[NUTCH-893] DataStore.put() silently loses records when executed from multiple processes

• Type: Bug
• Status: Closed
• Priority: Blocker
• Resolution: Invalid
• Affects Version/s: nutchgora
• Fix Version/s: nutchgora

Environment

Gora HEAD, SqlStore, MySQL 5.1, Ubuntu 10.4 x64, Sun JDK 1.6

• Assignee: Unassigned
• Reporter: Andrzej Bialecki
• Created: 25/Aug/10 21:59
• Updated: 13/Sep/10 14:18
• Resolved: 13/Sep/10 14:18

Description

In order to debug the issue described in NUTCH-879 I created a test to simulate multiple clients appending to webtable (please see the patch), which is the situation that we have in distributed map-reduce jobs.

There are two tests there: one that uses multiple threads within the same JVM, and another that uses single thread in multiple JVMs. Each test first clears webtable (be careful!), and then puts a bunch of pages, and finally counts that all are present and their values correspond to keys. To make things more interesting each execution context (thread or process) closes and reopens its instance of DataStore a few times.

The multithreaded test passes just fine. However, the multi-process test fails with missing keys, as many as 30%.