[ https://issues.apache.org/jira/browse/SPARK-8890?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14628620#comment-14628620 ]
Ilya Ganelin commented on SPARK-8890: ------------------------------------- [~yhuai] That makes sense, thank you. Wouldn't we still need to close/delete output buffers for keys that have been completely written? Thus, would we, for example, write all values associated with key=1, then close that output buffer, write the next one etc. Operational flow would become: 1) Attempt to create new outputWriter for each possible key 2) When maximum is exceeded, stop outputting rows. 3) Sort all remaining data by key (and persist this sorted set of {code}InternalRow{code} objects in memory. 4) One key at a time, create an outputWriter and write all rows associated with that key 5) Close outputWriter for that key and open a new outputWriter, continue from step 4. > Reduce memory consumption for dynamic partition insert > ------------------------------------------------------ > > Key: SPARK-8890 > URL: https://issues.apache.org/jira/browse/SPARK-8890 > Project: Spark > Issue Type: Sub-task > Components: SQL > Reporter: Reynold Xin > Priority: Critical > > Currently, InsertIntoHadoopFsRelation can run out of memory if the number of > table partitions is large. The problem is that we open one output writer for > each partition, and when data are randomized and when the number of > partitions is large, we open a large number of output writers, leading to OOM. > The solution here is to inject a sorting operation once the number of active > partitions is beyond a certain point (e.g. 50?) -- This message was sent by Atlassian JIRA (v6.3.4#6332) --------------------------------------------------------------------- To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org For additional commands, e-mail: issues-h...@spark.apache.org