isolating variable namespaces.
  5. The session terminates on explicit shutdown or after a configurable idle timeout.

For step-by-step guides, see Run a Spark Job e Offerte di lavoro interattive.

Tipi di comunicazione

Ilum supports two forms of communication between Spark jobs and ilum-core: Apache Kafka and gRPC.

Comunicazione Apache Kafka

L'integrazione di Ilum con Apache Kafka facilita una comunicazione affidabile e scalabile, supportando tutte le funzionalità di Ilum, tra cui l'alta disponibilità (HA) e la scalabilità. Tutti gli scambi di eventi vengono condotti tramite argomenti creati automaticamente utilizzando i broker Apache Kafka.

Comunicazione gRPC (impostazione predefinita)

In alternativa, gRPC può essere utilizzato per la comunicazione. Questa opzione semplifica il processo di distribuzione eliminando la necessità di Apache Kafka durante l'installazione. gRPC stabilisce connessioni dirette tra i processi ilum-core e Spark, eliminando la necessità di un broker di messaggi separato. Tuttavia, l'uso di gRPC non supporta la disponibilità elevata per ilum-core nell'implementazione corrente. Anche se l'ilum-core può essere ridimensionato, i processi Spark esistenti continueranno a comunicare con le stesse istanze dell'ilum-core.

Comparison

CaratteristicaApache KafkagRPC
Alta disponibilitàYes - ilum-core replicas share state via topicsNo - direct point-to-point connections
ScalabilitàHorizontally scalable with partitioned topicsLimited to single ilum-core affinity
Deployment ComplexityRequires Kafka cluster (3+ brokers recommended)Zero additional infrastructure
Recommended ForProduction, multi-replica, HA environmentsDevelopment, testing, single-node setups
mancia

Start with gRPC for development and testing. Switch to Kafka when deploying to production or enabling HA. See Production Deployment for HA configuration.

Multi-Engine Query Architecture

Ilum provides four execution engines through a unified SQL gateway built on Apache Kyuubi. This architecture allows users to submit SQL queries via standard JDBC/ODBC interfaces, the in-product SQL Editor, or the REST API, and have them routed to the most appropriate engine, either by explicit selection or by the automatic engine router.

┌──────────────────────┐
│ JDBC / ODBC │ BI tools, CLI, applications
│ Clients │
└─────────┬────────────┘
│ Thrift Binary Protocol
┌─────────▼────────────┐
│ Apache Kyuubi │ Session management, engine routing,
│ SQL Gateway │ authentication, connection pooling,
│ + Auto Router │ automatic engine selection
└──┬──────┬─────┬───┬──┘
│ │ │ │
┌──▼──┐ ┌─▼──┐ ┌▼──┐ ┌▼─────┐
│Spark│ │Trino│ │Duck│ │Flink │ Engine selection per query
│ SQL │ │ │ │ DB │ │(Beta)│ (auto or explicit)
└──┬──┘ └──┬──┘ └─┬──┘ └──┬───┘
│ │ │ │
┌──▼───────▼──────▼───────▼─────┐
│ Catalog Layer │ Hive / Nessie / Unity / DuckLake
├───────────────────────────────┤
│ Storage Layer │ MinIO / S3 / GCS / WASBS / HDFS
└───────────────────────────────┘

Automatic Engine Router

The automatic engine router observes incoming queries and selects the engine best suited to each workload, based on:

  • Data size: Estimated scan size of the query against catalog statistics.
  • Workload type: Streaming, interactive, batch ETL, or ad-hoc exploration.
  • Locality: Whether the data lives in DuckLake (local-first) or remote object storage.
  • Engine availability: Which engines are deployed and currently healthy.

User override is available for every query through the Engine Selector in the SQL Editor and the engine field on the REST API.

Spark SQL

On-demand sessions with DAG-based parallel execution. Spark SQL is the most versatile engine - it handles ETL, batch processing, ML pipelines, and complex analytical queries. Executors are provisioned dynamically via Kubernetes and can scale from zero to hundreds of pods. Best suited for heavy transformations, large-scale joins, and workloads that benefit from distributed shuffle.

