README.md

---
license: cc-by-4.0
pretty_name: trace-bench
viewer: false
tags: [causal-discovery, root-cause-analysis, microservices, synthetic, benchmark]
---

# trace-bench v0.4.0

## Executive summary

Each named instance simulates a SaaS system at one of five sizes. A
backend-for-frontend (BFF) gateway sits in front of the backend services, and
every instance also has one external service. The table below gives each
instance's size, as set in `configs/instances/<instance>.yaml`.

| instance | services | endpoints / service | backend endpoints | BFF endpoints | scenarios | clients | simulated time | where it runs |
|---|---|---|---|---|---|---|---|---|
| xs | 3 | 3 | 9 | 4 | 2 | 20 | 1 hour | local (test fixture) |
| s | 10 | 5 | 50 | 8 | 4 | 200 | 1 day | local |
| m | 50 | 10 | 500 | 20 | 8 | 2,000 | 2 days | local |
| l | 150 | 10 | 1,500 | 30 | 12 | 6,000 | 3 days | local |
| xl | 350 | 12 | 4,200 | 60 | 20 | 20,000 | 5 days | cloud only |

## About

A simulated microservice trace benchmark with ground-truth causal graphs.
Each named instance ships a raw heterogeneous log feed (no parent pointers),
the correlated sequence views, the mechanism graph that generated the data
(latent state variables flagged), its latent projection over
`(operation, outcome)` tokens as the scoring target, the deployment call
topology as a candidate structural prior, injected-fault labels and the
oracle linkage that makes correlation loss a measured quantity. Every corpus
is wholly synthetic and reproduces byte-identically from its configuration,
seed, tool version and constants version (see each `manifest.json`).

Three variants per instance and seed, one simulation. `latent`: no state
variable is observable. `twin`: every latent value is written onto the records
it influenced, with a state-change log; its target is a DAG over event and state
tokens. `metrics`: the latent corpus plus a method-readable channel `metrics/`
of 30-second categorical time series of traffic intensity, per-service load and
pool state and per-endpoint health — what a real cluster exports — while cache
contents and per-session network and auth state stay hidden; it is derived from
the latent corpus of the same seed (`manifest.derived_from`), and its target is
the partial projection: state tokens of the exposed groups as sources, bidirected
edges only through the hidden groups. The score gap across the three isolates
what each degree of observability costs a method.

Generator (MIT): https://github.com/alex-chadyuk/trace-bench — tool version 0.3.0, 0.4.0.
Corpus licence: CC BY 4.0. Fitted realism constants: 1.
Machine-readable metadata: `croissant.json` at the repository root (Croissant 1.0; every file class as a file set, the tabular files as record sets; validated with `mlcroissant`).

