[IMPALA-8656] Support for eagerly fetching and spooling all query result rows

• Type: Improvement
• Status: Resolved
• Priority: Critical
• Resolution: Fixed
• Affects Version/s: Impala 2.12.0, Impala 3.2.0
• Fix Version/s: Impala 3.4.0
• Component/s: Backend
• Epic Color: ghx-label-1
• Assignee: Sahil Takiar
• Reporter: Michael Ho
• Created: 12/Jun/19 02:22
• Updated: 03/Mar/21 00:38
• Resolved: 08/Oct/19 14:57

Description

Impala's current interaction with clients is pulled-based: it relies on clients to fetch results to trigger the generation of more result row batches until all the result rows have been produced. If a client issues a query without fetching all the results, the query fragments will continue to consume the resources until the query hits is cancelled and unregistered for whatever reasons. This is undesirable as resources are held up by misbehaving clients and other queries may wait for extended period of time in admission control due to this.

The high level idea for this JIRA is for Impala to have a mode in which result sets of queries are eagerly fetched and spooled somewhere (preferably some persistent storage). In this way, the cluster's resources are freed up once all result rows have been fetched and stored in the spooling location. Incoming client fetches can be returned from this spooled locations.

cc'ing stakiar, twm378, joemcdonnell, lv

Issue Links

• is related to: IMPALA-9210 Query timeline should include entry when all query results are spooled
• is related to: IMPALA-9339 Revise explanation of RowMaterializationTimer
• is related to: IMPALA-4268 Rework coordinator buffering to buffer more data
• is related to: IMPALA-9818 Add fetch size as option to impala shell
• is related to: IMPALA-8925 Consider replacing ClientRequestState ResultCache with result spooling
• is related to: IMPALA-9856 Enable result spooling by default
• relates to: IMPALA-10180 Add average size of fetch requests in runtime profile

Sub-Tasks

• Add RowBatchQueue interface with an implementation backed by a std::queue: Resolved
• Implementation of BufferedPlanRootSink where FlushFinal blocks until all rows are fetched: Resolved
• Add additional tests in test_result_spooling.py and validate cancellation logic: Resolved
• Implement a RowBatchQueue backed by a BufferedTupleStream: Resolved
• Coordinator should release admitted memory per-backend rather than per-query: Resolved
• Replace deque queue with spillable queue in BufferedPlanRootSink: Resolved
• BufferedPlanRootSink should handle non-default fetch sizes: Resolved
• Close ExecNode tree prior to calling FlushFinal in FragmentInstanceState: Resolved
• DCHECK(!page->attached_to_output_batch) in SpillableRowBatchQueue::AddBatch: Resolved
• Add additional counters to PlanRootSink: Resolved
• DCHECK(!closed_) in SpillableRowBatchQueue::IsEmpty: Resolved
• Add failpoint tests to result spooling code: Resolved
• Profile fetch performance when result spooling is enabled: Resolved
• BufferedPlanRootSink should directly write to a QueryResultSet if one is available: Resolved
• Impala Doc: Add docs for PLAN_ROOT_SINK and result spooling: Closed