digitalhen/us-airport-wait-times

US Airport Security & Immigration Wait Times

Minute-resolution TSA security checkpoint wait times for 30 US airports, plus hourly CBP immigration hall wait times for arriving international passengers.

Airports publish their current wait time and then overwrite it. Nobody keeps the history. This dataset is that history: a continuous archive collected by polling each airport's public feed roughly once a minute.

Collection began 1 April 2026 with the New York, Philadelphia and Florida airports; the rest were added later. Coverage therefore starts on different dates per airport — see the From column below, and do not read a short history as a quiet period.

As far as I know there is no other open historical dataset of US airport checkpoint wait times.

Collected for, and used in production by, the live wait-time boards at nycwaittimes.com and sister sites. The archive is a byproduct of running those services, which is why it keeps growing.

Coverage

Airport	Terminals	Observations	From	To
EWR — Newark Liberty	3	3,094,855	2026-04-01	2026-08-19
JFK — John F. Kennedy	5	2,785,266	2026-04-01	2026-08-19
PHL — Philadelphia	6	2,469,787	2026-04-01	2026-08-19
MCO — Orlando	3	1,856,904	2026-04-01	2026-08-19
MIA — Miami	3	1,741,502	2026-04-01	2026-08-19
LGA — LaGuardia	3	1,333,304	2026-04-01	2026-08-19
BOS — Boston Logan	7	153,648	2026-08-14	2026-08-19
SFO — San Francisco	6	138,072	2026-08-14	2026-08-19
DFW — Dallas/Fort Worth	5	128,050	2026-08-14	2026-08-19
IAH — George Bush Intercontinental	7	115,254	2026-08-14	2026-08-19
LAS — Harry Reid	4	102,416	2026-08-14	2026-08-19
BWI — Baltimore/Washington	4	102,256	2026-08-14	2026-08-19
PHX — Phoenix Sky Harbor	5	64,030	2026-08-14	2026-08-19
SEA — Seattle–Tacoma	6	61,923	2026-08-14	2026-08-19
CLT — Charlotte Douglas	3	51,224	2026-08-14	2026-08-19
MSP — Minneapolis–Saint Paul	4	51,224	2026-08-14	2026-08-19
SAN — San Diego	2	51,224	2026-08-14	2026-08-19
DEN — Denver	2	51,224	2026-08-14	2026-08-19
HOU — William P. Hobby	1	25,612	2026-08-14	2026-08-19
DTW — Detroit Metropolitan	2	25,480	2026-08-14	2026-08-19
DCA	3	21,520	2026-08-17	2026-08-19
PDX	2	18,924	2026-08-17	2026-08-19
CVG	1	9,462	2026-08-17	2026-08-19
PIT	1	9,462	2026-08-17	2026-08-19
CLE	3	9,192	2026-08-17	2026-08-19
JAX	1	5,688	2026-08-17	2026-08-19
IAD	1	4,731	2026-08-17	2026-08-19
CMH	1	4,731	2026-08-17	2026-08-19
STL	1	4,731	2026-08-17	2026-08-19
SLC	1	4,726	2026-08-17	2026-08-19

152 distinct lanes (airport × terminal × queue type), 14.5M raw observations. Currently growing by about 430,000 rows per day; this snapshot was taken 2026-08-19.

Subsets

Config	Rows	What it is
waits	7.1M	Start here. One row per distinct reading, with first_seen/last_seen.
waits_raw	14.5M	Every individual scrape, unfiltered.
hourly	131k	Per lane per hour: mean, p50, p75, p95, max. Open lanes only.
cbp	26k	CBP immigration hall waits, with passenger and booth throughput.

from datasets import load_dataset

ds = load_dataset("digitalhen/us-airport-wait-times", "waits", split="train")

Or query the Parquet directly without downloading the whole thing:

-- DuckDB
SELECT terminal, queue_type, round(avg(avg_wait), 1) AS avg_wait
FROM 'https://huggingface.co/datasets/digitalhen/us-airport-wait-times/resolve/main/data/hourly/hourly.parquet'
WHERE airport = 'JFK' AND hour_of_day BETWEEN 5 AND 8
GROUP BY 1, 2 ORDER BY 3 DESC;

Why waits rather than waits_raw

The scraper polls every ~60 seconds, but the upstream feeds refresh much more slowly — so most consecutive scrapes re-record an identical reading. waits collapses each run of unchanged readings into a single row spanning first_seen → last_seen, with n_scrapes recording how many polls saw it. That is 49% of the rows and none of the information.

Use waits_raw only if you specifically care about polling artifacts.

Schema

waits / waits_raw

Column	Type	Notes
airport	string	IATA code. Hive partition key.
terminal	string	As published by the airport (4, C, A-East, South).
queue_type	string	standard or precheck (TSA PreCheck).
wait_minutes	int16	Posted wait in minutes.
status	string	open, no_wait, closed, unknown.
status_raw	string	The upstream string, unmodified.
is_open	bool	True when status is open or no_wait. Filter on this.
wait_is_estimated	bool	See "Estimated values" below.
reported_at_local	string	Upstream's own "last updated" stamp, local clock time.
first_seen,last_seen	timestamptz	waits only — span over which this reading stood.
n_scrapes	int32	waits only — polls that saw this reading.
scraped_at	timestamptz	waits_raw only — when this poll happened.

hourly

Open lanes only, bucketed in US/Eastern (the operating timezone of every airport covered).

