Data Validator

A tool to validate data in Spark

Usage

Retrieving official releases via direct download or Maven-compatible dependency retrieval, e.g. spark-submit

You can make the jars available in one of two ways for the example run invocations below:

1. Get the latest version from GitHub Packages for the project. Place the jars somewhere and pass their path to --jars when running spark-submit.

2. You can pull in the dependency using spark-submit's --repositories, --packages, and --mainClass options, but it requires setting spark.jars.ivySettings and providing this file, populated with a valid personal access token having the read:packages scope enabled. N.b. it can be a challenge to secure this file on shared clusters; consider using a public GitHub service account instead of a token from your own personal GitHub account.

   <ivysettings>
     <settings defaultResolver="thechain">
       <credentials host="maven.pkg.github.com" realm="GitHub Package Registry"
                    username="${GITHUB_PACKAGES_USER}" passwd="${GITHUB_PACKAGES_USER_TOKEN}" />
     </settings>
     <resolvers>
       <chain name="thechain">
         <ibiblio name="central" m2compatible="true" root="https://repo1.maven.org/maven2" />
         <!-- add any other repositories here -->
         <ibiblio name="ghp-dv" m2compatible="true" root="https://maven.pkg.github.com/target/data-validator"/>
       </chain>
     </resolvers>
   </ivysettings>

   See also How do I add a GitHub Package repository when executing spark-submit --repositories?

Building locally

See CONTRIBUTING for development environment setup.

Assemble fat jar: make build or sbt clean assembly

spark-submit --master local data-validator-assembly-0.14.1.jar --help

data-validator v0.14.1
Usage: data-validator [options]

  --version
  --verbose                Print additional debug output.
  --config <value>         required validator config .yaml filename, prefix w/ 'classpath:' to load configuration from JVM classpath/resources, ex. '--config classpath:/config.yaml'
  --jsonReport <value>     optional JSON report filename
  --htmlReport <value>     optional HTML report filename
  --vars k1=v1,k2=v2...    other arguments
  --exitErrorOnFail true|false
                           optional when true, if validator fails, call System.exit(-1) Defaults to True, but will change to False in future version.
  --emailOnPass true|false
                           optional when true, sends email on validation success. Default: false
  --help                   Show this help message and exit.

If you want to build with Java 11 or newer, set the "MODERN_JAVA" environment variable. This may become the default in the future.

Example Run

With the JAR directly:

spark-submit \
  --num-executors 10 \
  --executor-cores 2 \
  data-validator-assembly-0.14.1.jar \
  --config config.yaml \
  --jsonReport report.json

Using packages loading, having created dv-ivy.xml as suggested above and having replaced the placeholders in the example:

touch empty.file && \
spark-submit \
  --class com.target.data_validator.Main \
  --packages com.target:data-validator_2.11:0.14.1 \
  --conf spark.jars.ivySettings=$(pwd)/dv-ivy.xml \ 
  empty.file \
  --config config.yaml \
  --jsonReport report.json

See the Example Config below for the contents of config.yaml.

Config file Description

The data-validator config file is yaml based and it has 3 sections, Global Settings, Table Sources, and Validators. The Table Sources, and Validators have the ability to use variables in the configuration. These variables are replaced at runtime with the values set via Global Settings section or the --vars option on the command line. Variables start with $ and must contain a word starting with a letter (A-Za-z) and followed by zero or more letters (A-Za-z), numbers(0-9), or underscore. Variables can optionally be wrapped in { }. i.e. $foo, ${foo} See the code for the regular expression used to find them in a string. All the table sources, and all but one validator (rowCount) supports variables in their configuration parameters. Note: Care must be taken for some of the substitutions, some possible values might require quoting the variables in the config.

Global Settings

The first section is the global settings that are used throughout the program.

Variable	Type	Required	Description
numKeyCols	Int	Yes	The number of columns from the table schema to use to uniquely identify a row in the table.
numErrorsToReport	Int	Yes	The number of detailed errors to include in Validator Report.
detailedErrors	Boolean	Yes	If a check fails, run a second pass and gather numErrorToReport examples of failure.
email	EmailConfig	No	See Email Config.
vars	Map	No	A map of (key, value) pairs used for variable substitution in tables config. See next section.
outputs	Array	No	Describes where to send .json report. See Validator Output.
tables	List	Yes	List of table sources used to load tables to validate.

