[
https://issues.apache.org/jira/browse/HIVE-23082?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17068482#comment-17068482
]
Zoltan Haindrich commented on HIVE-23082:
-----------------------------------------
there are 3 things I've found:
* there is some wierd distinct count is being calculated for that 5K rows or
so....there should be a bug in there; I guess we are using a 512 or something
sized hll sketch ; keeping track of 70 elements shouldn't be a problem - I
tried to repro it from a simple test; but so far it keep returning the right
value all the time.
* the primary key identification right now sees the numNulls as well during
evaluation; which may prevent it from setting it if the table contains nulls
** note that this would be purely an improvement; not related to the bug
** [~jcamachorodriguez] have suggested to get rid of the `numNulls` during the
`is null` evaluation - which would be more naturally placed; it could achieve
the same and could probably further improve things for other parts as well
* and there is the real cause of this issue: which is almost the same as the
previous improvement:
** the `sr_reason_sk` column is used in a join => `is not null` filter is
inserted into the plan; and it's not removed because the table does contain
`null` values
** the `is not null` is evaluated during stat annotation; and the referenced
column is marked as `filtered`
** but rescaling a filtered column with a primary key column may get us into
underestimation trouble - so the estimator decides to get a bit more
conservative with it
The problem is caused by that `isFiltered` is set for because of the
`isNotNull`.
* opt#1:
** introduce a model in `StatsRulesProcFactory` which could describe the
statistics of a whole subexpression - right now we only have the "numRows" (and
there are some other tweaks)
*** this could clear things up a bit
** use the model to handle `isNotNull` to zero out the `numNulls` field for the
column - and because that's all the operator does; it may skip setting the
`isFiltered` for the column
** (`isFiltered` will enable pkfk independency => which will then rescale the
join output)
* opt#2:
** intorduce some conditionals in `StatsRulesProcFactory`; to not do an
`updateStats` call for `isNotNull` functions
I had a patch for the refactor a while ago - I was just experimenting with it ;
but I somehow left it behind - and got forgotten....IIRC it wasn't that hard -
but it was moderately larger; I think if we keep it separetly that will work...
after that fixing the issue (and adding the improvement) will be easy as pie...
:)
I'll start digging in this direction :D
> PK/FK stat rescale doesn't work in some cases
> ---------------------------------------------
>
> Key: HIVE-23082
> URL: https://issues.apache.org/jira/browse/HIVE-23082
> Project: Hive
> Issue Type: Bug
> Reporter: Zoltan Haindrich
> Assignee: Zoltan Haindrich
> Priority: Major
>
> As a result in Joins may retain the original estimate; see MAPJOIN_33 in this
> plan ; which retained the estimate of SEL_32
> {code}
> +----------------------------------------------------+
> | Explain |
> +----------------------------------------------------+
> | Plan optimized by CBO. |
> | |
> | Vertex dependency in root stage |
> | Map 1 <- Map 2 (BROADCAST_EDGE) |
> | |
> | Stage-0 |
> | Fetch Operator |
> | limit:12 |
> | Stage-1 |
> | Map 1 vectorized |
> | File Output Operator [FS_36] |
> | Limit [LIM_35] (rows=12 width=4) |
> | Number of rows:12 |
> | Select Operator [SEL_34] (rows=5040 width=4) |
> | Output:["_col0"] |
> | Map Join Operator [MAPJOIN_33] (rows=5040 width=8) |
> | Conds:SEL_32._col0=RS_30._col0(Inner) |
> | <-Map 2 [BROADCAST_EDGE] vectorized |
> | BROADCAST [RS_30] |
> | PartitionCols:_col0 |
> | Select Operator [SEL_29] (rows=1 width=8) |
> | Output:["_col0"] |
> | Filter Operator [FIL_28] (rows=1 width=108) |
> | predicate:((r_reason_id = 'reason 66') and r_reason_sk
> is not null) |
> | TableScan [TS_3] (rows=2 width=108) |
> |
> default@rx0,reason,Tbl:COMPLETE,Col:COMPLETE,Output:["r_reason_id","r_reason_sk"]
> |
> | <-Select Operator [SEL_32] (rows=5040 width=7) |
> | Output:["_col0"] |
> | Filter Operator [FIL_31] (rows=5040 width=7) |
> | predicate:sr_reason_sk is not null |
> | TableScan [TS_0] (rows=5112 width=7) |
> |
> default@sr0,store_returns,Tbl:COMPLETE,Col:COMPLETE,Output:["sr_reason_sk"] |
> | |
> +----------------------------------------------------+
> {code}
> repro:
> {code}
> set hive.query.results.cache.enabled=false;
> set hive.explain.user=true;
> drop table if exists default.rx0;
> drop table if exists default.sr0;
> create table rx0 (r_reason_id string, r_reason_sk bigint);
> create table sr0 (sr_reason_sk bigint);
> insert into rx0 values ('AAAAAAAAAAAAAAAA',1),('AAAAAAAAGEAAAAAA',70);
> insert into sr0 values (NULL),(1),(2),(3),(4),(5),(6),(7),(8),(9),(10),
> (11),(12),(13),(14),(15),(16),(17),(18),(19),(20),(21),(22),(23),(24),(25),
> (26),(27),(28),(29),(30),(31),(32),(33),(34),(35),(36),(37),(38),(39),(40),
> (41),(42),(43),(44),(45),(46),(47),(48),(49),(50),(51),(52),(53),(54),(55),
> (56),(57),(58),(59),(60),(61),(62),(63),(64),(65),(66),(67),(68),(69),(70);
> insert into sr0 select a.* from sr0 a,sr0 b;
> -- |sr0| ~ 5112
> explain select 1
> from default.sr0 store_returns , default.rx0 reason
> where sr_reason_sk = r_reason_sk
> and r_reason_id = 'reason 66'
> limit 12;
> {code}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
