Upstrace

When a data quality check fails, trace it upstream to the change that caused it.

Upstrace profiles every column of a dbt project on every run, compares each day against its own history, walks the lineage graph to name the one node where a problem started, and asks an LLM to explain it in a sentence a human can act on.

Contents

• The problem it exists for
• How it works
• Tech stack
• Quick start — the demo, from zero
• Command reference
• Two ways to judge a value
• What the metrics mean
• The dashboard
• Run it on your own dbt project
• Configuration reference
• Slack alerts
• Environment variables
• How well does it work
• What it does not do
• Repo layout

---

The problem it exists for

A pipeline that crashes is not the expensive kind. The expensive kind is the one that succeeds.

An upstream vendor starts sending trip_distance in kilometres instead of miles. Nothing is null. Nothing violates a constraint. Every dbt test passes, every model builds green, and the number flows into three downstream marts, a dashboard, and whatever decision someone makes on Monday morning. The failure is silent, and by the time anyone notices, the question is not what broke but where did this start — a question usually answered by an analyst reading SQL for two hours.

Upstrace answers that question mechanically.

How it works

  dbt project
      │
      │  manifest.json  ────────────────┐
      ▼                                 │
  ┌────────────┐                        ▼
  │  profile   │  every column, whole-table AND per-day
  └─────┬──────┘  → upstrace_meta.column_profiles
        │           upstrace_meta.partition_profiles
        ▼
  ┌────────────┐
  │   drift    │  this run vs the previous run, per metric, per day
  └─────┬──────┘  → upstrace_meta.drift_signals
        │
        ▼
  ┌────────────┐
  │    rca     │  a node is a ROOT if it drifted and no ancestor drifted
  └─────┬──────┘  → incidents, each with evidence and blast radius
        │
        ▼
  ┌────────────┐
  │  explain   │  numbers + lineage → LLM → ranked hypothesis
  └────────────┘  (never the rows; cached by prompt hash)

Four steps, each doing one thing:

1. Profile. For every column in every dbt node, record row count, null rate, distinct count, min, max and mean — both for the table as a whole and once per day. The per-day part is the difference between working and not; see the numbers below.
2. Detect drift. Compare this run's profile to the previous one. A metric that moves more than its threshold becomes a signal, with severity scaled by how far past the threshold it went.
3. Find the root. Read the dbt manifest.json, build the lineage graph, and apply one rule: a node is a root cause if it drifted and none of its upstream nodes drifted. Everything downstream of a root is collateral.
4. Explain. Send the numbers and the lineage — never the rows — to an LLM and get back a ranked hypothesis with its reasoning and blast radius.

Tech stack

Layer	Choice	Why
Warehouse	DuckDB	Single file on disk, no server, no cost. Analytical SQL fast enough to profile millions of rows in seconds
Transformation	dbt (dbt-core 1.10, dbt-duckdb 1.9)	The lineage graph comes free inmanifest.json — that graph is what makes root-cause analysis possible at all
Profiling / detection	Python 3.10+, duckdb, pandas	Profiles written as oneGROUP BY per column, inserted in bulk from a DataFrame
CLI	Typer + Rich	Subcommands,--help, formatted tables
Config	PyYAML (upstrace.yml)	Nothing project-specific lives in Python
API	FastAPI + Uvicorn	One process serves JSON and the built SPA
Dashboard	React 18 + Vite	Hand-written SVG charts and lineage graph, no chart library
Alerting	Slack incoming webhooks (stdlib urllib)	Root-cause summary in a channel, with no HTTP dependency added
LLM	Groq (qwen/qwen3.8-27b), with Gemini, Ollama and a mock provider	Free tier, structured JSON output, responses cached by prompt hash and committed
Packaging	Hatchling, Docker (two build targets)	pip install -e . or docker build --target runtime
Dataset	NYC TLC yellow taxi, Jan–Mar 2024 (9.5M rows)	Public, messy, and has real date grain

Everything in this project runs on free tiers or on your own machine. There is no paid dependency anywhere.

