When TechVenia asked me to migrate SciVenia and Mubazar to AWS, the brief was simple: "make it more reliable, ship faster, don't break anything." The brief was simple. The work was not.
Here is what I would tell myself if I started over.
Treat the pipeline as the product
The first instinct is to wire up CI/CD just well enough that builds pass and deploys happen. That is a trap. Every minute spent on the pipeline pays back tenfold in operator hours saved.
I rewrote the GitLab CI from scratch with four jobs: lint → test → build → deploy. Each runs in isolation, fails fast, and produces a versioned image tag. Rollback is redeploy --tag previous. That single capability — the rollback button you trust — paid for the whole migration.
Secrets do not belong in the repo
The first hour of the migration was moving every credential out of .env files into AWS Secrets Manager. The code reads from a single getSecret adapter. Six months later when we needed to rotate the Stripe keys, it took 90 seconds.
The lesson: refactoring secrets out later costs more than doing it on day one. There is no later.
The dashboard ships with the deploy
Every deployment now creates its own CloudWatch dashboard linked from Slack. Engineers see exactly what they shipped, what regressed, and what stayed flat. Mean time to detect a regression dropped from "next morning" to "next 15 minutes".
The rule I would put on the wall: a deploy without observability is a guess.
What I would change
I optimized too early on a few things — multi-region failover before we had multi-region traffic, Redis cluster mode before we needed it. The quote that fits: "Premature optimization is the root of all evil." Knuth was right then; he is still right now.
If I started today, I would defer those decisions until the first real incident demanded them. Optionality is cheap; complexity is not.
Closing
You do not need a platform team to ship reliably. You need a small set of disciplined choices: the pipeline you trust, secrets you can rotate, dashboards you can read, and the courage to defer everything you cannot justify yet.
That is most of what I learned. The rest is muscle memory built in production.