[ https://issues.apache.org/jira/browse/HIVE-17417?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16242708#comment-16242708 ]
Prasanth Jayachandran commented on HIVE-17417: ---------------------------------------------- [~ashutoshc] Here is the difference between Joda, JDK8 and SDF {code} Benchmark Mode Cnt Score Error Units TimestampFormatBenchmark.benchmarkJDK8TimestampFormatter avgt 10 274.533 ± 2.663 ns/op TimestampFormatBenchmark.benchmarkJodaTimeTimestampFormatter avgt 10 230.096 ± 5.089 ns/op TimestampFormatBenchmark.benchmarkSimpleDateFormatter avgt 10 543.846 ± 12.677 ns/op {code} On bigger scale this slightly difference between joda vs jdk8 will not matter much. Moving to native JDK8 will be better too. Will update the patch with JDK8 formatter. > LazySimple Timestamp is very expensive > -------------------------------------- > > Key: HIVE-17417 > URL: https://issues.apache.org/jira/browse/HIVE-17417 > Project: Hive > Issue Type: Bug > Components: Serializers/Deserializers > Affects Versions: 3.0.0, 2.4.0 > Reporter: Prasanth Jayachandran > Assignee: Prasanth Jayachandran > Priority: Critical > Attachments: HIVE-17417.1.patch, HIVE-17417.2.patch, > HIVE-17417.3.patch, HIVE-17417.4.patch, date-serialize.png, > timestamp-serialize.png, ts-jmh-perf.png > > > In a specific case where a schema contains array<struct> with timestamp and > date fields (array size >10000). Any access to this column very very > expensive in terms of CPU as most of the time is serialization of timestamp > and date. Refer attached profiles. >70% time spent in serialization + > tostring conversions. -- This message was sent by Atlassian JIRA (v6.4.14#64029)