[SPARK-53792] Fix rocksdbPinnedBlocksMemoryUsage when bounded memory … #52527
Conversation
| // In unbounded memory mode, each instance has its own memory | ||
|
|
||
| // Question: Should we return 0L if the instance is not registed in instanceMemoryMap? | ||
| // but when it is not registed? |
There was a problem hiding this comment.
If not registered, we currently just return the totalMemory passed in. Why do you think 0L is better?
| requestedPinnedUsage: Long): Long = { | ||
| if (!instanceMemoryMap.containsKey(uniqueId)) { | ||
| // Instance not registered, return 0 | ||
| return 0L |
There was a problem hiding this comment.
why not just return the requestedPinnedUsage instead of 0L?
| if (numBoundedInstances > 0) { | ||
| globalPinnedUsage / numBoundedInstances | ||
| } else { | ||
| 0L |
There was a problem hiding this comment.
when would this happen? Given that we already found a boundedMemory instance for this when we did instanceMemoryMap.get(uniqueId) above. So don't think this is necessary.
| // Use RocksDBMemoryManager to calculate the memory usage accounting | ||
| val memoryUsage = RocksDBMemoryManager.getInstanceMemoryUsage(instanceUniqueId, getMemoryUsage) | ||
|
|
||
| val requestedPinnedBlocksMemUsage = getDBProperty("rocksdb.block-cache-pinned-usage") |
There was a problem hiding this comment.
do we actually need the getDBProperty("rocksdb.block-cache-pinned-usage") here? Because in unbounded mode, the getDBProperty("rocksdb.block-cache-pinned-usage") will be equal to lruCache.getPinnedUsage()
|
|
||
| def listFiles(file: String): Seq[File] = listFiles(new File(file)) | ||
|
|
||
| test("SPARK-53792: RocksDBMemoryManager getInstancePinnedBlocksMemUsage") { |
There was a problem hiding this comment.
nit: move this closer to the other memory manager test case
…usage is enabled
What changes were proposed in this pull request?
Changing the way we calculate
pinnedBlocksMemUsagewhen an instance has bounded memory enabled.Before the change, we always report back
getDBProperty("rocksdb.block-cache-pinned-usage")which returns that size of pinned block requested by an instance. This is not accurate when instances share the same cached, because instances might share the same pinned blockAfter this change, when isMemoryBounded is enabled for an instance, we call
lruCache.getPinnedUsage()to get the total memory usage of SHARED pinned blocks, and divide the global usage with the number of IS_MEMORY_BOUNDED instances. This is because RocksDBMemoryManager will return the same cache for all the instance that has isMemoryBounded = trueUnit test:
build/mvn -Dtest=none -DwildcardSuites=org.apache.spark.sql.execution.streaming.state.RocksDBSuite test
Why are the changes needed?
See above for why this is a bug.
This fix prevents us from over-reporting
pinnedBlocksMemUsagewhen isBoundedMemory is enabled for an instanceDoes this PR introduce any user-facing change?
No
How was this patch tested?
See added unit tests on the correctness of the calculation.
Was this patch authored or co-authored using generative AI tooling?
Yes. Generated-by cursor and 'claude-4-sonnet'