This patch adds support for having multiple aggregate functions in a
single SELECT block that use DISTINCT over different sets of columns.
Planner design:
- The existing tree-based plan shape with a two-phased
aggregation is maintained.
- Existing plans are not changed.
- Aggregates are grouped into 'aggregation classes' based on their
expressions in the distinct portion which may be empty for
non-distinct aggregates.
- The aggregation framework is generalized to simultaneously process
multiple aggregation classes within the tree-based plan. This
process splits the results of different aggregation classes into
separate rows, so a final aggregation is needed to transpose the
results into the desired form.
- Main challenge: Each aggregation class consumes and produces
different tuples, so conceptually a union-type of tuples flows
through the runtime. The tuple union is represented by a TupleRow
with one tuple per aggregation class. Only one tuple in such a
TupleRow is non-NULL.
- Backend exec nodes in the aggregation plan will be aware of this
tuple-union either explicitly in their implementation or by relying
on expressions that distinguish the aggregation classes.
- To distinguish the aggregation classes, e.g. in hash exchanges,
CASE expressions are crafted to hash/group on the appropriate slots.
Deferred FE work:
- Beautify/condense the long CASE exprs
- Push applicable conjuncts into individual aggregators before
the transposition step
- Added a few testing TODOs to reduce the size of this patch
- Decide whether we want to change existing plans to the new model
Execution design:
- Previous patches separated out aggregation logic from the exec node
into Aggregators. This is extended to support multiple Aggregators
per node, with different grouping and aggregating functions.
- There is a fast path for aggregations with only one aggregator,
which leaves the execution essentially unchanged from before.
- When there are multiple aggregators, the first aggregation node in
the plan replicates its input to each aggregator. The output of this
step is rows where only a single tuple is non-null, corresponding to
the aggregator that produced the row.
- A new expr is introduced, ValidTupleId, which takes one of these
rows and returns which tuple is non-null.
- For additional aggregation nodes, the input is split apart into
'mini-batches' according to which aggregator the row corresponds to.
Testing:
- Added analyzer and planner tests
- Added end-to-end queries tests
- Ran hdfs/core tests
- Added support in the query generator and ran in a loop.
Change-Id: I055402eaef6d81e5f70e850d9f8a621e766830a4
Reviewed-on: http://gerrit.cloudera.org:8080/10771
Reviewed-by: Impala Public Jenkins <impala-public-jenkins@cloudera.com>
Tested-by: Impala Public Jenkins <impala-public-jenkins@cloudera.com>
df53ec2385