investigate slow adls

[twuebi]

Creating a warehouse in adls is approximately 5 times slower than creating a warehouse in aws s3 (sts-enabled).

Creating 10 warehouses (times are millis):

s3 took: '13386', mean: '1338', all: '[1687, 1215, 1204, 2282, 1204, 1131, 1159, 1140, 1122, 1242]
adls took: '66873', mean: '6687', all: '[6824, 6644, 6545, 6673, 6895, 6682, 6546, 6597, 6643, 6824]

It seems that 7 times more connections are started to azure than to aws:

cat log | grep  "starting new connection" | rev | cut -f 1 -d" " | rev | sort | uniq -c
  30 https://ht0cht0tabular0eastus.blob.core.windows.net/
  50 https://ht0cht0tabular0eastus.dfs.core.windows.net/
  60 https://login.microsoftonline.com/
  20 https://s3.eu-north-1.amazonaws.com/

Recreating FileIOs means clients are also recreated every time which may mean that connection pooling is not really working.

[twuebi]

#425 reuses http clients against azure and reduces fetching sas from 800ms to 200ms for subsequent calls, also makes creation of table faster, down from 1700ms to 1200ms