| instance | variant | seed | exposed / hidden latent groups | alphabet (realized, train) | mechanism nodes | directed / bidirected at floor | edges on a Monte-Carlo estimate (max SE) |
|---|---|---|---|---|---|---|---|
| l | latent | 0 | none / all | 7435 | 8382 | 73837 / 50291 | 0 / 23984 (0.002499367) |
| l | latent | 1 | none / all | 7513 | 8382 | 73107 / 58068 | 0 / 23703 (0.002497846) |
| l | latent | 2 | none / all | 7533 | 8382 | 72764 / 60578 | 0 / 24375 (0.002487043) |
| l | latent | 3 | none / all | 7477 | 8382 | 68885 / 56822 | 0 / 24311 (0.002447562) |
| l | latent | 4 | none / all | 7430 | 8382 | 71266 / 58899 | 0 / 24374 (0.002498424) |
| l | metrics | 0 | health, intensity, load, pool / auth, cache, net | 7435 | 8382 | 96939 / 1085 | 9 / 4571 (0.002499367) |
| l | metrics | 1 | health, intensity, load, pool / auth, cache, net | 7513 | 8382 | 96153 / 2506 | 9 / 3988 (0.002497846) |
| l | metrics | 2 | health, intensity, load, pool / auth, cache, net | 7533 | 8382 | 96129 / 1772 | 9 / 3861 (0.002487043) |
| l | metrics | 3 | health, intensity, load, pool / auth, cache, net | 7477 | 8382 | 92252 / 1361 | 9 / 4275 (0.002447562) |
| l | metrics | 4 | health, intensity, load, pool / auth, cache, net | 7430 | 8382 | 94403 / 1755 | 9 / 4232 (0.002498424) |
| l | twin | 0 | all / none | 7435 | 8382 | 98678 / 0 | 6 / 0 (0.002499133) |
| l | twin | 1 | all / none | 7513 | 8382 | 98141 / 0 | 6 / 0 (0.002497846) |
| l | twin | 2 | all / none | 7533 | 8382 | 97944 / 0 | 6 / 0 (0.002487043) |
| l | twin | 3 | all / none | 7477 | 8382 | 93963 / 0 | 6 / 0 (0.002447562) |
| l | twin | 4 | all / none | 7430 | 8382 | 96285 / 0 | 6 / 0 (0.002476752) |
| m | latent | 0 | none / all | 2644 | 2942 | 24639 / 11878 | 0 / 6879 (0.002495237) |
| m | latent | 1 | none / all | 2650 | 2942 | 24205 / 14597 | 0 / 7149 (0.001746237) |
| m | latent | 2 | none / all | 2654 | 2942 | 25758 / 13387 | 0 / 8588 (0.002343441) |
| m | latent | 3 | none / all | 2650 | 2942 | 24527 / 13172 | 0 / 7231 (0.001926874) |
| m | latent | 4 | none / all | 2653 | 2942 | 26436 / 14993 | 0 / 27761 (0.002130168) |
| m | metrics | 0 | health, intensity, load, pool / auth, cache, net | 2644 | 2942 | 32442 / 666 | 0 / 0 (0.0) |
| m | metrics | 1 | health, intensity, load, pool / auth, cache, net | 2650 | 2942 | 32126 / 1010 | 0 / 0 (0.0) |
| m | metrics | 2 | health, intensity, load, pool / auth, cache, net | 2654 | 2942 | 33703 / 470 | 0 / 1688 (0.0) |
| m | metrics | 3 | health, intensity, load, pool / auth, cache, net | 2650 | 2942 | 32415 / 609 | 0 / 0 (0.0) |
| m | metrics | 4 | health, intensity, load, pool / auth, cache, net | 2653 | 2942 | 34390 / 878 | 0 / 0 (0.0) |
| m | twin | 0 | all / none | 2644 | 2942 | 33265 / 0 | 0 / 0 (0.0) |
| m | twin | 1 | all / none | 2650 | 2942 | 32952 / 0 | 0 / 0 (0.0) |
| m | twin | 2 | all / none | 2654 | 2942 | 34420 / 0 | 6 / 0 (0.0) |
| m | twin | 3 | all / none | 2650 | 2942 | 33207 / 0 | 0 / 0 (0.0) |
| m | twin | 4 | all / none | 2653 | 2942 | 35209 / 0 | 0 / 0 (0.0) |
| s | latent | 0 | none / all | 322 | 430 | 2064 / 796 | 0 / 0 (0.0) |
| s | latent | 1 | none / all | 322 | 430 | 2545 / 859 | 0 / 0 (0.0) |
| s | latent | 2 | none / all | 322 | 430 | 2555 / 851 | 0 / 0 (0.0) |
| s | latent | 3 | none / all | 322 | 430 | 2517 / 791 | 0 / 0 (0.0) |
| s | latent | 4 | none / all | 322 | 430 | 2576 / 997 | 0 / 0 (0.0) |
| s | metrics | 0 | health, intensity, load, pool / auth, cache, net | 322 | 430 | 2930 / 68 | 0 / 0 (0.0) |
| s | metrics | 1 | health, intensity, load, pool / auth, cache, net | 322 | 430 | 3427 / 268 | 0 / 0 (0.0) |
| s | metrics | 2 | health, intensity, load, pool / auth, cache, net | 322 | 430 | 3436 / 109 | 0 / 0 (0.0) |
| s | metrics | 3 | health, intensity, load, pool / auth, cache, net | 322 | 430 | 3393 / 48 | 0 / 0 (0.0) |
| s | metrics | 4 | health, intensity, load, pool / auth, cache, net | 322 | 430 | 3461 / 69 | 0 / 0 (0.0) |
