Measured across the formats and paths teams actually deploy.

The guarded matrix exercises nested ecommerce events, six rule coverage tiers, payloads through 10KB, malformed input, lookups, windows, regular expressions, model scoring, vector distance, and decision reduction on a local Apple Silicon system.

5.80M

peak direct Arrow IPC rec/s

2.89M

peak HTTP Arrow IPC rec/s

400

test combinations

6-504

rules per tier

668

measured runs

input formats

100B-10KB

requested payload range

1M-10M

large-run record counts

HTTP agent throughput

End-to-end HTTP rows include request handling, format decode, rule evaluation, and configured result construction. Peak rows use the six-rule checkout ceiling workload; medians span all rule tiers and payload widths.

Format	Median rec/s	Peak rec/s	Median MiB/s	Max RSS
Arrow IPC	330,025	2,887,494	606.2	858MB
NDJSON	240,228	2,467,259	791.5	949MB
JSON array	173,720	2,563,311	573.5	959MB

HTTP does not currently accept Avro or Protobuf content types. Those formats are measured through the direct IO decoder path below.

Direct IO throughput

These rows isolate decoder and engine work without HTTP. Arrow IPC provides the highest ceiling because typed columns avoid JSON parsing and repeated scalar conversion.

Format	Runs	Median rec/s	Peak rec/s	Median MiB/s	Max RSS
Arrow IPC	102	302,304	5,803,305	501.8	9.8GB
NDJSON	102	149,750	996,964	419.7	1.2GB
JSON array	78	23,231	271,586	108.9	9.2GB
Avro	13	289,415	613,500	273.9	1.3GB
Protobuf	13	308,435	1,012,041	317.9	1.7GB

Rule count changes the planning number.

A six-rule ceiling run answers how fast the transport and columnar path can move. A 504-rule run is the better capacity-planning reference for a broad production policy.

Sustained 500-rule workloads generally landed near 150K-200K records/sec at million-row scale, depending on format, payload width, and signal mix.

HTTP Arrow IPC, 504 rules257,957 median rec/s

HTTP NDJSON, 504 rules213,652 median rec/s

HTTP JSON array, 504 rules160,682 median rec/s

What changes as payloads get wider

JSON pays to parse bytes

NDJSON remains streamable, but wider nested records increase parsing and projection work even when only referenced fields become columns.

Arrow trades CPU for memory

Typed buffers lift the throughput ceiling, while large combined batches raise resident memory. Standard Arrow string and binary offsets also impose a practical per-column size limit.

Queues shape tail latency

The largest HTTP combinations produced p99 spikes when requests queued behind evaluation and output. Throughput and request latency must be capacity-planned together.

Benchmark method

Numbers are useful only with the workload attached. The matrix keeps format, payload, nesting, rule coverage, record count, and output path visible.

Data

Deterministic nested ecommerce records with flat, nested, array-heavy, and wide shapes.

Signals

Strings, regex, lookups, windows, SQL expressions, ONNX output, vector distance, and decisions.

Scale

50K through 10M records, requested payloads from 100 bytes through 10KB, six policy tiers.

Environment

Local Apple Silicon arm64 Release build with the NEON backend. One run per combination.

Use the numbers as a capacity range, not a universal promise.

Hardware, output detail, model complexity, state locality, payload shape, and concurrent transport load all move the result.

Performance guide