New Data Center purchases closed on 30 March 2026. Expansion ends 30 March 2028. End of life is 28 March 2029. Every remaining self-hosted Atlassian customer is being pushed to Cloud — and for a large number of European organisations, that is the one direction they cannot go.
Atlassian has published all three. There is no ambiguity left to plan around.
For most teams this is an inconvenience. For a regulated, air-gapped or residency-bound organisation it is a control you no longer hold.
Self-hosted, data residency is settled by where the server is. On Cloud it becomes a clause — a region commitment, a tier you have to be on to get it, and a sub-processor list that changes without your involvement.
For European organisations answering to a supervisory authority, the difference between “our data is in this rack” and “our vendor commits to a region” is the entire conversation.
Defence, critical infrastructure, industrial control, classified environments and segregated production networks do not have a Cloud option. There is no tier of a SaaS product that runs on a network with no route out.
For these organisations Data Center's end of life is not a migration project — it is the removal of the only product that was ever viable.
Self-hosted, you decide when a version lands, you test it against your integrations first, and you can stay on a known-good build through an audit or a peak trading period.
On Cloud, the vendor's release calendar is the release calendar. For an organisation with change-advisory discipline, that is a governance gap rather than a convenience.
Data Center was already priced per user — roughly $100,000 a year at 1,000 users and $182,000 at 2,000 on July 2026 list prices — and Cloud keeps that shape while adding Jira Service Management per agent and Atlassian Guard per user for SCIM and SAML.
Marketplace apps bill on total site users, not on the people who use the feature. Twelve people need timesheets and you licence all 1,000.
Self-hosting is not a legacy deployment option here. It is the headline capability, and the product is built around it.
A node's role lives in the database, in cluster_nodes, not in a properties file. The modes are ALL_IN_ONE (UI, REST and scheduler), FRONTEND_ONLY (Thymeleaf UI calling an external backend URL), BACKEND_ONLY (REST API, CORS-configured for the frontend) and SCHEDULER_ONLY (mail polling, SLA sweeps, garbage collection, report delivery, materialized-view refresh and OAuth token refresh).
Reassign a node's mode from the admin UI; with autoApply the node restarts itself to adopt it. Scheduler leadership is arbitrated so background work runs exactly once across the cluster.
Nodes self-register on startup and heartbeat, and are marked dead after three missed beats. You can forget, probe, enable and disable them individually, and designate a backup node.
The health check does not just report status — it returns a remedial action per verdict, across node registration, per-node liveness, frontend coverage, backend and API coverage, scheduler coverage, backup node presence, version skew across nodes, CORS correctness and mode source.
Java 25 on Spring Boot 4, GraalVM native-image capable, Flyway-managed schema, HikariCP pooling and Caffeine caching. The database is PostgreSQL only — deliberately one supported engine rather than a matrix of half-tested ones.
The first-run wizard walks six steps: test the database connection, install the schema, create the tenant, organisation and admin account in one transaction, set application properties, install the systemd service, done. Scheduling stays gated until setup is marked complete.
Backups come in organisation and tenant flavours, run asynchronously, stream a JSON export element by element, copy attachment files, package a ZIP and email the requester. Per-tenant settings cover enabled, cron and retention days, with an hourly scheduler deciding whether one is due and a daily cleanup of expired ones. Downloads are path-traversal-safe and access-scoped.
Software update checks an SFTP manifest for a newer release, downloads the JAR and SQL migrations with checksum verification — or you upload an update ZIP by hand. Applying copies the migrations into the Flyway directory and exits the JVM so the swap script replaces the JAR and restarts the service, with paginated update history.
The SFTP host, credentials and known-hosts entries used for updates are yours, stored in your database and testable from the admin UI. There is no seat metering to report, because the licence is priced on revenue rather than headcount.
AI is the usual air-gap casualty, and here it is not: alongside Anthropic and Gemini, fourteen self-hosted OpenAI-compatible runtimes are supported — Ollama, LM Studio, vLLM, llama.cpp, LocalAI, Jan AI, KoboldCpp, TGI, Tabby, LiteLLM and more, each with a default localhost base URL.
A curated configuration editor over an allow-list of application.properties keys with masked secrets and a confirm-with-password save. A migrations page that lists bundled migration files, shows the SQL, and can execute, mark or unmark them.
A log viewer with pattern and level filters, live tailing, download and clear. Staggered garbage collection across sessions, notifications, activity logs, orphaned attachments and audit logs. Outbound webhooks that are SSRF-guarded both at save time and at delivery time.
The stack a self-hosted Atlassian customer actually runs — Jira Data Center, Jira Service Management and the apps that fill the gaps. Atlassian list prices as published July 2026.
WAS, WAS IN, WAS NOT and CHANGED — for years, and anyone claiming otherwise should not be trusted with the rest of the table. TQL matches them and adds signed relative dates (-7d, -1q, +3d), a GIN full-text index over summary and description, live autocomplete on fields and values, and filters that can be shared organisation-wide and favourited — in one syntax, with validation that reports the character position of the error rather than silently dropping a clause.
