Measured results

Environment

Property	Value
distributions.pydepin	0.17.1
distributions.pytest	9.1.1
distributions.pytest-benchmark	5.3.0
host.available_processors	4
host.cpu_model	Intel(R) Xeon(R) CPU E5-2683 v4 @ 2.10GHz
host.load_average	2.41, 2.91, 2.83
host.machine	x86_64
host.processor	x86_64
host.processors	4
host.release	6.8.0-138-generic
host.system	Linux
interpreter.compiler	Clang 22.1.3
interpreter.free_threading	no
interpreter.hash_randomization	yes
interpreter.implementation	CPython
interpreter.recursion_limit	1000
interpreter.version	3.12.13

Latency

Workload	Repetitions	Rounds	Median	Spread across repetitions
test_latency[build_the_graph_view-depin]	5	392	4.232 ms	6.0%
test_latency[call_through_an_inject_wrapper-depin]	5	40729	5.626 µs	2.8%
test_latency[call_through_an_inject_wrapper-direct]	5	76576	122.610 ns	0.5%
test_latency[call_through_an_inject_wrapper_with_explicit_arguments-depin]	5	30824	6.720 µs	6.1%
test_latency[call_through_an_inject_wrapper_with_explicit_arguments-direct]	5	197004	176.795 ns	0.8%
test_latency[construct_a_singleton_for_the_first_time-depin]	5	44223	5.417 µs	5.9%
test_latency[construct_a_singleton_for_the_first_time-direct]	5	36925	296.300 ns	4.9%
test_latency[explain_a_deep_chain-depin]	5	259	6.457 ms	4.7%
test_latency[explain_a_deep_chain_with_every_node_decorated-depin]	5	102	16.902 ms	7.5%
test_latency[explain_a_layered_dag-depin]	5	343	5.003 ms	3.2%
test_latency[explain_an_unbound_key_of_16-depin]	5	208	7.408 ms	3.7%
test_latency[explain_an_unbound_key_of_20-depin]	5	207	7.387 ms	2.6%
test_latency[export_a_large_graph_as_dot-depin]	5	743	2.452 ms	5.3%
test_latency[freeze_a_chain_missing_a_provider_of_100-depin]	5	144	11.115 ms	4.9%
test_latency[freeze_a_chain_missing_a_provider_of_50-depin]	5	172	9.183 ms	4.3%
test_latency[freeze_a_chain_of_10-depin]	5	1000	379.478 µs	3.3%
test_latency[freeze_a_chain_of_100-depin]	5	466	3.546 ms	2.3%
test_latency[freeze_a_chain_of_1000-depin]	5	46	35.834 ms	6.1%
test_latency[freeze_a_decorated_chain_of_10-depin]	5	1000	922.849 µs	2.5%
test_latency[freeze_a_decorated_chain_of_100-depin]	5	190	8.699 ms	2.9%
test_latency[freeze_a_decorated_chain_of_1000-depin]	5	19	90.192 ms	3.4%
test_latency[freeze_a_generic_key_chain_of_10-depin]	5	1000	686.537 µs	2.6%
test_latency[freeze_a_generic_key_chain_of_100-depin]	5	248	6.688 ms	7.0%
test_latency[freeze_a_generic_key_chain_of_1000-depin]	5	20	85.924 ms	17.6%
test_latency[open_a_request_shaped_scope-depin]	5	8348	27.704 µs	2.1%
test_latency[open_a_request_shaped_scope-direct]	5	180151	796.863 ns	12.4%
test_latency[open_and_close_a_scope-depin]	5	2588	326.873 µs	4.9%
test_latency[open_and_close_a_scope-direct]	5	88842	3.133 µs	6.2%
test_latency[resolve_a_collection_of_10-depin]	5	12440	18.729 µs	1.4%
test_latency[resolve_a_collection_of_10-direct]	5	193163	202.132 ns	1.0%
test_latency[resolve_a_collection_of_100-depin]	5	5980	133.175 µs	2.4%
test_latency[resolve_a_collection_of_100-direct]	5	47547	417.099 ns	0.7%
test_latency[resolve_a_generic_key-depin]	5	60321	5.931 µs	2.3%
test_latency[resolve_a_generic_key-direct]	5	101823	95.370 ns	1.4%
test_latency[resolve_a_sync_resource_with_teardown-depin]	5	10424	12.901 µs	3.4%
test_latency[resolve_a_sync_resource_with_teardown-direct]	5	149656	1.577 µs	1.3%
test_latency[resolve_a_transient_chain-depin]	5	9228	34.752 µs	2.9%
test_latency[resolve_a_transient_chain-direct]	5	165452	2.275 µs	2.7%
test_latency[resolve_an_async_singleton-depin]	5	17542	16.582 µs	6.6%
test_latency[resolve_an_async_singleton-direct]	5	21515	13.499 µs	3.0%
test_latency[resolve_cached_singleton-depin]	5	42635	1.817 µs	1.9%
test_latency[resolve_cached_singleton-direct]	5	102219	96.030 ns	0.2%
test_latency[resolve_cached_singleton_through_an_alias-depin]	5	71933	3.747 µs	5.9%
test_latency[resolve_cached_singleton_through_an_alias-direct]	5	102596	95.180 ns	1.7%
test_latency[resolve_singleton_through_a_two_deep_decoration_chain-depin]	5	130839	1.781 µs	4.6%
test_latency[resolve_singleton_through_a_two_deep_decoration_chain-direct]	5	102187	95.180 ns	0.8%
test_latency[resolve_through_an_active_override-depin]	5	91711	3.844 µs	1.9%
test_latency[resolve_through_an_active_override-direct]	5	96759	99.581 ns	4.1%
test_latency[resolve_with_no_active_override-depin]	5	119561	1.815 µs	1.8%
test_latency[resolve_with_no_active_override-direct]	5	23303	95.030 ns	1.0%
test_latency[warmup_a_cold_singleton_graph-depin]	5	142	11.945 ms	5.3%
test_latency[warmup_a_cold_singleton_graph-direct]	5	2981	262.381 µs	10.1%

