[MAPREDUCE-3234] Locality scheduling broken due to mismatch between IPs and hosts

• Type: Bug
• Status: Resolved
• Priority: Blocker
• Resolution: Duplicate
• Affects Version/s: 0.23.0
• Component/s: mrv2
• Assignee: Unassigned
• Reporter: Todd Lipcon
• Created: 20/Oct/11 16:58
• Updated: 22/Oct/11 06:20
• Resolved: 22/Oct/11 06:20

Description

I noticed that, on a single-rack cluster, I wasn't getting hardly any data locality. The issue appears to be the code in RMContainerAllocator which changes the resource requests to use IP addresses instead of hostnames:

           //host comes from data splitLocations which are hostnames. Containers
           // use IP addresses.
           //TODO Temporary fix for locality. Use resolvers from h-common. 
           // Cache to make this more efficient ?

However, at least on my cluster, the resource manager sees node resources as hostnames, not IPs. Removing this code fixed data locality.

Issue Links

• is duplicated by: MAPREDUCE-2693 NPE in AM causes it to lose containers which are never returned back to RM