Sources

Everything on this site is somebody else's open data. Nothing here is scraped, nothing is behind a key, and every feed below permits republication. This page is the whole list, with what each one costs to fetch — because a public service's bandwidth is a real thing to spend.

What IMGW-PIB publishes

Poland's Institute of Meteorology and Water Management is, on the evidence of this scan, one of the more open meteorological services in Europe. The legal basis is stated on its own portal: Poland's open data act of 11 August 2021, and Commission Implementing Regulation (EU) 2023/138, which places meteorological model output, radar and station observations in the category of high-value datasets that must be available free, in bulk, machine-readable and under an open licence. IMGW complies with it properly.

Feed	What it is	Format	Cadence	Cost per cycle	Used for
COSMO 2.8 km COSMO_HVD_<HH>_<00,01>	The national mesoscale model. 380 × 405 on a rotated grid, 50 hybrid levels, hourly to T+60.	GRIB1	4 runs/day	9.7 GB	All the 2.8 km maps, and every field on this site that does not exist elsewhere
POLRAD composite COMPO_CMAX_250, COMPO_SRI	National composite of ten Doppler radars — column-max reflectivity 900 × 900, surface rain rate 800 × 800.	ODIM HDF5	every 5 min	0.14 MB/frame	Radar, and the radar overlay in the explorer
synop	62 synoptic stations, hourly.	JSON	hourly	14 kB	Model verification, halny
hydro	913 river gauges with each gauge's own warning and alarm stage.	JSON	~hourly	550 kB	Rivers
warningsmeteo / warningshydro	Warnings, degree 1–3, each with the TERYT codes of the powiats it covers.	JSON	on issue	120 kB	Warnings, front page
meteo	788 automatic telemetric stations, 10-minute precipitation.	JSON	10 min	520 kB	Not yet used

Two things IMGW's own documentation gets wrong, both found by reading the files rather than the page, and both worth stating because anyone else building on this feed will hit them.

The COSMO readme.txt describes a 0.0625° grid of 415 × 460 points with the rotated pole at 32.5 °N and 40 levels. The GRIB says 0.025°, 380 × 405, southern pole at −40 °N and 50 levels. The readme describes the 7 km run this feed used to carry. Believing it builds a sampler onto a domain three times too large.
The readme's field names are not the file's. The land mask is lsm and not FR_LAND; the surface height is HSURF and not HH; the accumulated snowfall is lssf and not SNOW_GSP. Asking for a name eccodes does not use finds nothing and renders an empty field with no error anywhere.

A third, in the radar: the ODIM files state both a pixel size and their four corners, and the two disagree — the scales give 906.8 × 895.5 km where the corners give exactly 900 × 900. The corners are right.

The other model feeds

Source	Model	Transport	Cost per cycle	Licence
DWD opendata.dwd.de	ICON-EU 7 km, T+120	One bzip2 file per variable per step, ~1 MB each	~2.2 GB	CC BY 4.0
Met Office on AWS Open Data	Global 10 km, T+168	One NetCDF per variable per step, whole-globe, cropped on arrival	~2.6 GB	British Crown copyright, CC BY 4.0
ECMWF data.ecmwf.int	IFS 0.25°, T+240	HTTP byte ranges against the published JSON-lines index — a step costs the 12 records this site draws, about 6 MB against 94	~0.3 GB	CC BY 4.0
GIOŚ api.gios.gov.pl	Air quality index, ~290 stations	JSON-LD, one request per station	~1 MB	Polish open data act
Natural Earth	Coast, borders, voivodeships, rivers, lakes	Shapefiles, 10 m, built once into pre-projected JSON	once	Public domain

On bandwidth, and being a good guest

COSMO is by far the most expensive thing here, and the reason is worth stating plainly. IMGW's download endpoint does not support HTTP byte ranges — a Range request returns the whole object — so a lead time cannot be subsetted by asking for the fifteen records this site draws out of the 523 in the file. It has to be streamed.

What this site does instead is hang up. A GRIB file is a sequence of self-delimiting messages, so the ingest reads the socket, cuts messages out of the buffer as they arrive, and closes the connection the moment the last wanted record has gone past. The useful records end at about 90 % of the file, so that saves roughly a tenth — measured on a real cycle, 9,740 MB pulled against 10,494 MB of file. The larger benefit is that the resident set stays at a few hundred kilobytes per stream instead of 161 MB.

At four cycles a day that is about 39 GB a day from a public service. It is well within what the regulation contemplates, and it is still somebody's bandwidth. If IMGW ever asks, the axis is one constant in ingest/sources.py and halving the cycles halves the bill with no other change.

What this site is not

It is not an official source and it issues no warnings. It re-projects and draws other people's data, and it can be stale, incomplete or wrong. For anything that matters:

meteo.imgw.pl — IMGW's own forecasts and meteorological warnings
hydro.imgw.pl — river gauges and hydrological warnings
powietrze.gios.gov.pl — GIOŚ air quality, with concentrations rather than index values
danepubliczne.imgw.pl — the open data portal everything Polish here comes from

The inversion, smog and halny fields are computed here and are ours, not IMGW's. Their definitions are given in full on the smog page and the halny page so they can be argued with.