Tools we will defend, and the ones we stopped using.

The default for analytical reads at scale. Column codecs, materialised views and TTL tiering give us three separate levers on cost before anyone talks about bigger machines.

Still the right answer for transactional workloads and anything that needs real constraints. We keep it and stop asking it to do analytics.

Fine to arrive on, expensive to stay on when the workload is a small number of queries run very often. We migrate off it when the profile says so, not by default.

Partitioning that survives a spike and consumer lag you can see before a customer does. Most of the pain we get called about is configuration, not the broker.

Worth it when the connector already exists and the transformation is simple. Not worth it when the pipeline logic belongs in code you can test.

The scheduler is the point. We autoscale on request load rather than CPU, because CPU is a proxy that lies during a spike.

Provisions the node the pod actually needs instead of the node group somebody sized last year. The savings are real and they are also easy to lose to bad consolidation policy.

Throughput per GPU is where the inference bill is decided. Continuous batching moves that number more than a bigger instance does.

Predictable latency without garbage collection pause spikes, which is usually why the gateway rewrite happens in the first place.

Every environment reproducible, every change reviewable. The alternative is a console click nobody remembers making.

Versioned manifests that a new engineer can read on their first Monday. Boring on purpose.

Answers the question that matters after an incident: which model version served that request, and what trained it.