I was initially very excited because this data is not nearly as it should be, especially historical forecasts. However, your pricing model seems to seriously limit the potential uses.
I would imagine that most people who have a serious interest in weather forecasting and would be target users of this service don’t think in terms of number of points but rather in lat/lon bounds, resolution, and number of hours & days for the predictions. I imagine they would also like to download a GRIB and not a CSV.
Your pricing for any large enough area to be useful presumably gets somewhat prohibitive, eg covering the North Pacific (useful for West Coast modeling) at 0.25 deg resolution might be ~300k data points per hour if I am doing my math right?
This is really great feedback. The truth is that the pricing model is being figured out, so if you have a specific type of use in mind maybe we could figure out what works best and it might become the actual default pricing model.
I tried to tie the pricing with the amount of processing that the API needs to do, which is closely related to the number of grib2 files that the API needs to download and process in order to create the response. And it doesn't change as much wether I extract 1 point or 1000 points. But I thought I had to draw the line somewhere or nobody would ever pay for anything because the freetier is enough.
But I might make it same price for maybe chunks of 5000 or more points.
From the line of business I come from the main usage is actually to extract scattered coordinates (think weather where specific assets are, like hotels or solar panels or wind farms) and not whole boundaries at full resolution but it makes a lot of sense that for other types of usage that is not the case.
It is definitely in the roadmap to be able to select based on lat/lon bounds and even shapes. Also to return data not as timeseries but the gridded data itself, either as grib2 or netCDF or parquet or a plain matrix of floats or png or even mp4 video.
Shoot me an email and I'll reach out when I can implement selecting based on bounds.
Out of the top of my head, re-slicing into grib files for the response is probably a big lift but some of the other formats like maybe netCDF or geoTIFF or just compressed array of floats might be a nice MVP.
In my experience, NWP data is big. Like, really big! Data over HTTP calls seems to limit the use case a bit; have you considered making it possible to mount storage directly (fsspec), and use fx the zarr format? In this way, querying with xarray would be much more flexible
The point of GribStream is that you don't need to download the grided data (cause as you say it is huge! and it grows many Gb every single hour for years).
This is an API that will do streaming extraction and only have you download what you actually need.
When you make the http request to the API, the API will be processing up to terabytes of data only to respond to you with maybe a few Kb of csv.
> It looks like e.g. sqlite-parquet-vtable implements shadow tables to memoize row group filters. How does JOIN performance vary amongst sqlite virtual table implementations?
datasette can JOIN across multiple SQLite databases.
Perhaps datasette and datasette-lite could support xarray and thus NetCDF-style multidimensional arrays in WASM in the browser with HTTP Content Range requests to fetch and cache just the data requested
> The header can also be used to verify the order of dimensions that a variable is saved in (which you will have to know to use, unless you’re using a tool like xarray that lets you refer to dimensions by name) - for a 3-dimensional variable, `lon,lat,time` is common, but some files will have the `time` variable first.
> Why is weather data stored in netcdf instead of tensors or sparse tensors?
NetCDF is a "tensor", at least in the sense of being a self-describing multi-dimensional array format. The bigger problem is that it's not a Cloud-Optimized format, which is why Zarr has become popular.
> Also, SQLite supports virtual tables that can be backed by Content Range requests
The multi-dimensional equivalent of this is "virtual Zarr". I made this library to create virtual Zarr stores pointing at archival data (e.g. netCDF and GRIB)
> xarray and thus NetCDF-style multidimensional arrays in WASM in the browser with HTTP Content Range requests to fetch and cache just the data requested
Pretty sure you can do this today already using Xarray and fsspec.
In theory it could be done. It is sort of analogous to what GribStream is doing already.
The grib2 files are the storage. They are sorted by time in the path and so that is used like a primary index. And then grib2 is just a binary format to decode to extract what you want.
I originally was going to write this as a plugin for Clickhouse but in the end I made it a Golang API cause then I'm less constrained to other things. Like, for example, I'd like to create and endpoint to live encode the gribfiles into MP4 so the data can be served as Video. And then with any video player you would be able to playback, jump to times, etc.
I might still write a clickhouse integration though because it would be amazing to join and combine with other datasets on the fly.
> It is sort of analogous to what GribStream is doing already.
The difference is presumably that you are doing some large rechunking operation on your server to hide from the user the fact that the data is actually in multiple files?
Cool project btw, would love to hear a little more about how it works underneath :)
I basically scrape all the grib index files to know all the offsets into all variables for all time. I store that in clickhouse.
