[CASSANDRA-12083] Race condition during system.roles column family creation

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 3.0.8, 3.8
• Component/s: Legacy/Core
• Severity: Normal
• Assignee: Sam Tunnicliffe
• Reporter: Sharvanath Pathak

Authors

Sam Tunnicliffe

Reviewers

Aleksey Yeschenko

• Created: 23/Jun/16 20:54
• Updated: 16/Apr/19 09:30
• Resolved: 27/Jun/16 15:11

Description

There is an issue where Cassandra fails with the following exception on startup:

DEBUG [InternalResponseStage:2] 2016-06-20 09:43:00,651 Schema.java:465 - Adding org.apache.cassandra.config.CFMetaData@2882b66d[cfId=5bc52802-de25-35ed-aeab-188eecebb090,ksName=system_auth,cfName=roles,flags=[COMPOUND],params=TableParams{comment=role definitions, read_repair_chance=0.0, dclocal_read_repair_chance=0.0, bloom_filter_fp_chance=0.01, crc_check_chance=1.0, gc_grace_seconds=7776000, default_time_to_live=0, memtable_flush_period_in_ms=3600000, min_index_interval=128, max_index_interval=2048, speculative_retry=99PERCENTILE, caching={'keys' : 'ALL', 'rows_per_partition' : 'NONE'}, compaction=CompactionParams{class=org.apache.cassandra.db.compaction.SizeTieredCompactionStrategy, options={max_threshold=32, min_threshold=4}}, compression=org.apache.cassandra.schema.CompressionParams@e6e0212, extensions={}},comparator=comparator(),partitionColumns=[[] | [can_login is_superuser salted_hash member_of]],partitionKeyColumns=[ColumnDefinition{name=role, type=org.apache.cassandra.db.marshal.UTF8Type, kind=PARTITION_KEY, position=0}],clusteringColumns=[],keyValidator=org.apache.cassandra.db.marshal.UTF8Type,columnMetadata=[ColumnDefinition{name=role, type=org.apache.cassandra.db.marshal.UTF8Type, kind=PARTITION_KEY, position=0}, ColumnDefinition{name=salted_hash, type=org.apache.cassandra.db.marshal.UTF8Type, kind=REGULAR, position=-1}, ColumnDefinition{name=member_of, type=org.apache.cassandra.db.marshal.SetType(org.apache.cassandra.db.marshal.UTF8Type), kind=REGULAR, position=-1}, ColumnDefinition{name=can_login, type=org.apache.cassandra.db.marshal.BooleanType, kind=REGULAR, position=-1}, ColumnDefinition{name=is_superuser, type=org.apache.cassandra.db.marshal.BooleanType, kind=REGULAR, position=-1}],droppedColumns={},triggers=[],indexes=[]] to cfIdMap
INFO  [InternalResponseStage:2] 2016-06-20 09:43:00,653 ColumnFamilyStore.java:381 - Initializing system_auth.roles
DEBUG [InternalResponseStage:2] 2016-06-20 09:43:00,664 MigrationManager.java:556 - Gossiping my schema version c2a2bb4f-7d31-3fb8-a216-00b41a643650
DEBUG [InternalResponseStage:1] 2016-06-20 09:43:00,669 ColumnFamilyStore.java:831 - Enqueuing flush of keyspaces: 1566 (0%) on-heap, 0 (0%) off-heap
DEBUG [MemtableFlushWriter:2] 2016-06-20 09:43:00,669 Memtable.java:372 - Writing Memtable-keyspaces@650010305(0.437KiB serialized bytes, 3 ops, 0%/0% of on/off-heap limit)
ERROR [main] 2016-06-20 09:43:00,670 CassandraDaemon.java:692 - Exception encountered during startup
java.lang.IllegalArgumentException: Unknown CF 5bc52802-de25-35ed-aeab-188eecebb090
        at org.apache.cassandra.db.Keyspace.getColumnFamilyStore(Keyspace.java:206) ~[apache-cassandra-3.0.6.jar:3.0.6]
        at org.apache.cassandra.db.Keyspace.getColumnFamilyStore(Keyspace.java:199) ~[apache-cassandra-3.0.6.jar:3.0.6]
        at org.apache.cassandra.cql3.restrictions.StatementRestrictions.<init>(StatementRestrictions.java:168) ~[apache-cassandra-3.0.6.jar:3.0.6]
        at org.apache.cassandra.cql3.statements.SelectStatement$RawStatement.prepareRestrictions(SelectStatement.java:874) ~[apache-cassandra-3.0.6.jar:3.0.6]
        at org.apache.cassandra.cql3.statements.SelectStatement$RawStatement.prepare(SelectStatement.java:821) ~[apache-cassandra-3.0.6.jar:3.0.6]
        at org.apache.cassandra.cql3.statements.SelectStatement$RawStatement.prepare(SelectStatement.java:809) ~[apache-cassandra-3.0.6.jar:3.0.6]
        at org.apache.cassandra.auth.CassandraRoleManager.prepare(CassandraRoleManager.java:446) ~[apache-cassandra-3.0.6.jar:3.0.6]
        at org.apache.cassandra.auth.CassandraRoleManager.setup(CassandraRoleManager.java:144) ~[apache-cassandra-3.0.6.jar:3.0.6]
        at org.apache.cassandra.service.StorageService.doAuthSetup(StorageService.java:1084) ~[apache-cassandra-3.0.6.jar:3.0.6]
        at org.apache.cassandra.service.StorageService.joinTokenRing(StorageService.java:1032) ~[apache-cassandra-3.0.6.jar:3.0.6]
        at org.apache.cassandra.service.StorageService.initServer(StorageService.java:755) ~[apache-cassandra-3.0.6.jar:3.0.6]
        at org.apache.cassandra.service.StorageService.initServer(StorageService.java:620) ~[apache-cassandra-3.0.6.jar:3.0.6]
        at org.apache.cassandra.service.CassandraDaemon.setup(CassandraDaemon.java:333) [apache-cassandra-3.0.6.jar:3.0.6]
        at org.apache.cassandra.service.CassandraDaemon.activate(CassandraDaemon.java:551) [apache-cassandra-3.0.6.jar:3.0.6]
        at org.apache.cassandra.service.CassandraDaemon.main(CassandraDaemon.java:679) [apache-cassandra-3.0.6.jar:3.0.6]

