A multi-core wire-protocol load generator: one BEAM process per connection, so N connections spread across all schedulers (unlike a single-threaded event loop). It drives produce/fetch/mixed/stream load and control-plane (user/acl) load against a running Malachi server, measuring throughput and latency.

run/1 is the entry point (the mix malachi.loadtest task is a thin wrapper). Metrics are collected lock-free: an :counters array for ops/records/errors plus backpressure events (dropped connections, server-shed overloaded produces, quota-refused rate_limited produces, and reconnects) and a Malachi.Histogram for latency. A genuine error (any refusal other than those two) is also counted under the reason the server gave, in a table the report returns as error_reasons, so a run that records errors says which ones; the reasons always add up to errors. A worker connects and authenticates, waits at a barrier so all connections start together, runs its scenario for warmup + duration, and records only during the measured window. It is resilient: a shed produce backs off and continues, and a dropped connection reconnects (capped) rather than aborting.

How the connections are OPENED is a strategy (:connect_strategy), because each one pays a full credential verification on the server and opening hundreds at once is a self-inflicted auth storm:

• :bounded (default) - at most :connect_concurrency (default 32) connections inside connect+auth at once; the next starts when one finishes. Keeps each auth's individual latency near concurrency / server_verify_rate instead of the whole queue's drain time.
• :stagger - connection i starts i * :connect_stagger_ms (default 100) after the first, with no in-flight bound. A time ramp that does not adapt to the server's verify rate.
• :all_at_once - every connection starts simultaneously (the storm, kept on purpose for reproducing it and for comparing with historical runs).

A connection that fails to open under any strategy fails the run with a SetupError naming how many of how many failed, rather than crashing a linked worker.