As of May 13, 2022, access to the NYC Taxi data has changed. Parquet has now become the new default file format, instead of CSV. Practically, this means you will need to change two things in your code:
# read in 2012 Parquet data
ddf = dd.read_parquet(
This post provides context on the change, explains the benefits of the Parquet file format and demonstrates running computations on 11 years of NYC Taxi data using a Coiled cluster. It also shows you can still access the old CSV files and explains why you probably don't want to do this.
The NYC TLC dataset is one of the most well-known public datasets. It's one of the few public datasets that is both large (>100GBs) and relatively clean. Because of this, many companies use it for demos and to run internal tests. The dataset has been a reliable feature of the big data landscape for well over a decade.
Over the weekend, this suddenly changed. If you now try to run your familiar read_csv call, you will run into an IndexError: list index out of range:
# read in 2012 CSV data
ddf = dd.read_csv(
IndexError Traceback (most recent call last)
Input In , in <cell line: 2>()
1 # read in 2012 CSV data
----> 2 ddf = dd.read_csv(
3 "s3://nyc-tlc/trip data/yellow_tripdata_2012-*.csv",
File ~/mambaforge/envs/dask-dataframes/lib/python3.9/site-packages/dask/dataframe/io/csv.py:741, in make_reader.<locals>.read(urlpath, blocksize, lineterminator, compression, sample, sample_rows, enforce, assume_missing, storage_options, include_path_column, **kwargs)
--> 741 return read_pandas(
File ~/mambaforge/envs/dask-dataframes/lib/python3.9/site-packages/dask/dataframe/io/csv.py:520, in read_pandas(reader, urlpath, blocksize, lineterminator, compression, sample, sample_rows, enforce, assume_missing, storage_options, include_path_column, **kwargs)
515 paths = get_fs_token_paths(urlpath, mode="rb", storage_options=storage_options)[
519 # Infer compression from first path
--> 520 compression = infer_compression(paths)
522if blocksize == "default":
523 blocksize = AUTO_BLOCKSIZE
IndexError: list index out of range
Parquet has become the new default for the NYC TLC data. To access the data, you will need to:
Dask is the best way to read the new NYC Taxi data at scale. Dask enables you to maximise the parallel read/write capabilities of the Parquet file format.
You can also use pandas with pd.read_parquet() but this would mean you are limited to using only a single CPU core to process your data. This will make your workflow slower and less scalable.
Read this post to learn more about how Dask can help speed up your data analyses.
You can still access the CSV data in the csv_backup directory:
Note that these CSV files will suffer from slower parallel I/O, less compression options and no column pruning or predicate pushdown than the new Parquet files. If you are working at scale - and unless you have a very strong reason to use CSV - you should generally use Parquet rather than CSV. Read this blog to learn more about how to write Parquet files with Dask.
The only drawback of the NYC TLC Parquet files is that downloading these parquet files takes a long time, as there are 12 very large partitions per year. It's best to repartition the dataset to a more optimal size for parallel IO and faster computations.
The code below demonstrates performing a groupby computation on the NYC TLC data for the years 2011 to 2021. This is more than 200GB of data uncompressed on disk. It's unlikely that your local machine has the RAM to run this analysis. We'll be running our computations in the cloud on a Coiled cluster with 50 workers and 16GB RAM each. Read our documentation to get started with Coiled.
Now we can load in our dataset:
And now run our groupby computation:
Yes, this access change is painful and means you may have to update some legacy code. But the change is made for good reason: Parquet is a much more efficient file format, especially when working with datasets of this scale that are best read in parallel. Parquet enables you to perform time-saving operations like parallel IO, column pruning, and predicate pushdown.