[SPARK-17980] [SQL] Fix refreshByPath for converted Hive tables #15521
Conversation
|
Test build #67090 has finished for PR 15521 at commit
|
| /** | ||
| * Invalidate and refresh all the cached data (and the associated metadata) for any dataframe that | ||
| * contains the given data source path. | ||
| * contains the given data source path. Path matching is by prefix, i.e. "/" would invalidate |
There was a problem hiding this comment.
So we are changing the semantic of REFRESH PATH right?
|
Before #14690, users can refresh the whole table(including all partitions) by REFRESH the table path right? even the partition path is not under table path. |
|
That's correct, refresh table has always worked. There was just a bug introduced that broke refreshByPath, since it doesn't invalidate the lazy val. |
|
The new prefix matching semantics makes sense to me |
| private val baseLocation = catalogTable.storage.locationUri | ||
|
|
||
| // Populated on-demand by calls to cachedAllPartitions | ||
| private var allPartitions: ListingFileCatalog = null |
There was a problem hiding this comment.
nit: according to the existing name style, we should name this var cachedAllPartitions, and name the public method allPartitions
| sql("select * from test").count() | ||
| } | ||
| assert(e2.getMessage.contains("FileNotFoundException")) | ||
| spark.catalog.refreshByPath(dir.getAbsolutePath) |
There was a problem hiding this comment.
Note: before #14690, users need to refresh one of the partition paths to invalide the cache, but now they have to refresh the table path, because TableFileCatalog.rootPaths only contains table path while ListingFileCatalog.rootPaths only contains partition paths.
I think it's better than before, but it's still a breaking change, should we docuement it in the 2.1 release notes?
There was a problem hiding this comment.
Makes sense. To get the old behavior, they can also disable the feature flag.
| .map(_.makeQualified(fs.getUri, fs.getWorkingDirectory)) | ||
| .contains(qualifiedPath) | ||
| .map(_.makeQualified(fs.getUri, fs.getWorkingDirectory).toString) | ||
| .exists(_.startsWith(prefixToInvalidate)) |
There was a problem hiding this comment.
why do we need the prefix resolution? I think it's useful, so that users can refresh the table path to invalidate cache for partitioned data source table, but it's not related to this PR right?
There was a problem hiding this comment.
You actually need this when metastore partition pruning is disabled for converted hive tables. Otherwise, the unit test below would fail on that case.
(but yeah, we could also leave that alone)
|
Test build #67137 has finished for PR 15521 at commit
|
|
thanks, merging to master! |
## What changes were proposed in this pull request? There was a bug introduced in apache#14690 which broke refreshByPath with converted hive tables (though, it turns out it was very difficult to refresh converted hive tables anyways, since you had to specify the exact path of one of the partitions). This changes refreshByPath to invalidate by prefix instead of exact match, and fixes the issue. cc sameeragarwal for refreshByPath changes mallman ## How was this patch tested? Extended unit test. Author: Eric Liang <[email protected]> Closes apache#15521 from ericl/fix-caching.
## What changes were proposed in this pull request? There was a bug introduced in apache#14690 which broke refreshByPath with converted hive tables (though, it turns out it was very difficult to refresh converted hive tables anyways, since you had to specify the exact path of one of the partitions). This changes refreshByPath to invalidate by prefix instead of exact match, and fixes the issue. cc sameeragarwal for refreshByPath changes mallman ## How was this patch tested? Extended unit test. Author: Eric Liang <[email protected]> Closes apache#15521 from ericl/fix-caching.
4 participants
What changes were proposed in this pull request?
There was a bug introduced in #14690 which broke refreshByPath with converted hive tables (though, it turns out it was very difficult to refresh converted hive tables anyways, since you had to specify the exact path of one of the partitions).
This changes refreshByPath to invalidate by prefix instead of exact match, and fixes the issue.
cc @sameeragarwal for refreshByPath changes
@mallman
How was this patch tested?
Extended unit test.