Hi all,

I'd like to report a weird slow read_parquet behavior between dask versions.
On all examples I'll be using:

• CentOS 7.7
• Python 3.6
• Fastparquet 0.4.0 (Installed with pip)

Dask versions were also installed with pip.
pip3 install dask=={version} dask[dataframe]=={version}

The I'm using is:

filters = [('x', '>=', time_from), ('x', '<', time_to)]
ddf = dd.read_parquet('path', columns=['x', 'y'], filters=filters, engine='fastparquet')
s = datetime.datetime.now().timestamp()
df = ddf.compute()
e = datetime.datetime.now().timestamp()
print('took', e-s, 'secs')

Test 1 <--- slow
dask==2.20.0
fastparquet==0.4.0
the output is:
took 62.25287580490112 secs

Test 2 <--- slow
dask==2.18.1
fastparquet==0.4.0
output:
took 62.15111708641052 secs

Test 3 <--- fast
dask==1.2.2
fastparquet==0.4.0
output:
took 1.4967741966247559 secs

The data schema is as shown below:
df.head()

df.dtypes

x    datetime64[ns]
y           float64
dtype: object

What happened:
read_parquet function is slow on recent dask versions > 1min

What you expected to happen:
faster read_parquet function.
dask=1.2.2 took 1sec

Regards,
C.