When the API gets a request for a time range, set of coordinates and a set of weather parameters, first I pre-compute the mapping of (lat,lon) into the 1 dimensional index in the gridded data. That is a constant across the whole dataset. Then I query the clickhouse table to find out all the files+offset that need to be processed and all of them are queued into a multi-processing pool. And then processing each parameter implies parsing a grib file. I wrote a grib2 parser from scratch in golang so as to extract the data in a streaming fashion. As in... I don't extract the whole grid only to lookup the coordinates in it. I already pre-computed the index, so I can just decode every value in the grid in order and when I hit and index that I'm looking for, I copy it to a fixed size buffer with the extracted data. When You have all the pre-computed indexes then you don't even need to finish downloading the file, I just drop the connection immediately.
It is pretty cool. It is running in very humble hardware so I'm hoping I'll get some traction so I can throw more money into it. It should scale pretty linearly.
I've tested doing multi-year requests and the golang program never goes over 80Mb of memory usage. The CPUs get pegged so that is the limiting factor.
Grib2 complex packing (what the NBM dataset uses) implies lots of bit-packing. So there is a ton more to optimize using SIMD instructions. I've been toying with it a bit but I don't want to mission creep into that yet (fascinating though!).
That's pretty cool! Quite specific to this file format/workload, but this is an important enough problem that people might well be interested in a tailored solution like this :)
There used to be a really nice website that allowed you to scroll and scale years worth of historical weather for any location. Also allowed you to prepare average annual graphs.
It fell into disrepair after Flash was killed off, and the maintainer wasn't able to commit time to porting it over to a new platform.
I figure that the commodification of weather data is the real reason why it hasn't been replaced with a viable alternative.
Wow. This site is just amazing. Nice and useful (!) graphs especially for comparing cities between each other. And the small things like how they store all the state in the url so I could just share a comparison with a friend by just copying the url. Perfect.
Just a heads up that there is a similar Python package for this that is free called Herbie [1], though the syntax of this one looks a little easier to use.
This dataset is is part of the AWS Open Data program and so it is freely available from S3. By running the API within AWS then you get a massive latency and bandwidth advantage.
So with GribStream you are pushing the computation closer to where the big data is and only downloading the small data. And GribStream uses a custom grib2 parser that allows it to extract the data in a streaming fashion, using very little memory.
It makes a huge difference if you need to extract timeseries of a handful of coordinates for months at a time.
Grib2 files are much more compact than netCDF, just less convenient to use. But GribStream takes care of that and just returns you the timeseries for the coordinates you need.
Besides using the usual index files to only do http range requests for weather parameters of interest, GribStream also avoids creating big memory buffers to decode/decompress the whole grid. It does the decoding in a streaming fashion and only accumulates the values that are being looked for so it can do so very efficiently. It doesn't even finish downloading the partial grib file, it early aborts. And it also skips ahead many headers and parts of the grib2 format that are not really required or that can be assumed for being constant in the whole dataset. In other words, it cuts all possible corners and the parse is (currently) specifically optimized for the NBM and GFS datasets.
Although I intend to support several others, like the Rapid Refresh (RAP) model.
And the fact that this process runs close to the data (AWS), it can do so way faster than you can run it anywhere else.
Definitely more models right now. GribStream.com will be supporting many other models soon.
But open-meteo free access is only for non-commercial use. GribStream.com allows any use.
Also, can open-meteo query forecasts 10 days out at hourly resolution for 150.000 coordinates in a single request and take just 8 seconds? At what price?
Every hour these models forecast every hour 10 or 15 days ahead. The further ahead they become less accurate of course.
Having all historical weather forecasts is useful to backtest how accurate the model is because you can compare what was being forecasted at the time with what actually happened.
It is also useful to backtest derived models. If you want to figure out how accurate your model is forecasting something like solar power generation while using one of these weather datasets as features for cloud cover and solar radiation, you will want to run the model with the forecast that was current at that point in time. Otherwise you would be "cheating" by using data you wouldn't have had if you ran the model at that time.
In other words, you'd want to "see" the weather as it was being predicted at the time in order to find out what your model would predict for solar power generation.
I would imagine that most people who have a serious interest in weather forecasting and would be target users of this service don’t think in terms of number of points but rather in lat/lon bounds, resolution, and number of hours & days for the predictions. I imagine they would also like to download a GRIB and not a CSV.
Your pricing for any large enough area to be useful presumably gets somewhat prohibitive, eg covering the North Pacific (useful for West Coast modeling) at 0.25 deg resolution might be ~300k data points per hour if I am doing my math right?