---

Quick start — the demo, from zero

Prerequisites

• Python 3.10 or newer
• Node 18 or newer (only for the dashboard)
• git
• About 1.5 GB of free disk for the full dataset

1. Clone and install

git clone https://github.com/ashg2099/upstrace
cd upstrace

python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate

pip install -e .
pip install "dbt-core==1.10.*" "dbt-duckdb==1.9.*"

pip install -e . installs the package and puts the upstrace command on your PATH. Check it:

upstrace --help

2. Add an API key (optional)

Only needed for upstrace explain. Everything else works without it.

cp .env.example .env

Open .env and paste a free Groq key:

GROQ_API_KEY=gsk_...
UPSTRACE_LLM_PROVIDER=groq
UPSTRACE_LLM_MODEL=qwen/qwen3.8-27b

No key? Set UPSTRACE_LLM_PROVIDER=mock. The explanations already committed in cache/llm/ still replay, because responses are cached by prompt hash.

3. Get the data

python scripts/download_data.py

Downloads three months of NYC taxi parquet files (~150 MB) into data/. Want a different range:

python scripts/download_data.py 2024-04 2024-05

4. Load it into DuckDB

python scripts/load_duckdb.py

Creates warehouse/upstrace.duckdb with raw.yellow_trips — 9,554,778 rows — and prints a hand-written data quality report. That report is the manual work Upstrace automates; it is worth reading once to see how dirty the data already is before anything is injected.

5. Build the dbt models

cd transform
dbt run --profiles-dir .
dbt test --profiles-dir .
cd ..

This builds four nodes on top of the source:

yellow_trips (source) → stg_trips (view) → fct_trips (table) → agg_daily_revenue
                                                             → agg_daily_quality

dbt run also writes transform/target/manifest.json, which is where Upstrace reads the lineage from. Nothing works before this step.

Two dbt test failures are expected — they are real problems in the raw data, kept rather than filtered.

6. Take a baseline

upstrace models        # what Upstrace can see
upstrace profile       # baseline
upstrace profile       # current
upstrace drift         # should say: nothing moved

Drift always compares the two most recent profile runs, so you need at least two. Two runs over unchanged data must produce zero signals — if they do not, something is non-deterministic, which is exactly the bug the negative controls later caught.

7. Break something on purpose

upstrace fault unit-change

trip_distance silently switches from miles to kilometres in the source, and the whole pipeline is rebuilt. dbt reports success. That is the point.

upstrace profile
upstrace drift
upstrace rca
upstrace explain

What comes back:

yellow_trips   CRITICAL   source table
  no upstream node drifted, so this is where it started

  column          metric        baseline   current   change   days
  trip_distance   mean_value      3.2496    5.2298    60.9%     31
  trip_distance   mean_value      4.6339    5.7401    23.9%      -
  trip_distance   distinct_count   3,214     3,906    21.5%      -

  carried downstream to
    stg_trips.trip_distance_miles
    fct_trips.trip_distance_miles
    agg_daily_revenue.avg_distance_miles

Two rows for the same metric, and the difference between them is the whole argument for per-partition profiling. The whole-table mean moved 23.9% — suspicious, but not obviously anything. The per-day mean on the 31 affected days moved 60.9%, a ratio of 1.6094. The true miles-to-kilometres factor is 1.60934. Diluting a fault across 500,000 rows hides its signature; measuring each day against itself preserves it.

8. Put it back

upstrace fault reset

9. What the model says

The LLM sees the evidence table and the lineage. It never sees a single row.

The source data for trip_distance likely switched from miles to kilometers, causing a ~1.6x increase in values that propagated through the pipeline.

HIGH CONFIDENCE — The upstream source system changed the unit of measurement for trip_distance from miles to kilometers. The mean value increased by a factor of approximately 1.61 (3.25 to 5.23). The conversion factor from miles to kilometers is 1.60934. This ratio is too close to be coincidental. Additionally, the downstream column is named trip_distance_miles, which explicitly claims the unit is miles. If the values are now in kilometers, the name contradicts the data, confirming a unit mismatch.

