[
https://issues.apache.org/jira/browse/KYLIN-3122?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16344161#comment-16344161
]
Vsevolod Ostapenko commented on KYLIN-3122:
-------------------------------------------
I think I found one place in the code that is at least partially responsible
for the behavior observed.
The convertFilterColumnsAndConstants method in the GTUtil.java rewrites
statement filter, after static values in the WHERE clause were checked against
tri-dictionary.
There seems to be multiple issues with this approach:
1) Filtering on the partitioning key is treated the same as filtering on a
non-partitioning columns, which is incorrect, as presence or absence of a lower
or upper range bound for partitioning column in the dictionary in a specific
segment provides no guarantees that this segment is or is not a candidate for
further scan.
2) As the side effect of the #1, it looks like after first candidate segment is
hit (the lower bound date-time value is found in dictionary), the filter is
modified in place (rewritten) to exclude the upper bound condition (if upper
bound condition is not found in the segment, which is always the case in our
scenario).
Partitioning keys require special handling, they need to be checked against
segment range meta-data, and excluded from dictionary-based checks.
> Partition elimination algorithm seems to be inefficient and have serious
> issues with handling date/time ranges, can lead to very slow queries and
> OOM/Java heap dump conditions
> -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: KYLIN-3122
> URL: https://issues.apache.org/jira/browse/KYLIN-3122
> Project: Kylin
> Issue Type: Bug
> Components: Query Engine
> Affects Versions: v2.2.0
> Environment: HDP 2.5.6, Kylin 2.2.0
> Reporter: Vsevolod Ostapenko
> Assignee: hongbin ma
> Priority: Critical
> Attachments: partition_elimination_bug_single_column_test.log
>
>
> Current algorithm of cube segment elimination seems to be rather inefficient.
> We are using a model where cubes are partitioned by date and time:
> "partition_desc":
> { "partition_date_column": "A_VL_HOURLY_V.THEDATE", "partition_time_column":
> "A_VL_HOURLY_V.THEHOUR", "partition_date_start": 0, "partition_date_format":
> "yyyyMMdd", "partition_time_format": "HH", "partition_type": "APPEND",
> "partition_condition_builder":
> "org.apache.kylin.metadata.model.PartitionDesc$DefaultPartitionConditionBuilder"
> }
> ,
> Cubes contain partitions for multiple days and 24 hours for each day. Each
> cube segment corresponds to just one hour.
> When a query is issued where both date and hour are specified using equality
> condition (e.g. thedate = '20171011' and thehour = '10') Kylin sequentially
> integrates over all the segment cubes (hundreds of them) only to skip all
> except for the one that needs to be scanned (which can be observed by looking
> in the logs).
> The expectation is that Kylin would use existing info on the partitioning
> columns (date and time) and known hierarchical relations between date and
> time to locate required partition much more efficiently that linear scan
> through all the cube partitions.
> Now, if filtering condition is on the range of hours, behavior of the
> partition pruning and scanning becomes not very logical, which suggests bugs
> in the logic.
> If filtering condition is on specific date and closed-open range of hours
> (e.g. thedate = '20171011' and thehour >= '10' and thehour < '11'), in
> addition to sequentially scanning all the cube partitions (as described
> above), Kylin will scan HBase tables for all the hours from the specified
> starting hour and till the last hour of the day (e.g. from hour 10 to 24,
> instead of just hour 10).
> As the result query will run much longer that necessary, and might run out
> of memory, causing JVM heap dump and Kylin server crash.
> If filtering condition is on specific date by hour interval is specified as
> open-closed (e.g. thedate = '20171011' and thehour > '09' and thehour <=
> '10'), Kylin will scan all HBase tables for all the later dates and hours
> (e.g. from hour 10 and till the most recent hour on the most recent day,
> which can be hundreds of tables and thousands of regions).
> As the result query execution will dramatically increase and in most cases
> Kylin server will be terminated with OOM error and JVM heap dump.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