I tried to tie the pricing with the amount of processing that the API needs to do, which is closely related to the number of grib2 files that the API needs to download and process in order to create the response. And it doesn't change as much wether I extract 1 point or 1000 points. But I thought I had to draw the line somewhere or nobody would ever pay for anything because the freetier is enough.
But I might make it same price for maybe chunks of 5000 or more points.
From the line of business I come from the main usage is actually to extract scattered coordinates (think weather where specific assets are, like hotels or solar panels or wind farms) and not whole boundaries at full resolution but it makes a lot of sense that for other types of usage that is not the case.
It is definitely in the roadmap to be able to select based on lat/lon bounds and even shapes. Also to return data not as timeseries but the gridded data itself, either as grib2 or netCDF or parquet or a plain matrix of floats or png or even mp4 video.
Out of the top of my head, re-slicing into grib files for the response is probably a big lift but some of the other formats like maybe netCDF or geoTIFF or just compressed array of floats might be a nice MVP.
[email protected]
This is an API that will do streaming extraction and only have you download what you actually need.
When you make the http request to the API, the API will be processing up to terabytes of data only to respond to you with maybe a few Kb of csv.
Also, SQLite supports virtual tables that can be backed by Content Range requests; https://www.sqlite.org/vtab.html
sqlite-wasm-http, sql.js-httpvfs; HTTP VFS: https://www.npmjs.com/package/sqlite-wasm-http
sqlite-parquet-vtable: https://github.com/cldellow/sqlite-parquet-vtable
Could there be a sqlite-netcdf-vtable or a sqlite-gribs-vtable, or is the dimensionality too much for SQLite?
From https://news.ycombinator.com/item?id=31824578 :
> It looks like e.g. sqlite-parquet-vtable implements shadow tables to memoize row group filters. How does JOIN performance vary amongst sqlite virtual table implementations?
https://news.ycombinator.com/item?id=42264274
SpatialLite does geo vector search with SQLite.
datasette can JOIN across multiple SQLite databases.
Perhaps datasette and datasette-lite could support xarray and thus NetCDF-style multidimensional arrays in WASM in the browser with HTTP Content Range requests to fetch and cache just the data requested
"The NetCDF header": https://climateestimate.net/content/netcdfs-and-basic-coding... :
> The header can also be used to verify the order of dimensions that a variable is saved in (which you will have to know to use, unless you’re using a tool like xarray that lets you refer to dimensions by name) - for a 3-dimensional variable, `lon,lat,time` is common, but some files will have the `time` variable first.
"Loading a subset of a NetCDF file": https://climateestimate.net/content/netcdfs-and-basic-coding...
From https://news.ycombinator.com/item?id=42260094 :
> xeus-sqlite-kernel > "Loading SQLite databases from a remote URL" https://github.com/jupyterlite/xeus-sqlite-kernel/issues/6#i...
NetCDF is a "tensor", at least in the sense of being a self-describing multi-dimensional array format. The bigger problem is that it's not a Cloud-Optimized format, which is why Zarr has become popular.
> Also, SQLite supports virtual tables that can be backed by Content Range requests
The multi-dimensional equivalent of this is "virtual Zarr". I made this library to create virtual Zarr stores pointing at archival data (e.g. netCDF and GRIB)
https://github.com/zarr-developers/VirtualiZarr
> xarray and thus NetCDF-style multidimensional arrays in WASM in the browser with HTTP Content Range requests to fetch and cache just the data requested
Pretty sure you can do this today already using Xarray and fsspec.
The grib2 files are the storage. They are sorted by time in the path and so that is used like a primary index. And then grib2 is just a binary format to decode to extract what you want.
I originally was going to write this as a plugin for Clickhouse but in the end I made it a Golang API cause then I'm less constrained to other things. Like, for example, I'd like to create and endpoint to live encode the gribfiles into MP4 so the data can be served as Video. And then with any video player you would be able to playback, jump to times, etc.
I might still write a clickhouse integration though because it would be amazing to join and combine with other datasets on the fly.
The difference is presumably that you are doing some large rechunking operation on your server to hide from the user the fact that the data is actually in multiple files?
Cool project btw, would love to hear a little more about how it works underneath :)
I basically scrape all the grib index files to know all the offsets into all variables for all time. I store that in clickhouse.