Column	Type	Notes
airport,terminal,queue_type	string
hour_et	timestamp	Hour bucket, US/Eastern.
day_of_week	int8	0 = Sunday.
hour_of_day	int8	0–23.
avg_wait	double
p50_wait,p75_wait,p95_wait,max_wait	int16
n_samples	int32	Scrapes behind the bucket — see the resolution caveat.

cbp

CBP immigration halls, already hourly at source. Unlike the TSA feed this carries throughput, which makes staffing analysis possible.

Column	Type	Notes
airport,terminal,site_id	string	site_id is CBP's own identifier.
flight_date	date
hour_range,hour_of_day	string, int8	e.g. 0500 - 0600 → 5.
avg_wait,max_wait	int32	Minutes, all passengers.
us_citizen_avg_wait,us_citizen_max_wait,us_citizen_count	int32
non_us_citizen_avg_wait,non_us_citizen_max_wait,non_us_citizen_count	int32
total_passengers,global_entry_count	int32
booths_used,flight_count	int32	Open booths and arriving flights that hour.

Caveats — please read before drawing conclusions

These are posted wait times, not measured transit times. Every value is what the airport chose to display. It is an operator's estimate, produced by methods that differ per airport and are not published. Treat this as a dataset about what airports told passengers, which may or may not match what passengers experienced.

Effective resolution is not one minute, and it varies by airport. The scraper polls every ~60s, but each operator refreshes on its own schedule, and those differ by more than an order of magnitude — the Port Authority feeds behind JFK, EWR and LGA update far less often than most. n_scrapes in waits is the honest signal: a high value means the reading sat unchanged, not that the queue was stable. Do not read sub-refresh-interval structure as real, and be careful comparing volatility across airports — you may be measuring publishing cadence rather than queues.

Estimated values. ~12,000 rows arrived as free text (Greater than 30 minutes, Less than 5 mins, At least 30 mins) rather than a number. These were converted at collection time — lower bound for "greater than", midpoint for "less than" — and flagged with wait_is_estimated = true. status_raw always preserves the original. Drop them if you need exact readings.

closed and unknown. 484k rows report a closed checkpoint and 624k carry no usable status. Both are kept, because checkpoint closures are themselves interesting, but wait_minutes is meaningless for them. Filter on is_open.

Terminal identifiers are the airport's own and are not consistent between airports. Some terminals expose several physical checkpoints under one label, so multiple readings share a terminal and timestamp. EWR Terminal B is the clearest case: the operator reports three checkpoints under the single label "B", so each tick carries three rows per queue type with differing waits. Average them, or treat the set as the terminal's spread — do not assume one is canonical.

Gaps happen. Upstream outages, deploys and network failures all leave holes. There is no backfill and no interpolation — a missing interval is simply absent.

An example of what's in here

TSA PreCheck's value is usually asserted, rarely measured. With 2.7M paired observations (same terminal, same moment, standard lane vs PreCheck lane) it can be measured directly. Mean minutes saved over the trailing 60 days, for airports with at least 14 days of paired history:

Airport	Standard	PreCheck	Saved
JFK	18.2	4.2	14
EWR	13.7	2.9	10.8
PHL	7.3	2	5.3
MCO	7.8	3.1	4.8
LGA	5	1.4	3.6
MIA	4.5	3.1	1.4

The spread is the interesting part, and it widens sharply at peak: the worst terminal-hours run several times the airport-wide average above. Note the resolution caveat before comparing airports too finely.

Provenance

Collected by the scraper behind nycwaittimes.com, which polls each airport's public wait-time feed roughly once a minute and appends to Postgres. Every source is the airport operator's own public endpoint — the Port Authority of NY & NJ for JFK/EWR/LGA, and each other operator's public feed or published page — plus U.S. Customs and Border Protection's public airport wait-time service for the cbp subset. There is no official nationwide TSA feed; each airport is integrated separately.

A few operators publish no timestamp of their own. For those, notably DTW, reported_at_local records when the reading was collected rather than when the airport says it was measured. Everywhere else it is the operator's own stamp.

No personal data is collected. Every value is an aggregate the operator already publishes to the travelling public.

Files are regenerated by scripts/export-open-data.sh, which pins a single snapshot timestamp so all four subsets describe one consistent moment. That timestamp is recorded in snapshot.json.

License

CC BY 4.0 — use it for anything, including commercially, with attribution.

@misc{us_airport_wait_times,
  title  = {US Airport Security and Immigration Wait Times},
  author = {Williams, Henry},
  year   = {2026},
  url    = {https://huggingface.co/datasets/digitalhen/us-airport-wait-times},
  note   = {Minute-resolution TSA checkpoint and CBP immigration wait times
            for major US airports, collected from public airport feeds}
}

If you build something with this, I'd like to hear about it.