[OAK-2621] Too many reads for child nodes

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.0
• Fix Version/s: 1.7.11, 1.8.0
• Component/s: mongomk
• Labels: performance
• Epic Link: Improve DocumentMK performance
• Assignee: Marcel Reutegger
• Reporter: Marcel Reutegger
• Created: 12/Mar/15 20:17
• Updated: 13/Nov/17 13:55
• Resolved: 24/Oct/17 11:19

Description

The DocumentNodeStore issues a lot of reads when sibling nodes are deleted, which are also index with a property index.

The following calls will become a hotspot:

	at org.apache.jackrabbit.oak.plugins.document.mongo.MongoDocumentStore.query(MongoDocumentStore.java:406)
	at org.apache.jackrabbit.oak.plugins.document.DocumentNodeStore.readChildDocs(DocumentNodeStore.java:846)
	at org.apache.jackrabbit.oak.plugins.document.DocumentNodeStore.readChildren(DocumentNodeStore.java:788)
	at org.apache.jackrabbit.oak.plugins.document.DocumentNodeStore.getChildren(DocumentNodeStore.java:753)
	at org.apache.jackrabbit.oak.plugins.document.DocumentNodeState.getChildNodeCount(DocumentNodeState.java:194)
	at org.apache.jackrabbit.oak.plugins.memory.ModifiedNodeState.getChildNodeCount(ModifiedNodeState.java:198)
	at org.apache.jackrabbit.oak.plugins.memory.MutableNodeState.getChildNodeCount(MutableNodeState.java:265)
	at org.apache.jackrabbit.oak.plugins.memory.MemoryNodeBuilder.getChildNodeCount(MemoryNodeBuilder.java:293)
	at org.apache.jackrabbit.oak.plugins.index.property.strategy.ContentMirrorStoreStrategy.prune(ContentMirrorStoreStrategy.java:456)

I think the code triggering this issue is in ModifiedNodeState.getChildNodeCount(). It keeps track of already deleted children and requests max += deleted. The actual max is always 1 as requested from ContentMirrorStoreStrategy.prune(), but as more nodes get deleted, the higher max gets passed to DocumentNodeState.getChildNodeCount(). The DocumentNodeStore then checks if it has the children in the cache, only to find out the cache entry has too few entries and it needs to fetch one more.

It would be best to have a minimum number of child nodes to fetch from MongoDB in this case. E.g. when NodeState.getChildNodeEntries() is called, the DocumentNodeState fetches 100 children.