[
https://issues.apache.org/jira/browse/SPARK-40499?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17806201#comment-17806201
]
Joey Pereira commented on SPARK-40499:
--------------------------------------
I've found another app of ours with similarly unbearable regressions. The app
previously performed in <2hr run-time prior to upgrading, and is now >24hr on
Spark3. It's a lot simpler, so this is more-or-less the SQL it's doing (with a
number of other internal group-by fields removed for brievity).
{code:sql}
with t as (
select
method,
response_code,
if(response_code >= 200 and response_code < 300, duration_ms, null) as
duration_ms_2xx,
from logs
)
select
method,
count(*) as cnt,
sum(case when response_code >= 200 and response_code < 300 then 1 else 0 end)
as cnt_2xx,
sum(case when response_code >= 300 and response_code < 400 then 1 else 0 end)
as cnt_3xx,
sum(case when response_code >= 400 and response_code < 500 then 1 else 0 end)
as cnt_4xx,
sum(case when response_code >= 500 then 1 else 0 end) as cnt_5xx,
approx_percentile(duration_ms_2xx, 0.5) as latency_p50,
approx_percentile(duration_ms_2xx, 0.9) as latency_p90,
approx_percentile(duration_ms_2xx, 0.99) as latency_p99,
approx_percentile(duration_ms_2xx, 0.999) as latency_p999,
from t
group by 1
{code}
This is simple enough that I'm a bit more convinced it could be the CaseWhen
change.
> Spark 3.2.1 percentlie_approx query much slower than Spark 2.4.0
> ----------------------------------------------------------------
>
> Key: SPARK-40499
> URL: https://issues.apache.org/jira/browse/SPARK-40499
> Project: Spark
> Issue Type: Bug
> Components: Shuffle
> Affects Versions: 3.2.1
> Environment: hadoop: 3.0.0
> spark: 2.4.0 / 3.2.1
> shuffle：spark 2.4.0
> Reporter: xuanzhiang
> Priority: Major
> Attachments: Screenshot 2024-01-05 at 3.51.52 PM.png, Screenshot
> 2024-01-05 at 3.53.10 PM.png, spark2.4-shuffle-data.png,
> spark3.2-shuffle-data.png
>
>
> spark.sql(
> s"""
> |SELECT
> | Info ,
> | PERCENTILE_APPROX(cost,0.5) cost_p50,
> | PERCENTILE_APPROX(cost,0.9) cost_p90,
> | PERCENTILE_APPROX(cost,0.95) cost_p95,
> | PERCENTILE_APPROX(cost,0.99) cost_p99,
> | PERCENTILE_APPROX(cost,0.999) cost_p999
> |FROM
> | textData
> |""".stripMargin)
> * When we used spark 2.4.0, aggregation adopted objHashAggregator, stage 2
> pull shuffle data very quick . but , when we use spark 3.2.1 and use old
> shuffle , 140M shuffle data cost 3 hours.
> * If we upgrade the Shuffle, will we get performance regression？
> *
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