Trino

Always-on MPP (Massively Parallel Processing) engine with a coordinator-worker topology. Trino uses pipelined execution to deliver sub-second interactive query latency on large datasets. Workers remain running for instant query response. Best suited for interactive analytics, dashboarding, and ad-hoc exploration.

DuckDB

Embedded analytical engine running inside the ilum-core process. DuckDB provides zero-overhead local query execution with no pod provisioning or network latency, and pairs naturally with the DuckLake catalog. Best suited for lightweight analytics, small dataset queries, and rapid prototyping.

Distributed stream-processing engine for low-latency, event-driven workloads. Available for Enterprise deployments through the Kyuubi SQL gateway. Best suited for continuous data pipelines, real-time enrichment, and event-time analytics.

Engine Selection Guide

Caso d'usoRecommended Engine
ETL / large-scale batch processingSpark SQL
Interactive dashboards and BI queriesTrino
Complex ML pipelinesSpark SQL
Ad-hoc exploration on large datasetsTrino
Quick queries on small datasets / DuckLakeDuckDB
Streaming ingestion (Structured Streaming)Spark SQL
Low-latency stream processingCoraggioso

All engines access the same data through the shared catalog layer and object storage, enabling users to choose the right engine per workload without data movement. See SQL Editor for engine configuration and Performance for optimization details.

Data Catalog Layer

Data catalogs provide the persistent metadata layer that enables SQL access, schema management, and multi-engine data sharing. When a Spark job or Trino query creates a table, the catalog records its schema, location, and format so that any engine can access it later.

Ilum supports four catalog implementations:

Hive Metastore (default)

PostgreSQL-backed metadata service, automatically deployed and configured by Helm. Compatible with Spark, Trino, and Superset out of the box. Provides the broadest ecosystem compatibility and is the recommended default for most deployments.

Nessie

Git-like catalog for Apache Iceberg tables. Supports branching, tagging, and merging of table metadata - enabling CI/CD workflows for data. Create a branch, experiment with schema changes or data transformations, and merge only when validated.

Catalogo di Unity

Three-level namespace model (catalog, schema, table) with governance-focused access control. Provides a familiar structure for organizations migrating from Databricks or requiring fine-grained catalog-level permissions.

DuckLake

Lightweight embedded catalog optimized for DuckDB. Stores metadata directly in a DuckDB database file, ideal for local development and single-engine DuckDB workloads.

Catalog Integration

Spark jobs launched through ilum are automatically configured with catalog connection parameters at startup - no manual configuration required. Trino and Superset also receive catalog configuration via Helm values.

Table Format Support

CatalogoLago DeltaApache IcebergApache HudiParquet
Hive Metastore
NessieNoNoNo
Catalogo di UnityNo
DuckLakeNoNoNo

Vedere Cataloghi dati for setup details and Ilum Tavolo for the unified table abstraction.

Tipi di cluster

Ilum semplifica la configurazione del cluster Spark e, una volta configurato, il cluster può essere utilizzato per vari processi, indipendentemente dal tipo o dalla quantità. Ilum supporta attualmente tre tipi di cluster: Kubernetes, Yarn e Local.

nota

Durante l'avvio iniziale, Ilum crea automaticamente un cluster predefinito, utilizzando lo stesso cluster Kubernetes in cui è installato Ilum. Se questo cluster predefinito viene eliminato accidentalmente, è possibile ricrearlo facilmente riavviando il pod ilum-core.

Kubernetes Cluster

L'obiettivo principale di Ilum è quello di facilitare la facile integrazione tra Spark e Kubernetes. Semplifica la configurazione e l'avvio delle applicazioni Spark su Kubernetes. Per connettersi a un cluster Kubernetes esistente, gli utenti devono fornire informazioni di configurazione predefinite, ad esempio l'URL dell'API Kubernetes e i parametri di autenticazione. Ilum supporta sia i metodi di autenticazione basati su utente/password che su certificati. Più cluster Kubernetes possono essere gestiti da Ilum, a condizione che siano accessibili. Questa funzionalità consente la creazione di un hub per la gestione di numerosi ambienti Spark da un'unica posizione.

