[ML] Client-Side Job Validation v6.3

[elasticmachine]

Original comment by @walterra:

(Client-Side means x-pack-kibana instead of x-pack-elasticsearch in this context)

This is a follow up issue to LINK REDACTED to keep track of the actual implementation.

UI/Job Validation Framework

• General boilerplate/framework to support validation checks including UI prototype LINK REDACTED
• Refactor UI to use EUI/React LINK REDACTED LINK REDACTED
• Enable Buttons in jobs list and wizards LINK REDACTED LINK REDACTED
• Change UI to display "success" messages for test categories LINK REDACTED
• Include job validation message status in messages.json LINK REDACTED
• Deduplicate messages LINK REDACTED
• Loading indicator / status icon for job validation button LINK REDACTED
• Add check for duplicated detectors to population wizard LINK REDACTED
• "Learn more" links should link to current version LINK REDACTED
• Fix "undefined" fieldname LINK REDACTED

Bugs

• supress result message when model memory checks are not executed [ML] Job Validation - supress result message when model memory checks are not executed #18517
• check for valid format of model memory limit [ML] Job Validation - check for valid format of model memory limit #18518

Checks v6.3

• Basic checks (previously advanced job wizard form checks) moved to server LINK REDACTED
• Check cardinality of fields
  • Basic checks LINK REDACTED
  • by fields and other tweaks LINK REDACTED
  • Rules for skipping checks LINK REDACTED
• Check bucket_span
  • invalid if zero LINK REDACTED
  • test if > 1 day LINK REDACTED
  • test if different from bucket-estimation (trigger INFO level message) LINK REDACTED (disabled for now)
• Check influencers
  • are there any? LINK REDACTED
  • Rules for skipping checks LINK REDACTED
• Improve start/end time handling
  • Add optional support for start/end timestamps to the node API endpoint. Use this for the job validation checks triggered from wizards instead of hijacking data_count. LINK REDACTED
  • If start/end is not set, try to fallback to timespan of original data LINK REDACTED
  • Check for data before the unix epoch (see LINK REDACTED)
  • Check if time range is 25 times the bucket span LINK REDACTED or at least 2h LINK REDACTED
• Check against model memory limit LINK REDACTED
• Check job does not contain duplicate detectors (same function and field names) LINK REDACTED
• Check for non-aggregatable fields LINK REDACTED LINK REDACTED
• If model_plot_config enabled and cardinality of by/over/partition >100 then warn this is a resource intensive job LINK REDACTED

Unit Tests

They should cover all possible message ids being returned.

• Basic tests LINK REDACTED LINK REDACTED
• Cardinality LINK REDACTED
• Bucket Span LINK REDACTED
• Influencers LINK REDACTED
• Time Range LINK REDACTED
• Model Memory Limit LINK REDACTED

[elasticmachine]

Original comment by @sophiec20:

Awesome first review... feedback coming...

A1: Repeated error for each detector - any way to de-dup?
Double over_field value check.

and also

A2: Links to blogs
I don't think we should link to blogs from inside the UI. It seems a bit too temporary and unable to version control. We could link to a page in the docs, which could then link to a blog - however should we do this either? @lcawl thoughts? is there precedence in say upgrade wiz or add data?

A3: bucket span test is very quick (and always seems to be 15m) - is this actually plugged in?

[elasticmachine]

Original comment by @walterra:

Yes those blog links are temporary, Lisa is already working on replacing them here: LINK REDACTED

[elasticmachine]

Original comment by @jgowdyelastic:

Just making a note here so I don't forget.
The current way that the validation button has been added to the job wizards might need addressing.
The job config is passed to the mlValidateJob directive by a function call:
job="getJobFromConfig()"
this means this function is called on every digest cycle.
Also as a side effect it means the job id validation is called every digest cycle. So the user gets this error message as soon as the page loads:

Perhaps a reference to getJobFromConfig could be passed across instead and only called when the Validate job button is pressed.

[elasticmachine]

Original comment by @walterra:

• there's now a check in place to detect duplicate detectors LINK REDACTED (already in master/6.x)
• job validation message deduplication is added in LINK REDACTED
• the above problem in the wizards described by James is solved in LINK REDACTED

[elasticmachine]

Original comment by @walterra:

I cleaned up the issue's description an moved validation checks planned for v6.4 to #18368.

[walterra]

Created a follow up issue to track further progress for v6.3.x #18898