| s | twin | 0 | all / none | 322 | 430 | 3114 / 0 | 0 / 0 (0.0) |
| s | twin | 1 | all / none | 322 | 430 | 3655 / 0 | 0 / 0 (0.0) |
| s | twin | 2 | all / none | 322 | 430 | 3646 / 0 | 0 / 0 (0.0) |
| s | twin | 3 | all / none | 322 | 430 | 3559 / 0 | 0 / 0 (0.0) |
| s | twin | 4 | all / none | 322 | 430 | 3649 / 0 | 0 / 0 (0.0) |
| xl | latent | 0 | none / all | 19677 | 22772 | 198148 / 181365 | 0 / 63115 (0.002396905) |
| xl | latent | 1 | none / all | 19772 | 22772 | 197410 / 201988 | 0 / 64834 (0.002407846) |
| xl | latent | 2 | none / all | 19873 | 22772 | 193085 / 204374 | 0 / 64891 (0.001854232) |
| xl | latent | 3 | none / all | 19788 | 22772 | 193375 / 203924 | 0 / 63779 (0.002499959) |
| xl | latent | 4 | none / all | 19713 | 22772 | 192821 / 208923 | 0 / 64139 (0.002360767) |
| xl | metrics | 0 | health, intensity, load, pool / auth, cache, net | 19677 | 22772 | 263307 / 2900 | 9 / 10894 (0.002396905) |
| xl | metrics | 1 | health, intensity, load, pool / auth, cache, net | 19772 | 22772 | 262500 / 3458 | 9 / 10339 (0.002407846) |
| xl | metrics | 2 | health, intensity, load, pool / auth, cache, net | 19873 | 22772 | 257880 / 3672 | 9 / 10135 (0.001854232) |
| xl | metrics | 3 | health, intensity, load, pool / auth, cache, net | 19788 | 22772 | 258758 / 3066 | 9 / 9943 (0.002499959) |
| xl | metrics | 4 | health, intensity, load, pool / auth, cache, net | 19713 | 22772 | 257847 / 3447 | 9 / 9798 (0.002360767) |
| xl | twin | 0 | all / none | 19677 | 22772 | 267210 / 0 | 6 / 0 (0.002396905) |
| xl | twin | 1 | all / none | 19772 | 22772 | 266771 / 0 | 6 / 0 (0.002407251) |
| xl | twin | 2 | all / none | 19873 | 22772 | 262139 / 0 | 6 / 0 (0.001852761) |
| xl | twin | 3 | all / none | 19788 | 22772 | 262667 / 0 | 6 / 0 (0.002488808) |
| xl | twin | 4 | all / none | 19713 | 22772 | 261924 / 0 | 6 / 0 (0.002360767) |
| xs | latent | 0 | none / all | 78 | 128 | 453 / 111 | 0 / 0 (0.0) |
| xs | latent | 1 | none / all | 77 | 128 | 479 / 107 | 0 / 0 (0.0) |
| xs | latent | 2 | none / all | 76 | 128 | 410 / 120 | 0 / 0 (0.0) |
| xs | latent | 3 | none / all | 79 | 128 | 241 / 185 | 0 / 0 (0.0) |
| xs | latent | 4 | none / all | 79 | 128 | 434 / 165 | 0 / 0 (0.0) |
| xs | metrics | 0 | health, intensity, load, pool / auth, cache, net | 78 | 128 | 654 / 6 | 0 / 0 (0.0) |
| xs | metrics | 1 | health, intensity, load, pool / auth, cache, net | 77 | 128 | 677 / 20 | 0 / 0 (0.0) |
| xs | metrics | 2 | health, intensity, load, pool / auth, cache, net | 76 | 128 | 601 / 14 | 0 / 0 (0.0) |
| xs | metrics | 3 | health, intensity, load, pool / auth, cache, net | 79 | 128 | 445 / 6 | 0 / 0 (0.0) |
| xs | metrics | 4 | health, intensity, load, pool / auth, cache, net | 79 | 128 | 644 / 43 | 0 / 0 (0.0) |
| xs | twin | 0 | all / none | 78 | 128 | 720 / 0 | 0 / 0 (0.0) |
| xs | twin | 1 | all / none | 77 | 128 | 743 / 0 | 0 / 0 (0.0) |
| xs | twin | 2 | all / none | 76 | 128 | 664 / 0 | 0 / 0 (0.0) |
| xs | twin | 3 | all / none | 79 | 128 | 506 / 0 | 0 / 0 (0.0) |
| xs | twin | 4 | all / none | 79 | 128 | 716 / 0 | 0 / 0 (0.0) |

To score a method: read only `raw/`, `views/` and (metrics variant) `metrics/`;
score against `graphs/scoring-target.json` (request grain) with
`python -m tracebench.score`.

Scoring-target strengths are exact marginalisations of the mechanism, except
where the exact frontier of one (source, destination) chain exceeds the tool's
particle cap: those effects are counter-keyed Monte-Carlo estimates with common
random numbers (the sample count and the largest standard error are recorded
per edge as `mc: {n, se}` and per target as `n_effects_mc` / `mc_se_max`).
The last column counts the request-grain edges that carry the flag; a reader
who wants only exact strengths can drop them. The scorer reads strengths only.

The dataset viewer is disabled: a corpus is a tree of gzipped JSON lines and
parquet files with several schemas, not one table.