Application tier

Tail quantiles and CPU time are published for the application tier only. An end-to-end request has a tail a caller meets; a microbenchmark round is a calibrated loop, so its p99 describes the calibration rather than the operation. CPU is reported and not gated: process CPU on a shared runner carries the runner's noise, and the deterministic metrics already carry what can be gated exactly.

Workload	Repetitions	Rounds	Median	p95	p99	CPU	Spread across repetitions
test_latency[fastapi_application_startup-depin]	5	751	2.462 ms	2.816 ms	3.977 ms	2.347 ms	2.6%
test_latency[fastapi_application_startup-direct]	5	1000	1.611 ms	1.907 ms	3.047 ms	1.554 ms	4.6%
test_latency[fastapi_async_resource_teardown-depin]	5	1281	758.806 µs	852.651 µs	1.014 ms	760.685 µs	5.8%
test_latency[fastapi_async_resource_teardown-direct]	5	1486	642.123 µs	709.753 µs	869.443 µs	640.911 µs	2.4%
test_latency[fastapi_cpu_light_endpoint-depin]	5	1127	726.118 µs	824.394 µs	968.206 µs	744.329 µs	4.8%
test_latency[fastapi_cpu_light_endpoint-direct]	5	1340	646.634 µs	733.901 µs	885.088 µs	634.950 µs	3.5%
test_latency[fastapi_endpoint_with_work-depin]	5	1000	970.261 µs	1.073 ms	1.234 ms	985.039 µs	1.2%
test_latency[fastapi_endpoint_with_work-direct]	5	1000	867.000 µs	963.845 µs	1.119 ms	854.845 µs	2.9%
test_latency[fastapi_request_scoped_graph-depin]	5	1000	763.957 µs	867.273 µs	1.017 ms	746.031 µs	3.3%
test_latency[fastapi_request_scoped_graph-direct]	5	1425	648.046 µs	739.813 µs	889.659 µs	642.627 µs	4.7%
test_latency[fastapi_singletons_and_transients-depin]	5	1288	736.115 µs	857.765 µs	999.305 µs	729.399 µs	4.5%
test_latency[fastapi_singletons_and_transients-direct]	5	1376	642.044 µs	728.115 µs	884.272 µs	657.609 µs	3.6%

Work

Workload	Python calls per operation
allocations_of_a_cached_singleton_resolution	9
allocations_of_a_request_shaped_scope	105
allocations_of_a_scope_cycle	396
allocations_of_a_transient_chain	202
allocations_of_an_inject_call	29

Allocations

Workload	Blocks per operation	Bytes per operation	Peak bytes
allocations_of_a_cached_singleton_resolution	13	1168	2056
allocations_of_a_request_shaped_scope	27	2360	4360
allocations_of_a_scope_cycle	76	6240	14496
allocations_of_a_transient_chain	53	4968	5752
allocations_of_an_inject_call	17	1360	2064

Retained memory

Workload	Bytes held
retained_by_a_frozen_container_of_100	34568
retained_by_a_frozen_container_of_1000	326000
retained_by_a_warm_singleton_cache_of_1000	397560
retained_by_an_open_scope_of_20	10512

