The HYROX results dataset¶

Pyrox publishes HYROX race results as structured, analysis-ready data. One row is one entry (an athlete, or a pair in doubles divisions) in one race, with every timed segment of that race as its own numeric column.

Coverage¶

Measured from the published manifest on 30 July 2026. Coverage grows every week, so treat these as floor values rather than fixed totals.

Dimension	Coverage
Race results	1,235,605 rows
Races	316 (season × location × year)
Locations	134
Seasons	1–9, spanning 2018–2026
Genders	`male`, `female`, `mixed`
Format	Parquet on a CDN, ~105 MB compressed

Per season, the dataset has grown sharply with the sport:

Season	Races	Locations	Results
1	9	9	4,750
2	11	11	7,271
3	4	4	1,433
4	24	21	15,538
5	38	34	43,607
6	53	49	144,114
7	69	65	289,842
8	104	97	698,163
9	4	4	30,887

Results, not athletes

1,235,605 counts result rows, not unique people. One athlete racing three times is three rows, and a doubles row represents a pair. Don't quote it as an athlete count.

How to see current coverage¶

Ask an AIPythonREST

Using Pyrox, what seasons, divisions, genders, locations and age groups
are available?

import pyrox

client = pyrox.PyroxClient()
client.list_seasons()
client.list_locations(season=8)
client.list_races(season=8)

curl https://pyrox-api.fly.dev/api/filter-options

Schema¶

Each race arrives as a pandas.DataFrame. All time columns are floats in minutes, parsed once on load, so there's no HH:MM:SS handling left in your analysis code.

Identity and totals¶

Column	Type	Meaning
`name`	string	Athlete name as shown on the official results site
`gender`	string	`male`, `female` or `mixed`
`division`	string	Division label, lowercase (see below)
`age_group`	string	Age bracket, e.g. `30-34`
`nationality`	string	Nationality as recorded upstream
`event_id`	string	Identifier for the race
`event_name`	string	Human-readable race name
`source_result_id`	string	Upstream identifier for the row
`total_time`	float (min)	Total race time
`work_time`	float (min)	Total time spent on stations
`run_time`	float (min)	Total running time across all eight legs
`roxzone_time`	float (min)	Total transition time between segments

That's 12 columns; with the 16 segment columns below, 28 in total.

The eight stations¶

Column	Station
`skiErg_time`	SkiErg
`sledPush_time`	Sled push
`sledPull_time`	Sled pull
`burpeeBroadJump_time`	Burpee broad jumps
`rowErg_time`	RowErg
`farmersCarry_time`	Farmers carry
`sandbagLunges_time`	Sandbag lunges
`wallBalls_time`	Wall balls

The eight runs¶

run1_time through run8_time: each 1 km leg gets its own column, so you can watch pacing decay across the race instead of staring at one aggregate running number.

Two surfaces, two schemas

The columns above describe the Parquet files the Python client reads. The hosted reporting service (REST and MCP) uses canonical *_time_min columns internally and resolves friendly metric aliases at its boundary. Don't assume a column name from one surface exists on the other; see Data model and the Client API.

Schema drift¶

Upstream data evolves between seasons. Check df.columns and adapt rather than assuming a column is always present.

Divisions¶

The supported division vocabulary is open, pro, doubles, pro_doubles, relay and adaptive. Which of them appear in any given race depends on what that event ran, so check with list_filters or by inspecting df["division"].unique() rather than assuming all six are present.

Pyrox doesn't pool divisions by default: an open athlete and a pro athlete aren't doing the same work, so a combined average benchmarks nothing and a combined percentile misleads.

How often it updates¶

New results land on a weekly cadence, published each Tuesday.

That's the server side. The Python client also caches locally, on top of that: races and the manifest for 2 hours, a full season for 1 hour. If you're after this week's publish and it isn't showing up yet, don't wait out the TTL — clear the cache:

client.clear_cache()

See Caching for the full set of defaults and how to opt out per call.

Three ways to query it¶

Access path	Best for	Cost
Python client	Notebooks, modelling, bulk analysis	`pip install pyrox-client`, free
MCP server	Conversational analysis with Claude, Codex or another MCP client	Public endpoint, no key
REST API	Dashboards and applications	Public endpoint, rate-limited

The Python client reads Parquet from the CDN directly and has no rate limit, which makes it the right choice for bulk work. The hosted service is limited to 60 requests per minute per IP, shared across REST and MCP calls, and returns 429 past that. The service also scales to zero when idle, so the first request after a quiet period can take a few minutes to answer while it boots and verifies the current database artifact.

Licence and attribution¶

The pyrox-client code is MIT licensed: use, modify and redistribute it freely, including commercially.

The underlying results are public HYROX data. Pyrox claims no ownership of it and is not affiliated with or endorsed by HYROX. A link back to this project is appreciated if you publish analysis built on the dataset, though not required; what you may do with the underlying race data is yours to check independently.

Pyrox is built on public source data and may inherit issues present in it. Where a cohort is small, the reporting surfaces say so rather than reporting a percentile from a handful of finishers as if it meant something.