The Problem
A logistics technology company was running a nightly batch job that consolidated records from six upstream systems into a single reporting database. The job had grown to over four hours in execution time, had no retry logic, and a single failure anywhere in the sequence would invalidate the entire night’s run. Deployments required a maintenance window because the monolith held shared state that could not be safely updated while the batch was running. On-call engineers were waking up to failures two to three times a week.
The Approach
The batch process was decomposed into a set of independently deployable Azure Functions, each responsible for a single upstream system. Events were published to a shared topic in Cosmos DB change-feed format, allowing downstream consumers to process records as they arrived rather than waiting for an overnight window. Each function was designed to be idempotent, with structured retry policies and a dead-letter path that preserved failed records for inspection without halting the rest of the pipeline. Application Insights correlation IDs tied every step back to the originating source record.
The Outcome
The nightly four-hour batch was replaced by a continuous pipeline that keeps reporting data within fifteen minutes of the upstream source at all times. Deployments became zero-downtime by default, as individual functions can be updated independently without coordinating a maintenance window. On-call incident volume halved within the first month, and the remaining pages were all handled within minutes rather than requiring multi-hour investigations. The team reported a significant reduction in deployment anxiety and an improvement in confidence when pushing changes.