[
https://issues.apache.org/jira/browse/HIVE-20332?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Jesus Camacho Rodriguez updated HIVE-20332:
-------------------------------------------
Description:
Currently, we do not expose stats over {{ROW\_\_ID.writeId}} to the optimizer.
Even if we did, we always assume uniform distribution of the column values,
which can easily lead to overestimations on the number of rows read when we
filter on {{ROW\_\_ID.writeId}} for materialized views (think about a large
transaction for MV creation and then small ones for incremental maintenance).
This overestimation can lead to incremental view maintenance not being
triggered as cost of the incremental plan is overestimated (we think we will
read more rows than we actually do). This could be fixed by introducing
histograms that reflect better the column values distribution.
Till that moment, we will use a config variable that will set the selectivity
for filter condition on {{ROW\_\_ID}} during the cost calculation. Setting that
variable to a low value will favour incremental rebuild over full rebuild.
was:
Currently, we do not expose stats over {{ROW__ID.writeId}} to the optimizer.
Even if we did, we always assume uniform distribution of the column values,
which can easily lead to overestimations on the number of rows read when we
filter on {{ROW__ID.writeId}} for materialized views (think about a large
transaction for MV creation and then small ones for incremental maintenance).
This overestimation can lead to incremental view maintenance not being
triggered as cost of the incremental plan is overestimated (we think we will
read more rows than we actually do). This could be fixed by introducing
histograms that reflect better the column values distribution.
Till that moment, we will use a config variable that will set the selectivity
for filter condition on {{ROW\_\_ID}} during the cost calculation. Setting that
variable to a low value will favour incremental rebuild over full rebuild.
> Materialized views: Introduce heuristic on selectivity over ROW__ID to favour
> incremental rebuild
> -------------------------------------------------------------------------------------------------
>
> Key: HIVE-20332
> URL: https://issues.apache.org/jira/browse/HIVE-20332
> Project: Hive
> Issue Type: Improvement
> Components: Materialized views
> Reporter: Jesus Camacho Rodriguez
> Assignee: Jesus Camacho Rodriguez
> Priority: Major
>
> Currently, we do not expose stats over {{ROW\_\_ID.writeId}} to the
> optimizer. Even if we did, we always assume uniform distribution of the
> column values, which can easily lead to overestimations on the number of rows
> read when we filter on {{ROW\_\_ID.writeId}} for materialized views (think
> about a large transaction for MV creation and then small ones for incremental
> maintenance). This overestimation can lead to incremental view maintenance
> not being triggered as cost of the incremental plan is overestimated (we
> think we will read more rows than we actually do). This could be fixed by
> introducing histograms that reflect better the column values distribution.
> Till that moment, we will use a config variable that will set the selectivity
> for filter condition on {{ROW\_\_ID}} during the cost calculation. Setting
> that variable to a low value will favour incremental rebuild over full
> rebuild.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
