flight leg record

A flight leg record is one observation of one flight segment flown by one carrier on one date between two airports - never a daily aggregate or a route average. BTS TranStats states the granularity directly: "On-time data: one row per flight", with the Airline On-Time Performance series reaching back to October 1987. The Kaggle mirror of 2015 U.S. DOT on-time data keeps the same grain as flights.csv, one of three CSVs holding 5,819,079 rows and 31 columns covering carrier code, flight number, tail number, origin and destination, scheduled versus actual times, delay minutes and cancellation reasons. AviationStack preserves the identity but nests the payload: one record per flight leg with embedded departure, arrival, aircraft, airline and status objects. FlightAware works at the same per-flight level, publishing per-flight position and event records alongside airport-level arrival/departure boards.

Grain errors are the most expensive mistake in aviation analytics. If a vendor hands you monthly totals when your model expects legs, every delay distribution, cancellation rate and connection-time study is wrong before it starts.

• Aggregates cannot be un-aggregated. BTS T-100 traffic files arrive as one row per carrier-aircraft-service-class-segment-month, so per-flight questions need the on-time files instead - the two live in the same portal with different grains.
• Nested JSON is still one leg. AviationStack's embedded departure, arrival, aircraft and status objects describe one leg per record, but flattening them badly silently drops or duplicates legs.
• Boards are not archives. FlightAware's airport arrival/departure tables are views over per-flight records; its database spans over 699 million flights with roughly 201,000 arrivals tracked per day, and licensed history via AeroAPI goes back to 1 January 2011.

1. Read the stated granularity before modeling. BTS publishes "On-time data: one row per flight" for its on-time database; T-100 segment files are instead one row per carrier-aircraft-service-class-segment-month.
2. Reconcile row counts against the advertised universe. The 2015 Kaggle snapshot holds 5,819,079 rows for every domestic flight that year - a supplier's extract should reconcile to that order of magnitude.
3. Locate every attribute before flattening. In flights.csv the 31 columns sit flat; in AviationStack responses the departure, arrival, aircraft, airline and status blocks are nested objects inside one leg record.

Entries that sit next to flight leg record in this glossary:

• airline-on-time-performance
• ml-ready-flight-delay-snapshot
• realtime-flight-status-api
• socrata soda api
• grain truck rates
• refrigerated truck rates
• realtime flight status api