Ilum is compatible with any CNCF-compliant Kubernetes distribution:

  • Managed cloud: Google Kubernetes Engine (GKE), Amazon EKS, Azure AKS, DigitalOcean Kubernetes
  • Azienda: Red Hat OpenShift / OKD
  • Lightweight: k3s, Rancher, MicroK8s
  • Bare metal: Self-managed Kubernetes installations
  • Local development: Minikube, k3d, Docker Desktop

Cluster di filati

Ilum supporta anche i cluster Apache Yarn, che possono essere facilmente configurati utilizzando i file di configurazione di Yarn presenti nell'installazione di Yarn.

Cluster locale

Il tipo di cluster locale esegue applicazioni Spark in cui viene distribuito ilum-core, ovvero esegue applicazioni Spark all'interno del contenitore ilum-core quando viene distribuito in Docker/Kubernetes o nel computer host quando viene distribuito senza un agente di orchestrazione. Questo tipo di cluster è adatto a scopo di test a causa delle limitazioni delle risorse.

Comparison

CaratteristicaKubernetesFiloLocale
Production ReadyNo (testing only)
Alta disponibilitàYes (via YARN RM HA)No
Auto-ScalingYes (K8s autoscaler)Yes (YARN node managers)No
Multi-ClusterN/A
Dynamic Executor AllocationLimited

A single ilum instance can manage multiple clusters of different types simultaneously, enabling hybrid K8s + Yarn deployments - useful for organizations migrating from Hadoop to Kubernetes. See Cluster e storage for configuration details.

Decoupled Compute-Storage Architecture

Ilum follows a decoupled compute-storage architecture where processing engines and data storage are independent, horizontally scalable layers. This separation is fundamental to ilum's design and enables independent scaling, multi-engine access, and cost-efficient resource utilization.

┌─────────────────────────────────────────────────────────────┐
│ Compute Layer │
│ ┌──────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Spark │ │ Trino │ │ DuckDB │ Engines │
│ │ Executors│ │ Workers │ │(embedded)│ scale │
│ └────┬─────┘ └─────┬────┘ └──────┬───┘ independently │
│ │ │ │ │
├───────┼──────────────┼──────────────┼───────────────────────┤
│ │ Catalog Layer (Hive Metastore / Nessie) │
├───────┼──────────────┼──────────────┼───────────────────────┤
│ │ │ │ │
│ ┌────▼──────────────▼──────────────▼─────┐ │
│ │ Storage Layer │ │
│ │ MinIO / S3 / GCS / WASBS / HDFS │ Storage │
│ │ (Iceberg / Delta / Hudi / Parquet) │ scales │
│ └────────────────────────────────────────┘ independently │
└─────────────────────────────────────────────────────────────┘

Benefits of this architecture:

  • Independent scaling: Add more Spark executors or Trino workers without provisioning additional storage, and vice versa
  • Multi-engine access: Multiple compute engines can read from and write to the same datasets concurrently, mediated by table formats (Iceberg, Delta) that provide ACID guarantees
  • Cost efficiency: Compute resources can be released when not in use (e.g., Spark dynamic allocation, auto-pause) while data remains persistently available in object storage. Scale compute to zero during off-hours - storage costs persist, but expensive compute does not
  • Engine flexibility: Choose the right engine for each workload - Spark for ETL, Trino for interactive analytics, DuckDB for lightweight queries - all accessing the same data
  • Catalog-mediated consistency: The catalog layer ensures all engines see a consistent view of table metadata, enabling safe concurrent reads and writes across Spark, Trino, and DuckDB

Scalabilità

ilum-core è stato progettato pensando alla scalabilità. Essendo completamente stateless, ilum-core può ripristinare l'intero stato esistente dopo un arresto anomalo, semplificando la scalabilità verso l'alto o verso il basso in base ai requisiti di carico.

Ilum supports multiple scaling dimensions:

  • ilum-core horizontal scaling - Deploy multiple stateless replicas behind a load balancer. Requires Kafka for inter-replica coordination
  • Spark dynamic executor allocation - Executors scale up and down automatically based on workload. Idle executors are released after a configurable timeout
  • Kubernetes HPA - Horizontal Pod Autoscaler can manage always-on services (Trino workers, ilum-core replicas) based on CPU/memory metrics
  • Cluster autoscaler - Node-level scaling via Kubernetes Cluster Autoscaler or cloud provider auto-scaling groups, triggered when pending pods cannot be scheduled
  • Resource quotas - Kubernetes namespaces and ilum resource controls enforce per-tenant resource limits, preventing any single workload from monopolizing cluster resources

Vedere Controllo delle risorse e Production Deployment for configuration details.