When the API gets a request for a time range, set of coordinates and a set of weather parameters, first I pre-compute the mapping of (lat,lon) into the 1 dimensional index in the gridded data. That is a constant across the whole dataset. Then I query the clickhouse table to find out all the files+offset that need to be processed and all of them are queued into a multi-processing pool. And then processing each parameter implies parsing a grib file. I wrote a grib2 parser from scratch in golang so as to extract the data in a streaming fashion. As in... I don't extract the whole grid only to lookup the coordinates in it. I already pre-computed the index, so I can just decode every value in the grid in order and when I hit and index that I'm looking for, I copy it to a fixed size buffer with the extracted data. When You have all the pre-computed indexes then you don't even need to finish downloading the file, I just drop the connection immediately.
It is pretty cool. It is running in very humble hardware so I'm hoping I'll get some traction so I can throw more money into it. It should scale pretty linearly.
I've tested doing multi-year requests and the golang program never goes over 80Mb of memory usage. The CPUs get pegged so that is the limiting factor.
Grib2 complex packing (what the NBM dataset uses) implies lots of bit-packing. So there is a ton more to optimize using SIMD instructions. I've been toying with it a bit but I don't want to mission creep into that yet (fascinating though!).
I'm tempted to port this https://github.com/fast-pack/simdcomp to native go ASM.
It fell into disrepair after Flash was killed off, and the maintainer wasn't able to commit time to porting it over to a new platform.
I figure that the commodification of weather data is the real reason why it hasn't been replaced with a viable alternative.
https://weatherspark.com/ is awesome.
It seems more static now.
Can you describe how it worked? Who knows, maybe I could make it happen.
I've been thinking of creating a more involved "demo" to entice traffic/links/SEO. Maybe something like this could be it.
[1]: https://herbie.readthedocs.io/en/stable/
This is a little different though.
This dataset is is part of the AWS Open Data program and so it is freely available from S3. By running the API within AWS then you get a massive latency and bandwidth advantage.
So with GribStream you are pushing the computation closer to where the big data is and only downloading the small data. And GribStream uses a custom grib2 parser that allows it to extract the data in a streaming fashion, using very little memory.
It makes a huge difference if you need to extract timeseries of a handful of coordinates for months at a time.
Cheers!
How does this compare to using Xarray on a netCDF dataset?
Besides using the usual index files to only do http range requests for weather parameters of interest, GribStream also avoids creating big memory buffers to decode/decompress the whole grid. It does the decoding in a streaming fashion and only accumulates the values that are being looked for so it can do so very efficiently. It doesn't even finish downloading the partial grib file, it early aborts. And it also skips ahead many headers and parts of the grib2 format that are not really required or that can be assumed for being constant in the whole dataset. In other words, it cuts all possible corners and the parse is (currently) specifically optimized for the NBM and GFS datasets.
Although I intend to support several others, like the Rapid Refresh (RAP) model.
And the fact that this process runs close to the data (AWS), it can do so way faster than you can run it anywhere else.
> Download numerical weather prediction datasets (HRRR, RAP, GFS, IFS, etc.) from NOMADS, NODD partners (Amazon, Google, Microsoft), ECMWF open data, and the Pando Archive System
The Herbie docs mention a "GFS GraphCast" but not yet GenCast? https://herbie.readthedocs.io/en/stable/gallery/noaa_models/...
"GenCast predicts weather and the risks of extreme conditions with state-of-the-art accuracy" (2024) https://deepmind.google/discover/blog/gencast-predicts-weath...
"Probabilistic weather forecasting with machine learning" (2024) ; GenCast paper https://www.nature.com/articles/s41586-024-08252-9
google-deepmind/graphcast: https://github.com/google-deepmind/graphcast
Are there error and cost benchmarks for these predictive models?
But open-meteo free access is only for non-commercial use. GribStream.com allows any use.
Also, can open-meteo query forecasts 10 days out at hourly resolution for 150.000 coordinates in a single request and take just 8 seconds? At what price?
I'll do a benchmark soon.
Having all historical weather forecasts is useful to backtest how accurate the model is because you can compare what was being forecasted at the time with what actually happened.
It is also useful to backtest derived models. If you want to figure out how accurate your model is forecasting something like solar power generation while using one of these weather datasets as features for cloud cover and solar radiation, you will want to run the model with the forecast that was current at that point in time. Otherwise you would be "cheating" by using data you wouldn't have had if you ran the model at that time.
In other words, you'd want to "see" the weather as it was being predicted at the time in order to find out what your model would predict for solar power generation.
I hope that makes sense.
Fixed!
Thank you!