A quick glance at code suggests that this race is possible:

In Keyspace.java
    public ColumnFamilyStore getColumnFamilyStore(String cfName)
    {
        UUID id = Schema.instance.getId(getName(), cfName);
        if (id == null)
            throw new IllegalArgumentException(String.format("Unknown keyspace/cf pair (%s.%s)", getName(), cfName));
        return getColumnFamilyStore(id);
    }

    public ColumnFamilyStore getColumnFamilyStore(UUID id)
    {
        ColumnFamilyStore cfs = columnFamilyStores.get(id);
        if (cfs == null)
            throw new IllegalArgumentException("Unknown CF " + id);
        return cfs;
    }

In Schema.java:
 public void addTable(CFMetaData cfm)
    {
        assert getCFMetaData(cfm.ksName, cfm.cfName) == null;

        // Make sure the keyspace is initialized
        Keyspace.open(cfm.ksName);
        // Update the keyspaces map with the updated metadata
        update(cfm.ksName, ks -> ks.withSwapped(ks.tables.with(cfm)));
        // Update the table ID <-> table name map (cfIdMap)
        load(cfm);

        // init the new CF before switching the KSM to the new one
        // to avoid races as in CASSANDRA-10761
        Keyspace.open(cfm.ksName).initCf(cfm, true);
        MigrationManager.instance.notifyCreateColumnFamily(cfm);
    }

    public void load(CFMetaData cfm)
    {
        Pair<String, String> key = Pair.create(cfm.ksName, cfm.cfName);

        if (cfIdMap.containsKey(key))
            throw new RuntimeException(String.format("Attempting to load already loaded table %s.%s", cfm.ksName, cfm.cfName));

        logger.debug("Adding {} to cfIdMap", cfm);
        cfIdMap.put(key, cfm.cfId);
    }

So something can be added to the cfIdMap and not be present in the columnFamilyStores (which is added in Keyspace.open(cfm.ksName).initCf(cfm, true); step).

Issue Links

• relates to: CASSANDRA-11950 dtest failure in consistency_test.TestAvailability.test_network_topology_strategy_each_quorum