Observability
High-cardinality metric rollup
Counters, gauges and histogram buckets from thousands of agents, rolled up exactly per series.
Measured 2026-06-21 on ALINX/HamGeek AX7020 (XC7Z020-2CLG484); reproduce with abench 65537 0xDEADBEEF0BADF00D. Throughput is the engine's accumulation rate (the aggregation step); end-to-end depends on your ingest path, which ATOMiK's order-independence frees from global ordering / locks.
This is you if
- You run a Prometheus / Datadog / StatsD-shape metrics pipeline.
- You shard or lock to keep per-series counters exact under load.
- High-cardinality / hot series is your scaling pain.
The workload
- Thousands of agents emit per-series counter & gauge deltas and histogram-bucket increments.
- Deltas arrive out of order over lossy transport, concurrently across the fleet.
- Each series must roll up to an exact value per scrape / flush window.
- Millions of active series — high cardinality.
Today — serialized behind a lock
To keep per-series values exact under concurrent, out-of-order arrival, the rollup serializes per series (locks or a sharded hot path). Hot series contend; throughput is capped by the lock, not the silicon.
With ATOMiK — order-independent, lock-free
Agents feed deltas into a shared accumulator in any order. The merge is order-independent and lock-free, and the rolled-up value is byte-identical regardless of arrival order or how many banks process it.
- No per-series lock or hot-shard contention — agents never coordinate.
- Linear scaling with allocated banks (1 / 2 / 4 / 8×).
- Exact, byte-identical rollups under reordering and out-of-order arrival.
- No global sequencing or backpressure on ingest — order is irrelevant.