Time-Windowed Data

Info

Before you read this section, make sure that you understand the concepts described in the following sections:

• Datum
• Distributed Computing
• Developer Workflow

If you are analyzing data that is changing over time, you might need to analyze historical data. For example, you might need to examine the last two weeks of data, January’s data, or some other moving or static time window of data.

HPE Machine Learning Data Management provides the following approaches to this task:

• Fixed time windows: for rigid, fixed time windows, such as months (Jan, Feb, and so on) or days—01-01-17, 01-02-17, and so on.
• Moving time windows: for rolling time windows of data, such as three-day windows or two-week windows.

Fixed Time Windows #

Datum is the basic unit of data partitioning in HPE Machine Learning Data Management. The glob pattern property in the pipeline specification defines a datum. When you analyze data within fixed time windows, such as the data that corresponds to fixed calendar dates, HPE Machine Learning Data Management recommends that you organize your data repositories so that each of the time windows that you plan to analyze corresponds to a separate file or directory in your repository, and therefore, HPE Machine Learning Data Management processes it as a separate datum.

Organizing your repository as described above, enables you to do the following:

• Analyze each time window in parallel.
• Only re-process data within a time window when that data, or a corresponding data pipeline, changes.

For example, if you have monthly time windows of sales data stored in JSON format that needs to be analyzed, you can create a sales data repository with the following data:

sales
├── January
|   ├── 01-01-17.json
|   ├── 01-02-17.json
|   └── ...
├── February
|   ├── 01-01-17.json
|   ├── 01-02-17.json
|   └── ...
└── March
    ├── 01-01-17.json
    ├── 01-02-17.json
    └── ...

When you run a pipeline with sales as an input repository and a glob pattern of /*, HPE Machine Learning Data Management processes each month’s worth of sales data in parallel if workers are available. When you add new data into a subset of the months or add data into a new month, for example, May, HPE Machine Learning Data Management processes only these updated datums.

More generally, this structure enables you to create the following types of pipelines:

• Pipelines that aggregate or otherwise process daily data on a monthly basis by using the /* glob pattern.
• Pipelines that only analyze a particular month’s data by using a /subdir/* or /subdir/ glob pattern. For example, /January/* or /January/.
• Pipelines that process data on daily by using the /*/* glob pattern.
• Any combination of the above.

Moving Time Windows #

In some cases, you need to run analyses for moving or rolling time windows that do not correspond to certain calendar months or days. For example, you might need to analyze the last three days of data, the three days of data before that, or similar. In other words, you need to run an analysis for every rolling length of time.

For rolling or moving time windows, there are a couple of recommended patterns:

1. Bin your data in repository folders for each of the moving time windows.

2. Maintain a time-windowed set of data that corresponds to the latest of the moving time windows.

Bin Data into Moving Time Windows #

In this method of processing rolling time windows, you create the following two-pipeline DAGs to analyze time windows efficiently:

By splitting this analysis into two pipelines, you can benefit from using parallelism at the file level. In other words, Pipeline 1 can be easily parallelized for each file, and Pipeline 2 can be parallelized per bin. This structure enables easy pipeline scaling as the number of files increases.

For example, you have three-day moving time windows, and you want to analyze three-day moving windows of sales data. In the first repo, called sales, you commit data for the first day of sales:

sales
└── 01-01-17.json

In the first pipeline, you specify to bin this data into a directory that corresponds to the first rolling time window from 01-01-17 to 01-03-17:

binned_sales
└── 01-01-17_to_01-03-17
    └── 01-01-17.json

When the next day’s worth of sales is committed, that data lands in the sales repository:

sales
├── 01-01-17.json
└── 01-02-17.json

Then, the first pipeline executes again to bin the 01-02-17 data into relevant bins. In this case, the data is placed in the previously created bin named 01-01-17 to 01-03-17. However, the data also goes to the bin that stores the data that is received starting on 01-02-17:

binned_sales
├── 01-01-17_to_01-03-17
|   ├── 01-01-17.json
|   └── 01-02-17.json
└── 01-02-17_to_01-04-17
    └── 01-02-17.json

As more and more daily data is added, your repository structure starting to looks as follows:

binned_sales
├── 01-01-17_to_01-03-17
|   ├── 01-01-17.json
|   ├── 01-02-17.json
|   └── 01-03-17.json
├── 01-02-17_to_01-04-17
|   ├── 01-02-17.json
|   ├── 01-03-17.json
|   └── 01-04-17.json
├── 01-03-17_to_01-05-17
|   ├── 01-03-17.json
|   ├── 01-04-17.json
|   └── 01-05-17.json
└── ...

The following diagram describes how data accumulates in the repository over time:

Your second pipeline can then process these bins in parallel according to the glob pattern of /* or as described further. Both pipelines can be easily parallelized.

In the above directory structure, it might seem that data is duplicated. However, under the hood, HPE Machine Learning Data Management deduplicates all of these files and maintains a space-efficient representation of your data. The binning of the data is merely a structural re-arrangement to enable you to process these types of moving time windows.

It might also seem as if HPE Machine Learning Data Management performs unnecessary data transfers over the network to bin files. However, HPE Machine Learning Data Management ensures that these data operations do not require transferring data over the network.

Maintaining a Single Time-Windowed Data Set #

The advantage of the binning pattern above is that any of the moving time windows are available for processing. They can be compared, aggregated, and combined in any way, and any results or aggregations are kept in sync with updates to the bins. However, you do need to create a process to maintain the binning directory structure.

There is another pattern for moving time windows that avoids the binning of the above approach and maintains an up-to-date version of a moving time-windowed data set. This approach involves the creation of the following pipelines:

For example, you have three-day moving time windows, and you want to analyze three-day moving windows of sales data. The input data is stored in the sales repository:

sales
├── 01-01-17.json
├── 01-02-17.json
├── 01-03-17.json
└── 01-04-17.json

When the January 4th file, 01-04-17.json, is committed, the first pipeline pulls out the last three days of data and arranges it in the following order:

last_three_days
├── 01-02-17.json
├── 01-03-17.json
└── 01-04-17.json

When the January 5th file, 01-05-17.json, is committed into the sales repository:

sales
├── 01-01-17.json
├── 01-02-17.json
├── 01-03-17.json
├── 01-04-17.json
└── 01-05-17.json

the first pipeline updates the moving window:

last_three_days
├── 01-03-17.json
├── 01-04-17.json
└── 01-05-17.json

The analysis that you need to run on the moving windowed dataset in moving_sales_window can use the / or /* glob pattern, depending on whether you need to process all of the time-windowed files together or if they can be processed in parallel.

Warning

When you create this type of moving time-windowed data set, the concept of now or today is relative. You must define the time based on your use case. For example, by configuring to use UTC. Do not use functions such as time.now() to determine the current time. The actual time when this pipeline runs might vary.