[MAPREDUCE-6685] LocalDistributedCacheManager can have overlapping filenames

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Duplicate
• Affects Version/s: 3.0.0-alpha1
• Assignee: Ray Chiang
• Reporter: Ray Chiang
• Created: 26/Apr/16 18:46
• Updated: 24/Aug/16 16:04
• Resolved: 01/Jun/16 05:34

Description

LocalDistributedCacheManager has this setup:

AtomicLong uniqueNumberGenerator = new AtomicLong(System.currentTimeMillis());

to create this temporary filename:

new FSDownload(localFSFileContext, ugi, conf, new Path(destPath, Long.toString(uniqueNumberGenerator.incrementAndGet())), resource);

when using LocalJobRunner. When two or more start on the same machine, then it's possible to end up having the same timestamp or a large enough overlap that two successive timestamps may not be sufficiently far apart.

Given the assumptions:

1) Assume timestamp is the same. Then the most common starting random seed will be the same.
2) Process ID will very likely be unique, but will likely be close in value.
3) Thread ID is not guaranteed to be unique.

A unique ID based on PID as a seed (in addition to the timestamp) should be a better unique identifier for temporary filenames.

Issue Links

• duplicates: MAPREDUCE-6441 Improve temporary directory name generation in LocalDistributedCacheManager for concurrent processes
• relates to: MAPREDUCE-6766 Concurrent local job failures due to uniqueNumberGenerator = new AtomicLong(System.currentTimeMillis())