`read_deltalake` vs `read_parquet` performance

[j-bennet]

`read_deltalake` vs `read_parquet` performance

[j-bennet]

Second attempt, after merging #57.

Cluster: https://cloud.coiled.io/clusters/245853?account=dask-engineering

dataset	computation	timing (read_parquet)	timing (read_deltalake)
ds20f_100M	ddf["int1"].sum().compute()	CPU times: user 41.2 ms, sys: 10.6 ms, total: 51.8 ms, Wall time: 9.48 s	CPU times: user 181 ms, sys: 42.3 ms, total: 224 ms, Wall time: 59.8 s
ds20f_100M	ddf.describe().compute()	CPU times: user 243 ms, sys: 27.9 ms, total: 271 ms, Wall time: 23 s	CPU times: user 308 ms, sys: 51.7 ms, total: 360 ms, Wall time: 1min 1s
ds25f_250M	ddf["int1"].sum().compute()	CPU times: user 63.8 ms, sys: 15.9 ms, total: 79.7 ms, Wall time: 16.6 s	CPU times: user 716 ms, sys: 182 ms, total: 897 ms, Wall time: 3min 51s
ds25f_250M	ddf.describe().compute()	CPU times: user 623 ms, sys: 71 ms, total: 694 ms, Wall time: 1min 9s	CPU times: user 986 ms, sys: 189 ms, total: 1.17 s, Wall time: 3min 52s
ds50f_500M	ddf["int1"].sum().compute()	CPU times: user 199 ms, sys: 47.5 ms, total: 246 ms, Wall time: 1min	CPU times: user 2.89 s, sys: 799 ms, total: 3.69 s, Wall time: 16min 7s
ds50f_500M	ddf.describe().compute()	CPU times: user 3.45 s, sys: 383 ms, total: 3.83 s, Wall time: 5min 37s	CPU times: user 5.36 s, sys: 832 ms, total: 6.19 s, Wall time: 16min 24s

Looks like dask-deltalake is doing something very inefficient.