Email Config

Variable	Type	Required	Description
smtpHost	String	Yes	The smtp host to send email message through.
subject	String	Yes	Subject for email message.
from	String	Yes	Email address to appear in from part of message.
to	Array[String]	Yes	Must specify at least one email address to send the email report to.
cc	Array[String]	No	Optional list of email addresses to send message to via cc field in message.
bcc	Array[String]	No	Optional list of email addresses to send message to via bcc field in message.

Note that Data Validator only sends email on failure by default. To send email even on successful runs, pass --emailOnPass true to the command line.

Defining Variables

There are 4 different types of variables that you can specify, simple, environment, shell and SQL.

Simple Variable

Simple variables are specified by the name and value pairs and are very straight forward.

vars:
  - name: ENV
    value: prod

This sets the variable ENV to the value prod

Environment Variable

Environment variables import the value from the operating system

vars:
  - name: JAVA_DIR
    env: JAVA_HOME

This will set the variable JAVA_DIR to the value returned by the System.getenv("JAVA_HOME") If JAVA_HOME does not exist in the system environment, the data-validator will stop processing and exit with an error.

Shell Variable

Shell variable will take the first line of output from a shell command and store it a variable.

vars:
  - name: NEXT_SATURDAY
    shell: date -d "next saturday" +"%Y-%m-%d"

This will set the variable NEXT_SATURDAY to the first line of output from the shell command date -d "next saturday" +"%Y-%m-%d".

SQL Variable

SQL variable will take the first column from the first row of the results from a Spark SQL statement.

vars:
  - name: MAX_AGE
    sql: select max(age) from census_income.adult

This runs the sql command that gets the max value from the column age from the table adult in the census_income database and stores it in MAX_AGE.

Validator Output

In addition to the --jsonReport command line option, the .yaml has a outputs section that directs the .json event report to a file or pipes it to a program. There is no current limit on the number of outputs.

Filename

outputs:
  - filename: /user/home/sample.json
    append: true

If the filename specified begins with a / or local:/// it is written to the local filesystem. If the filename begins with hdfs:// the report is written to the hdfs path. An optional append boolean can be specified, and if it is true the current report will be appended to the end of the specified file. The default is append: false and the filename is overwritten. The filename supports variable substitution, the optional append does not. Before the validator starts processing tables, it checks to verify that it can create or append to the filename, if it cannot, the data validator will exit with an error (non-zero value).

Pipe

outputs:
  - pipe: /path/to/program
    ignoreError: true

A pipe is used to send the .json event report to another program for processing. This is a very powerful feature, and can enable the data-validator to be integrated with virtually any other system. An optional ignoreError boolean can also be specified, if true the exit value of the program will be ignored. If false (default) and the program exits with a non-zero status, the data-validator will fail. The pipe supports variable substitution, the optional ignoreError does not.

Before the validator starts processing tables, it checks to see if the pipe program is executable, if it is not, the data-validator will exit with an error (non-zero value). The program must be on a local filesystem to be executed.

Table Sources

Table sources are used to specify how to load the tables to be validated. Currently supported sources are HiveTable, and OrcFile. Each table source has 3 common arguments, keyColumns, condition, checks, and its own source specific argument(s). The keyColumns are list of columns that can be used to uniquely identify a row in the table for the detailed error report when a validator fails. The condition enables the user to specify a snippet of sql to pass to the where clause. The checks argument is a list of validators to run on this table.

HiveTable

To validate a Hive table, specify the db and the table, see below.

- db: $DB
  table: table_name
  condition: "col1 < 100"
  keyColumns:
    - col1
    - col2
  checks:

OrcFile

To validate an .orc file, specify orcFile and the path to the file, see below.

- orcFile: /path/to/orc/file
  keyColumns:
    - col1
    - col2
  checks:

Parquet File

To validate an .parquet file, specify parquetFile and the path to the file, see below.

- parquetFile: /path/to/parquet/file
  keyColumns:
    - col1
    - col2
  checks:

Core spark.read fluent API specified format loader

To validate data loadable by the Spark DataFrameReader Fluent API, use something like this:

  # Some systems require a special format
  format: llama
  # You can also pass any valid options
  options:
    maxMemory: 8G
  # This is a string passed to the varargs version of DataFrameReader.load(String*)
  # If omitted, then DV will call DataFrameReader.load() without parameters.
  # The DataSource that Spark loads is expected to know how to handle this.
  loadData:
    - /path/to/something/camelid.llama
  keyColumns:
    - col1
    - col2
  condition: "col1 < 100"
  checks:

Under the hood the above would be like loaded a DataFrame with:

spark.read
  .format("llama")
  .option("maxMemory", "8G")
  .load("/path/to/something/camelid.llama")

Validators

The third section are the validators. To specify a validator, you first specify the type as one of the validators, then specify the arguments for that validator. Some of the validators support an error threshold. This options allows the user to specify the number of errors or percentage of errors they can tolerate. In some use cases, it might not be possible to eliminate all errors in the data.

