[HBASE-14177] Full GC on client may lead to missing scan results

• Type: Bug
• Status: Closed
• Priority: Critical
• Resolution: Won't Fix
• Affects Version/s: 0.98.12, 0.98.13, 1.0.2
• Component/s: Client
• Labels: dataloss
• Assignee: Unassigned
• Reporter: James Estes
• Created: 31/Jul/15 19:05
• Updated: 01/Jul/22 20:52
• Resolved: 17/Jun/16 01:03

Description

After adding a large row, scanning back that row winds up being empty. After a few attempts it will succeed (all attempts over the same data on an hbase getting no other writes).

Looking at logs, it seems this happens when there is memory pressure on the client and there are several Full GCs that happen. Then messages that indicate that region locations are being removed from the local client cache:

2015-07-31 12:50:24,647 [main] DEBUG org.apache.hadoop.hbase.client.HConnectionManager$HConnectionImplementation - Removed 192.168.1.131:50981 as a location of big_row_1438368609944,,1438368610048.880c849594807bdc7412f4f982337d6c. for tableName=big_row_1438368609944 from cache

Blaming the GC may sound fanciful, but if the test is run with -Xms4g -Xmx4g then it always passes on the first scan attempt. Maybe the pause is enough to remove something from the cache, or the client is using weak references somewhere?

More info http://mail-archives.apache.org/mod_mbox/hbase-user/201507.mbox/%3CCAE8tVdnFf%3Dob569%3DfJkpw1ndVWOVTkihYj9eo6qt0FrzihYHgw%40mail.gmail.com%3E

Test used to reproduce:
https://github.com/housejester/hbase-debugging#fullgctest

I tested and had failures in:

0.98.12 client/server
0.98.13 client 0.98.12 server
0.98.13 client/server
1.1.0 client 0.98.13 server
0.98.13 client and 1.1.0 server
0.98.12 client and 1.1.0 server

I tested without failure in:

1.1.0 client/server