[ 
https://issues.apache.org/jira/browse/SPARK-26841?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16849454#comment-16849454
 ] 

Tomas Bartalos commented on SPARK-26841:
----------------------------------------

Hi [~Yohan123],

Those queries are useful when you want to query data directly from Kafka. Let's 
say Kafka have retention of 1 week but you're interested in data for the last 4 
hours to create a live report. The current spark implementation will do 
"full-table" scan of the whole week and filter the last 4 hours in spark, which 
is very inefficient. My proposal is to push the timestamp filter down to Kafka 
during initial query and get back only 4 hours of data, to make queries fast.

> Timestamp pushdown on Kafka table
> ---------------------------------
>
>                 Key: SPARK-26841
>                 URL: https://issues.apache.org/jira/browse/SPARK-26841
>             Project: Spark
>          Issue Type: Improvement
>          Components: Input/Output
>    Affects Versions: 2.4.0
>            Reporter: Tomas Bartalos
>            Priority: Major
>              Labels: Kafka, pushdown, timestamp
>
> As a Spark user I'd like to have fast queries on Kafka table restricted by 
> timestamp.
> I'd like to have quick answers on questions like:
>  * What was inserted to Kafka in past x minutes
>  * What was inserted to Kafka in specified time range
> Example:
> {quote}select * from kafka_table where timestamp > 
> from_unixtime(unix_timestamp() - 5 * 60, "YYYY-MM-dd HH:mm:ss")
> select * from kafka_table where timestamp > $from_time and timestamp < 
> $end_time
> {quote}
> Currently timestamp restrictions are not pushdown to KafkaRelation and 
> querying by timestamp on a large Kafka topic takes forever to complete.
> *Technical solution*
> Technically its possible to retrieve Kafka's offsets by provided timestamp 
> with org.apache.kafka.clients.consumer.Consumer#offsetsForTimes(..) method. 
> Afterwards we can query Kafka topic by retrieved timestamp ranges.
> Querying by timestamp range is already implemented so this change should have 
> minor impact.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org

Reply via email to