Indian-boult opened a new pull request, #20739:
URL: https://github.com/apache/hudi/pull/20739

   
   ### Describe the issue this Pull Request addresses
   
   Closes #19893
   
   `Registry.getRegistryOfClass` (hudi-io) logs an error and returns the 
existing registry when the registry under the key has a different class. 
`SparkHoodieBackedTableMetadataWriter#initRegistry` and its `TableVersionSix` 
copy cast `Registry.getRegistry("HoodieMetadata", DistributedRegistry)` to 
`DistributedRegistry`. The key `::HoodieMetadata` is JVM-wide, so once a writer 
with metrics on and executor metrics off has created it as a `LocalRegistry`, 
every later metadata writer with executor metrics on (and a reporter other than 
INMEMORY) fails in its constructor:
   
   ```
   java.lang.ClassCastException: class 
org.apache.hudi.common.metrics.LocalRegistry cannot be cast to class 
org.apache.hudi.metrics.DistributedRegistry
       at 
org.apache.hudi.metadata.SparkHoodieBackedTableMetadataWriter.initRegistry(SparkHoodieBackedTableMetadataWriter.java:122)
   ```
   
   Reproduced on master with the order from the issue 
(`-Dsurefire.runOrder=reversealphabetical 
-Dtest='TestMetricsReporter,TestMetadataTableWithSparkSQL'`, one fork): 
TestMetricsReporter passes, then both TestMetadataTableWithSparkSQL tests fail 
with this exception.
   
   The registry those lines create is never read: HUDI-7391 (#10635) moved 
`HoodieMetadataMetrics` to `Metrics`, and the local `registry` variable has 
been unused since.
   
   ### Summary and Changelog
   
   The Spark metadata writers no longer create a registry they do not use, and 
a registry class mismatch fails with a message naming the key and both classes 
instead of returning a registry of the wrong class.
   
   - `SparkHoodieBackedTableMetadataWriter`, 
`SparkHoodieBackedTableMetadataWriterTableVersionSix`: `initRegistry` keeps 
only the `HoodieMetadataMetrics` setup, as in 
`JavaHoodieBackedTableMetadataWriter` and 
`FlinkHoodieBackedTableMetadataWriter`. Four unused imports are removed.
   - `Registry#getRegistryOfClass`: throws `HoodieException("Registry 
<table>::<name> already exists with class <existing>, but class <requested> was 
requested")` instead of logging and returning the existing registry.
   - Callers of `getRegistryOfClass`, checked one by one:
     - `HoodieEngineContext#getMetricRegistry` (LocalRegistry): no change.
     - `HoodieSparkEngineContext#getMetricRegistry` casts the result to 
`DistributedRegistry`. Over a table registry created earlier as a 
`LocalRegistry` it failed with `ClassCastException`; it now fails with the 
`HoodieException` above.
     - `DistributedRegistryUtil#createWrapperFileSystemRegistries` with 
executor metrics off asked for the class of the process-wide 
`HoodieWrapperFileSystem` registry (always `LocalRegistry`), and got back the 
table's `DistributedRegistry` when a writer of the same table with executor 
metrics on had created one in this JVM. The strict check would fail that write 
client, so this path now reuses the table's registry of any class and creates a 
`LocalRegistry` otherwise. It no longer creates the process-wide 
`HoodieWrapperFileSystem` and `HoodieWrapperFileSystemMetaFolder` registries, 
which nothing reads.
     - `Registry#getRegistry(name)` and `getRegistry(name, clazz)`: the 
remaining callers (timeline service, tests) use one class per name.
   - Tests:
     - `TestRegistry#testGetRegistryOfClassRejectsAnotherClass`: the same class 
returns the same instance; another class throws with the exact message and 
leaves the registry in place. On master nothing is thrown.
     - 
`TestDistributedRegistry#testEngineContextRejectsRegistryOfAnotherClass`: 
`HoodieSparkEngineContext#getMetricRegistry` over an existing `LocalRegistry` 
throws the `HoodieException`. On master it throws `ClassCastException`.
     - 
`TestDistributedRegistry#testWrapperFileSystemRegistriesReuseTheTableRegistryOfAnyClass`:
 after a config with executor metrics on, a config with them off reuses both 
`DistributedRegistry` instances of the table. It passes on master; with the 
strict check and the old `DistributedRegistryUtil` it fails with "Registry 
testWrapperFileSystemRegistries::HoodieWrapperFileSystem already exists with 
class org.apache.hudi.metrics.DistributedRegistry, but class 
org.apache.hudi.common.metrics.LocalRegistry was requested".
   
   ### Impact
   
   A JVM that runs Spark metadata writers both with and without executor 
metrics no longer fails with `ClassCastException`, whatever order its test 
classes or writers run in. `Registry#getRegistryOfClass` now throws on a class 
mismatch. In Hudi that is reachable only through 
`HoodieSparkEngineContext#getMetricRegistry`, when a writer with executor 
metrics on follows a writer of the same table with them off in one JVM; that 
case already failed, with a `ClassCastException`. No metric that a reporter 
publishes changes: the removed registries were never written to.
   
   ### Risk Level
   
   low. The removed code had no reader. The strict check changes one failure 
from `ClassCastException` to `HoodieException`, and the only lenient caller in 
Hudi keeps its behavior.
   
   ### Documentation Update
   
   none
   
   ### Contributor's checklist
   
   - [x] Read through [contributor's 
guide](https://hudi.apache.org/contribute/how-to-contribute)
   - [x] Enough context is provided in the sections above
   - [x] Adequate tests were added if applicable
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

  • [PR] fix(metrics): fail on a re... Indian-boult (via GitHub) via commits

Reply via email to