##### The fields in azure.{vcpus,memory}.1hour.dedupe:

This file provides the time series.  Most of the meta data is unused
in the shared codebase.

1. linux epoch timestamp, here always 2588400

2. aggregation in {stopincludedmax,stopreleasedmin}

3. availability zone, here a dummy value of "az-azure-1a", or ""

4. series type, in {all_host_aggregate,flavor_type_payment,tenant_usage}

5. series subtype, e.g., flavor-type-ratio, or a dummy tenant ID or "All"

6. series sub-subtype, vcpus + flavor-type-ratio or "All" or ""

7. vm category, in {Delay-insensitive,Interactive,Unkown,""}

8. region, here a dummy value of "az-azure-1" or ""

9. resource in {vcpus,memory} (depending on file)

10. sampling period, here 3600

11. time series name, in {1.0,All,""}

12. Time series values, as 719 comma-separated values (~30 days of
hourly readings)


##### The fields in azure.1hour~168~168~24.{train,dev,test}

This file defines the sequence of windows (conditioning+prediction
ranges) via "offsets" or "indices" from which we create examples from
the time series.  Note that a portion of a training prediction range
may span beyond the edge of the training time period (first 20 days),
in which case we ignore the prediction target values occurring beyond
the training time period cutoff.  Similarly, we ignore parts of
dev/test prediction ranges that span beyond the dev/test time periods.

1. LAST DATESTAMP - timestamp of the last element of the window
extracted at that offset.

2. Line number, line number of file to get window from.

3. Target index (in line), where to begin the window in the line of
values.

4. Resource file to use, in {0,1} for "vcpus" or "memory"
