The owned keyspace

EchoMQ 2.0 drops BullMQ wire compatibility on the first rung, and the most visible part of that break is the keyspace. The v1 line — echo/apps/echomq v1.3.0, the line the fork freezes — speaks the emq: prefix and interpolates the queue name verbatim: the shipped EchoMQ.Keys.base/1 returns "#{prefix}:#{name}" (keys.ex:97-98), key_prefix/1 returns "#{prefix}:#{name}:" (keys.ex:85-86) — no wrapping, no validation, no placement. A queue named payments produces emq:payments:wait, emq:payments:active, and the rest: flat strings with nothing for a thread-per-shard datastore to balance on.

emq.1 ships the owned keyspace. Every per-queue key becomes emq:{q}:<type> — the emq: prefix replacing emq:, the queue name brace-wrapped transparently by the core (a caller still writes Queue("orders") and the core places emq:{orders}:*). The wrap is idempotent (an already-braced caller name is not double-wrapped) and validated (no colon inside a hashtag; emq is rejected as a queue name). The braced base {emq}: is reserved for the core's own cross-queue keys — so the core's namespace and any queue's cannot collide (INV3). Every queue is slot-local by construction.

Take the fixed queue payments. The v1 line, today, produces the emq: families — the verbatim-interpolated builders in EchoMQ.Keys: the LIST keys :wait / :active / :paused, the ZSET keys :delayed / :prioritized / :completed / :failed / :waiting-children, the SET :stalled, the STREAM :events, the HASH :meta, and the per-job emq:payments:<jobId>:lock via EchoMQ.Keys.lock/2. These are flat: no hashtag, nothing to place across threads. This is the line the fork freezes at 1.3.0.

emq.1 ships the v2 form of each. emq:payments:wait becomes emq:{payments}:wait, emq:payments:active becomes emq:{payments}:active, … emq:payments:<jobId>:lock becomes emq:{payments}:<jobId>:lock. The emq: prefix replaces emq:; the queue name carries its {…} hashtag, applied transparently by the core; the field set is otherwise the same. The hero figure shows each line's key set side by side — toggle it and watch the prefix flip and the hashtag appear.

emq.1 ships the transparent hashtag so a queue's keys are placed precisely. A datastore that locks on the hashtag routes a key by the substring inside the first {…} — the hashtag. So every v2 key of the payments queue — emq:{payments}:wait, emq:{payments}:active, emq:{payments}:<jobId>:lock — carries the same tag, payments, and lands in the same slot. That is what lets a v2 Lua script touch several of a queue's keys in one atomic multi-key transition — the R2 atomicity floor, now placement-native.

The core's own cross-queue keys live under the reserved {emq}: base — {emq}:version, {emq}:nodes — and group into a second, distinct slot: the queue's and the core's. They cannot collide, because emq is rejected as a queue name (INV3). A property test over arbitrary queue names checks the rule holds: wrap°wrap = wrap, and every per-queue key shares exactly the queue's one hashtag (emq.1-D1). The figure below extracts the {…} tag from each key and groups by tag — the queue's keys to one slot, the core's to the other.

The v1 line interpolates the queue name verbatim under emq: (EchoMQ.Keys, keys.ex:85-86) — flat, unplaced, frozen at 1.3.0. emq.1 ships emq:{q}:<type>: the emq: prefix replacing emq:, the per-queue {q} hashtag applied transparently (idempotent, validated), the braced base {emq}: reserved for the core's own cross-queue keys, and emq rejected as a queue name so the namespaces cannot collide (INV3). Every queue is slot-local by construction, so a hashtag-locking datastore places one queue per thread and distinct queues across cores.

Where the owned keyspace is grounded: the BullMQ wire the v1 line speaks and the fork leaves, the Dragonfly hashtag-locking the transparent brace targets, the cluster hashtag rule, and the chapter's own pages.