Metadata Requirements and Examples

Field level metadata tagging:
1. Capabilities to add and retrieve metadata or tags at field level for dataset schemas.
2. The capability needs to be exposed to data pipelines and plugins as well as enterable through the UI.
3. It should be possible to add tags at field level both automated and manual way. Andreas: not quite clear what that means - automated/manual? Sagar: My understanding is automated way meaning tags are emitted programmatically (by plugins or CDAP programs) while manual way is assigning tags through UI (or REST API). For example, plugin can be implemented to mark the user address field as 'PII' in automated fashion, however once the data is landed in the data lake, data governance officer decides that another field say, phone number should also have been marked as 'PII'. He should be able to do that through UI which is manual way. We can confirm this requirement though.
  
  Example 1: Metadata based processing: Consider a CDAP data pipeline which reads a user profile dataset. A user profile contains several PII fields such as social security number, phone number, etc. A stage in the data pipeline is responsible for the obfuscating the PII data. This stage looks at all the fields and if a field is tagged as 'PII', it obfuscates it.
  
  Example 2. Metadata based discovery: Data governance officer might want to get the list of PII fields in the user profile dataset present in the data lake. In this case he would like to get the fields social security number, phone number, and address as well.
  
  Example 3. Metadata based discovery: A Data Scientist wants to use a user profile dataset. To comply with regulations, he wants to use a dataset where the phone numbers are anonymized. He searches for "profile phone:anonymized".
Finer granularity (File/Partition/Table in a database) metadata:
1. Capabilities to annotate the object level metadata. (we need to define what different types of objects are and also is it possible to specify the custom objects).
  For example:
  1. How is the directory on HDFS is created?
  2. Which files are responsible for creating this partition in a dataset
  3. Run id of the pipeline that creates a specific file in a dataset
  4. Checksum, umber of errors, and other operational meta data for a file in a dataset
2. This helps answering the compliance related questions such as how the file got created. This also gives a traceability so that If the file is bad we know whats the origin of the file.
  
  Example 1: Organization can have multiple business units such as 'HR', 'Finance' etc. SFTP is commonly used mechanism for sharing data within the business unit in an organization. Since these files are used within the single business unit only the format in which the files are stored might not be consistent in the organization. Files from the HR maybe stored as CSV, while files from Finance may be stored as rich XML format. In order to perform analysis on the file, they must be imported into the HDFS by using separate CDAP data pipelines, or different configurations for the same pipeline. While importing, normalization is done on the files to store them in HDFS in common format.
  
  Data governance officer still need to look at the data in HDFS at file level rather than abstracted CDAP dataset level. By looking at the file he would want to know the checksum associated with the file, which business unit the file belongs to, when the file was created etc.
  
  Example 2: As an extension to the above example, multiple different services (such as CDAP, Custom Hadoop stack etc.) are responsible for pumping files in the data lake. If we tag the file with the name of the process such as 'CDAP', then the data governance officer will know which service is bringing the file in.
  
  Example 3: A pipeline ingests data into files. It gathers metrics about the data quality, such as how many schema errors, how many syntax errors, etc. were encountered during processing. At the end of the pipeline, a custom action computes a data quality score (between 0 and 100) and assigns it as meta data to the output file (or directory) in the sink dataset. Downstream pipelines that consume this dataset will decide based on each file's data quality score whether to process or to ignore that file.
Store metadata along with the record: Capabilities are required to store the metadata supplied to the data pipeline as a part of record itself.

Example 1: Consider a data pipeline which processes files containing news viewership feeds from multiple publishers. Files are tagged with the id of the publisher. In a data lake we want to store the viewership from all publishers in a single dataset, so that we can perform analysis such as which news got highest number of views. However it would be still useful to get the information such as for a particular publisher which news is the most popular. Since the files are tagged with the publisher id, we would like to store the publisher id in the record itself for the analysis in the data lake.

Example 2: A dataset contains data from many publishers. For each publisher, the same pipeline is used for ingestion, with different runtime arguments, one of which provides the publisher id for the data being ingested. This publisher id as well as the source file name are added as extra fields to every output record, in order to be able to identify the publisher if one of the records is later reported as inaccurate, offensive or otherwise non-compliant. This will help trace every record back to its publisher and source file.

Field level lineage: Andreas: I think this example is implementation, not a use case. Can you rewrite this from the perspective of the user? The Data Officer (or Business Analyst) wants to understand the meaning of a field, by reading how it was produced. For example, the "fullName" fiel was created by concatenating the "firstName" and "lastName" field, both of which were extracted as positional fields from a CSV record in the source named "passengerList". The Operations team (and sometimes a developer) wants to identify how a field was computed, typically in a triage/debugging scenario where some fields show up with wrong values. Sagar (Updated the example) Example: Consider a CDAP data pipeline which imports the airline passenger information stored on the SFTP server into HDFS. Since the passenger information is coming from different sources, it needs to be normalized into the standard format while importing. Consider few sample records from the SFTP file (shown as table for better understanding) -

FullName	FirstName	LastName	Date of Birth	Gender	Country of Residence	Place of Birth
	Bruce	Wayne	12/10/1975	1	United States of America	San Jose
Kent, clark			Nov 11, 1980	M	Canada	Toronto
	Kathey	doe	11/11/1987	Female	England	Manchester
	brian	smith	April 05, 1992	Male	America	Palo Alto

