Loading...
Modernization of COBOL batch applications to TypeScript, Node.js architecture, and cloud. ATLAS methodology, proven functional parity.
Migrating COBOL to TypeScript is a trajectory increasingly chosen by IT departments looking to move from a monolithic mainframe to a cloud-native serverless architecture. TypeScript offers two decisive advantages over Java or .NET on the right perimeters: a drastically reduced execution cost through serverless runtimes (Cloudflare Workers, Azure Functions, AWS Lambda, Deno), and a higher velocity of evolution thanks to the npm ecosystem and a structural typing more flexible than Java.
TypeScript fits particularly well with COBOL batch perimeters (overnight processing, reporting, ETL, extracts) and moderate-volume transactions (under one thousand transactions per second). Typical sectors: growing fintechs that do not want to invest in heavy Java infrastructure, telecom and media oriented serverless, digitized administrations that have adopted cloud-first APIs, niche banks looking to quickly exit a costly mainframe. For very high-volume transactions (more than ten thousand TPS), look instead at COBOL to Java or COBOL to .NET Core.
Our internal POCs have converted more than 15,000 lines of COBOL to TypeScript on Cloudflare Workers across three representative cases (CardDemo 7,650 lines, Portfolio 7,972 lines, CBSA 7,900 lines). The observed ratio is 10:1 (10 lines of COBOL per 1 line of TypeScript) — higher than the Java ratio (7:1) because TypeScript avoids Spring boilerplate. Vibe coding assisted by AI (Claude Code, GitHub Copilot) accelerates conversion by 2 to 3 times in this modern language. See the ATLAS methodology applied on these POCs.
Serverless architecture, moderate volumes, minimal operating cost, global edge distribution. The default choice for agile fintechs and telecoms.
Classic containerized architecture, need for fine control over infrastructure, on-premise or multi-cloud compatible.
Modern ecosystem, security by default, native TypeScript without compilation. Suited to greenfield projects or migrations where an emerging stack is acceptable.
Very high transactional volumes, strong in-house Java skills. See [COBOL to Java](/parcours/cobol-to-java).
Microsoft ecosystem dominant. See [COBOL to .NET Core](/parcours/cobol-to-dotnet-core).
A COBOL to TypeScript migration is typically structured over **eight to eighteen months** depending on volume and criticality. The typical cell combines a **serverless cloud architect**, a senior **TypeScript tech lead**, three to five **TypeScript developers** ideally with a Java or C# background, a QA specialized in characterization tests, and a delivery director. For **twenty to sixty thousand lines of COBOL**, plan **six to ten months** with a cell of **four to six people** — TypeScript allows a more compact cell than Java or .NET thanks to the 10:1 ratio.
Using the JavaScript number type to port COBOL PIC S9(n)V9(n) and COMP-3. Floating-point rounding errors are guaranteed on any financial calculation.
Systematic mapping of PIC and COMP-3 to the decimal.js library or BigDecimal via BigInt. Financial calculations are encapsulated in a Money class with explicit scale and rounding mode. Unit parity tests on overflow, underflow, division by zero, comparison with COBOL outputs on representative datasets.
Reproducing sequential COBOL batch processing in TypeScript without exploiting native Node.js parallelization. The result is a slow batch that does not benefit from cloud advantages.
Explicit parallelization with Promise.all, worker_threads, or job queue (BullMQ, Cloudflare Queues). Splitting large batches into idempotent jobs orchestrated by a scheduler. Observability via structured logs (Pino, Winston) and metrics (Prometheus, Cloudflare Analytics).
Neglecting the difference in cloud billing model. COBOL batch bills on infrastructure, serverless bills per execution. A poorly optimized batch can become expensive at scale.
Cloud cost benchmark integrated in the target architecture phase. For very large batches, arbitration between Cloudflare Workers (edge, low latency) and Azure Container Apps or AWS Fargate (long-running, predictable cost). Monthly cost estimate validated before deployment.
