DataFocus · Journal
OT Flash vs OT Heavy: Three Questions That Decide Which DataFocus Edition You Install
Flash runs the ontology on one Apache Doris engine; Heavy runs it over Trino, Hive, Iceberg and Hadoop with DataSpring ETL and AutoML. Three questions, a worked example and the licence facts that decide between them.
DataFocus OT Flash and OT Heavy run the same ontology layer on different data foundations, and that one difference decides the choice. Flash puts Apache Doris underneath as both compute engine and warehouse: you install one engine, import the tables for one business domain and start, with no ETL platform and no AutoML. Heavy runs on Trino, Hive, Iceberg and Hadoop with DataSpring ETL (built on Flink) and AutoML built in, for enterprises that already operate, or need, that stack. Three questions settle it: do you already run Hadoop, Hive, Trino or Iceberg in production; can the first domain's data sit in one analytical engine; and do you need multi-step transformation pipelines or modelling rather than scheduled imports. When two answers point the same way, that is your edition.
What both editions share
OT stands for Ontology. Both editions ship ontology management: the layer that binds business objects, definitions, rules and actions to the tables that hold the data. The OT product brief states the shared design principle in one line: "a real compute engine sits underneath (Apache Doris in Flash; Trino and others in Heavy)", so an answer is the result of SQL that actually ran, not a figure estimated by a language model. Both editions also expose the ontology to external AI agents over MCP.
The ontology concept, the agent interface and the SQL-backed answer are therefore constant. What changes is the engine, and everything that follows from the engine: how data arrives, how far the system scales, and who has to operate it.
The architectural difference, stated precisely
| OT Flash | OT Heavy | |
|---|---|---|
| Positioning on the editions page | "One engine. A fast start." | Continues the Hadoop, Hive, Trino and Iceberg architecture, with ETL and AutoML built in |
| Query engine | Apache Doris | Trino (distributed SQL queries) |
| Storage and warehouse | Apache Doris internal tables | Apache Hive (data warehouse), Apache Iceberg (open table format), Apache Hadoop (distributed storage and compute) |
| Ontology management | Yes | Yes |
| Data pipelines | Not offered; table-level import tools instead | DataSpring ETL, built on Flink |
| AutoML and AI-driven modelling research | Not offered | Yes |
| AI agent access (MCP) | Yes | Yes |
| Recommended for | One business domain running quickly; data that fits a single analytical engine | Large data volumes, an existing big-data platform, a need for pipelines and modelling |
Sources: the editions and download page and the OT product brief. The GitHub repositories say the same thing in fewer words: the Flash repository describes itself as "a lighter and faster version of DataFocus, without ETL and AutoML capabilities", and the Heavy repository as "a more comprehensive version of DataFocus, featuring ontology management, ETL, AutoML, and full-stack data warehouse capabilities".
The engines explain why the two editions have different shapes:
- Apache Doris is, in the project's own overview (accessed November 2026), "an MPP-based real-time data warehouse". Frontend nodes handle query planning and metadata; backend nodes handle storage and execution; data is split into shards with replicas across backend nodes. Storage and compute live in the same system, which is what "one engine" means operationally.
- Trino is a query engine without storage of its own. Its concepts page describes a coordinator "responsible for parsing statements, planning queries, and managing Trino worker nodes", and connectors that adapt Trino "to a data source such as a data lake using Hadoop/Hive or Apache Iceberg". Its use-cases page adds that "Trino was not designed to handle Online Transaction Processing (OLTP)": it is an analytic layer over storage you already own.
- Apache Iceberg calls itself "a high-performance format for huge analytic tables" on its project site, and its partitioning docs explain hidden partitioning, where consumers "don't need to know how the table is partitioned", against Hive, where partitions are explicit columns a query must filter on.
In Flash, the whole footprint is visible on one monitoring page: the r57 manual's System Manager lists one running container each for the web server, search, BI server, Doris FE, Doris BE, PostgreSQL, Redis and Nginx. In Heavy, the data layer is a distributed stack with a coordinator, workers, a warehouse and a storage tier, which is the right shape when the data already lives there and the wrong shape when it does not. System requirements and supported platforms for each edition are in its release notes; plan resources around data volume, concurrency and model requirements.
"No ETL in Flash" does not mean "no external data in Flash"
The editions table lists ETL and AutoML for Heavy and not for Flash. That describes what each edition offers; it says nothing about whether Flash can reach your databases. It can.
The Flash r57 user manual documents a built-in source, "DataFocus system Doris", plus external sources grouped as Database, API and HTTP. The database list includes PostgreSQL, MySQL and MariaDB, Oracle, SQL Server, IBM DB2, ClickHouse, Hive, Iceberg, Delta Lake, TiDB, SAP HANA and others. Connecting one takes a source name, server, port, database, optional schema and JDBC parameters, and credentials; the manual recommends a dedicated read-only database account limited to the tables you intend to analyse. Import is a three-step wizard (choose the source, choose the tables, set import type and an optional schedule), and a Table Task tab holds the scheduled import jobs. Local files arrive as CSV, Excel or JSON. Once tables are in Doris, Flash offers table-level shaping: Join Worksheet, Merge Worksheet for stacking same-structure tables, SQL views with an editor, datasets that combine related tables, and formulas.
What Flash does not have is DataSpring, which the Heavy manual calls "the unified batch and stream ETL platform built into OT Heavy". In DataSpring, logic is a task flow of processors on a canvas, run manually or on a schedule. Every processor output must carry a primary key or the flow will not run (a source table without one gets a generated flink_uuid column), and a Lineage button shows where each column came from across the flow. That is a different class of tool from an import wizard and a SQL view.
