YARN multitenant application separation and impersonation at YARN level

Current situation

• A single Analytic Server (AS) instance can only create one spark context at YARN level.
• Analytic Server does not support Impersonation at YARN level.

Bosch is planning to use the Analytic Server in a multiuser environment. In such an environment, different users create different streams that can have different priorities. One example for such priorities are “production” and “development”.  Streams that are needed in production (shop floor data) are more important (priority “production”) than streams that are created by developers that are developing new applications (priority “development”).

When a user first executes a modeler stream (submits a job), the analytic server generates a spark job. This spark job initializes a new YARN application, owned by the “as_user”. All subsequent jobs submits (by arbitrary users) are submitted into this previously started YARN application.

Consequently, SPARK jobs are not separated from each other. They are all part of one single YARN application. This can result in critical situations. If a single stream shows unwanted behavior it can compromise other streams. For example if a stream stalls and needs to be killed, the whole application needs to be killed. If the whole application fails, all streams in execution will fail. This is a very serious problem when running production streams next to development streams on the same cluster, as there is no separation of concerns.

In order to realize such a separation of concerns, the Analytic server provides multi tenancy concept. With this concept multiple Analytic Servers are started each AS creates its own YARN application. This concept, however, comes with the huge and unacceptable overhead of multiple Analytic Server instances.

Furthermore, Bosch also needs to provide SPSS to a wide range of different customers as a shared service. Therefore, Analytic Server must be capable of scheduling jobs from different customers into their specific resource pools. In this context, Bosch requires impersonation. In the best case, each stream that is submitted by a user is executed as an individual YARN application, owned by the respective user.

Bosch requirements to realize the use case described above:

• Separate YARN (spark-)application for every job/stream which is being executed
  (separate application for every user, not tenant)
• Impersonation of spark job at YARN level (not “as_user” but real “Bosch user”) as the “User” of the application (queue placement policy)
• Definition of a specific resource pool (queues) for every tenant
  (Tenant <-> Queue mapping is necessary)
• All functionality with only one Analytic Server instance per cluster