Backstop control-path overhead — 2026-10-20

The only committed overhead figures for this project, with the host, the seed, the transport and both payload sizes.

The only committed overhead figures for this project. Measured with PYTHONPATH=src python3 benchmarks/local_overhead.py --requests 2000.

Conditions

The previous snapshot recorded the date, the seed, the mock transport and the request count, and nothing else. This one records the host, the OS, the interpreter and the package versions, because a latency figure without the conditions it was measured under is not a fact.

Host: Linux-6.6.87.2-microsoft-standard-WSL2-x86_64-with-glibc2.39 (x86_64)
Python: CPython 3.12.3
Packages: httpx 0.28.1, openai 2.46.0, anthropic 0.116.0, prometheus-client 0.26.0
Transport: local httpx.MockTransport, no network
Requests per payload: 2,000
Seed: none — the request bodies are deterministic per index

The numbers

Payload	Direct p50	Backstop p50	Overhead p50	Overhead p95	Overhead p99
94 B	0.1084 ms	0.2056 ms	0.0972 ms	0.1278 ms	0.1318 ms
4096 B	0.1184 ms	0.2196 ms	0.1012 ms	0.1342 ms	0.1105 ms

Reading these

Two payload sizes, deliberately. The 94-byte one is the body the previously published 0.07 ms figure used. The 4 KB one is closer to a real prompt. The overhead is body-proportional — the guardrail reads, parses and inspects every request, so a larger request costs more to guard — and a snapshot that only reported the small body would have made the guardrail look cheaper than it is.

These are not the old 0.07 ms. They are larger, and the reason is recorded: the previous figure was measured on a different host, without the host, the OS, the interpreter or the package versions, so it could not be checked on another machine or compared with anything. A number nobody can reproduce is not a better number for being smaller.

The overhead is the cost of the guardrail doing its job: reserving and reconciling a budget, checking the admission gate, recording the outcome, publishing the gauges, and (with the ledger on) pricing the request. Turning the ledger off removes its share; the rest is the ceiling itself.

What is in the 0.097 ms, measured on this host:

JSON body parsed	once (it used to be twice)
Five gauges published	twice, ~3.3 µs each time — the in-flight reading and the final one
One labelled counter + one histogram	~1.5 µs each
Quota header scan	only when a provider quota header is present (it used to rebuild a dict of every header, on every response)
Spend event + pricing	only when the ledger or the detector is on — ~17 µs, and it is off by default

Re-measure it yourself. The numbers above are one machine on one day:

backstop benchmark
PYTHONPATH=src python3 benchmarks/local_overhead.py --requests 2000

The overhead check inside backstop verify passes below a 5 ms p99 bound. That proves the control path stayed in the sub-millisecond class; it does not pin a figure.