LOW CONFIDENCE — A data quality issue introduced duplicate or outlier records that skewed the mean upward. While the distinct count increased by 21.5%, this is a much smaller change than the 60.9% increase in the mean. A simple increase in row count or distinct values would not typically cause the mean to jump by a factor of 1.6 unless the new data was systematically larger, which points back to a unit change rather than random noise or duplication.

IMPACT — All downstream consumers of agg_daily_revenue, fct_trips, and stg_trips are seeing trip_distance values that are ~60% higher than expected. Specifically, avg_distance_miles in agg_daily_revenue and trip_distance_miles in fct_trips/stg_trips are mislabeled; they contain kilometer values but are reported as miles, leading to overestimated distances and any metrics derived from them (e.g., speed, fuel consumption estimates).

The column name used as evidence, and the explicit rejection of the alternative hypothesis, are what make this worth more than a threshold alert.

---

Command reference

Every command takes --help.

upstrace models     List the dbt models Upstrace can see, and what each depends on.
upstrace profile    Measure every column of every model and append to the metric history.
upstrace history    Show how one column's measurements have moved across runs.
upstrace fault      Break the raw data on purpose, or put it back.
upstrace drift      Compare the two most recent profile runs and report what moved.
upstrace rca        Turn the drift signals from the latest run into root causes.
upstrace explain    Ask a language model what the root-cause incidents actually mean.
upstrace eval       Inject every seeded defect in turn and score the root-cause analysis.
upstrace init       Write a starter upstrace.yml in the current directory.
upstrace config     Show the resolved configuration and where it came from.
upstrace report     Write a self-contained HTML report of the latest analysis.
upstrace slack-test Send a dummy message to confirm a Slack webhook works.

upstrace init / upstrace config

upstrace init                 # writes upstrace.yml here
upstrace init --force         # overwrite an existing one
upstrace config               # what resolved, and which file it came from

config is the first thing to run when something looks wrong. If it prints (defaults, no upstrace.yml found), you are in the wrong directory — configuration is discovered by walking up from the working directory, the way git finds .git.

upstrace models

Lists every node from the manifest with its dependencies. Empty output means the manifest is missing (run dbt run first) or include/exclude filtered everything out.

upstrace profile

upstrace profile              # every included node
upstrace profile -m fct_trips # one model only

Writes one whole-table row per column and one row per column per day into upstrace_meta. Prints the run id and how many profiles it wrote. Run it twice before drift — comparison needs two runs.

upstrace history

upstrace history -m yellow_trips -c trip_distance

Both flags are required. Shows how that one column's metrics have moved across every recorded run — the CLI version of the dashboard's metric chart.

upstrace fault

upstrace fault list                          # what can be injected
upstrace fault unit-change                   # inject, from 2024-03-01 onward
upstrace fault null-spike --since 2024-02-15 # inject from a different date
upstrace fault reset                         # put the raw data back

Key	What it simulates
unit-change	trip_distance silently switches from miles to kilometres
null-spike	vendor 2 stops sendingpassenger_count
enum-drift	an undocumentedpayment_type 7 starts appearing
case-change	store_and_fwd_flag starts arriving lowercase
reset	rebuild the pipeline from the untouched source

--since defaults to 2024-03-01, so a fault touches the last 31 days of the three-month dataset. That is where the days 31 column in the evidence table comes from — and shrinking the window is how the benchmark builds its one-week and one-day scenarios, which are the hard ones.

Injection changes the raw table only — rebuild with dbt yourself afterwards, and watch it report success either way. The dashboard's fault buttons rebuild and reprofile in one step; the CLI leaves that to you so you can see dbt pass.

upstrace drift

upstrace drift                        # report only
upstrace drift --baseline rolling     # override the configured baseline mode
upstrace drift --fail-on critical     # exit 1 if any critical signal
upstrace drift --fail-on warning      # exit 1 if anything at all moved
upstrace drift --slack-on warning     # also post to Slack

Compares the two most recent profile runs, writes signals to upstrace_meta.drift_signals, and prints them worst-first. Two runs over unchanged data must print No drift above threshold.