Atlassian list prices as published July 2026, USD converted at approximately 1.08 USD/EUR. Trakr priced on the published revenue schedule. All figures ex. VAT.
Not every Data Center customer is at 1,000 users. Take a €100M-turnover organisation with 175 users and 35 service-desk agents, costed line by line on July 2026 list prices: Jira Software Standard at €13,000, JSM Standard at €7,800, Atlassian Guard at €7,800, Tempo Timesheets at €9,700, eazyBI at €3,100, ScriptRunner at €6,300 and Email This Issue at €3,100 — about €50,800 a year across seven vendors, seven renewal dates and seven data processing agreements.
The same organisation pays €13,000 for Trakr — a 74% reduction, with one contract and one sub-processor to assess. Structure, BigPicture and Advanced Roadmaps are deliberately excluded from that comparison: they solve portfolio and hierarchy planning problems that inflate the number and hand the incumbent an easy rebuttal.
Native importers, not a third-party ETL step and not a professional services engagement.
The Jira XML importer streams an RSS/XML backup with StAX and brings across items, comments and work logs, with DTD support and external entity resolution disabled — XXE and billion-laughs hardening is a property of the parser configuration, not a promise. The Jira CSV importer reads standard Jira export column names. A generic CSV importer covers everything else, auto-detecting field mappings from the header row across summary, description, type, priority, status, resolution, assignee, reporter, parent, due date, original estimate, time spent, created, updated, resolved, labels, components and external key, with 13 accepted date formats.
Three dates matter. New customers have not been able to purchase Data Center since 30 March 2026. Existing customers can expand their licences through 30 March 2028. End of life is 28 March 2029, after which Cloud is the only Atlassian option.
Not as a new customer — that door closed on 30 March 2026. If you already hold a Data Center licence you can expand it until 30 March 2028, which means any capacity you will need before end of life has to be bought inside that window.
It means the ticket system stops being infrastructure you control and becomes a third-party processor. Data residency becomes a contractual assurance rather than a physical fact, air-gapped and segregated networks stop being possible at all, and every upgrade happens on the vendor's schedule rather than yours.
If none of those matter to you, Cloud is a perfectly reasonable destination and you should take it. This page is for the organisations where at least one of them is non-negotiable.
A node's role lives in the database rather than a properties file, so one binary serves four deployment modes: ALL_IN_ONE, FRONTEND_ONLY, BACKEND_ONLY and SCHEDULER_ONLY. Nodes self-register on startup and heartbeat, and are considered dead after three missed beats.
The cluster page runs a health check covering node registration, per-node liveness, frontend coverage, backend and API coverage, scheduler coverage, backup node presence, version skew across nodes, CORS correctness and mode source — each verdict carrying a remedial action. The only external dependency is PostgreSQL, and the build is GraalVM native-image capable.
There is no seat metering to report, because Trakr is priced on your revenue rather than your headcount. Software update checks are a pull against an SFTP host whose address, credentials and known-hosts entries you configure and store yourself — and you can skip that entirely by uploading an update ZIP by hand.
AI is the usual air-gap casualty, and here it is not: fourteen self-hosted OpenAI-compatible runtimes are supported, including Ollama, LM Studio, vLLM and llama.cpp, each with a default localhost base URL.
Yes. There are native Jira XML and Jira CSV importers alongside a generic CSV importer. The Jira XML path streams an RSS/XML backup and brings in items, comments and work logs, with DTD and external entities disabled against XXE and billion-laughs attacks.
External users are mapped to Trakr users with suggested matches, the preview lets you select row by row, and every job is kept in history so a staged migration is straightforward.
Jira Data Center is roughly $100,000 a year at 1,000 users and roughly $182,000 a year at 2,000 users on July 2026 list prices. Trakr is priced on revenue: a group with around 1,000 Trakr users is roughly a €1B-turnover business on blended EU figures, which is €52,000 a year for unlimited users — and that figure already includes the service desk, SCIM, timesheets and reporting that Data Center bills separately.
No. Trakr's Support Desk and guest portal are part of the product rather than a separate SKU — passwordless portal sign-in over an emailed access code, a three-step submit flow, a ticket conversation thread, SLA policies with working calendars and breach escalation, and inbound email-to-ticket over Microsoft 365 Graph, the Gmail API, IMAP and Zoho.
Service requests, incidents, problems and changes are all tickets carrying a request kind, so there is no second record type, key format or permission model to learn.
From today there are roughly nineteen months until expansion closes on 30 March 2028 and thirty-one months until end of life on 28 March 2029. The constraint is rarely the import itself — it is procurement, security review and the parallel-running period, all of which are easier to schedule now than in 2028.
If self-hosting is not negotiable for your organisation, the Data Center clock is a procurement deadline rather than a technology one. Start now and it is a project. Start in 2028 and it is a scramble.