Comparison
Big bang migrationvsParallel run
Big bang migration
everything switched over on one weekend, and the rollback plan was a sentence in a document.
Moving everything from the old system to the new one at a single moment. It is sometimes unavoidable — some data models and some vendors do not allow a gradual path — and it should be recognised as the highest-risk option available, because every problem arrives at once and the rollback window is the same weekend. Where it cannot be avoided, the mitigation is rehearsal: run the cutover against production data more than once before the real one.
Full entry →Parallel run
both systems processed every transaction for a month and you only looked at where they disagreed.
Running the old and new implementations side by side on real input, with only the old one's output used, and comparing results. It is the most convincing evidence available that a replacement is correct, because it tests against real data rather than the data you imagined. It costs double compute and a genuine comparison mechanism, and the discipline is deciding in advance what disagreement rate means 'switch'.
Full entry →