--fail-on is what makes this usable in CI. Without an exit code the job is green whatever the tool found, and "monitoring" means someone remembering to read logs.

It also warns when the two runs measured very different row volumes — comparing a 500k-row sample against a 9.5M-row load makes every metric look like it moved, and the tool would otherwise report that nonsense confidently.

--slack-on posts the run to Slack at or above a severity. It is independent of --fail-on, and alerts are sent before the exit-code check — so a run that fails CI still notifies. See Slack alerts.

upstrace rca

Groups the latest run's signals into incidents and names a root for each, using one rule: a node is a root if it drifted and none of its upstream nodes drifted. Prints the evidence and the downstream columns the change was carried into.

upstrace explain

upstrace explain              # uses the cached response if the prompt matches
upstrace explain --no-cache   # force a fresh model call

Sends the evidence table and the lineage — never the rows — to the configured LLM. Responses are cached on disk by prompt hash, so the same incident always produces the same explanation and the committed cache replays without a key.

upstrace eval

upstrace eval                                          # 56 scenarios + 4 controls
upstrace eval --limit 3                                # quick check
upstrace eval --sample 100000                          # faster, smaller baseline
upstrace eval --sample 500000 --report docs/eval-results.md

Flag	Default	Meaning
--sample	500000	Rows to sample per run. The sample is seeded andREPEATABLE, so runs are comparable
--limit	all	Run only the first N scenarios
--report	docs/eval-results.md	Where to write the markdown report

A full run takes roughly 15 minutes at --sample 500000. It resets to a clean baseline before every scenario, so it is safe to interrupt.

--report overwrites its target. Commit the previous report before rerunning with a smaller --limit or --sample, or you will replace a full result with a partial one.

Two ways to judge a value

A metric moved — compared to what? Upstrace supports two baselines.

previous_run (default) compares each partition against its own value in the previous profile run. It is exact, needs no history, and catches a change the moment it appears. It cannot see a fault that was already in place before you started profiling — that value simply becomes "normal".

rolling compares each partition against a trailing window of its own history using a median/MAD robust z-score, so one bad day does not move the bar. A signal must clear both the statistical test and the configured practical threshold, so a 0.4% move on a very stable column is not called an incident.

baseline:
  mode: rolling             # or previous_run
  window: 28                # trailing partitions to compare against
  z: 3.5                    # robust z-score threshold
  min_history: 14           # skip a partition with less history than this
  min_partition_share: 0.2  # skip partitions far smaller than the window median

Or per run: upstrace drift --baseline rolling.

Rolling is experimental. On clean taxi data it surfaced real anomalies nothing had planted — stray rows dated 2002, 2008, 2009 and 2023, a snowstorm on 2024-02-13, and a source-format change on 2024-02-15 — none of which previous_run can structurally see. But it shares a weakness with every trailing-window method: a fault in place for the whole window becomes the baseline. Catching that needs changepoint detection over the full history, tracked in #1.

What the metrics mean

Six metrics are recorded per column, per run, and per day.

Metric	What it is	How change is measured	Default threshold
row_count	rows in the table (or in that day)	relative:abs(new − old) / old	2%
null_rate	fraction of values that are null	absolute, in percentage points	1 pp
distinct_count	exact count of distinct values	relative	10%
mean_value	average, numeric columns only	relative	5%
min_value	smallest value, as text	any change at all	—
max_value	largest value, as text	any change at all	—

null_rate is absolute on purpose. A relative change from 0.001 to 0.002 is 100% and means nothing; a move from 0.1% to 1.1% is one percentage point and means a vendor stopped sending a field.

min_value and max_value are reported with their actual values, not just as "changed" — a move from 'casual' to 'Casual' is a diagnosis, while "changed" is a shrug. They are the one place Upstrace handles real cell values rather than metrics, so profile.mask_values withholds them for columns that should never leave the warehouse.

distinct_count is counted exactly, and floats are rounded to six decimals first. Both choices are explained in determinism.

