[
https://issues.apache.org/jira/browse/CASSANDRA-20709?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Michael Semb Wever updated CASSANDRA-20709:
-------------------------------------------
Fix Version/s: 5.1
(was: 6.x)
> SAI queries can miss a memtable/sstable if flush happens during query
> ---------------------------------------------------------------------
>
> Key: CASSANDRA-20709
> URL: https://issues.apache.org/jira/browse/CASSANDRA-20709
> Project: Apache Cassandra
> Issue Type: Bug
> Components: Feature/SAI
> Reporter: Michael Marshall
> Assignee: Michael Marshall
> Priority: Normal
> Fix For: 5.0.5, 5.1
>
> Attachments: ci_summary-1.html, ci_summary.html,
> result_details.tar-1.gz, result_details.tar.gz
>
> Time Spent: 1h 50m
> Remaining Estimate: 0h
>
> The IndexSearchResultIterator takes sstable index references to query and
> then gets the memtable index references within the searchMemtableIndexes
> method. This ordering is not safe because a concurrent flush can lead to
> missing rows that were in the memtable when the QueryView was created but in
> an sstable when the searchMemtableIndexes is called.
> The solution is to get the memtables before getting the sstables. I propose
> that we add the memtables to the QueryView to make the abstraction a bit
> simpler to follow. We already do the correct thing in hybrid ANN queries, but
> my proposed solution would simplify that code too by storing the memtable
> index references in the QueryView.
> Here is a test to reproduce the bug:
> {code:java}
> public class FlushIndexWhileQueryingTest extends SAITester
> {
> @Test
> public void testFlushDuringEqualityQuery() throws Throwable
> {
> createTable("CREATE TABLE %s (k text PRIMARY KEY, x int)");
> createIndex("CREATE CUSTOM INDEX ON %s(x) USING
> 'StorageAttachedIndex'");
> waitForTableIndexesQueryable();
> execute("INSERT INTO %s (k, x) VALUES (?, ?)", "a", 0);
> execute("INSERT INTO %s (k, x) VALUES (?, ?)", "b", 0);
> execute("INSERT INTO %s (k, x) VALUES (?, ?)", "c", 1);
> // We use a barrier to trigger flush at precisely the right time
> InvokePointBuilder initialInvokePoint =
> InvokePointBuilder.newInvokePoint()
>
> .onClass(IndexSearchResultIterator.class)
> .onMethod("build")
> .atEntry();
> Injections.Barrier initialBarrier =
> Injections.newBarrier("pause_query", 2, false)
> .add(initialInvokePoint)
> .build();
> InvokePointBuilder secondInvokePoint =
> InvokePointBuilder.newInvokePoint()
>
> .onClass(MemtableIndexManager.class)
>
> .onMethod("searchMemtableIndexes")
> .atEntry();
> Injections.Barrier secondBarrier =
> Injections.newBarrier("resume_query", 2, false)
> .add(secondInvokePoint)
> .build();
> Injections.inject(initialBarrier, secondBarrier);
> // Flush in a separate thread to allow the query to run concurrently
> ForkJoinPool.commonPool().submit(() -> {
> try
> {
> initialBarrier.arrive();
> flush();
> secondBarrier.arrive();
> }
> catch (InterruptedException t)
> {
> throw new RuntimeException(t);
> }
> });
> assertRowCount(execute("SELECT k FROM %s WHERE x = 0"), 2);
> assertEquals("Confirm that we hit the barrier (helps in case method
> name changed)", 0, initialBarrier.getCount());
> assertEquals("Confirm that we hit the barrier (helps in case method
> name changed)", 0, secondBarrier.getCount());
> }
> }
> {code}
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
