Within one cluster a stream has three copies, but the leader counts only two as caught up — what does that distinction mean?

## Two questions that sound like one When someone says a stream is "kept in three copies", they have answered only one question. Within a single cluster, a stream's data is stored more than once: one copy is the **leader**, holding the authoritative sequence of records that writers append to, and the others are **follower copies** on other nodes, each continuously pulling records from the leader so that it holds the same sequence. That gives an operator two separate facts, and on a running cluster they drift apart constantly: - **How many copies exist?** This is the **copy count** — a placement decision, chosen when the stream is created and changed deliberately. It is stable; nothing about the day's traffic moves it. - **How many of those copies are current right now?** This is the **caught-up set** — the copies the leader presently treats as close enough to itself to be holding the stream. It moves minute by minute, without anyone touching a setting. ## Why a copy falls behind A follower copy is not a magic mirror; it is a process reading from another machine over a network and writing to a disk. Anything that slows that loop opens a gap between **the position that follower has replicated to** and the leader's newest record: - a disk that has begun retrying, or a volume shared with a noisy workload; - a saturated link, especially when the copy sits in a different failure domain from the leader; - a long pause on the node — a stall, a reclaim of memory, a scheduling hiccup; - a node that has just restarted and has minutes of records to make up; - a leader accepting records faster than that particular follower can consume them. None of these corrupts the follower. It is simply out of date, and how far out of date changes second by second. ## What the caught-up set is for The leader keeps this set because durability claims have to be made against copies that actually hold the newest records. A copy that is four minutes behind cannot stand behind a record written ten seconds ago — if the leader's volume died right now, that record would not be on it. | Fact | What it tells you | What moves it | |---|---|---| | Copy count | how many stored replicas of the stream exist, and where | an operator changing it, or copies being moved between nodes | | The caught-up set | how many of those are current enough to be counted right now | a slow disk, a saturated link, a pause, a restart, a traffic burst | | A copy outside the set | still exists, still holds everything it pulled so far | closing its gap and being re-admitted, or falling further behind | The practical consequence is that a stream configured for three copies can be running, perfectly healthily in every dashboard sense, with a caught-up set of two — or of one. The configured number is what you asked for; the set is what you have. ## Entering and leaving the set Leaving is not a manual act and not a failure alarm. Platforms that maintain such a set define an allowance — **the catch-up window** — expressed either as a number of records or as elapsed time since the copy was last current. A copy past that allowance is dropped from the set; a copy that closes its gap back inside the allowance is re-admitted. Both events are ordinary on a busy cluster. A dropped copy is not deleted, not fenced off, and not broken. In designs where a follower pulls continuously it keeps pulling; whether it re-enters depends on whether it can read faster than the leader is writing, which is not guaranteed. ## Designs differ on what the set even is This picture — a leader maintaining a list of current followers — is the shape used by platforms that replicate from a leader to followers. It is not universal. Some platforms commit a write when a majority of copies has answered, with no standing list of who is current; there, being behind is decided per write rather than by membership. Others put durability in an underlying shared store, so brokers hold no per-record copies of their own and the question does not arise in this form. When you move between platforms, check which shape you are looking at before you read a copy count as a durability promise. ## What a junior should carry away Three copies is a statement about storage. Two caught up is a statement about **now**. If a write is supposed to be safe because several machines hold it, the number that matters is the second one, and it is the one nobody configured.