Severity

A signal's severity comes from how far past its threshold it went:

Severity	Trigger
warning	at or past 1× the threshold
high	at or past 2×
critical	at or past 5×

Both the thresholds and these multipliers are configurable per node.

Where it is all stored

Everything Upstrace records lives in a upstrace_meta schema inside the same DuckDB file — so it is queryable with plain SQL:

Table	Contents
profile_runs	one row per run: id, start, end, node count
column_profiles	one row per column per run — whole-table metrics
partition_profiles	one row per columnper day per run
drift_signals	every signal, with baseline, current, change, severity, and how many days moved

-- what moved most in the latest run
SELECT model_name, column_name, metric, change, severity, partitions
FROM upstrace_meta.drift_signals
ORDER BY change DESC
LIMIT 10;

---

The dashboard

Development (two processes, hot reload)

# terminal 1 - API
uvicorn upstrace.api:app --reload --reload-dir src

# terminal 2 - SPA
cd app && npm install && npm run dev

Open the URL Vite prints (usually http://localhost:5173). It proxies /api to port 8000.

Single process

cd app && npm run build && cd ..
uvicorn upstrace.api:app --port 8000

FastAPI serves the built SPA from app/static. Open http://localhost:8000. Interactive API docs at /api/docs.

Four views

• Incident feed and RCA detail — every incident, its root node, the columns that moved, the evidence table, and where the change was carried downstream
• Lineage graph — read left to right; red is where the analysis says it started, amber is what carried it
• Metric history — per-day baseline and current overlaid, with metric tabs and a hover tooltip showing the ratio. The y-axis clips to p2–p98 and marks what falls outside; one freak day in the taxi data reaches 61.7 and would otherwise flatten the step change the chart exists to show
• Fault panel — inject a fault live and watch the incident appear

Demo order matters: press Reset warehouse first, then inject. Resetting after injecting would show the repair itself as drift.

API endpoints

Method	Path	Returns
GET	/api/health	liveness
GET	/api/overview	node count, rows profiled, signal count, incident count
GET	/api/lineage	nodes and edges, each flagged root / affected / clean
GET	/api/incidents	every incident with its severity and summary
GET	/api/incidents/{root}?explain=true	one incident, optionally with the LLM explanation
GET	/api/metrics/{model}/{column}?metric=mean_value	per-day baseline and current series
GET	/api/faults	the faults available to inject
POST	/api/faults/{key}	inject a fault, rebuild, reprofile, re-analyse
POST	/api/reset	restore the clean baseline

---

Run it on your own dbt project

The engine names no table and no dataset. It reads whatever dbt manifest you point it at.

Requirements: a dbt project with a DuckDB target, and at least one date or timestamp column on the tables you care about.

1. Install Upstrace

git clone https://github.com/ashg2099/upstrace
cd upstrace && pip install -e .

2. Build your own project once

Upstrace reads target/manifest.json, which only exists after a dbt build:

cd /path/to/your-dbt-project
dbt run

3. Create a config

upstrace init

This writes upstrace.yml in the current directory. Open it and set the two paths that matter:

project: my-project
warehouse: ./warehouse/analytics.duckdb   # your .duckdb file
dbt_project_dir: .                        # where dbt_project.yml lives

Both are resolved relative to upstrace.yml itself, so relative paths are fine.

4. Check what resolved

upstrace config
upstrace models

upstrace config prints which file it read and every resolved value. If it says (defaults, no upstrace.yml found), you are in the wrong directory — configuration is discovered by walking up from the working directory, the way git finds .git.

upstrace models should list your nodes. If it is empty, the manifest is missing or include/exclude filtered everything out.

5. Profile twice, then look

upstrace profile      # baseline
# ... let real changes happen, or run your pipeline again ...
upstrace profile      # current
upstrace drift
upstrace rca

The first two runs over unchanged data should produce zero signals. If they do not, something in your pipeline is non-deterministic — which is itself worth knowing.

