[
https://issues.apache.org/jira/browse/IGNITE-13270?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17178561#comment-17178561
]
Tanmay Ambre commented on IGNITE-13270:
---------------------------------------
[~zstan] - yes 2.8.0 is good. There was this one query which spiked the CPU. We
have broken it down into 2 queries and the performance is acceptable. when we
moved to 2.8.0 - 2.8.1 was not available. However, in the coming months - we
would upgrade.
Yes - I do agree for having the test case for avoiding potential issues in the
future. What I dont understand is even with 2.7.5 we were getting accurate
results with lower CPU and with 2.8.0 we are getting the same results (in our
case we are looking for just one record). I think what you found out was when
the result set holds multiple records - 2.7.5 had issues.
> Massive CPU burn in 2.8.0 and 2.8.1 after upgrading from 2.7.5
> --------------------------------------------------------------
>
> Key: IGNITE-13270
> URL: https://issues.apache.org/jira/browse/IGNITE-13270
> Project: Ignite
> Issue Type: Bug
> Components: sql
> Affects Versions: 2.8, 2.8.1
> Reporter: Tanmay Ambre
> Assignee: Stanilovsky Evgeny
> Priority: Major
> Attachments: IDX13270.java
>
>
> After upgrading to ignite 2.8.0 from 2.7.5 the following query is causing
> massive CPU burn. We have a 8 node server cluster for Ignite (with 32 CPUs on
> each node). with 2.8.0, 2 nodes are maxed out on CPU. with 2.8.1 all nodes
> are above 80%. On 2.7.5 the same query was performing very nicely.
> The query is a join with 3 tables having same affinity key. however, we dont
> know the affinity key when we are querying and there this results in a merge
> scan.
> Query:
> select A.A_ID, A.AFFINITYKEY, B.B_ID from A, B, C
> where
> A.A_ID = B.A_ID AND
> B.B_ID = C.B_ID AND
> A.AFFINITYKEY = B.AFFINITYKEY AND
> B.AFFINITYKEY = C.AFFINITYKEY AND
> B.B_ID = ? AND
> C.C_ID = ?
>
> Performance:
> 2.7.5 ~ 2 to 5 ms
> 2.8.0 ~ 6 to 12 seconds
> 2.8.1 ~ within 3 seconds
>
> Volume (we have a EDA based system where messages come in burst. One burst
> has approx. 150K events - each event correspond to one query).
>
> throughput of our java based processor (that fires this query):
> 2.7.5 ~ 750 transactions per second
> 2.8.0 ~ 3 transactions per second
> 2.8.1 ~ 6 transactions per second
>
> CPU burn
> 2.7.5 ~ 10 to 15% on each node
> 2.8.0 ~ 100% on 2 nodes other nodes are idling
> 2.8.1 ~ 80 to 90% on each node
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
