[
https://issues.apache.org/jira/browse/HIVE-19985?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16568707#comment-16568707
]
Gopal V commented on HIVE-19985:
--------------------------------
Tested with/without flag=true, with {{select count(1), sum(ss_net_profit) from
store_sales;}}.
with: 31.383 seconds (cold run), 15.278 seconds (hot run)
without: 35.525 seconds (cold run), 22.934 seconds (hot run)
Latest patch has a big impact on the cached runs.
I can see there's L1 cache miss hotspot in the double System.arrayCopy to make
AcidWrapper and then to copy it back in copyBase().
{code}
if (isAcidScan) {
+ int acidColCount = acidReader.includeAcidColumns() ?
OrcInputFormat.getRootColumn(false) - 1 : 0;
...
+ int ixInVrb = includes.getPhysicalColumnIds().get(ixInReadSet) -
+ (acidReader.includeAcidColumns() ? 0 : OrcRecordUpdater.ROW);
{code}
Can that be changed to if(isAcidScan && innerReader.includeAcidColumns()) to
skip that entirely, because the offsets fall back the same way to the non-acid
impl?
> ACID: Skip decoding the ROW__ID sections for read-only queries
> ---------------------------------------------------------------
>
> Key: HIVE-19985
> URL: https://issues.apache.org/jira/browse/HIVE-19985
> Project: Hive
> Issue Type: Improvement
> Components: Transactions
> Reporter: Gopal V
> Assignee: Eugene Koifman
> Priority: Major
> Labels: Branch3Candidate
> Attachments: HIVE-19985.01.patch, HIVE-19985.04.patch
>
>
> For a base_n file there are no aborted transactions within the file and if
> there are no pending delete deltas, the entire ACID ROW__ID can be skipped
> for all read-only queries (i.e SELECT), though it still needs to be projected
> out for MERGE, UPDATE and DELETE queries.
> This patch tries to entirely ignore the ACID ROW__ID fields for all tables
> where there are no possible deletes or aborted transactions for an ACID split.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
