Details
-
Bug
-
Status: Patch Available
-
Normal
-
Resolution: Unresolved
-
None
-
Correctness - Transient Incorrect Response
-
Normal
-
Normal
-
User Report
-
All
-
None
Description
Bug Description
When using Cassandra 3.11.16 to load legacy data from 2.2.10, I noticed that the byte representation of the column identifier is incorrect.
The legacy data contain two tables, and the schema is as follows.
cqlsh> desc test.alpha ; CREATE TABLE test.alpha ( key text PRIMARY KEY, foo text ) WITH COMPACT STORAGE AND bloom_filter_fp_chance = 0.01 AND caching = '{"keys":"ALL", "rows_per_partition":"NONE"}' AND comment = '' AND compaction = {'min_threshold': '4', 'class': 'org.apache.cassandra.db.compaction.SizeTieredCompactionStrategy', 'max_threshold': '32'} AND compression = {'sstable_compression': 'org.apache.cassandra.io.compress.LZ4Compressor'} AND dclocal_read_repair_chance = 0.1 AND default_time_to_live = 0 AND gc_grace_seconds = 864000 AND max_index_interval = 2048 AND memtable_flush_period_in_ms = 0 AND min_index_interval = 128 AND read_repair_chance = 0.0 AND speculative_retry = 'NONE'; cqlsh> DESC test.foos ; CREATE TABLE test.foos ( key text PRIMARY KEY, "666f6f" text ) WITH COMPACT STORAGE AND bloom_filter_fp_chance = 0.01 AND caching = '{"keys":"ALL", "rows_per_partition":"NONE"}' AND comment = '' AND compaction = {'min_threshold': '4', 'class': 'org.apache.cassandra.db.compaction.SizeTieredCompactionStrategy', 'max_threshold': '32'} AND compression = {'sstable_compression': 'org.apache.cassandra.io.compress.LZ4Compressor'} AND dclocal_read_repair_chance = 0.1 AND default_time_to_live = 0 AND gc_grace_seconds = 864000 AND max_index_interval = 2048 AND memtable_flush_period_in_ms = 0 AND min_index_interval = 128 AND read_repair_chance = 0.0 AND speculative_retry = 'NONE'; CREATE INDEX idx_foo ON test.foos ("666f6f");
There exists a column in test.foo with name = "666f6f", the corresponding byte representation should be Hex(666f6f) == 363636663666. However, when 3.11.15 loads the data and creating the column, if we check the value in byteBuffer, the it still stores "666f6f".
// src/java/org/apache/cassandra/schema/SchemaKeyspace.java public static ColumnDefinition createColumnFromRow(UntypedResultSet.Row row, Types types) { String keyspace = row.getString("keyspace_name"); String table = row.getString("table_name"); ColumnDefinition.Kind kind = ColumnDefinition.Kind.valueOf(row.getString("kind").toUpperCase()); int position = row.getInt("position"); ClusteringOrder order = ClusteringOrder.valueOf(row.getString("clustering_order").toUpperCase()); AbstractType<?> type = parse(keyspace, row.getString("type"), types); if (order == ClusteringOrder.DESC) type = ReversedType.getInstance(type); // Injected log to check byteBuffer value logger.info(String.format("column_name = %s, column_name_bytes = %s" , row.getString("column_name"), new String(row.getBytes("column_name_bytes").array(), StandardCharsets.UTF_8))); ColumnIdentifier name = new ColumnIdentifier(row.getBytes("column_name_bytes"), row.getString("column_name")); return new ColumnDefinition(keyspace, table, name, type, position, kind); }
Logs
INFO [main] 2023-08-07 02:21:53,762 SchemaKeyspace.java:1136 - column_name = 666f6f, column_name_bytes = foo
It should be : column_name_bytes = 666f6f
INFO [main] 2023-08-07 02:21:53,722 StorageService.java:773 - Populating token metadata from system tables
INFO [main] 2023-08-07 02:21:53,736 StorageService.java:780 - Token metadata: Normal Tokens:
localhost/127.0.0.1:[95610762103941981519101009083045058398]INFO [main] 2023-08-07 02:21:53,756 SchemaKeyspace.java:1136 - column_name = column1, column_name_bytes = column1
INFO [main] 2023-08-07 02:21:53,756 SchemaKeyspace.java:1136 - column_name = foo, column_name_bytes = foo
INFO [main] 2023-08-07 02:21:53,756 SchemaKeyspace.java:1136 - column_name = key, column_name_bytes = key
INFO [main] 2023-08-07 02:21:53,756 SchemaKeyspace.java:1136 - column_name = value, column_name_bytes = value
INFO [main] 2023-08-07 02:21:53,762 SchemaKeyspace.java:1136 - column_name = 666f6f, column_name_bytes = foo // Incorrect!
