CITA Traffic Platform · Bachelor’s thesis · 2026
I built a collector, a database, and a dashboard to explore Luxembourg’s traffic over time. The work was as much about preserving the evidence as drawing the charts.
Archived evaluation dataset
23 July–3 September 2026
Seven roads · 246 measurement-site identifiers
The problem
The CITA feeds describe the latest road conditions. To compare hours, days, or roads later, I needed to collect those changing states myself. Seven measurement feeds sit alongside incidents, regulations, and static and dynamic parking data, each with its own structure and update rhythm.
I separated the collector from the dashboard. The collector keeps building the history while the interface is closed, and the dashboard can query stored observations when a feed is unavailable.
Inside the engineering
Each stage has a specific responsibility. Keeping those boundaries clear let me trace a result back through the query, parser, and original response.
Open a stage to inspect the decision and its consequences.
Continue to the measured results ↓Each feed has its own schedule: incidents every minute, measurements and dynamic parking every five minutes, regulations every fifteen minutes, and static parking once a day. A single scheduler processes due feeds sequentially, with bounded retries and timeouts.
Before parsing, I hash the response bytes with SHA-256 and preserve new contents as gzip files. A temporary file, fsync, and an atomic rename protect the archive write.
A parser can fail without destroying the input that explains the failure. If the database is unavailable, the collector still attempts filesystem archival.
Two byte-identical responses from the same feed share one snapshot. Every fetch still gets its own observation, so the history records when unchanged content was delivered again. The collector skips parsing a registered duplicate.
if duplicate:
return FeedCollectionResult(
feed.name, snapshot_id,
archived.sha256,
archived.path,
duplicate=True
)collector/collector.py · collect_feed
Excerpt, with line breaks adjusted. Snapshot registration has already recorded the fetch; the early return skips parsing.
49,488 fetches returned known content. Hashing detects byte equality; differently formatted XML can still describe the same information without sharing a hash.
Feed-specific parsers handle the different DATEX II structures, namespaces, nested records, and optional values. Missing values remain missing; an unknown coordinate system is preserved with a warning rather than guessed onto a map.
if snapshot_id is not None:
mark_snapshot_failed(
snapshot_id,
exc,
database_url
)collector/collector.py · collect_feed
Exception-handler excerpt, with line breaks adjusted. The snapshot is marked as failed; its already archived response remains available.
The evaluation recorded 46 failed parses. Their raw responses remained available for diagnosis. An offline reprocessing command can apply an updated parser to archived snapshots.
The detailed schema preserves measurements, versioned incident records, regulations, parking data, and their snapshot references. Flatter compatibility tables support the existing historical queries.
The parsed output and its compatibility projection are written in one transaction. A failed write rolls back that group while the archived source remains. Unique keys prevent uncontrolled duplication.
Raw bytes and relational projections serve different tasks, at the cost of extra storage and consistency work. The evaluation database occupied roughly 15 GB including indexes.
PostgreSQL filters and aggregates before results reach Polars and Streamlit. Historical series stay separated by road and stored direction. Timestamps are stored in UTC; hourly and weekday profiles use Europe/Luxembourg.
The dashboard provides current observations, a map, history, comparisons, collection integrity, and data-quality views. Automatic refresh is limited to live areas.
Early charts connected different directions into artificial zigzags. I fixed the series grouping and aggregation rather than smoothing the line.
Inspect the A3 hourly profileWhat the data showed
These results describe the frozen evaluation sample. They are historical observations, with collection gaps and filtering.
From my bachelor’s thesis
In my sample, weekday speeds dipped in the morning; the weekend profile stayed steadier.
Weekday vs. weekend mean speeds · A3 · stored direction outboundFromTown
Explore the day Hover or drag to inspect an hour. Keyboard: ← / →, Home / End; Escape resets to 08:00.
Source: my bachelor’s thesis, section 6.4.1 / figure 6.4, and its evaluation_hourly_profile.csv export. I grouped Monday–Friday and Saturday–Sunday and weighted each weekday/hour mean by its observation count, following the thesis plotting code. Source rows include positive speed, flow, and occupancy values; results are rounded to 0.1 km/h. The source window runs from 23 July 2026, 16:23 to 3 September 2026, 19:13 (Europe/Luxembourg).
