Fixes on hard legacy surfaces

The problem

Most of what I do is maintenance in the real sense. A customer hits something broken on a product that predates everyone currently working on it, and I own it from the first reproduction through to the merged fix. Finding where the fix belongs usually takes longer than writing it.

What I did

A few that were worth the time they took.

There was a crash in the record merge path of a two-decade-old desktop CRM, triggered by merging records in a particular state. It was native C++, and I worked it through crash dumps from customer machines rather than a reproduction I could attach a debugger to.

Calendar and mail connectors were creating duplicate records against Microsoft Exchange. The duplicates were the symptom. The cause sat in how the connector reconciled identity between the two systems, which meant the fix had to go where every sync path routed through instead of where the duplicates surfaced.

Spreadsheet exports were coming out corrupted. That one traced back to the transform layer rather than the export itself, which is where everyone had been looking.

On a high-traffic platform I rebuilt activity feed pagination on keyset pagination with pre-filtered common table expressions, correcting cursor uniqueness and index alignment, shipped by concurrent migration with no downtime.

The migrations were their own category: rewriting every polling worker in a Scala service for a new database driver, moving the datastore across two major versions, and replacing a decommissioned analytics vendor by rewriting its SDK calls as direct database queries.

What changed

580 of 591 assigned defects closed across six products, every change traceable to a tracked ticket and paired with a regression test that was observed failing first.