6. Tune it

Two things almost always need adjusting on a real project:

Which models to profile. Hundreds of models means slow runs and noise:

profile:
  include: ["stg_*", "fct_*", "agg_*"]
  exclude: ["*_scratch", "stg_pii_*"]

The partition axis. Auto-detection picks the first DATE column, else the first TIMESTAMP. That is right often, and silently wrong sometimes — a table whose first date column is customer_signup_date would be profiled along an axis that means nothing:

profile:
  partition_column:
    fct_orders: order_date
    dim_customers: null      # null = skip per-day, whole-table only

Then re-run upstrace profile and check days in the drift output looks sensible.

7. Thresholds

Start with the defaults. When a table alerts every day for no reason, override it rather than raising the global number:

thresholds:
  overrides:
    agg_marketing_events:
      row_count: 0.40      # this table genuinely swings 40% day to day

With Docker instead

docker build --target runtime -t upstrace .
docker run -v "$PWD":/project -w /project -p 7860:7860 upstrace

The runtime target contains the tool and no data. The demo target — the default if you build without --target — bakes in a 400,000-row sample and builds its own warehouse, so docker run -p 7860:7860 <image> gives a working demo with nothing to set up.

---

Configuration reference

Every key in upstrace.yml, with its default:

# Name shown by `upstrace config`. Cosmetic.
project: my-project

# Path to the DuckDB file. Relative paths resolve against this config file.
warehouse: warehouse/upstrace.duckdb

# Directory containing dbt_project.yml. The manifest is read from
# <dbt_project_dir>/target/manifest.json
dbt_project_dir: transform

# Where source parquet lives, for the demo scripts only.
data_dir: data

# Schema Upstrace writes its own tables into.
metrics_schema: upstrace_meta

profile:
  # Glob patterns matched against dbt node names. Exclude wins over include.
  include: ["*"]
  exclude: []

  # Per-day profiling axis.
  #   omitted  -> auto-detect (first DATE column, else first TIMESTAMP)
  #   a column -> use that column
  #   null     -> no per-day profiling for this node
  partition_column: {}

  # Refuse per-day profiling above this many distinct partitions. Guards against
  # a column that looks date-like but is effectively unique.
  max_partitions: 400
  # Columns whose min/max values must never leave the warehouse. min_value and
  # max_value are real cell values - the smallest value in an email column is a
  # real email address. Matched columns report <masked> instead.
  mask_values: []

thresholds:
  # Relative change, except null_rate which is absolute percentage points.
  default:
    row_count: 0.02
    null_rate: 0.01
    distinct_count: 0.10
    mean_value: 0.05

  # Per-node overrides, glob-matched, merged over the defaults.
  overrides: {}

baseline:
  # previous_run compares a partition to its own value in the last run.
  # rolling compares it to a trailing window of its own history.
  mode: previous_run
  window: 28
  z: 3.5
  min_history: 14
  min_partition_share: 0.2

# Multipliers of the threshold at which each severity begins.
severity:
  warning: 1.0
  high: 2.0
  critical: 5.0

Slack alerts

Upstrace can post a run's incidents to a Slack channel.

1. Create an incoming webhook for the channel you want.
2. export UPSTRACE_SLACK_WEBHOOK='https://hooks.slack.com/services/...'
3. upstrace slack-test — confirms the webhook before you rely on it.

Then any drift run posts automatically:

upstrace drift --slack-on warning --report-url https://you.github.io/upstrace/

The message leads with the root cause and its blast radius, then the evidence behind it — not a flat list of every correlated signal:

🔴 Upstrace: 1 root cause, 15 signals

Root cause — yellow_trips (source) · trip_distance → 3 downstream: agg_daily_revenue, fct_trips, stg_trips

That distinction is the point. Fifteen alerts across four models read as four problems; one root with a blast radius reads as one.

--slack-on takes critical, high or warning, and is independent of --fail-on: CI can alert on warnings while only failing the build on criticals. Alerts fire before the exit-code check, so a failing run still notifies.

The notifier uses only the standard library, so alerting adds no dependency. --report-url adds an Open report button pointing at a published upstrace report output.

This repo's own nightly workflow deliberately does not post to Slack: it injects a fault every night on purpose, and a channel that cries wolf nightly is the exact failure mode this tool exists to prevent.

