mix aurora_meter.bench (Aurora Meter v1.0.0)

Measures Aurora Meter and writes a machine-readable record of what it measured.

mix aurora_meter.bench <mode> [options]

Modes

kind is the column to read first. A micro mode isolates an in-memory path with stubbed or absent persistence; an end-to-end mode reaches Postgres through the real adapter on every write. The two differ by orders of magnitude, so a figure quoted without its kind says very little.

Mode	Kind	What it measures
`spread`	micro	`Counter.incr/4` on one distinct key per worker
`hot`	micro	`Counter.incr/4` on a single shared key
`reserve`	micro	`AuroraMeter.reserve/3`, admitting and denying
`with_quota`	micro	`AuroraMeter.with_quota/4` with a no-op callback
`cluster_2_sim`	micro	gossip apply cost in one VM (not a cluster result)
`record`	end_to_end	`AuroraMeter.record/4`, one event per call
`record_batch`	end_to_end	`AuroraMeter.record_batch/2` at 1, 10, 100 and 500
`correct`	end_to_end	`AuroraMeter.correct/4` against pre-recorded originals
`replay`	end_to_end	`AuroraMeter.Events.Replay.run/1` over a seeded population
`credits_debit`	end_to_end	`AuroraMeter.Credits.debit/4` across distinct wallets
`credits_hot_wallet`	end_to_end	the same debit against one shared wallet
`flush_1k`, `flush_10k`, `flush_100k`	end_to_end	one `Flusher.flush/0` with that many dirty keys
`db_delay`	end_to_end	sustained load with a delay in front of every storage call
`db_recovery`	end_to_end	sustained load across a real database outage, and the drain
`cluster_2`, `cluster_4`	end_to_end	convergence and overshoot on real peer nodes

Options

--procs N workers, default 8
--per N operations per worker, default 500,000 micro and 2,000 end-to-end
--warmup N discarded operations per worker, default a tenth of --per
--rounds N repetitions for the modes whose unit is a whole operation (replay, flush_*), default 5
--duration MS the load window for db_delay and db_recovery
--json PATH write the record here
--label TEXT a name for this run, default the mode
--repo MODULE the end-to-end repo
--history / --no-history whether day buckets are maintained
--seed N recorded in the record; random when not given
--sample-every N micro latency sampling rate, default 1,000
--delay-us N the per-callback delay for db_delay
--broadcast-interval MS, --flush-interval MS override the mode's own defaults. The cluster overshoot guarantee is stated in terms of one broadcast_interval, so being able to vary it is how that guarantee is measured rather than assumed
--allow-any-database run against a database whose name does not end in _bench
--quiet print nothing but the verdict

The database

End-to-end modes run against their own database, whose name must end in _bench, with an ordinary pooled connection. Set AURORA_BENCH=1 so this repository's test repo resolves to aurora_meter_bench, and create it with mix bench.setup. A repo configured with Ecto.Adapters.SQL.Sandbox is refused outright: the sandbox serialises every worker onto one owned connection, so the run would measure the sandbox.

The deprecated positional form

mix aurora_meter.bench 8 500000 still runs, as spread --procs 8 --per 500000, and prints a deprecation notice. It is removed in 2.0.