[SIP-11] Proposal for deprecating the native Druid NoSQL connector

[john-bodley]

Motivation

Superset currently supports two engine connectors for querying datasources; SQLAlchemy and the Druid REST API. The later was the initial use case for Superset, i.e., a UI for visualizing Druid datasources.

Since version 0.10.0 Druid has included a built-it SQL server which has a SQLAlchemy binding provided by the pydruid library (courtesy of @betodealmeida and @mistercrunch) and thus the proposed change is to deprecate the REST API interface in favor of having a single interface (SQLAlchemy) to all engines. Note all future engines (there has been mentioned of adding support for Elasticsearch) would require a SQLAlchemy dialect.

There is a non-insignificant amount of overhead in supporting both connectors including:

Code

From a code perspective each connector needs to define similar views and models. The Druid connector alone comprises of around 2,000 lines of code. There is additional frontend logic which needs to construct filters, metrics, etc. for both the Druid REST API and SQLAlchemy. Note there are 74 files (including documentation) which reference Druid in the repo.

Models

In addition to code overhead each connector defines its own models and database tables:

Druid:

• clusters
• datasources
• columns
• metrics

SQLAlchemy:

• dbs
• tables
• table_columns
• sql_metric

which complicates logic, i.e., the slices table does not have a SQLAlchemy relationship to a "datasource" table as the datasource type determines the association. This results in denormalized tables with potentially incorrect values, i.e., the slices table contains the datasource_name column for the FAB CRUD views, however this may not accurately reflect the underlying datasource name.

Proposed Change

The proposed change would be to deprecate all the Druid REST logic from the codebase. This significantly simplifies and streamlines a number of facets of Superset by ensuring that all engines connect via a SQLAlchemy dialect.

Currently there is support for syncing/refreshing Druid datasource associated with the REST API connector which I suspect is leveraged by a number of organizations. SIP-7 discussing "refreshing" of Superset datasources.

Note this would be a breaking change for any organizing using a Druid version less than 0.10.0. Also there may be some instances of post-aggregate Druid functions which are not supported in Druid SQL.

New or Changed Public Interfaces

There would be no new or changed public interfaces.

New dependencies

There would be no new dependencies.

Migration Plan and Compatibility

A non-trivial database migration would be required including:

• All records in the Druid tables listed above would need to be migrated to the SQLAlchemy equivalent table.
• Existing slices would need to be updated to reference the new SQLAlchemy representation of the Druid datasource.
• Re-normalize the slices table.
• Update chart data to remove the obsolete table__ or druid__ prefixes.

Rejected Alternatives

None.

to: @betodealmeida @graceguo-supercat @kristw @michellethomas @mistercrunch @timifasubaa

[kristw]

I am in favor of consolidating and standardizing in general. 👍
This should reduce bugs and improve maintainability.

[mistercrunch]

That will effectively mean deprecating support for pre-good-SQL-support druid versions, which I think is fine. If people want to use old Druid, they can use old Superset with it :)

[john-bodley]

@mistercrunch are you aware of any functionality that the REST API provides (such as phase II queries) that couldn’t be captured in Druid SQL?

[kkalyan]

How are Druid dimension extractions/lookup, filtered aggregations and javascript post-aggregators with Druid SQL? Druid users use them heavily. Some of these are not native to SQL, it would be good to support them.

[datametrics]

I use the REST API of Druid heavily. It makes it totally easy to make discovery of datasources as well as to implement Druid clients from other languages. The rest format is platform interchangeable and there is no need to implement any further sql parser / converter logic. One can just throw some model classes together and serialise that to Json. From my point of view it would be great when superset continues with REST support or at least leaves the opportunity for connector injection.

[SpyderRivera]

Druid SQL won't support core features like multi-value dimensions until 0.13.1

[akuckartz]

Is SQLAlchemy a REST API ?

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[john-bodley]

#8512 introduced a feature flag which enables the Druid NoSQL connector. Note by default this is disabled.