[CASSANDRA-14647] Reading cardinality from Statistics.db failed

• Type: Bug
• Status: Resolved
• Priority: Low
• Resolution: Fixed
• Fix Version/s: 4.0-alpha1, 4.0
• Component/s: Observability/Metrics

Environment

Clients are doing only writes with Local One, cluster consist of 3 regions with RF3.
Storage is configured wth jbod/XFS on 10 x 1Tb disks
IOPS limit for each disk 500 (total 5000 iops)
Bandwith for each disk 60mb/s (600 total)
OS is Debian linux.

• Severity: Low
• Since Version: 2.1 beta1

Test and Documentation Plan

adds test case

• Assignee: n.v.harikrishna
• Reporter: Vitali Djatsuk

Authors

n.v.harikrishna

Reviewers

Marcus Eriksson

• Created: 16/Aug/18 10:26
• Updated: 15/May/20 08:02
• Resolved: 09/May/19 12:37

Description

There is some issue with sstable metadata which is visible in system.log, the messages says:

WARN  [Thread-6] 2018-07-25 07:12:47,928 SSTableReader.java:249 - Reading cardinality from Statistics.db failed for /opt/data/disk5/data/keyspace/table/mc-big-Data.db.

Although there is no such file.

The message has appeared after i've changed the compaction strategy from SizeTiered to Leveled. Compaction strategy has been changed region by region (total 3 regions) and it has coincided with the double client write traffic increase.
I have tried to run nodetool scrub to rebuilt the sstable, but that does not fix the issue.

So very hard to define the steps to reproduce, probably it will be:

1. run stress tool with write traffic
2. under load change compaction strategy from SireTiered to Leveled for the bunch of hosts
3. add more write traffic

Reading the code it is said that if this metadata is broken, then "estimating the keys will be done using index summary". 
https://github.com/apache/cassandra/blob/cassandra-3.0.17/src/java/org/apache/cassandra/io/sstable/format/SSTableReader.java#L247