On a schedule

Upstrace has no scheduler — whatever runs your pipeline runs it. Add upstrace profile && upstrace drift --slack-on warning after dbt run and alerts arrive whenever that pipeline does.

A clean run posts nothing at all: the notifier returns without sending when no signal meets the threshold. The channel stays quiet until something actually moves, which is the only way an alert channel survives a month.

From cron, use a wrapper script and set everything explicitly — cron does not load your shell profile, so a job that works in your terminal will silently do nothing at 6am:

#!/usr/bin/env bash
set -euo pipefail
cd /srv/analytics && source .venv/bin/activate
export UPSTRACE_SLACK_WEBHOOK='https://hooks.slack.com/services/...'
dbt run && upstrace profile && upstrace drift --slack-on warning

In GitHub Actions, schedule.cron is UTC — 06:00 IST is 30 0 * * *. In Airflow, a BashOperator downstream of the dbt task: with --fail-on critical the task turns red and the channel gets the root cause, from one command.

Environment variables

Configuration lives in upstrace.yml; secrets and overrides live in the environment (or .env, which is gitignored).

LLM responses are cached on disk under cache/llm/, keyed by a hash of the prompt. The cache is committed, so every explanation in this repo reproduces without a key. Change the prompt or the evidence and the hash changes, so a cache hit always means the same question.

Variable	Purpose
GROQ_API_KEY	API key for the Groq provider
UPSTRACE_LLM_PROVIDER	groq, gemini, ollama or mock
UPSTRACE_LLM_MODEL	model id, e.g.qwen/qwen3.8-27b
UPSTRACE_CONFIG	path to a specificupstrace.yml, skipping discovery
UPSTRACE_WAREHOUSE	override the warehouse path
UPSTRACE_DBT_PROJECT_DIR	override the dbt project path
UPSTRACE_METRICS_SCHEMA	override the metadata schema name
UPSTRACE_BASELINE_MODE	previous_run or rolling, overriding upstrace.yml
UPSTRACE_SLACK_WEBHOOK	Slack incoming-webhook URL for alerts
UPSTRACE_DEMO_SAMPLE	rows the dashboard's reset button loads (demo only)

---

How well does it work

Claims about detection are cheap. This one is measured by a harness that injects known faults and checks whether the reported root cause is the node that was actually broken.

56 fault scenarios across four axes — fault type (unit change, fare inflation, null spike, enum drift, row drop, case change, precision loss, flag logic break, mart scaling), magnitude (small / medium / large), time window (one day / one week / one month) and layer (source / model) — plus 4 negative controls.

Drift detected	56 / 56
Correct root node	55 / 56
Correct column	55 / 56
Fully correct	55 / 56 (98%)
False positives on controls	0 / 4

Full per-scenario output: docs/eval-results.md. Model comparison: docs/model-comparison.md.

Verified on a second dataset

The benchmark above runs on the taxi data. To test the claim that the engine is dataset-agnostic, it was pointed at a Citi Bike dbt project it had never seen — different schema, different domain, a separate warehouse and config. It ran first attempt, produced zero signals on two identical runs, and named the correct root cause for an injected fault.

It also found a real bug that 56 taxi scenarios could not: docs/second-dataset.md.

Why per-partition profiling was necessary

The first working version profiled whole tables only. It scored 44/56, and every single failure named a downstream aggregate rather than the source. The reason was dilution: a fault confined to one day of ninety barely moves a table-wide mean, but it moves a small aggregate table hard, so the aggregate looked like the origin.

	whole-table only	per-partition
All scenarios	44 / 56	55 / 56
One-day window only	9 / 17	17 / 17

The one-day scenarios are the honest test — they are the ones a whole-table average is designed to miss.

The negative controls, and the bug they found

Four scenarios inject nothing, or something deliberately below threshold:

• control-null — the pipeline is rebuilt and reprofiled with no change at all
• control-subthreshold-tip — tips up 3%, under the 5% mean threshold
• control-subthreshold-extra — extras up 2%
• control-subthreshold-tip-near-limit — tips up 4.5%, just under the line

A control passes only if it raises no high or critical signal. When they were added, control-null failed: three high-severity signals on a run where nothing had changed.