INFO [main] 2023-08-07 02:21:53,762 SchemaKeyspace.java:1136 - column_name = column1, column_name_bytes = column1
Reproduce Method
Method1: load attached data file
I have attached the data tar file, if start up Cassandra 3.11.16 with it and inject a the log statement to print out the buffer value, we can notice that the value is incorrect in the log.
Method2: Generate data from the old version (2.1.19)
Start up Cassandra 2.1.19 version, using bin/cassandra-cli to construct the following data
create keyspace test with strategy_options = {replication_factor:1} and placement_strategy = 'org.apache.cassandra.locator.SimpleStrategy'; use test; create column family alpha with column_type = 'Standard' and comparator = 'UTF8Type' and key_validation_class = 'UTF8Type' and column_metadata = [{column_name: 'foo', validation_class: 'UTF8Type'}]; create column family foos with column_type = 'Standard' and comparator = 'BytesType' and key_validation_class = 'UTF8Type' and column_metadata = [{column_name: '666f6f', validation_class: 'UTF8Type'}];
Then load the data using 3.0.16 with the log statements injected, and you will encounter the logs mentioned above.
Thoughts
This is a transient bug which won't lead to exceptions or error logs. But the incorrect byte representation might lead to some issues.
This bug shares the same triggering method with CASSANDRA-14468. I believe this bug also shares the same root cause as CASSANDRA-14468. In CASSANDRA-14468, the incorrect byte representation could lead to an upgrade exception. It was partially fixed by avoiding the intern of ColumnIdentifier (which makes this bug transient). But the real root cause remains, and it's still possible to cause other problems.
Root Cause
== TL, DR ==
The new version (3.11.x) uses the comparator of the table to create ColumnIdentifier. If the old version table comparator is "BytesType", the new version assumes that the old regular column name is already in bytes format and thus it directly puts the string in the ByteBuffer.
This generates a ColumnIdentifier whose text and bytes are inconsistent
- ColumnIdentifier: {text = "666f6f", bytes = "666f6f"}.
- The correct ColumnIdentify should be {text = "666f6f", bytes = "363636663666"}.
== Full Version ==
In more detail, this is how it happens
1. In code, it tries to intern the ColumnIdentifer using the comparator. The comparator is BytesType and column name is "666f6f".
ColumnIdentifier.getInterned(comparator.fromString(row.getString("column_name")), comparator);
2. comparator.fromString(row.getString("column_name") directly returns a ByteBuffer containing "666f6f". The code below directly assumes that the source is in bytes format.
// BytesType.java public ByteBuffer fromString(String source) { try { return ByteBuffer.wrap(Hex.hexToBytes(source)); } catch (NumberFormatException e) { logger.info("running into MarshalException"); throw new MarshalException(String.format("cannot parse '%s' as hex bytes", source), e); } }
3. ColumnIdentifier.getInterned uses the returned ByteBuffer to create a new ColumnIdentifier object.
text = "666f6f" bytes = "666f6f"
Fix
This can be fixed in a simple way. If the comparator type is BytesType, we shouldn't use comparator to get the ByteBuffer. Instead, consider it as String format and directly use ByteBufferUtil.bytes to get the bytes.
Then the generated column identifier will be {text = "666f6f", bytes = "363636663666"}
The patch is here.
Attachments
Attachments
Issue Links
- links to