Friday, November 05, 2010

A Fix for "Exceeded MAX_FAILED_UNIQUE_FETCHES" in Hadoop

In a project I'm currently working on, we're moving a bunch of our back-end processing to Hadoop.  We started a two-node cluster: one master, one slave.  That seemed to work fine.  Then, we went to four nodes, and about the same time I was testing out a new Hadoop job.  The (single) reducer was hanging with this somewhat cryptic message:
Shuffle Error: Exceeded MAX_FAILED_UNIQUE_FETCHES; bailing-out

I went out to the slave node and looked through the job logs, and I could see that it was timing out trying to transfer data from one of the other slave nodes - an upstream mapper node. Upon closer scrutiny of the the log file I realized that Hadoop was trying to transfer from the other slave's public IP address, which is behind a firewall that blocks public access.

Key take-away number one: when you're just starting out with Hadoop, if you only have one slave, you've only demonstrated one real communication path: master-to-slave. Your cluster isn't doing any slave-to-slave transfers because everything was on the one slave. Also, our initial job had no reducer, so it ran fine on the new, 4-node cluster because it was still only master-slave communication.

For some reason, the mapper slave was advertising the location of the map output data via its public IP address. My first attempt at fixing this problem involved the dfs.datanode.dns.interface configuration parameter (and it's mapred equivalent). This tells Hadoop that when a process (mapred or dfs) wants to figure out it's host name, use the IP address associated with the given interface. (You could even have dfs and mapred using separate interfaces for additional throughput.)

This failed for me because I had one interface with two addresses, not two interfaces. I dug through the Hadoop DNS code ( - God, I love open-source: you can just look for yourself) and saw that if there is one interface, the code loops through the IP addresses and performs reverse DNS lookups and takes the first successful result. I was fortunate in that the private IP address was coming up first in that enumeration of the IPs on the interface, but it still wasn't working. I talked to our system admin/configuration guru. It turns out that our hosting provider doesn't provide reverse DNS for those private IP addresses. We could have set up our own DNS server for just these reverse lookups, but there was a brute-force option available to us.

You can bypass all of Hadoop's efforts to automatically figure out the slave's host name by specifying the parameter in the configuration files. If that is set, Hadoop will just take your word for it and use the name you provide. Now, in theory, this might be onerous - it means you have a different configuration file per-slave. However, our cluster is configured and maintained via Puppet. So, our puppet master just tweaked his Puppet script, and we never looked back.

Take-away number two: Exceeded MAX_FAILED_UNIQUE_FETCHES could mean a simple connectivity problem. I'm sure there are other possible causes, but an inability to connect between slaves is comparatively simple to troubleshoot.