View the 48-point aggregate JSON ↗ View the original thesis figure ↗
Scroll within the table to see all hours and columns. Speeds are in km/h; observation columns show sample counts.
| Hour | Weekday km/h | Observations | Weekend km/h | Observations |
|---|---|---|---|---|
| 00:00 | 93.4 | 4,357 | 94.7 | 1,731 |
| 01:00 | 92.6 | 4,236 | 95.9 | 1,730 |
| 02:00 | 94.1 | 4,217 | 95.4 | 1,651 |
| 03:00 | 97.8 | 4,334 | 96.3 | 1,644 |
| 04:00 | 99.0 | 4,404 | 96.4 | 1,632 |
| 05:00 | 94.7 | 4,503 | 99.2 | 1,641 |
| 06:00 | 82.7 | 4,502 | 99.6 | 1,645 |
| 07:00 | 69.8 | 4,580 | 99.5 | 1,675 |
| 08:00 | 68.3 | 4,600 | 101.2 | 1,706 |
| 09:00 | 78.4 | 4,559 | 99.0 | 1,645 |
| 10:00 | 93.9 | 4,540 | 98.3 | 1,682 |
| 11:00 | 94.7 | 4,531 | 96.4 | 1,699 |
| 12:00 | 94.3 | 4,611 | 95.5 | 1,726 |
| 13:00 | 94.4 | 4,543 | 94.3 | 1,693 |
| 14:00 | 95.3 | 4,578 | 91.9 | 1,671 |
| 15:00 | 95.8 | 4,581 | 94.7 | 1,685 |
| 16:00 | 95.3 | 4,614 | 96.9 | 1,669 |
| 17:00 | 96.3 | 4,680 | 97.5 | 1,702 |
| 18:00 | 98.0 | 4,685 | 96.9 | 1,662 |
| 19:00 | 99.3 | 4,500 | 99.4 | 1,659 |
| 20:00 | 98.2 | 4,460 | 98.8 | 1,732 |
| 21:00 | 97.3 | 4,441 | 96.0 | 1,751 |
| 22:00 | 94.7 | 4,364 | 94.9 | 1,749 |
| 23:00 | 95.1 | 4,405 | 94.9 | 1,697 |
Reproduce this result Start with the public export. No database or university account needed.
At 08:00 in the weekday outboundFromTown profile, A3 mean speed was 68.3 km/h at about 2,551 vehicles/hour. A6 carried a higher flow of about 2,921 vehicles/hour while averaging 95.6 km/h.
I therefore considered speed, flow, and occupancy together. This comparison does not establish why the roads differed or rank their capacity.
Thesis §6.5.2 · Figure 6.6
The rule requires speed ≤75 km/h, flow ≥900 vehicles/hour, and occupancy ≥10%, with at least eight observations from three dates per weekday/hour profile. At least two qualifying hours must be consecutive.
Seven windows qualified, all on the A3: mornings in one stored direction and afternoons in the other. These are project-specific rules without external ground-truth validation.
Thesis §6.6 · Table 6.5
Thesis table 6.5 · Europe/Luxembourg. End times are exclusive: 07:00–09:00 includes the hours beginning at 07:00 and 08:00. Values are observation-weighted means. Scroll horizontally on a narrow screen.
| Stored direction | Weekday | Local time | Speed km/h | Flow vehicles/h | Occupancy % |
|---|---|---|---|---|---|
outboundFromTown | Monday | 07:00–09:00 | 67.1 | 2,512 | 11.8 |
outboundFromTown | Tuesday | 07:00–10:00 | 67.5 | 2,476 | 11.7 |
outboundFromTown | Wednesday | 07:00–09:00 | 68.6 | 2,574 | 11.6 |
outboundFromTown | Thursday | 07:00–09:00 | 68.7 | 2,656 | 11.6 |
inboundTowardsTown | Monday | 15:00–19:00 | 69.5 | 2,059 | 11.2 |
inboundTowardsTown | Tuesday | 16:00–18:00 | 70.8 | 2,187 | 11.0 |
inboundTowardsTown | Thursday | 15:00–18:00 | 71.8 | 2,189 | 11.0 |
The direction labels are stored categories. Their geographic meanings were not independently verified for every road, so I do not translate them into compass or city-bound directions.
What failed · 25 July 2026
A high success rate can hide a long, continuous gap.
The evaluation reported successful outcomes for 98.14% of 161,746 feed-specific collection runs. Yet all seven measurement feeds shared a gap of roughly 18 hours and 10 minutes on 25 July.
Thesis §6.2.3 · Figure 6.1. Success rate is a share of logged runs, not uptime or complete temporal coverage.
What I would improve
About 10% of measurements lacked a normalized direction. The remaining labels still need independent geographic validation.
The dashboard computes quality dynamically. Historical fields remain unassessed; reproducing an earlier judgment needs its rule version and evaluation time.
Alert candidates, persistence, and lifecycle tests exist. Automatic synchronization was not wired into the main collector or dashboard flow.
43 days support weekly patterns, not seasonal claims. A longer dataset, independent traffic labels, load tests, and user studies would strengthen the evaluation.
