[FEATURE] Track query state in plugin with calling EMR-S jobRun API

[penghuo]

• Track all the query state in plugin. currently, we only track interactive queries. we should track batch and streaming queries.
• one requirement of this feature is to reduce getJobRun call on EMR-S service.

[penghuo]

Queries Journey

Based on job types, we categorize queries into three main types:

1. Interactive jobs. e.g. create table
2. Batch jobs. e.g. refresh table
3. Streaming jobs. e.g. create table with (auto_refresh=true)

Additionally, we have queries that run on the plugin side. e.g. drop index In following table, we explain how the queries is processed in the system

POST

POST async_query create 2 docs in flint_execution_request index

• statement. docId: statementId

type: statement queryId: // asssigned by Transport statementId: // same as queryId state: "waiting"

• session. docId: sessionId
  • based on the query, the sessionType will be different.

type: session sessionType: interactive / batch / streaming / local state: "NOT_STARTED"

FlintREPLJob

• get sessionId from spark.flint.job.sessionId conf.
• get sessionType of sessionId.
• take action based on sessionType
  • interactive
    • while (true)
      • read statement with sessionId and process.
      • write result to resultIndex
      • update statementState = success
    • quit if process done
  • batch
    • read statement with sessionId and process.
    • write result to resultIndex
    • update statementState = success
    • quit if process done
  • streaming
    • read statement with sessionId and process.
    • write result to resultIndex
    • update statementState = success
• Update session lastUpdateTime regularly
• Update sessionState to DEAD before quit

GET

GET async_query fetch state from statement doc. If the state is success, fetch result from query_result_indices.

CANCEL

CANCEL async_query.

• opt-1
  • Plugin (1) set statement state to cancelling state.
  • FlintJob (1) read statement state, if it cancelling. mark as cancelled state. force quite the job.
• opt-2
  • Plugin (1) set statement state to cancelling. (2) cancel the job associate with session. (3) set statement to cancelled state.

Limitation: DROP INDEX query can not be cancelled.