[
https://issues.apache.org/jira/browse/IGNITE-11891?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16856708#comment-16856708
]
João Fonseca commented on IGNITE-11891:
---------------------------------------
I used "explain sql ..." and verified that it uses the index_event_level index.
But from what I understood, it seems to only use it to fetch records where
level=WARNING, without taking into account that those results are already
ordered correctly.
A second stage in the execution will then try to order the results by
timestamp, which is unnecessary. Because the table contains millions of WARNING
records, the server dies with an OOM error.
In practice, Ignite behaves as though the index is just "(level)", not "(level,
timestamp desc)".
> Multi-column index - query out of memory
> ----------------------------------------
>
> Key: IGNITE-11891
> URL: https://issues.apache.org/jira/browse/IGNITE-11891
> Project: Ignite
> Issue Type: Bug
> Components: sql
> Affects Versions: 2.7
> Reporter: João Fonseca
> Priority: Major
>
> My application uses a table for logging events. Something like:
>
> {noformat}
> create table event (
> id bigint not null,
> level varchar(8) not null,
> timestamp bigint not null,
> message varchar(4096) not null,
> primary key (id)
> ) ;
> {noformat}
> I have two indexes:
>
> {noformat}
> create index index_event_timestamp on event (timestamp desc)
> create index index_event_level on event (level, timestamp desc)
> {noformat}
> The idea is to support both the following queries:
>
> {noformat}
> select * from event order by timestamp desc limit 25
> select * from event where level = 'WARNING' order by timestamp desc limit 25
> {noformat}
> Once the table size increases to several million records, the second query
> generates OOM on the server. From what I can see (from the explain results),
> the index_event_level is used to fetch records with WARNING level, but the
> timestamp column available with the index is not used in the "order by"
> clause. The server attempts to fetch all records and then sort them by
> timestamp, despite the index already doing this...
> I removed the second index as a work-around, and the query runs faster on the
> first index - it scans index_event_timestamp, and retrieves the records with
> level=WARNING. It's smart to realize that the scan results are already sorted
> correctly.
>
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
