[
https://issues.apache.org/jira/browse/IMPALA-11293?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17683474#comment-17683474
]
Zoltán Borók-Nagy commented on IMPALA-11293:
--------------------------------------------
Thanks for the ideas [~lipenglin].
RewriteFiles need the exact list of the old data files which might not be
trivial to get in the presence of concurrent operations.
We could use ReplacePartitions API instead, but that also needs to be protected
by concurrent writes. But if we track the original snapshot id, and add some
conflict validators then it seems doable.
At first it's probably enough to handle IDENTITY-partitions, dealing with
partition transforms can be a future task.
Partition-evolution can also cause some problems. But again, at first maybe we
can just deal with tables with unmodified partition specs.
> Add COMPACT command for Iceberg tables
> --------------------------------------
>
> Key: IMPALA-11293
> URL: https://issues.apache.org/jira/browse/IMPALA-11293
> Project: IMPALA
> Issue Type: Improvement
> Components: Frontend
> Reporter: Zoltán Borók-Nagy
> Assignee: Tamas Mate
> Priority: Major
> Labels: impala-iceberg
>
> Currently Impala cannot compact Iceberg tables.
> The following INSERT OVERWRITE statement could be used in the simple cases,
> i.e. when the following conditions meet:
> * all data files use the same partition spec (i.e. no partition evolution)
> * no bucket partitioning (we currently forbid INSERT OVERWRITE for bucket
> partitioning)
> {noformat}
> INSERT OVERWRITE t SELECT * FROM t;{noformat}
> We could have a command that compacts the Iceberg table (syntax needs to be
> the same with Hive), e.g.:
> {noformat}
> ALTER TABLE t EXECUTE compaction();{noformat}
> At first, the compact command could be just rewritten to the INSERT OVERWRITE
> command, but it would also check that there's no partition evolution.
> The "no bucket" partitioning condition could be relaxed in this case, because
> the result would be deterministic. I.e. the only condition we need to check
> is that there was no partition evolution.
> Later, we could do compaction by
> {noformat}
> TRUNCATE TABLE t;
> INSERT INTO t SELECT * FROM t FOR SYSTEM_TIME AS OF ...;{noformat}
> Currently time-travel queries are not optimized, but we could workaround it
> by doing planning at first of:
> {noformat}
> Create the plan for:
> TRUNCATE TABLE t;
> INSERT INTO t SELECT * FROM t;{noformat}
> Then execute them:
> {noformat}
> Actually execute:
> TRUNCATE TABLE t;
> INSERT INTO t SELECT * FROM t; (no need for time-travel, plan was created
> before TRUNCATE){noformat}
> This could workaround the planning overhead of time-travel queries.
> Also, we might add some locking for the table if possible.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
