When a stream is created without a declared definition, which values does it receive and who actually chose them?

## What a silent request still gets A stream cannot exist half-configured. Whatever the request omits, the platform supplies, so a stream created by a client call or by a bare name-only request is just as fully configured as one created from a careful declaration — it is simply configured by somebody else. The values stamped at birth are broadly the same across this product class, even though every platform names them differently: - **Copy count** — how many stored copies of a record the platform keeps. Some platforms express this as a number of replicas, some as a durability tier you buy, and some hide it entirely behind a managed offering. - **Retention** — the rule that decides how long a record survives, by age, by accumulated bytes, or by a latest-value-per-key rule. On some platforms this is a property of the stream; on others it is per subscriber; on a rented one it may have a ceiling you cannot raise. - **Parallelism count** — the number of parts a stream is divided into, on platforms that divide one. Platforms where consumers simply compete for a shared queue have no equivalent birth value at all. - **The owner record** — the team identity the stream carries. Where creation is unreviewed this is usually absent, because no field asked for it. ## Who chose them The honest answer is: whoever set the cluster up, at a moment when no workload existed. That is not carelessness — there is no better time available to them — but it has two predictable effects. First, the values tend to reflect **the situation of the installation**, not of production. A cluster first stood up to prove something works is configured to work on one machine, and a rented tier's values reflect what the vendor found safe to give everybody by default. Second, the values are **uniform across the estate** by construction. Every stream created without a declaration gets the same treatment, whether it carries one audit record an hour or the organisation's payment events. ## Why almost nothing prompts a revisit This is the part worth articulating in an interview, because it explains why a five-year-old estate looks the way it does. 1. **A value nobody chose produces no signal.** Alerts fire on symptoms — a full volume, a stalled reader, a failed write. None of them say "this number was never a decision". 2. **The stream works.** A stream running on a durability value that was never considered behaves identically to one running on a considered value, right up until the failure the value was supposed to survive. 3. **Nobody owns the question.** With no owner record, there is no team whose review would surface it, and platform teams review clusters rather than individual streams. 4. **Revisiting costs more than deciding did.** Whatever the change costs on your platform, it is strictly more than stating the value in the request would have cost, because at birth the stream had no readers, no writers and no stored records. That asymmetry is the whole reason this subject exists: **the birth moment is the only cheap moment.** ## Where platforms genuinely differ | Point | How it varies | |---|---| | Which values exist at birth | A parallelism count exists only where a stream is divided into parts | | Where the value comes from | A cluster-wide setting, a rented tier's fixed value, or a built-in constant | | How reversible each one is | Some values are adjusted freely afterwards; others are effectively fixed once traffic exists | | Whether an owner can be required | Some platforms can enforce a required field at creation; on others the requirement lives in the pipeline that applies the request | The safe way to state this in an interview is at the level of the mechanism rather than the number: say that every platform stamps something, that the set differs, and that the direction of the cost — cheap now, expensive later — is what is common to all of them. ## What to do about it The remedy is not to memorise good values. It is to make the birth moment carry a statement: a declared definition, holding at minimum the name, the owning team, and the values whose wrong setting is expensive on your platform, applied by a pipeline rather than typed by hand. Everything the definition states is a decision with a date and an author on it; everything it leaves out is a decision made on your behalf by someone who never saw your workload and will never hear that it was wrong.