Uploaded image for project: 'Cassandra'
  1. Cassandra
  2. CASSANDRA-15077

Dropping column via thrift renders cf unreadable via CQL, leads to missing data



    • Bug
    • Status: Open
    • Normal
    • Resolution: Unresolved
    • None
    • None
    • Correctness - Unrecoverable Corruption / Loss
    • Normal
    • Normal
    • User Report



      We have a lot of thrift/compact storage column families in production. We upgraded to 3.11.4 last week. This week we ran a (thrift) schema change to drop a column from a column family. Our CQL clients immediately starting getting a read error ("ReadFailure: Error from server: code=1300 ...") trying to read the column family. Thrift clients were still able to read the column family.

      We determined restarting the nodes "fixed" CQL reads, so we did that, but soon discovered that we were missing data because cassandra was skipping sstables it didn't like on startup. That exception looked like this:

      INFO  [main] 2019-04-04 20:06:35,676 ColumnFamilyStore.java:430 - Initializing test.test
      ERROR [SSTableBatchOpen:1] 2019-04-04 20:06:35,689 CassandraDaemon.java:228 - Exception in thread Thread[SSTableBatchOpen:1,5,main]
      java.lang.RuntimeException: Unknown column foo during deserialization
              at org.apache.cassandra.db.SerializationHeader$Component.toHeader(SerializationHeader.java:326) ~[apache-cassandra-3.11.4.jar:3.11.4]
              at org.apache.cassandra.io.sstable.format.SSTableReader.open(SSTableReader.java:522) ~[apache-cassandra-3.11.4.jar:3.11.4]
              at org.apache.cassandra.io.sstable.format.SSTableReader.open(SSTableReader.java:385) ~[apache-cassandra-3.11.4.jar:3.11.4]
              at org.apache.cassandra.io.sstable.format.SSTableReader$3.run(SSTableReader.java:570) ~[apache-cassandra-3.11.4.jar:3.11.4]
              at java.util.concurrent.Executors$RunnableAdapter.call(Executors.java:511) ~[na:1.8.0_121]
              at java.util.concurrent.FutureTask.run(FutureTask.java:266) ~[na:1.8.0_121]
              at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142) ~[na:1.8.0_121]
              at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617) [na:1.8.0_121]
              at org.apache.cassandra.concurrent.NamedThreadFactory.lambda$threadLocalDeallocator$0(NamedThreadFactory.java:81) [apache-cassandra-3.11.4.jar:3.11.4]
              at java.lang.Thread.run(Thread.java:745) ~[na:1.8.0_121]


      Below is a list of steps to reproduce the issue. Note that in production our column families were all created via thrift, but I thought it was simpler to create them using CQL for the reproduction script.

      ccm create test -v 3.11.4 -n 1
      ccm updateconf 'start_rpc: true'
      ccm start
      sleep 10
      ccm node1 cqlsh <<SCHEMA
      CREATE KEYSPACE test WITH REPLICATION = {'class': 'SimpleStrategy', 'replication_factor': 1};
      CREATE COLUMNFAMILY test.test (
        id text,
        foo text,
        bar text,
        PRIMARY KEY (id)
      INSERT INTO test.test (id, foo, bar) values ('1', 'hi', 'there');
      pip install pycassa
      python <<DROP_COLUMN
      import pycassa
      sys = pycassa.system_manager.SystemManager('')
      cf = sys.get_keyspace_column_families('test')['test']
      sys.alter_column_family('test', 'test', column_metadata=filter(lambda c: c.name != 'foo', cf.column_metadata))
      # this produces the "ReadFailure: Error from server: code=1300" error
      ccm node1 cqlsh <<QUERY
      select * from test.test;
      ccm node1 stop
      ccm node1 start
      sleep 10
      # this returns 0 rows (i.e. demonstrates missing data)
      ccm node1 cqlsh <<QUERY
      select * from test.test;

      We added the columns back via thrift and restarted cassandra to restore the missing data. Later we realized a secondary index on the affected column family had become out of sync with the data. We assume that was somehow a side effect of running for a period with data missing.




            Unassigned Unassigned
            muir Muir Manders
            1 Vote for this issue
            3 Start watching this issue