[ https://issues.apache.org/jira/browse/HIVE-3585?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13713370#comment-13713370 ]
Yin Huai commented on HIVE-3585: -------------------------------- I am concerned about the performance of Trevni. Please correct me if my explanation is not accurate. First, seems we store a column by multiple compression blocks and the size of a block is around 64 KiB. When the reader reads data from a stored table, we read a compression block at a time. Because the reader is row-oriented, in the worst case, the reader needs to seek to another column after reading every single compression block. As pointed out in AVRO-1208, the performance degradation is significant. Second, since we read a compression unit at a time, when the size of a compression unit is smaller than the size of buffer used in the BufferedInputStream inside BlockReader, we may not be able to efficiently read data. We will just use a portion of data fetched by BufferedInputStream and the rest of data cannot be used. I have experience that the performance of Trevni on HDFS without short circuit is not good. I think this issue may be the reason. But I have not dug very deep. > Integrate Trevni as another columnar oriented file format > --------------------------------------------------------- > > Key: HIVE-3585 > URL: https://issues.apache.org/jira/browse/HIVE-3585 > Project: Hive > Issue Type: Improvement > Components: Serializers/Deserializers > Affects Versions: 0.10.0 > Reporter: alex gemini > Assignee: Mark Wagner > Priority: Minor > Attachments: futurama_episodes.avro, HIVE-3585.1.patch.txt > > > add new avro module trevni as another columnar format.New columnar format > need a columnar SerDe,seems fastutil is a good choice.the shark project use > fastutil library as columnar serde library but it seems too large (almost > 15m) for just a few primitive array collection. -- This message is automatically generated by JIRA. If you think it was sent incorrectly, please contact your JIRA administrators For more information on JIRA, see: http://www.atlassian.com/software/jira