Stream: what it does in a system design

Stream

An ordered log split into partitions. Each consumer group reads every message at its own position, in order within a partition, and a message is kept until retention ends, not until it is read.

Settings

Instance tier: Sizing label for this stream instance (small to xlarge). It is shown in the metrics; in the current simulation it does not change append or read speed.
Partitions: How many ordered lanes the log is split into. A key always lands in the same lane, so order is kept per key. A group reads each lane one message at a time, so more partitions allow more parallel work.
Keep messages for (ms): How long the stream keeps messages. A group that falls further behind than this loses the messages it never read. Empty keeps everything for the run. Longer retention is safer but costs storage.
A group starts from: Where a consumer group begins reading: 'earliest' replays everything the stream still holds, 'latest' only sees messages published from the moment it joins. Each group can override it.
Attempts per message: How many times a message is tried before it is skipped and dead-lettered. While a message is being retried, everything behind it in the same partition waits. A group can override this.
First retry after (ms): How long to wait before retrying a failed message. Each later retry waits twice as long, and the whole partition is held up meanwhile (head-of-line blocking). A group can override this.
Consumer groups: Per-group overrides. Each consumer group reads every message at its own pace and position, so a slow group builds its own lag without affecting others. Unlisted groups use the stream defaults.