The decision rule follows directly. If the transformation you need can be written as "copy these tables into Doris on a schedule, then join or define a SQL view over them", Flash covers it. If you need multi-step flows with per-column lineage, stream processing, or a pipeline that several teams maintain, that is what DataSpring exists for, and it is in Heavy.
One piece of background: the Doris 4.x lakehouse overview (last updated May 2026) says Doris can connect to Hive, Iceberg, Hudi, Paimon and JDBC databases. That is an upstream Doris capability, not a documented Flash feature. Treat it as context, not as a promise.
A worked example, hypothetical
Consider a hypothetical mid-size distributor whose sales orders sit in MySQL and whose ERP runs on Oracle. The sales director wants margin by customer and region, and wants the sales team to ask for it in plain language.
The Flash path is short. A DBA creates read-only accounts on both databases scoped to the order, customer, product and cost tables. In Flash, an administrator adds the two sources, runs the import wizard for each with a nightly schedule, and builds a Join Worksheet or SQL view that produces one margin table in Doris. The sales ontology is modelled over that table and the related tables, and the agent answers from SQL run against Doris. The only data-engineering decision is refresh cadence, which the editions page tells Flash buyers to plan explicitly: how data reaches Doris and how often it is refreshed.
Now change one fact. Suppose the same distributor is part of a group that already keeps several years of order, shipment and inventory history in Hive tables on a Hadoop cluster, and finance wants a reconciled daily margin table built from four systems, with lineage it can show an auditor. Copying that history into a single Doris engine nightly would be the wrong design; the data already lives in a warehouse the group operates. Heavy keeps the ontology next to that estate, queries it through Trino, and builds the reconciled table as a DataSpring task flow whose Lineage button answers the auditor's question column by column. The same distributor, two different estates, two different editions.
Agent access in both editions
Both editions list MCP access for AI agents, and the documentation is organised around Flash.
- Flash has a published Agent/MCP manual (r56). It states that agent table queries are Doris SQL, and that the original Trino, Hive and Hadoop services and Heavy ETL are not among Flash's entry points: an agent must not call them.
- Heavy provides MCP access as well, and its administration surface for agents is more detailed in the manual. The Admin area has an External Agents tab that sets which kinds of write operations external agents may initiate, with agent change approvals and MCP data administrator access. An external agent's effective write scope is the intersection of three things: the External Agent operation setting, the exact OAuth grant range, and EDIT access on the system resource; the tab does not create a second permission system. The roles list includes dedicated MCP roles such as
MCP_EXECUTION_RUN_ADMINandMCP_DATA_ENTRY_AUTHORING_ADMIN. For the integration mechanics, the Heavy manual's documentation index points to the Flash manuals.
Two practical notes apply to both. Heavy is the edition that includes AutoML and AI-driven modelling research; if those are on your requirements list, the comparison is over. And DataFocus publishes no public hosted MCP endpoint and no live sandbox, so agent evaluation happens against your own installed instance, with its tool schema coming from that authenticated instance.
Licensing and distribution: the node question
Both installation images are distributed free on GitHub, without website registration or payment; the release notes tell you how to select a platform, verify checksums and install. Start from the editions and download page, then the Flash and Heavy repositories. A cloud option exists separately, with registration and activation through a console.
The download page sets the licence terms in three sentences: "The Community Edition runs on a single node with unlimited users and no license fee"; "the software is closed source, and its terms ship with each package"; and the Community licence "does not cover servers, model calls or operations". Enterprise subscriptions add multi-node and cluster deployment, private tickets, upgrade and migration help, and long-term security patches.
The node clause matters more for Heavy than for Flash. A Flash deployment is a single-node design by nature: one Doris FE and BE alongside the application containers. A Hadoop-class deployment is usually multi-node, and multi-node and cluster deployment sit on the Enterprise side of that page. Read the licence, service scope and node definitions in the documentation that ships with the edition you select before you size a Heavy cluster.
The three questions as a decision table
| Question | Yes | No |
|---|---|---|
| 1. Do you already run Hadoop, Hive, Trino or Iceberg in production? | Heavy: put the ontology on the estate you have | Go to 2 |
| 2. Can the first domain's data sit in one analytical engine? | Flash: one Doris engine, scheduled imports | Heavy |
| 3. Do you need multi-step pipelines with lineage, stream processing, or AutoML, beyond scheduled imports and SQL views? | Heavy: DataSpring and AutoML are there | Flash |
Two answers in the same column name the edition. A split verdict, which usually means "no Hadoop today, but pipelines or AutoML required", is a Heavy requirement with a Heavy operating cost: you will be standing up a Hadoop, Hive, Trino and Iceberg stack to get DataSpring and AutoML. Decide whether the pipelines justify the platform, or whether the first domain can start on Flash while that question is settled.
Trade-offs to accept either way
Flash is one engine. Its scale is the scale of a single Doris deployment, its data arrives by import rather than by pipeline, and refresh cadence is your design decision. Heavy is a platform. It assumes an operating team for Hadoop, Hive, Trino and Iceberg, and a Community licence that covers one node. On both: the sources available for this post publish no benchmarks, no prices and no compliance certifications, so the comparison has to be made on architecture and fit. Before moving between editions later, confirm target versions, source mappings and ontology export scope.
Next step
Read the editions and download page for both editions' positioning and licence terms, then the OT Flash user manual chapter on external data sources to see the import path before you size anything.