[
https://issues.apache.org/jira/browse/HIVE-13132?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15167639#comment-15167639
]
Anthony Hsu commented on HIVE-13132:
------------------------------------
Thanks for the review, [~alangates].
# I tested and unfortunately, HIVE-2573 does NOT solve this issue. A stacktrace
in jdb shows all functions are loaded during CliDriver start-up:
{noformat}
main[1] where
[1] org.apache.hadoop.hive.ql.metadata.Hive.reloadFunctions (Hive.java:173)
[2] org.apache.hadoop.hive.ql.metadata.Hive.<clinit> (Hive.java:166)
[3] org.apache.hadoop.hive.ql.session.SessionState.start
(SessionState.java:503)
[4] org.apache.hadoop.hive.cli.CliDriver.run (CliDriver.java:677)
[5] org.apache.hadoop.hive.cli.CliDriver.main (CliDriver.java:621)
[6] sun.reflect.NativeMethodAccessorImpl.invoke0 (native method)
[7] sun.reflect.NativeMethodAccessorImpl.invoke
(NativeMethodAccessorImpl.java:62)
[8] sun.reflect.DelegatingMethodAccessorImpl.invoke
(DelegatingMethodAccessorImpl.java:43)
[9] java.lang.reflect.Method.invoke (Method.java:483)
[10] org.apache.hadoop.util.RunJar.run (RunJar.java:221)
[11] org.apache.hadoop.util.RunJar.main (RunJar.java:136)
{noformat}
# My fix is a bit hacky and only works on the Hive 0.13.1 branch. I basically
changed the CliDriver initialization code to use
{{FunctionRegistry.getFunctionNames(String funcPatternStr)}} instead of
{{FunctionRegistry.getFunctionNames()}}. In the Hive 0.13.1 branch, the former
does NOT search the metastore while the latter does. This is no longer the case
in trunk.
# I don't have much experience with HS2 but I will take a look.
I'll work on a cleaner solution that removes the pre-loading of metastore
functions from the CliDriver initialization code path. Suggestions welcome.
> Hive should lazily load and cache metastore (permanent) functions
> -----------------------------------------------------------------
>
> Key: HIVE-13132
> URL: https://issues.apache.org/jira/browse/HIVE-13132
> Project: Hive
> Issue Type: Improvement
> Affects Versions: 0.13.1
> Reporter: Anthony Hsu
> Assignee: Anthony Hsu
> Attachments: HIVE-13132.1.patch
>
>
> In Hive 0.13.1, we have noticed that as the number of databases increases,
> the start-up time of the Hive interactive shell increases. This is because
> during start-up, all databases are iterated over to fetch the permanent
> functions to display in the {{SHOW FUNCTIONS}} output.
> {noformat:title=FunctionRegistry.java}
> private static Set<String> getFunctionNames(boolean searchMetastore) {
> Set<String> functionNames = mFunctions.keySet();
> if (searchMetastore) {
> functionNames = new HashSet<String>(functionNames);
> try {
> Hive db = getHive();
> List<String> dbNames = db.getAllDatabases();
> for (String dbName : dbNames) {
> List<String> funcNames = db.getFunctions(dbName, "*");
> for (String funcName : funcNames) {
> functionNames.add(FunctionUtils.qualifyFunctionName(funcName,
> dbName));
> }
> }
> } catch (Exception e) {
> LOG.error(e);
> // Continue on, we can still return the functions we've gotten to
> this point.
> }
> }
> return functionNames;
> }
> {noformat}
> Instead of eagerly loading all metastore functions, we should only load them
> the first time {{SHOW FUNCTIONS}} is invoked. We should also cache the
> results.
> Note that this issue may have been fixed by HIVE-2573, though I haven't
> verified this.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
