[CASSANDRA-9045] Deleted columns are resurrected after repair in wide rows

• Type: Bug
• Status: Resolved
• Priority: Urgent
• Resolution: Cannot Reproduce
• Severity: Critical
• Assignee: Marcus Eriksson
• Reporter: Roman Tkachenko

Authors

Marcus Eriksson

• Tester: Philip Thompson
• Created: 26/Mar/15 17:25
• Updated: 16/Apr/19 09:31
• Resolved: 24/Aug/15 15:51

Description

Hey guys,

After almost a week of researching the issue and trying out multiple things with (almost) no luck I was suggested (on the user@cass list) to file a report here.

Setup

Cassandra 2.0.13 (we had the issue with 2.0.10 as well and upgraded to see if it goes away)
Multi datacenter 12+6 nodes cluster.

Schema

cqlsh> describe keyspace blackbook;

CREATE KEYSPACE blackbook WITH replication = {
  'class': 'NetworkTopologyStrategy',
  'IAD': '3',
  'ORD': '3'
};

USE blackbook;

CREATE TABLE bounces (
  domainid text,
  address text,
  message text,
  "timestamp" bigint,
  PRIMARY KEY (domainid, address)
) WITH
  bloom_filter_fp_chance=0.100000 AND
  caching='KEYS_ONLY' AND
  comment='' AND
  dclocal_read_repair_chance=0.100000 AND
  gc_grace_seconds=864000 AND
  index_interval=128 AND
  read_repair_chance=0.000000 AND
  populate_io_cache_on_flush='false' AND
  default_time_to_live=0 AND
  speculative_retry='99.0PERCENTILE' AND
  memtable_flush_period_in_ms=0 AND
  compaction={'class': 'LeveledCompactionStrategy'} AND
  compression={'sstable_compression': 'LZ4Compressor'};

Use case

Each row (defined by a domainid) can have many many columns (bounce entries) so rows can get pretty wide. In practice, most of the rows are not that big but some of them contain hundreds of thousands and even millions of columns.

Columns are not TTL'ed but can be deleted using the following CQL3 statement:

delete from bounces where domainid = 'domain.com' and address = '[e-mail]';

All queries are performed using LOCAL_QUORUM CL.

Problem

We weren't very diligent about running repairs on the cluster initially, but shorty after we started doing it we noticed that some of previously deleted columns (bounce entries) are there again, as if tombstones have disappeared.

I have run this test multiple times via cqlsh, on the row of the customer who originally reported the issue:

• delete an entry
• verify it's not returned even with CL=ALL
• run repair on nodes that own this row's key
• the columns reappear and are returned even with CL=ALL

I tried the same test on another row with much less data and everything was correctly deleted and didn't reappear after repair.

Other steps I've taken so far

Made sure NTP is running on all servers and clocks are synchronized.

Increased gc_grace_seconds to 100 days, ran full repair (on the affected keyspace) on all nodes, then changed it back to the default 10 days again. Didn't help.

Performed one more test. Updated one of the resurrected columns, then deleted it and ran repair again. This time the updated version of the column reappeared.

Finally, I noticed these log entries for the row in question:

INFO [ValidationExecutor:77] 2015-03-25 20:27:43,936 CompactionController.java (line 192) Compacting large row blackbook/bounces:4ed558feba8a483733001d6a (279067683 bytes) incrementally

Figuring it may be related I bumped "in_memory_compaction_limit_in_mb" to 512MB so the row fits into it, deleted the entry and ran repair once again. The log entry for this row was gone and the columns didn't reappear.

We have a lot of rows much larger than 512MB so can't increase this parameters forever, if that is the issue.

Please let me know if you need more information on the case or if I can run more experiments.

Thanks!

Roman