The summary.
ORDERING IS GUARANTEED WITHIN A PARTITION, NOT A TOPIC
Events for one entity must share a partition, keyed by entity identifier, or they may be processed out of order.
Partition count sets maximum parallelism, and increasing it later is disruptive.
EXACTLY-ONCE REFERS TO THE EFFECT, NOT THE DELIVERY
The practical approach is at-least-once delivery with idempotent processing. Commit offsets late — duplication is recoverable, loss is not.
AGGREGATING BY PROCESSING TIME PRODUCES WRONG RESULTS WHEN DATA IS DELAYED
Use event time, set watermarks from observed lateness, and decide deliberately what happens to late records.
A DEAD LETTER DESTINATION IS ESSENTIAL
Without it, one unprocessable record blocks its partition indefinitely.
CONSUMER LAG IS THE DIRECT MEASURE OF WHETHER PROCESSING KEEPS UP
NEVER PUT SENSITIVE DATA IN AN EVENT THAT CONSUMERS DO NOT NEED
Events are retained and widely readable.
BATCH FOR MOST THINGS; STREAM ONLY WHERE LATENCY GENUINELY DEMANDS IT
And if both exist, check automatically that they agree — divergence destroys confidence in both.
