Caleb Rackliffe created CASSANDRA-21516:
-------------------------------------------
Summary: SAI Component Checksum Validation Should be Segment-Aware
Key: CASSANDRA-21516
URL: https://issues.apache.org/jira/browse/CASSANDRA-21516
Project: Apache Cassandra
Issue Type: Bug
Components: Feature/SAI
Reporter: Caleb Rackliffe
Assignee: Caleb Rackliffe
SAI index components are buffered in memory and then flushed to disk in
segments during index builds and compactions. Each segment writes its own
checksum. However, the underlying checksum validation logic assumes single
segments, a major oversight. This is fairly easy to reproduce. Throw the
following test in {{SegmentFlushTest}}:
{noformat}
@Test
public void multiSegmentBalancedTreeFailsPerIndexValidation() throws IOException
{
Path tmpDir = Files.createTempDirectory("SegmentFlushTest");
IndexDescriptor indexDescriptor = IndexDescriptor.create(new Descriptor(new
File(tmpDir.toFile()), "ks", "cf", new SequenceBasedSSTableId(1)),
Murmur3Partitioner.instance, SAITester.EMPTY_COMPARATOR);
ColumnMetadata column = ColumnMetadata.regularColumn("sai", "internal",
"ts", TimestampType.instance);
StorageAttachedIndex index = SAITester.createMockIndex(column);
SSTableIndexWriter writer = new SSTableIndexWriter(indexDescriptor, index,
V1OnDiskFormat.SEGMENT_BUILD_MEMORY_LIMITER, () -> true);
List<DecoratedKey> keys = Arrays.asList(dk("1"), dk("2"));
Collections.sort(keys);
writer.addRow(SAITester.TEST_FACTORY.create(keys.get(0)), createRow(column,
TimestampType.instance.decompose(new Date(1_000L))), 0L);
writer.addRow(SAITester.TEST_FACTORY.create(keys.get(1)), createRow(column,
TimestampType.instance.decompose(new Date(2_000L))),
SegmentBuilder.LAST_VALID_SEGMENT_ROW_ID + 1);
writer.complete(Stopwatch.createStarted());
// TODO: Should not throw, but does...
indexDescriptor.validatePerIndexComponents(index.termType(),
index.identifier(), IndexValidation.CHECKSUM, true, true);
}
{noformat}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
