On a mainframe, the programs, batch processes, files and their scheduling have grown together, on a single platform, sometimes over several decades. A mainframe migration therefore moves a whole, of which the screens that users see are only one part.
Five options are open to an organisation that is questioning its large system: stay on it, rehost the programs on another platform, convert the code to a mainstream language, rewrite the application, or replace it with a software package. Four criteria decide between them: the criticality of what the machine carries, the volume of code and batch processes, the skills available and the dependencies. The choice is made domain by domain, and several options can coexist on the same system.
A whole held together by the machine that carries it
The word mainframe, also referred to as a large system, refers to the central computers on which organisations run their transactions and mass processing. Programs, written in COBOL or PL/I for example, respond to the day’s transactions. Chains of batch processes, launched by a scheduler, run one after another, at night for example, within a bounded time window. The data reside in databases and files in formats specific to the platform.
Two features can weigh on an exit project. The data may be encoded in EBCDIC, an encoding whose character sort order may differ from that of ASCII or Unicode: the result of sorts and the numeric formats of amounts, rounding included, need checking during conversion. The programs, for their part, share files and databases: removing a single one requires knowing who reads and who writes what.
The options for a mainframe migration and what each involves
Staying is a decision in its own right. It involves a recurring cost, a succession plan for the people who know the system, and a review date. Transactions exposed through APIs can open the system to new uses without touching its core.
Rehosting recompiles the programs, with few or no modifications, to run them on standard servers or in the cloud, within an environment that reproduces the services of the large system. It frees you from the machine without changing the code: the language, the structure of the programs and the technical debt come along. It creates a dependency on the vendor of that runtime environment. When the target is a cloud provider, the workstream also follows the rules of cloud migration.
Code conversion translates the programs, using a tool, into a mainstream language such as Java. The AWS Transform documentation describes code analysis, documentation generation, business logic extraction and the conversion of COBOL programs into Java applications, with points where a person validates or completes. Google Cloud’s documentation presents an assessment tool, Mainframe Assessment Tool, which relies on generative AI to help understand the code and extract business rules from it, which a person reviews and validates, as well as a parallel execution set-up, Dual Run, which compares the old and new systems. Documentation consulted on 12 November 2026.
Converted code may retain the original procedural structure: it runs on a mainstream platform, without its readability for the team that will take it over being assured. What an assistant produces is reviewed and tested like any other code, according to the questions raised in our article on AI-assisted development.
Rewriting rebuilds the application from its business rules, which here may be spread across the programs, the batch processes and their scheduling. It makes it possible to revisit the data model and the functional breakdown, involves the business over the long term and implies a long coexistence of old and new.
Replacement with a software package abandons the code. The effort shifts to measuring what the product does not cover and to data migration; our article custom development or software package gives the criteria.
Rehosting, rewriting and replacing appear, outside the large system, among the options compared in our article on legacy application modernisation.
Criticality, volumes, skills, dependencies
Criticality sets the level of evidence. The more an outage or a wrong result weighs on the business, the more the demonstration of equivalence occupies the project. Rehosting and conversion aim for identical behaviour: the old and the new process the same files and the same transactions, and any difference in results is recorded. A rewrite or a software package change behaviour by design, and comparison gives way to acceptance testing.
Volume rules out options. Suppose two domains on the same machine: one has 200 programs and around ten batch processes, the other 2,500 programs and 600 processes chained every night. The first can be rewritten or replaced. For the second, a manual rewrite in a single piece would have two systems coexist for its entire duration: rehosting, tool-assisted conversion or a stepwise breakdown are to be examined first. Batch processes are counted separately: each has its input files, its rank in the chain and a duration to meet, on the target, within the night window.
Skills are read from both sides. Who can still explain the rules and the sequence of processes? Who will operate the target? Rehosting keeps the language, and therefore the need for developers who practise it. Converting shifts this need to the target language, provided that people understand both worlds during the transition. The signs of dwindling knowledge are described in our article on application obsolescence.
Dependencies draw the exit order. Files exchanged with other systems, databases read by several domains, flows to partners: a domain only leaves on its own if its links with those that remain hold during the coexistence.
An exit by domains, until the machine is shut down
None of these options requires a single cutover. An exit by domains follows the strangler fig pattern: a single entry point routes transactions between the large system and the target, domain after domain. On a mainframe, the dividing lines are better drawn along the data than along the screens: two domains that write to the same file leave together, or agree on which one is authoritative.
Each step is prepared with the non-regression evidence and the rollback described on our page on application modernisation; here they apply to transactions as well as to batch processes, execution time included.
The end point of the project is the shutdown of the machine. As long as a domain remains on it, the hardware, the software and the operations team must still be funded, in addition to those of the target. The fate of the last domain, the one nobody wants to touch, is therefore decided at the start, even if its exit comes last.
The cases where staying is defensible
Staying is defensible when the manufacturer maintains the machine and its system, skills succession is organised, the cost is known and accepted, and no change expected by the business runs up against the platform.
Postponing is also defensible as long as the inventory of dependencies has not been done, or the people able to explain the existing system cannot devote themselves to the project.
The first piece of work is therefore this inventory, drawn up by domain: programs, batch processes and their chains, data and formats, flows, people who know. Read against the four criteria, it gives each domain an option, an expected piece of evidence and an exit rank.