Task 1

Buber Fuel

Buber, a cab company, has hired a team of vibe coders to train a model that predicts the fuel a ride burns from the car, the trip distance and a bunch of other parameters. Buber is large, so its ride history is huge, and the vibe coders, being vibe coders, tried pd.read_csv() on it and failed miserably. You are the adult in the room. Build the service they query instead. There is no frontend: package it as a Docker container, push it to GitHub and submit the repository here. Our server builds it, runs it in a sandbox with the real dataset, hits it with a burst of queries and gives you a verdict.

1. What to build

A query holds the quantitative description of a ride, nothing else: no ride id, no pickup time, no driver, no zones. Your service returns the fuel burned (litres) by a ride in the dataset that matches the query, within the error the query allows. It is guaranteed that at least one ride in the dataset matches every query exactly.

POST /fuel HTTP/1.1
Content-Type: application/json

{"distance_km": 12.35, "engine_cc": 1600, "fuel_type": 0,
 "vehicle_weight_kg": 1250, "avg_speed_kmh": 41.5, "idle_min": 3.2,
 "ambient_temp_c": 27, "ac_on": 1, "passengers": 2, "gradient_pct": 0.4,
 "max_error": 0.05}

HTTP/1.1 200 OK
Content-Type: application/json

{"fuel_l": 1.874}

The features

The dataset keeps more digits than a query does. A query is the ride rounded to the precision above (halves round up), so a ride in the file is matched by a query when its features round to the query's. Several rides can share one query and then any of their fuel values is acceptable.

The dataset

A fixed ride log of roughly 30 million rows (about 2.5 GB of CSV). It is not in your repository and you cannot download it. A random sample of about 300,000 rows comes from the same log: rides_sample.csv (24 MB). Columns:

ride_id,pickup_ts,driver_id,start_zone,end_zone,distance_km,engine_cc,fuel_type,vehicle_weight_kg,
avg_speed_kmh,idle_min,ambient_temp_c,ac_on,passengers,gradient_pct,fuel_l

• Sample rows are never used as queries when you are judged, so a service that only knows the sample will not get far. The sample is for experimenting, not for answering.
• Queries are drawn uniformly over the whole log, so caching answers does not help.
• Your code must read the dataset from the path in the environment variable DATASET_PATH. Never hard-code a file name. In your experiments it points at the sample; at judging time we mount the full file and set DATASET_PATH to it. The same image runs in both cases. Mind the 512 MB: the full file is several times larger than your RAM.
• The starter kit has a Dockerfile, a tiny working server, a local accuracy checker and docker-example.sh, which shows exactly how the sample (dev) and the full dataset (judge) are mounted.

How the judge talks to you

After a few single-request checks, the judge sends 2,000 POST /fuel queries inside 1 second. It does not wait for answers before sending the next query, so a slow server simply falls behind. Each query carries its own max_error:

Accuracy and speed pull against each other: scanning for the closest rides is exact but slow, guessing is instant but sloppy. Pick your balance. Answers are only verified once the burst is over. You have 2 CPUs and 512 MB.

2. Container contract

Minimal Dockerfile shape

# Any base image works: node, python, go, java...
FROM python:3.12-slim
WORKDIR /app
COPY . .
RUN pip install --no-cache-dir -r requirements.txt   # the build has internet access
USER nobody
EXPOSE 8080
# Experiments: the sample. The judge replaces this with -e DATASET_PATH=<full dataset>.
ENV DATASET_PATH=/dataset/rides_sample.csv
CMD ["python", "-u", "server.py"]

3. How to submit

1. Push your project to a GitHub repository.
2. Register your team (once) and log in.
3. Open Submit, paste the repository URL, optionally a branch or tag (default branch if empty).
4. Private repository? Create a fine-grained personal access token limited to that one repository with "Contents: Read-only" and paste it in the token field. It is encrypted at rest and deleted as soon as judging finishes.
5. Watch the live status page. You can resubmit after each result (short cooldown applies, one submission at a time per team).

4. How you are judged

Every submission goes through the same automated pipeline:

1. Clone your repository at its latest commit (the commit hash is recorded with your result).
2. Build the Docker image.
3. Run it in an isolated sandbox with no outbound network and the full dataset mounted.
4. Check the contract with a handful of single requests: a valid query, a missing feature, a bad body.
5. Hit it with a burst of 2,000 POST /fuel queries in one second, record every answer, then check each one against the rides that match its query.
6. Run it again, with no dataset. The same image is restarted with DATASET_PATH pointing at a file that does not exist, and gets 1,000 more queries (half of them with max_error 1%). Whatever you can answer then, you can answer without reading the log.
7. Score the checks below and issue a verdict.

Verdicts

Scoring rubric

Total 148 points. Checks marked required must all pass for a PASSED verdict. The Model track is the hard part and is worth the most: a service that only looks answers up in the dataset cannot earn those points, so a perfect score is only reachable by another route. The sample is a window into how the whole log behaves.

5. Common mistakes

• Server listens on 127.0.0.1, or on a different port than the one in your EXPOSE line.
• Hard-coding rides_sample.csv instead of reading $DATASET_PATH.
• Loading the whole file into memory. It works on the sample and gets your container killed on the real thing (exit code 137).
• Starting to answer before you are ready, or taking so long to start that you miss the health budget.
• Rounding queries differently from the rules above, so your lookups miss.
• Writing files outside /tmp or /data (the root filesystem is read-only).
• Crashing or refusing to start when DATASET_PATH does not exist. The last run has no dataset.
• Trying to fetch data over the network at startup. There is no network at runtime.

Submit your solution