Thresholds

Thresholds can be specified as an absolute number of errors, or a percentage of the row count. If the threshold is >= 1 it is considered an absolute number of errors. For example 1000 would fail the check if there are more then 1000 rows that failed the check.

If the threshold is < 1 it is considered a fraction of the row count. For example 0.25 would fail the check if more then rowCount * 0.25 of the rows fail the check. If the threshold ends in a % its considered a percentage of the row count. For eample 33% would fail the check if more then rowCount * 0.33 of the rows fail the check.

Currently supported validators are listed below:

columnMaxCheck

Takes 2 parameters, the column name and a value. The check will fail if max(column) is not equal to the value.

Arg	Type	Description
column	String	Column within table to find the max from.
value	*	The column max should equal this value or the check will fail. Note: The type of the value should match the type of the column. If the column is a NumericType, the value cannot be a String.

negativeCheck

Takes a single parameter, the column name to check. The validator will fail if any rows with that column are negative.

Arg	Type	Description
column	String	Table column to be checked for negative values. If it contains a null validator will fail. Note: Column must be of a NumericType or the check will fail during the config check.
threshold	String	See above description of threshold.

nullCheck

Takes a single parameter, the column name to check. The validator will fail if any rows with that column are null.

Arg	Type	Description
column	String	Table column to be checked for null. If it contains a null validator will fail.
threshold	String	See above description of threshold.

rangeCheck

Takes 2 - 4 parameters, described below. If the value in the column doesn't fall within the range specified by (minValue, maxValue) the check will fail.

Arg	Type	Description
column	String	Table column to be checked.
minValue	*	lower bound of the range, or other column in table. Type depends on the type of the column.
maxValue	*	upper bound of the range, or other column in table. Type depends on the type of the column.
inclusive	Boolean	Include minValue and maxValue as part of the range.
threshold	String	See above description of threshold.

Note: To specify another column in the table, you must prefix the column name with a ` (backtick).

stringLengthCheck

Takes 2 to 4 parameters, described in the table below. If the length of the string in the column doesn't fall within the range specified by (minLength, maxLength), both inclusive, the check will fail. At least one of minLength or maxLength must be specified. The data type of column must be String.

Arg	Type	Description
column	String	Table column to be checked. The DataType of the column must be a String
minLength	Integer	Lower bound of the length of the string, inclusive.
maxLength	Integer	Upper bound of the length of the string, inclusive.
threshold	String	See above description of threshold.

stringRegexCheck

Takes 2 to 3 parameters, described in the table below. If the column value does not match the pattern specified by the regex, the check will fail. A value for regex must be specified. The data type of column must be String.

Arg	Type	Description
column	String	Table column to be checked. The DataType of the column must be a String
regex	String	POSIX regex.
threshold	String	See above description of threshold.

rowCount

The minimum number of rows a table must have to pass the validator.

Arg	Type	Description
minNumRows	Long	The minimum number of rows a table must have to pass.

See Example Config file below to see how the checks are configured.

uniqueCheck

This check is used to make sure all rows in the table are unique, only the columns specified are used to determine uniqueness. This is a costly check and requires an additional pass through the table.

Arg	Type	Description
columns	Array[String]	Each set of values in these columns must be unique.

columnSumCheck

This check sums a column in all rows. If the sum applied to the column doesn't fall within the range specified by (minValue, maxValue) the check will fail.

Arg	Type	Description
column	String	The column to be checked.
minValue	NumericType	The lower bound of the sum. Type depends on the type of the column.
maxValue	NumericType	The upper bound of the sum. Type depends on the type of the column.
inclusive	Boolean	Include minValue and maxValue as part of the range.

Note: If bounds are non-inclusive, and the actual sum is equal to one of the bounds, the relative error percentage will be undefined.

colstats

This check generates column statistics about the specified column.

Arg	Type	Description
column	String	The column on which to collect statistics.