We want to store this information in the HDFS in the following format:

PassengerId	Name	DOB	Gender	Place of Birth	Contry Code
PQR6248	Wayne, Bruce	12/10/1975	Male	San Jose	USA
XLM0912	Kent, Clark	11/11/1980	Male	Toronto	CAN
YQR8765	Doe, Kathey	11/11/1987	Female	Manchester	ENG
NULL	Smith, Brian	04/05/1992	Male	Palo Alto	NULL

Custom algorithm is reponsible for generating PassengerId in the above table using the normalized fields Name, DOB, and Country Code.
Now Data Officer(or Business Analyst) wants to understand which fields are critical for generating the PassengerId field. By looking at the field level lineage graph as given below, he would be able to know the fields from the source contribute to the PassengerId field.

            (Normalize)
FullName-------------------------------------------->
                                                     |
FirstName--> (concat)             (Normalize)        |-------------------->Name------------->
            |---->ConcatenatedName------------------>										|
LastName--->  																				|
													  (Date Normalizer)	         			|	
                						Date Of Birth----------------------->DOB----------->|---->PassengerId
																							|
															(Code Generator)				|
         								Country of Residence--------->Country Code--------->|

Software developer is notified that the PassengerId in the above dataset for user Smith, Brian is generated as NULL. By looking at the field level lineage graph for the PassengerId, developer can figure out that since Country Code itself is NULL, its generating NULL PassengerId.

Metadata provenance: According to Wiki (https://en.wikipedia.org/wiki/Provenance#Science) provenance allow us to assign metadata in sufficient details so that it allows us reproducibility of the data in the data lake.
1. This helps answering questions such as who changed what metadata. (we need both pieces who changed and what was the change)
2. Provenance information should be available through the REST api as well as through UI.
3. Technical and business (Andreas: all?) metadata should be tracked.
  
  Example 1: Debugging: Consider that a data governance officer is looking at the fields of a particular dataset and associated metadata for the fields. He notices that the 'User Address' field is missing a PII tag. He looks at the other provenance related tags such as 'Owner process', 'Creation date' and can notify appropriate parties to look into further. For example if the 'Owner Process' is 'Data import team' then he can notify them. Data import team then can drill further based on the 'Creation date' and adjust the pipeline accordingly.
  
  Example 2: Metadata change based processing:
  Only trigger processing if the metadata associated with the source is changed last time since the source processed. (Example not clear yet).
Metadata propagation: By default propagate the metadata at the field level. but this can be overridden by the pipeline developer.

Example: Consider a CDAP data pipeline which reads the user profile information which has "address" field and that has PII=true. The pipeline performs address normalization and stores both the original and the normalized form in output fields. For the original form, since it was simnply carried forward from the source, all metadata should automatically propagate (by the platform). For the normalized address field, the "address normalizer" plugin needs to have a way to either copy the "address" meta data to its output field "nomalizedAddress", or (better) have a programmatic way to signal that all metadata from "address" must be propagated to "normalilzedAddress:". Furthermore, if the plugin anonymizes rather than normalizes, it needs a way to either declare that its output field is not PII, or add an extra field "anonymized=true" to that field.

A big question here is when that should happen. For every run of a pipeline? What if different runs produce/propagate different meta data? How is that resolved?
Accessing the captured metadata:
1. The captured metadata should be accessible through API, plugins, and UI.
2. This will help applying rules based on the retrieved metadata.
  
  Example 1: If the field is marked as PII, then perform obfuscation process.
  
  Example 2: In flight metadata: Consider a CDAP data pipeline which reads the User profile information. The dataset contains the firstname and lastname as a field. These fields are tagged as 'User Identifier'. Now the pipeline perform Title casing on the field such that first character is capitalised to generate the fields UpperFirstName and UpperLastName. Next stage in the pipeline concatenates the UpperFirstName and LowerFirstName fields to generate the Name field. Since the UpperFirstName, UpperLastName, and Name field are in flight and not persisted to any dataset it, the stage in the pipeline when see these fields should still be able to access the tag associated with those fields which is 'User Identifier'.
Integration with Enterprise metadata systems:
1. Consume the metadata events from the external systems such as Atlas, Colibra etc. into the CDAP.
2. Push the metadata events from CDAP to the external systems such as Atlas, Colibra.
3. What metadata is pushed to the external systems from CDAP need to be configurable.
4. Integration should also allow maintain the referential integrity into the external systems. For example if I browse for a field in the UI corresponding tags should also be fetched from the external systems (not very clear about this requirement)
5. This helps achieving the complete automated metadata framework.
Operational metadata:
1. System should be able to generate the operational metadata.
  Example: A file was generated and during the processing of which 1000 Warning were generated. This information need to be captured.
  Andreas: see my example 3 under object-level meta data.

Following are just notes for Sagar:

Lower priority : FLL need to be available in the plugin so you can do conditional processing based on when the metadata is changed

Metadata need to be in sync with the Colibra and CDAP

possibility to add policies about how to rollup metadata from diff runs

Metadata propagation: By default propagate the metadata at the field level. but this can be overridden by the pipeline developer.

accumulators for the operational metadata - what about checksum?

8. In flight data? for example: change name to full name... any plugin which accesses the field should be able to access the metadata associated with it, even if the metadata is not persisted.