← Back to the journal

DataFocus · Journal

Self-hosting DataFocus OT Flash: seven decisions to settle before the install window

The seven decisions DataFocus asks you to settle before self-hosting OT Flash, each with an owner, the artifact it produces, the mechanism behind it, and a first-hour verification sequence taken from the r57 manual.

Before you install DataFocus OT Flash on your own servers, seven decisions need an owner and a written answer: the usage scope, the operating system and CPU architecture, capacity, data authorization, the model service, data egress, and who owns backups. That list is DataFocus's own, from its deployment documentation, which also says to take minimum resources from the installation guide of the version you choose. System requirements and supported platforms are in each release's notes, published on the download page and in the OT Flash GitHub repository. The rest of this post gives each item an owner, explains the mechanism behind it, and ends with a verification sequence for the first hour after install.

This is a preparation checklist, not an installation guide. It draws on DataFocus's published documentation, the r57 user manual and the OT Flash README; nothing here was measured on a running system.

Who owns which decision

OT Flash is self-hosted and an administrator configures the large-language-model service, so no single team can sign off on the whole list. Assign the seven items before kickoff, and ask each owner for one artifact rather than a verbal yes.

DecisionOwnerArtifact to produce
Usage scopeBusiness owner, platform leadOne named business domain and its first three questions
OS and architectureInfrastructureThe pinned release and its release-notes platform line
CapacityInfrastructureA sizing sheet with one row per resource and a pilot date
Data authorizationData owner, securityPer-source read scope and a read-only service account
Model servicePlatform lead, securityHosted-where decision, key custody, fallback
Data egressSecurity, complianceWhat may leave the network, under which policy
Backup ownershipOperationsSnapshot target, credential rotation, restore rehearsal date

Is the scope really one domain?

Write the usage scope as a single business domain, such as purchase approvals or order delivery, and list the first three questions users will ask of it. DataFocus's edition notes position Flash for a team that wants one domain running quickly, with data that can sit in a single analytical engine: Apache Doris serves as both the compute engine and the warehouse. OT Heavy continues the Hadoop, Hive, Trino and Iceberg architecture with ETL and AutoML built in, for large volumes, an existing big-data platform, and a need for pipelines and modelling.

That gives a decision rule for the first meeting. If the request is "one domain, data we can land in one engine", Flash fits. If it is "all of the company's data" or "we need pipelines first", you are describing Heavy, and the rest of this checklist changes. The sizing sheet, the connector list and the model-service decision all follow from this answer.

Which release, on which platform?

Supported operating systems and CPU architectures are listed in the release notes of the specific OT Flash version you install. Two consequences follow:

  1. Pin the release first, then read its notes for your platform, and record both in the change record. A sizing sheet that says "latest" is not a sizing sheet.
  2. Re-read the release notes on the morning of the change window; if the pinned release has moved, the window moves with it.

The README adds a step teams often discover mid-install: new installations require instance-specific product authorization and model configuration. Request the authorization before the window opens. Downloads need no website registration or payment; the release notes tell you how to check file integrity with checksums, and that check belongs in the runbook.

What actually consumes capacity?

The download page states resource requirements as "plan for your data scale and workload", and the release notes carry the system requirements for each release. Beyond that, DataFocus's guidance is to plan around data volume, concurrency and model requirements. To turn that into a sizing sheet, you need to know what the installed system is made of.

The r57 manual's System Manager monitoring page lists the service containers: df-webserver, df-search, df-biserver, df-doris-fe, df-doris-be, df-pgsql, df-redis and df-nginx, one running instance each in normal operation, with Doris, PostgreSQL, RabbitMQ and Redis as infrastructure services. The two Doris containers matter most for disk. In the Apache Doris documentation (accessed November 2026), frontend nodes handle requests, planning and metadata while backend nodes hold storage and execute queries. The README says Doris stores imported business data in internal tables, so the backend's data directory grows with what you import, not with the size of the source systems you connect to.

A sizing sheet for Flash therefore has at least these rows, each with a source for the number and a date to validate it:

  • Imported tables. Which tables, how many rows today, and the growth expected over the pilot. Size the data disk for these, not for the source databases.
  • Documents. The README describes document workers that parse, chunk, embed and index documents. Budget for the originals and their derived indexes, and ask about the installer's storage layout before you partition disks.
  • Concurrent users and scheduled jobs. DataFocus publishes no concurrency or latency figures, so none can be borrowed. Count expected simultaneous users and scheduled imports, then validate on a pilot.
  • Model calls. Admin > Device authorization shows daily and monthly AI model token usage after install. Leave a placeholder and fill it from that page during the pilot.

One licensing fact belongs on the same sheet. DataFocus's release policy describes one active DataFocus Node Unit per node, enforced by a local hard limit, with unlimited human users. The Community Edition runs on a single node with no license fee; multi-node and cluster deployment are listed under Enterprise subscriptions. If a capacity row says "more than one node", that is a commercial conversation to start early.

How does data get in, and who said it could?

Plan how data reaches Doris and how often it refreshes. The README lists connectors for MySQL, PostgreSQL, SQL Server, Oracle, DB2, ClickHouse, SAP HANA and OceanBase, with compatible TiDB, Doris and Greenplum, plus CSV import; the r57 manual adds Excel and JSON files, API sources and HTTP sources, and a Table Task tab for scheduled imports. For each source on your list:

  1. Name the data owner and write down which tables and columns the ontology may read.
  2. Create a read-only database account for OT Flash. The r57 manual recommends a dedicated read-only account with access only to the databases and tables to be analyzed, so this is the documented path, not extra caution.
  3. Decide the refresh schedule per table, since scheduled import is configured per table task.
  4. Import one real table from that source during the pilot before committing to the schedule.