These keys and their corresponding values will appear in the check's JSON summary when using the JSON report output mode:

Key	Type	Description
count	Integer	Count of non-null entries in the column.
mean	Double	Mean/Average of the values in the column.
min	Double	Smallest value in the column.
max	Double	Largest value in the column.
stdDev	Double	Standard deviation of the values in the column.
histogram	Complex	Summary of an equi-width histogram, counts of values appearing in 10 equally sized buckets over the range [min, max].

Example Config

---

# If keyColumns are not specified for a table, we take the first N columns of a table instead.
numKeyCols: 2

# numErrorsToReport: Number of errors per check show in "Error Details" of report, this is to limit the size of the email.
numErrorsToReport: 5

# detailedErrors: If true, a second pass will be made for checks that fail to gather numErrorsToReport examples with offending value and keyColumns to aide in debugging
detailedErrors: true

vars:
  - name: ENV
    value: prod

  - name: JAVA_DIR
    env: JAVA_HOME

  - name: TODAY
    shell: date + "%Y-%m-%d"

  - name: MAX_AGE
    sql: SELECT max(age) FROM census_income.adult

outputs:
  - filename: /user/home/sample.json
    append: true

  - pipe: /path/to/program
    ignoreError: true

email:
  smtpHost: smtp.example.com
  subject: Data Validation Summary
  from: data-validator-no-reply@example.com
  to:
    - person1@example.com
  cc:
    - person2@example.com, person3@example.com
  bcc:
    - person4@example.com

tables:
  - db: census_income
    table: adult
    # Key Columns are used when errors occur to identify a row, so they should include enough columns to uniquely identify a row.
    keyColumns:
      - age
      - occupation
    condition: educationNum >= 5
    checks:
      # rowCount - checks if the number of rows is at least minRows
      - type: rowCount
        minNumRows: 50000

      # negativeCheck - checks if any values are less than 0
      - type: negativeCheck
        column: age
      
      # colstats - adds basic statistics of the column to the output
      - type: colstats
        column: age
        
      # nullCheck - checks if the column is null, counts number of rows with null for this column.
      - type: nullCheck
        column: occupation

      # stringLengthCheck - checks if the length of the string in the column falls within the specified range, counts number of rows in which the length of the string is outside the specified range.
      - type: stringLengthCheck
        column: occupation
        minLength: 1
        maxLength: 5

      # stringRegexCheck - checks if the string in the column matches the pattern specified by `regex`, counts number of rows in which there is a mismatch.
      - type: stringRegexCheck
        column: occupation
        regex: ^ENGINEER$ # matches the word ENGINEER

      - type: stringRegexCheck
        column: occupation
        regex: \w # matches any alphanumeric string

Working with OOZIE Workflows

The data-validator can be used in an oozie workflow to halt the wf if a check doesn't pass. There are 2 ways to use the data-validator in oozie and each has their own drawbacks. The selection of the methods is determined by the --exitErrorOnFail {true|false} command line option.

Setting ExitErrorOnFail to True

The first option, enabled by --exitErrorOnFail=true, is to have the data-validator exit with a non-zero value when a check fails. This enables the workflow to decide how it wants to handle a failed check/error. The downsides of this method, is that you can never be sure if the data-validator exited with an error because bad check, or if there was a problem with the execution of the data-validator. This also pollutes the oozie workflow info with ERROR, which some might not like. This is currently the default but likely to change with v1.0.0.

Example oozie wf snippet:

<action name="RunDataValidator">
    <shell xmlns="uri:oozie:shell-action:0.2">
      <job-tracker>${jobTracker}</job-tracker>
      <name-node>${nameNode}</name-node>
      <exec>spark-submit</exec>
      <argument>--conf</argument>
      <argument>spark.yarn.maxAppAttempts=1</argument>
      <argument>--class</argument>
      <argument>com.target.data_validator.Main</argument>
      <argument>--master</argument>
      <argument>yarn</argument>
      <argument>--deploy-mode</argument>
      <argument>cluster</argument>
      <argument>--keytab</argument>
      <argument>${keytab}</argument>
      <argument>--principal</argument>
      <argument>${principal}</argument>
      <argument>--files</argument>
      <argument>config.yaml</argument>
      <argument>data-validator-assembly-0.14.1.jar</argument>
      <argument>--config</argument>
      <argument>config.yaml</argument>
      <argument>--exitErrorOnFail</argument>
      <argument>true</argument>
      <argument>--vars</argument>
      <argument>ENV=${ENV},EMAIL_REPORT=${EMAIL_REPORT},SMTP_HOST=${SMTP_HOST}</argument>
      <capture-output/>
    </shell>
    <ok to="ValidatorSuccess" />
    <error to="ValidatorErrorOrCheckFail" />
  </action>