Scaling

Curve	Size	Cost per operation	Growth over the previous size
scale_async_teardown	10	171.166 µs	—
scale_async_teardown	20	270.509 µs	1.58x
scale_async_teardown	40	467.681 µs	1.73x
scale_freeze_graph_size	100	3.557 ms	—
scale_freeze_graph_size	200	7.099 ms	2.00x
scale_freeze_graph_size	400	14.159 ms	1.99x
scale_override_nesting	8	2.299 µs	—
scale_override_nesting	32	3.932 µs	1.71x
scale_override_nesting	128	10.834 µs	2.76x
scale_resolve_collection	10	23.939 µs	—
scale_resolve_collection	100	191.340 µs	7.99x
scale_resolve_collection	200	380.021 µs	1.99x
scale_resolve_fan_out	10	18.662 µs	—
scale_resolve_fan_out	20	37.439 µs	2.01x
scale_resolve_fan_out	40	71.075 µs	1.90x
scale_resolve_transient_depth	10	19.710 µs	—
scale_resolve_transient_depth	40	203.929 µs	10.35x
scale_resolve_transient_depth	160	768.570 µs	3.77x
scale_scope_teardown	10	101.862 µs	—
scale_scope_teardown	20	191.945 µs	1.88x
scale_scope_teardown	40	372.958 µs	1.94x

Retired measurements

Measured once, no longer measured. A workload withdrawn without a record is indistinguishable from one that was never written.

Workload	What it claimed	Why it was retired	What covers the path now
scale_failing_freeze	The complexity class of the failing-freeze path, as the growth ratio between graph sizes.	The path is dominated by a constant that does not depend on graph size: suggest_candidates scans sys.modules when the error is built. Measured on the pull-request runner with both sides on identical code, the curve read 7.095, 7.021 and 7.026 ms at sizes 25, 50 and 100 — flat across a fourfold range — and the difference between the two identical revisions reached +23.61% against a 15% budget. The scan also depends on how many modules each process loaded, which is not a property of the revision under test. The curve was valid before the walk it watched was repaired; the repair is what left the constant in charge.	tests/unit/test_longest_chain.py::test_failing_freeze_does_not_grow_cubically_with_the_chain_length, which compares 200 providers against 400 — the sizes at which the walk overtakes the constant — and reads 1.80 repaired against 5.91 with the cubic walk restored. It replaced a half-second wall-clock budget that the same seeded walk passed at 0.42 s, on a host faster than the one that budget was written on. The fixed-size latency workloads freeze_a_chain_missing_a_provider_of_50 and _of_100 cover the path as well.
scale_explain_missing_key	The complexity class of the missing-key walk, as the growth ratio between graph sizes.	The same constant, reached through render. The published reference-host dataset already recorded the curve as flat — 5.479, 5.503 and 5.444 ms at sizes 10, 12 and 14, growth 1.00x and 0.99x — while the number of simple paths through those graphs grows Fibonacci in the size. A curve that does not move where the quantity it claims to track quadruples is not measuring that quantity.	tests/unit/test_longest_chain.py::test_explain_of_an_unbound_key_does_not_grow_with_the_path_count, which compares a 16-node fan-in-2 DAG against a 24-node one — eighteen times the simple paths — and reads 1.00 repaired against 24.94 with the enumerating walk restored. The fixed-size latency workloads explain_an_unbound_key_of_16 and _of_20 cover the path as well.

Refused measurements

Asked for by the performance proposal and not measured here, with what an honest measurement would need in its place.

Case	Why it is refused	What it would need
fastapi_no_injection	The installed no-injection route exists only in the FastAPI optimization head; the accepted dataset predates it, so an archived baseline cannot supply a paired observation. It remains a head-only diagnostic and is reported separately rather than silently treated as paired evidence.	A five-repetition head-only diagnostic collection alongside the next accepted paired FastAPI run.
fastapi_lazy_host_publication	The accepted dataset predates the lazy FastAPI host-publication component diagnostic.	A dedicated component collection in a new accepted FastAPI evidence dataset.
fastapi_lazy_frame_activation_and_drain	The accepted dataset predates the lazy FastAPI frame-activation component diagnostic.	A dedicated component collection in a new accepted FastAPI evidence dataset.
fastapi_endpoint_program_one_key	The accepted dataset predates the one-key compiled FastAPI endpoint-program diagnostic.	A dedicated component collection in a new accepted FastAPI evidence dataset.
fastapi_endpoint_program_many_keys	The accepted dataset predates the multi-key compiled FastAPI endpoint-program diagnostic.	A dedicated component collection in a new accepted FastAPI evidence dataset.
fastapi_request_seed_read	The accepted dataset predates the lazy FastAPI request-seed component diagnostic.	A dedicated component collection in a new accepted FastAPI evidence dataset.
fastapi_async_resource_close	The accepted dataset predates the async FastAPI resource-close component diagnostic.	A dedicated component collection in a new accepted FastAPI evidence dataset.
Long-running allocation and retention drift.	Retention here is a point-in-time reading. Drift is only visible over a soak, and how much runner time a soak may consume in a blocking pull-request gate is a budget decision rather than a methodological one.	A scheduled job with its own time budget, not a check on the pull-request path.