Inside the product, visibility follows roles. The r57 manual's preset roles are SYSTEM_MANAGER (global configuration, users and roles), SOURCE_MANAGER (all resources and their grants), DEPARTMENT_MANAGER, DEVELOPER and LOG_MANAGER, and administrators can add custom roles with per-resource permissions. Users can be created singly, imported in bulk from a template, or synchronized from AD/LDAP, and the menus a user sees depend on the roles assigned. Before install day, name the first SYSTEM_MANAGER, the first SOURCE_MANAGER and the first ten business users, and decide whether AD/LDAP sync is in the pilot or a later phase.

Which model, and what leaves the building?

Ontology agents work only after an administrator has configured a large language model service. DataFocus does not bundle or name a model, so this decision is yours, and it is the one most likely to stall an install if security hears about it for the first time during the window. Answer five questions in writing:

  • Is the model hosted inside your network, or by an external provider?
  • If external, what is sent to it? DataFocus's own advice is to confirm this against your data policy.
  • Who holds the API key, and where is it stored? The System Manager's Interface authentication page manages keys for system interfaces and ships with one entry, ADM_CENTER_SECRET; its rule that keys are never pasted into chats or documents should extend to the model key.
  • Who watches usage? Admin > Device authorization shows daily and monthly token consumption.
  • What happens when the provider is unreachable? ontology agents depend on the configured model service, so decide what the first domain does while it is down.

The egress question has a regulatory edge for some readers. Consider a hypothetical team in mainland China whose first domain includes customer records. Under Article 38 of China's Personal Information Protection Law (Stanford DigiChina translation, law effective 1 November 2021), personal information may be provided outside mainland China only after one of four conditions is met, such as a regulator-organized security assessment or a standard contract with the overseas recipient. A model endpoint outside the country is a cross-border transfer in that reading, which is why the hosted-where question needs compliance in the room rather than a note afterwards.

Then write down the cost side. DataFocus's own wording is direct: "You pay for servers, model calls and third-party services." The Community Edition license does not cover servers, model calls or operations.

Will agents connect, and what does the edge need?

If coding agents or other MCP clients will use the instance, add one row to the network plan. OT Flash exposes external agents to ontology data through Remote MCP at /mcp, with a separate Business Workspace at /mcp/business; authorization uses the OAuth authorization code flow with PKCE, and logged-in users can see and revoke their connections at /oauth/grants. The Agent/MCP manual's proxy guidance is specific: preserve the Authorization, Content-Type, MCP-Protocol-Version, Mcp-Method and Mcp-Name headers and approved Origins, and never forward the public /mcp path to admin ports or another product. The r56 Agent manual says requests whose MCP headers do not match the message body are rejected with HTTP 400, so hand the network team this list before they write the configuration. If no agents will connect in the first phase, record that too.

What is the backup and upgrade plan?

Name a backup owner before install, not after the first import. The README describes Doris online snapshots to S3-compatible storage and warns that a restore interrupts service. If that is your route, decide where the bucket lives, who rotates its credentials, and when you will rehearse a restore on a non-production host. A snapshot nobody has restored is a hypothesis.

For upgrades, the r57 System Manager has an Upgrade Management page showing the product name, current version and latest version. With no online update source configured it tells you to use a verified release package. It offers automatic upgrades daily, weekly or monthly at a set time, or "Do not upgrade", which the manual marks not recommended. Decide now whether automatic upgrades fit your change process; for a production node most teams will want manual, verified-package upgrades and a rollback plan. If agents are connected, the Agent/MCP manual adds a step after each upgrade: rediscover tools and versions and compare schemas, because old grants do not automatically gain new scopes. That line belongs in the upgrade runbook.

The first hour after install

The manual gives you enough instrumented surfaces to verify an install without inventing tests. Run these in order and record each result.

  1. Reach the login page. Users open an address given by the administrator, for example https://<server address>:31600, in the latest Chrome, which the manual recommends. The first administrator login forces a password change; complete it and store the credential in your vault.
  2. Open the System Manager at /manager.html. It has its own operations account, separate from Admin, and its first login also requires changing the initial password. No vendor default should survive the first hour.
  3. Read System metrics. Node status shows disk, CPU (user, system, idle, I/O wait), memory and per-mount disk usage; confirm the figures match the host in your sizing sheet and that each service container shows one running instance.
  4. Set alarm thresholds for nodes, pods, Doris, PostgreSQL, RabbitMQ and Redis, choose in-app or email delivery, and note that the email cooldown default is 60 minutes.
  5. Configure mail. Password recovery needs a configured mail server, and so do email alarms; do this before the first user forgets a password.
  6. Configure the model service and ask the ontology agent one question about a small imported table. The manual says an answer usually takes tens of seconds, longer for complex questions.
  7. Check Operate Logs, visible to LOG_MANAGER or SYSTEM_MANAGER, for your own login and the import. Entries show host, type, status, detail, user and time, which is also what a support ticket needs: DataFocus asks for the product version, the steps taken and the expected result.

When Flash is not the right fit

Flash on a single Doris engine suits one domain with data you can land in one place. If the first request needs data pipelines or AutoML, DataFocus lists those for Heavy, not Flash, and Heavy is the right starting point. For installers and release notes, see the download page and the OT Flash manuals before you commit hardware.