This looks good to me. Merging later today if there are no further comments.

@petioptrv if you have any interest, it would be nice to see the performance improvement that this provides. Do you have any interest in running the small example provided in the original issue both before and after this change and reporting the difference? This isn't required, but I suspect that people here would enjoy seeing the results of this work.

Merging this in. Thanks @petioptrv !

Also, I notice that this is your first code contribution to this repository. Welcome!

Apologies for the late reply. I've been very busy at work lately. Thanks for the welcome @mrocklin! We started using Dask at my workplace recently, so I thought it would be the perfect opportunity to get a feel for open-source contributions.

As for the benchmark, here's the original problem with pre- and post-benchmarks:

import time

import pandas as pd
import dask.dataframe as dd

data = {}
for i in range(10000):
   data["col"+str(i)] = [1.0] * 10
df = pd.DataFrame(data)
ddf = dd.from_pandas(df, npartitions=1)

t0 = time.time()
dloc = ddf.loc[0]
t1 = time.time()
print(t1-t0)  # 2.118 pre vs 0.001 post

t0 = time.time()
dmeta = ddf._meta_nonempty
t1 = time.time()
print(t1-t0)  # 2.628 pre vs 0.984 post