[
https://issues.apache.org/jira/browse/PHOENIX-1498?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14232686#comment-14232686
]
James Taylor commented on PHOENIX-1498:
---------------------------------------
Only the tests that rely on the flashback functionality need
KEEP_DELETED_CELLS=true. Perhaps a better approach (albeit more work) for the
patch would be to turn it off, then for failed test add KEEP_DELETED_CELLS=true
for the DDL statement.
We should keep it on for SYSTEM.CATALOG, SYSTEM.SEQUENCE, and SYSTEM.STATS -
just add it to the DDL statements for those.
> Turn KEEP_DELETED_CELLS off by default
> --------------------------------------
>
> Key: PHOENIX-1498
> URL: https://issues.apache.org/jira/browse/PHOENIX-1498
> Project: Phoenix
> Issue Type: Bug
> Affects Versions: 4.0.0, 5.0.0
> Reporter: Jeffrey Zhong
> Assignee: Jeffrey Zhong
> Attachments: PHOENIX-1498.patch
>
>
> Phoenix table is created with "KEEP_DELETED_CELLS" enabled by default, this
> is only used to allow for flashback queries to work correctly. While
> flashback query isn't used often in field and we found that query performance
> degraded with the option on. This is likely a hbase scan issue though(will
> create a JIRA once having more info).
> Anyway Keeping deleted cells will add performance penalty and it's not used
> often. Therefore, I'm suggesting to set it off by default.
> We have a test where a table is loaded with > 5m rows and then some are
> deleted/reinserted. The count ( * ) performance became worse & worse:
> {code}
> +------------+
> | COUNT(1) |
> +------------+
> | 5078242 |
> +------------+
> 1 row selected (33.273 seconds)
> +------------+
> | COUNT(1) |
> +------------+
> | 5078242 |
> +------------+
> 1 row selected (174.771 seconds)
> +------------+
> | COUNT(1) |
> +------------+
> | 5078242 |
> +------------+
> 1 row selected (458.251 seconds)
> {code}
> I think we can provide a table property in CREATE TABLE & ALTER TABLE
> statement for people to enable KEEP_DELETED_CELLS if there is a need but by
> default it should be turned off.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