The cause was real and worth writing down. Distinct counts were computed with approx_count_distinct, a HyperLogLog estimate that varied by up to 30% between identical runs. Compounding it, sum() and avg() run in parallel and floating-point addition is not associative, so two mathematically equal groups could differ in their last bits and change which values counted as distinct. The fix was to count exactly, and to round floats first:

def _distinct_expr(col: str, data_type: str) -> str:
    if _is_float(data_type):
        return f"count(distinct round({col}, 6))"
    return f"count(distinct {col})"

Two identical runs now produce zero signals. Fifty-six fault scenarios never caught this; one negative control did — which is the argument for having them.

The one that fails

fare-inflation-small-day — a 3% fare increase on a single day out of ninety. Drift is detected, but the root is reported as agg_daily_quality rather than yellow_trips. On that day the source column's own movement stayed just under threshold while a downstream aggregate amplified it, so the rule that defines a root — drifted, with no drifted ancestor — legitimately named the aggregate. It is documented rather than tuned away, because moving the threshold to catch it would cost false positives elsewhere.

One more thing worth knowing

Partition profiles were originally inserted row by row with executemany: 8.2 seconds for 1,729 rows, growing with table size, and slow enough to time the benchmark out. Building a DataFrame and running one INSERT INTO ... SELECT * FROM frame brought it to 0.01 seconds.

---

What it does not do

Stated plainly, because scope questions get asked:

• dbt is required. Lineage comes from manifest.json. Without it there is no basis for root-cause analysis, only alerts.
• DuckDB only. The profiler's SQL is DuckDB dialect. A Snowflake or BigQuery adapter would touch exactly one file, profiler.py, but that file has not been written.
• Cannot detect a fault that predates its baseline. previous_run compares against the last run and rolling against a trailing window, so a defect that was already there before either window began reads as normal. Changepoint detection over the full history is the fix, and is not built
• No seasonality model. previous_run uses fixed per-table thresholds; rolling uses a robust z-score but knows nothing about weekday/weekend or holiday effects, so it will flag a quiet Sunday on a weekday-shaped table.
• No scheduler of its own. Profiling runs when something runs it. The intended production shape is a step after dbt run in CI or Airflow — see running it on a schedule..github/workflows/nightly.yml does exactly that against the demo data every night.

---

Repo layout

src/upstrace/        the engine - names no table, no dataset
  settings.py        upstrace.yml: paths, globs, thresholds, partition axis
  config.py          resolved paths, a thin façade over settings
  warehouse.py       DuckDB connection and the upstrace_meta tables
  manifest.py        read dbt manifest.json into nodes and edges
  profiler.py        whole-table and per-partition column profiles
  drift.py           threshold comparison and severity
  notify.py          Slack notifications - stdlib only
  report.py          the self-contained HTML report
  lineage.py         ancestors / descendants over the dbt graph
  rca.py             the root-cause rule
  llm.py             providers (groq / gemini / ollama / mock) + prompt-hash cache
  explain.py         the prompt
  faults.py          fault injection - the evaluation fixture
  scenarios.py       56 faults + 4 negative controls
  evaluate.py        the harness and its scoring
  api.py             FastAPI: JSON API + SPA
  cli.py             every upstrace command

transform/           the demo dbt project - staging and marts over NYC taxi data
app/                 React + Vite dashboard
scripts/             data download, warehouse load, demo sample, Docker build
cache/llm/           committed LLM responses, keyed by prompt hash
docs/                eval results, model comparison, screenshots
upstrace.yml         configuration for the demo project
Dockerfile           two targets: runtime (tool only), demo (tool + data)

transform/, faults.py and scenarios.py are the demo pipeline and the evaluation fixture. The engine never references them.

Data

NYC TLC yellow taxi trip records, January–March 2024 — 9.5M rows. A 400,000-row reservoir sample spanning 91 days is committed under data/demo/ so the Docker image can build its own warehouse without downloading anything.

License

MIT