Alta disponibilità

ilum-core and its required components support High Availability (HA) deployments. An HA deployment necessitates the use of Apache Kafka as the communication type, as gRPC does not support HA.

ComponentMinimum ReplicasNotes
ilum-core3Stateless; requires Kafka for HA coordination
ilum-api2Stateless module-management microservice
PostgreSQL3Primary metadata store; primary + standby with streaming replication
MongoDB (legacy)3Only for legacy deployments; replica set with automatic failover
Apache Kafka3KRaft quorum or ZooKeeper-based
MinIO4Erasure coding for data durability

Failure Domain Mitigation

  • Pod anti-affinity: Spread replicas of each component across different nodes to survive single-node failures.
  • Namespace isolation: Deploy Ilum infrastructure (Kafka, PostgreSQL, MongoDB, MinIO) in dedicated namespaces separated from user workloads.
  • Zone-aware scheduling: In multi-zone clusters, use topology spread constraints to distribute pods across availability zones.

Vedere Production Deployment for the full HA deployment guide.

Observability Architecture

Ilum provides three observability pillars - metrics, logs, and lineage - through optional Helm-deployed modules.

Metrics

Spark exposes execution metrics via the built-in PrometheusServlet. Ilum deploys PodMonitor resources that Prometheus scrapes automatically. Pre-configured Grafana dashboards visualize executor utilization, job duration, shuffle I/O, and GC pressure.

Spark Executor → PrometheusServlet → PodMonitor → Prometheus → Grafana

Registri

Container logs from all Spark driver and executor pods flow through the standard Kubernetes logging pipeline. When Loki is enabled, a Promtail DaemonSet collects logs from each node and ships them to Loki for centralized LogQL querying.

Container stdout → Promtail DaemonSet → Loki → LogQL queries

Derivazione dei dati

The OpenLineage Spark Listener captures table-level and column-level lineage events during job execution. These events are sent to the Marquez API, which stores them and exposes lineage graphs through the ilum UI (ERD and directed graph views).

Spark Job → OpenLineage Listener → Marquez API → Lineage UI

Server della cronologia delle scintille

For deep post-execution analysis, Spark History Server reads event logs from object storage and provides detailed DAG visualizations, stage breakdowns, and task-level metrics.

All observability components are optional and enabled via Helm flags. See Monitoraggio for metrics and log configuration, and Derivazione dei dati for lineage setup.

Sicurezza

Ilum provides a comprehensive security architecture covering authentication, authorization, data access control, and network security.

Autenticazione

Ilum supports multiple authentication methods:

  • Internal LDAP - Embedded OpenLDAP server deployed via Helm for standalone deployments
  • External LDAP/AD - Connect to existing Active Directory or LDAP directories
  • OAuth2/OIDC - Integration via ORY Hydra supporting Okta, Azure AD, Google, and Keycloak as identity providers

Autorizzazione

  • RBAC - Role-based access control with two modes: unrestricted (development - all users have full access) and restricted (production - permissions are enforced per role)
  • ABAC - Attribute-based access control using data classification tags for fine-grained policy decisions

Data Access Control

  • Row-level filters - Restrict query results based on user attributes (e.g., region, department)
  • Column-level masking - Redact or hash sensitive columns (PII, financial data) based on user roles
  • Hierarchical privileges - Grant access at catalog, schema, table, or column level

Network Security

  • TLS/mTLS - Encrypted communication between ilum-core, Spark jobs, and integrated services
  • Kubernetes Network Policies - Restrict pod-to-pod communication to authorized paths

Secrets Management

  • Kubernetes Secrets - Native secret storage for credentials, certificates, and API keys
  • External vault integration - Support for mounting secrets from external KMS providers

ilum as Identity Provider

Ilum can act as an OAuth2 provider for integrated services. When enabled, Airflow, Superset, Grafana, Gitea, and MinIO authenticate users through ilum's OAuth2 endpoints - providing single sign-on across the entire platform.

Vedere Sicurezza for the full security guide, Data Access Control for row/column policies, and OAuth2 for OIDC configuration.