 <action name="ValidatorErrorOrCheckFail">
  <!-- Check or data-validator failed  -->
  </action>

  <action name="ValidatorSuccess">
  <!-- Everything is wonderful!  -->
  </action>

Setting ExitErrorOnFail to False

The second option, enabled by --exitErrorOnFail=false, is to have the data-validator output to stdout DATA_VALIDATOR_STATUS=PASS or DATA_VALIDATOR_STATUS=FAIL and System.exit(0) when it completes. This enables the workflow to distinguish between a failed check, and a runtime error. The downside is that you must use the oozie shell action, with the capture output option, and run the validator via Spark's client mode. This will likely become the default behavior in v1.0.0.

Example oozie wf snippet:

<action name="RunDataValidator">
  <shell xmlns="uri:oozie:shell-action:0.2">
    <job-tracker>${jobTracker}</job-tracker>
    <name-node>${nameNode}</name-node>
    <exec>spark-submit</exec>
    <argument>--conf</argument>
    <argument>spark.yarn.maxAppAttempts=1</argument>
    <argument>--class</argument>
    <argument>com.target.data_validator.Main</argument>
    <argument>--master</argument>
    <argument>yarn</argument>
    <argument>--deploy-mode</argument>
    <argument>client</argument>
    <argument>--keytab</argument>
    <argument>${keytab}</argument>
    <argument>--principal</argument>
    <argument>${principal}</argument>
    <argument>data-validator-assembly-0.14.1.jar</argument>
    <argument>--config</argument>
    <argument>config.yaml</argument>
    <argument>--exitErrorOnFail</argument>
    <argument>false</argument>
    <argument>--vars</argument>
    <argument>ENV=${ENV},EMAIL_REPORT=${EMAIL_REPORT},SMTP_HOST=${SMTP_HOST}</argument>
    <capture-output/>
  </shell>
  <ok to="ValidatorDecision" />
  <error to="VaildatorError" />
</action>

<decision name="ValidatorDecision">
  <switch>
    <case to="ValidatorCheckFail">${wf:actionData('RunDataValidator')['DATA_VALIDATOR_STATUS'] eq "FAIL"}</case>
    <case to="ValidatorCheckPass">${wf:actionData('RunDataValidator')['DATA_VALIDATOR_STATUS'] eq "PASS"}</case>
    <default to="ValidatorNeither"/>
  </switch>
</decision>

<action name="ValidatorCheckFail">
  <!-- Handle Failed Check -->
</action>

<action name="ValidatorCheckPass">
  <!-- Everything is Wonderful! -->
</action>

<action name="ValidatorFailure">
  <!-- Notify devs of validator failure -->
</action>

Other tools included

Configuration parser check

com.target.data_validator.ConfigParser has an entrypoint that will check that the configuration file is parseable. It does not validate variable substitutions since those have runtime implications.

spark-submit \
  --class com.target.data_validator.ConfigParser \
  --files config.yml \
  data-validator-assembly-0.14.1.jar \
    config.yml

If there is an error, DV will print a message and exit non-zero.

Development Tools

Generate testing data with GenTestData or sbt generateTestData

Data Validator includes a tool to generate a sample .orc file for use in local development. This repo's SBT configuration wraps the tool in a convenient SBT task: sbt generateTestData
If you run this program or task, it will generate a file testData.orc in the current directory. You can then use the following config file to test the data-validator. It will generate a report.json and report.html.

spark-submit \
  --master "local[*]"  \
  data-validator-assembly-0.14.1.jar \
  --config local_validators.yaml \
  --jsonReport report.json  \
  --htmlReport report.html

local_validators.yaml

---
numKeyCols: 2
numErrorsToReport: 5
detailedErrors: true

tables:
  - orcFile: testData.orc

    checks:
      - type: rowCount
        minNumRows: 1000

      - type: nullCheck
        column: nullCol

History

This tool is based on methods described in Methodology for Data Validation 1.0 by Di Zio et al., published by Esset Validat Foundation in 2016. You can download the paper here.