Porting sequential COBOL files (VSAM, QSAM) to naive relational tables without rethinking the model. Positioned access becomes SQL full scans, catastrophic performance.
Relational modeling adapted to target queries. VSAM indexed files become tables with composite indexes suited to the access patterns identified during discovery. Sequential COBOL queries are ported to TypeScript streams with cursor-based pagination to avoid full memory loading.
Declaring the migration complete without validating end-to-end workflows in production with real volumes. On our Portfolio POC, the first delivery was validated through API tests but the browser dashboard did not work because the remote D1 database was not initialized.
Principle E7 — browser and workflow validation mandatory before delivery. Deployment checklist including: schema initialized in target environment, seed data loaded, API smoke tests passed, browser smoke tests passed, complete workflow replayed. This checklist emerged from honest self-criticism of our POCs. See the ATLAS methodology.
Internal Access POC: migration of 7,650 lines of CICS COBOL (AWS CardDemo, Apache 2.0) to TypeScript on Cloudflare Workers. Six COBOL programs, twenty copybooks, eight D1 tables, eighteen REST endpoints, HTML dashboard with 8 tabs. Functional parity validated on 6/6 V7 browser workflows, three discrepancies detected including one business-critical (overlimit/expiry cascade).
Internal Access POC: migration of 7,972 lines of COBOL (37 programs + 20 copybooks) to TypeScript on Cloudflare Workers with Hono and D1. Eighteen REST endpoints, HTML dashboard with 9 tabs, five discrepancies detected including two business-critical (SELL cost_basis and position delete).
Yes, subject to volume. TypeScript via Node.js or Cloudflare Workers is perfectly suited to moderate-volume transactions (up to one thousand TPS) with sub-second latencies. For higher volumes, look at COBOL to Java or COBOL to .NET Core. The real sweet spot of TypeScript for COBOL: batch and event-driven workloads where the serverless model shines.
Cloudflare Workers runs in 325 cities worldwide at the edge, with a cold start latency of a few milliseconds (vs several hundred for AWS Lambda). It offers a simple pricing model (per request) and a native D1 integration (distributed SQLite) for relational storage. Azure Functions is preferable if your IT is already in the Microsoft ecosystem. AWS Lambda if the AWS ecosystem dominates. Cloudflare Workers if you start from scratch and seek the best performance/cost ratio for serverless edge.
Never use the JavaScript `number` type for financial calculations. Use decimal.js (mature, widely adopted) or BigDecimal via BigInt (more performant but less ergonomic). Encapsulate any financial operation in a Money or Decimal class with explicit scale, currency, and rounding mode. Unit parity tests on each calculation with comparison to COBOL outputs on representative datasets. decimal.js covers 95 percent of cases.
Yes, generally even better. A sequential COBOL batch can be ported to TypeScript with parallelization via worker_threads or Promise.all, and executed on a cluster (Kubernetes, Cloudflare Queues). On our POCs, migrated batches are 2 to 5 times faster than their COBOL equivalent thanks to parallelization and co-location with the database. Exception: batches with strong sequential dependency (accounting with cumulative totals) where linearity is intrinsic.
For thirty thousand lines of batch COBOL in nearshore co-delivery, plan five hundred thousand to seven hundred thousand euros parity tests and documentation included — cheaper than Java or .NET thanks to the 10:1 ratio that reduces target code volume. For one hundred thousand lines, budget one to two million euros. See the delivery models for contractual formats.
Vibe coding designates legacy code conversion assisted by AI (Claude Code, GitHub Copilot, Cursor) framed by our ATLAS methodology. AI accelerates pattern-by-pattern translation by 2 to 3 times but does not replace discovery or characterization tests. We measured these gains on our 10 POCs, documented in our article Vibe coding: AI-augmented consultants.
We frame the trajectory, the budget, and the deliverables in a first thirty-minute conversation. A short POC can be proposed before committing to the full program.
Start this path →