[
https://issues.apache.org/jira/browse/HIVE-15181?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15685192#comment-15685192
]
Eugene Koifman edited comment on HIVE-15181 at 11/22/16 1:48 AM:
-----------------------------------------------------------------
I think
{noformat}if (inList == null || inList.size() == 0) {
return;
}
{noformat}
should raise an error instead and force the caller to not call this at all with
an empty list. Such a call is certainly a mistake but as this bug indicates,
returning an empty list of queries can cause really bad problems.
in TxnUtils.needNewQuery() "sizeInBytes / 1024 > queryMemoryLimit" doesn't do
the right thing.
If the user sets METASTORE_DIRECT_SQL_MAX_QUERY_LENGTH to 1K, they most likely
want each SQL string to be at most 1K.
But if sizeInBytes=2047, this still returns false.
It should include length of "suffix" in computation of sizeInBytes
Along the same lines: the check for max query length is done after each batch
is already added to the query. Suppose there are 1000 9-digit txn IDs in each
IN(...). That's, conservatively, 18KB of text. So the length of each query is
increasing in 18KB chunks.
I think the check for query length should be done for each item in IN clause.
If some DB has a limit on query length of X, then any query > X will fail. So
I think this must ensure not to produce any queries > X, even by 1 char.
For example, case 3.1 of the UT generates a query of almost 4000 characters -
this is clearly > 1KB.
was (Author: ekoifman):
I think
{noformat}if (inList == null || inList.size() == 0) {
return;
}
{noformat}
should raise an error instead and force the caller to not call this at all with
an empty list. Such a call is certainly a mistake but as this bug indicates,
returning an empty list of queries can cause really bad problems.
in TxnUtils.needNewQuery() "sizeInBytes / 1024 > queryMemoryLimit" doesn't do
the right thing.
If the user sets METASTORE_DIRECT_SQL_MAX_QUERY_LENGTH to 1K, they most likely
want each SQL string to be at most 1K.
But if sizeInBytes=2047, this still returns false.
It should include length of "suffix" in computation of sizeInBytes
Along the same lines: the check for max query length is done after each batch
is already added to the query. Suppose there are 1000 9-digit txn IDs in each
IN(...). That's, conservatively, 18KB of text. So the length of each query is
increasing in 18KB chunks.
I think the check for query length should be done for each item in IN clause.
If some DB has a limit on query length of X, then any query > X will fail. So
I think this must ensure not to produce any queries > X, even by 1 char.
> buildQueryWithINClause didn't properly handle multiples of
> ConfVars.METASTORE_DIRECT_SQL_MAX_ELEMENTS_IN_CLAUSE
> ---------------------------------------------------------------------------------------------------------------
>
> Key: HIVE-15181
> URL: https://issues.apache.org/jira/browse/HIVE-15181
> Project: Hive
> Issue Type: Bug
> Components: Hive, Transactions
> Affects Versions: 1.2.1, 2.1.0
> Reporter: Wei Zheng
> Assignee: Wei Zheng
> Priority: Critical
> Attachments: HIVE-15181.1.patch, HIVE-15181.2.patch
>
>
> As there is a bug, we can still work around the issue by using the settings
> below (making sure the second setting is always at least 1000 times of the
> first setting):
> set hive.direct.sql.max.query.length=1;
> set hive.direct.sql.max.elements.in.clause=1000;
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)