Details
-
Improvement
-
Status: Open
-
Major
-
Resolution: Unresolved
-
Impala 3.0
-
None
-
ghx-label-8
Description
Currently, the operators of a compute [incremental] stats' subquery rely on combined selectivities - as usual - to estimate cardinality, e.g. during aggregation. For example, note the expected cardinality of the aggregation on this subquery:
F00:PLAN FRAGMENT [RANDOM] hosts=1 instances=4
Per-Host Resources: mem-estimate=305.20GB mem-reservation=136.00MB
01:AGGREGATE [STREAMING]
| output: [...]
| group by: col_a, col_b, col_c
| mem-estimate=76.21GB mem-reservation=34.00MB spill-buffer=2.00MB
| tuple-ids=1 row-size=104.83KB cardinality=693000
|
00:SCAN HDFS [default.test, RANDOM]
partitions=1/554 files=1 size=109.65MB
stats-rows=1506374 extrapolated-rows=disabled
table stats: rows=821958291 size=unavailable
column stats: all
mem-estimate=88.00MB mem-reservation=0B
tuple-ids=0 row-size=2.06KB cardinality=1506374
This was generated as a result of compute incremental stats on a single partition, so the output of that aggregation is a single row. Due to the width of the intermediate rows, such overestimations lead to bloated memory estimates. Since the amount of partitions to be updated is known at plan-time, Impala could use that to set the aggregation's cardinality.