Because Redpanda gives each core its own shard, sizing mistakes show up as skew, not as a general slowdown. These are the questions we put to a team, and the reason for each.
How many cores per broker, and how many partitions per core? Partitions are spread across cores. Too many partitions per core increases Raft heartbeat and memory overhead and slows leader elections after a restart. Too few leaves cores idle. We compare your partition count with the core count and the target throughput, and suggest the headroom to leave for growth.
What is the disk, and is it tuned? Redpanda is designed for local NVMe with XFS. Network-attached volumes with low IOPS caps or burst credits behave well in a test and badly in a sustained load. We run a disk benchmark with rpk tooling and the tuners (rpk redpanda tune) before sign-off, and record the result.
What replication factor and what acks? A replication factor of 3 is typical for production. acks=all waits for a Raft majority. Mixed settings across producers explain much of the "it was fast in testing" variance. We check the producers against the topic's replication and the cluster's failure domains (rack awareness).
What is the retention plan, local and tiered? Local retention should be sized to the longest consumer outage you want to ride out from local disk. Tiered retention extends history in object storage. Setting the local target too high fills disks; too low forces reads from object storage during recovery. We size both from ingest rate and recovery needs.
How do clients connect, and what do they assume? Advertised addresses matter for clients inside and outside the cluster network. TLS and SASL settings, and the client's metadata refresh interval, decide how it behaves when a broker restarts. We test a rolling restart with real clients.
Who watches the cluster, and what pages them? We help set the first alerts: unavailable partitions, disk free, under-replicated partitions, leadership change rate, consumer lag, and tiered storage upload backlog.