[
https://issues.apache.org/jira/browse/HIVE-20198?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16549010#comment-16549010
]
Peter Vary commented on HIVE-20198:
-----------------------------------
+1 to the general idea.
Some thoughts (all of them just general ideas, and stating here only to start
the discussion)
* I would vote for a separate new column signaling that the table is already
soft deleted.
** I think that is the most straight-forward, method handling this type of
situations. Helps to have a clean, easily readable code.
** We need several queries where we have to filter out the already deleted
tables. The backend DB can do this for us better if there is a separate column.
** We would like to revise the indexes on the TBLS table, so the above queries
can have faster results
* Not sure about the TBLS.TBL_ID - If we want to follow up on [~alangates]'
federated MetaStore idea, where separate catalogs can be served by different
MetaStore instances we might end up with TBL_ID collisions. I would prefer to
keep open this direction if we want to move there. Also when
replicating/recovering the tables we might end up with different TBL_IDs.
* Constant time table drops, and renames should work for external tables too,
since we "by definition" do not manage the filesystem for them (only create
directories when necessary)
* I would love to see asynchronous metadata drop for managed tables too. We
can mark the metadata to be dropped (and later removed asynchronously), and
remove the directories synchronously.
** This would work in case when impersonation is enabled
** This would work for every existing managed tables too
** During my previous tests I have seen that the lion share of the time spent
on dropping tables is the metadata changes. Removing files from HDFS is
relatively cheap. In case of S3 or ADLS the user can still chose to "opt-in"
for the new table handling system.
* How to signal that this table is "rename" optimized:
** Currently we have TableType - Managed / External
** Currently we have properties - ACID, etc
** I think we should decide which is a way to go, and at least with the new
types do accordingly
* Renaming tables
** Currently we change only this for partitions:
{code:java}
String newPath = oldUri.getPath().replace(oldTblLocPath, newTblLocPath);
Path newPartLocPath = new Path(oldUri.getScheme(), oldUri.getAuthority(),
newPath);
{code}
** It might be worth to consider some DirectSQL change for this - might be
challenging to find a method that works on all supported databases, but we
might get very fast results - not O(1), but fast anyway :)
Thanks for reading through this long post!
Peter
> Constant time table drops/renames
> ---------------------------------
>
> Key: HIVE-20198
> URL: https://issues.apache.org/jira/browse/HIVE-20198
> Project: Hive
> Issue Type: Improvement
> Components: Metastore
> Affects Versions: 4.0.0
> Reporter: Alexander Kolbasov
> Priority: Major
>
> Currently table drops and table renames have O(P) performance (where P is the
> number of partitions). When a managed table is deleted, the implementation
> deletes table metadata and then deletes all partitions in HDFS. HDFS
> operations are optimized and only do a sequential deletes for partitions
> outside of table prefix. This operation is O(P)where Pis the number of
> partitions.
> Table rename goes through the list of partitions and modifies table name (and
> potentially db name) in each partition. It also modifies each partition
> location to match the new db/table name and renames directories (which is a
> non-atomic and slow operation on S3). This is O(P) operation where P is the
> number of partitions.
> Basic idea is to do the following:
> # Assign unique ID to each table
> # Create directory name based on unique ID rather then the name
> # Table rename then becomes metadata-only operation - there is no need to
> change any location information.
> # Table drop can become an asynchronous operation where the table is marked
> as "deleted". Subsequent public metadata APIs should skip such tables. A
> background cleaner thread may then go and clean up directories.
> Since the table location is unique for each table, new tables will not reuse
> existing locations. This change isn't compatible with the current behavior
> where there is an assumption that table location is based on table name. We
> can get around this by providing "opt-in" mechanism - special table property
> that tells that the table can have such new behavior, so the improvement will
> initially work for new tables created with this feature enabled. We may later
> provide some tool to convert existing tables to the new scheme.
> One complication is there in case where impersonation is enabled - the FS
> operations should be performed using client UGI rather then server's, so the
> cleaner thread should be able to use client UGIs.
> Initially we can punt on this and do standard table drops when impersonation
> is enabled.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
