[CASSANDRA-16807] Weak visibility guarantees of Accumulator lead to failed assertions during digest comparison

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 4.0, 4.0.0, 4.1-alpha1, 4.1
• Component/s: Consistency/Coordination
• Bug Category: Correctness - Consistency
• Severity: Critical
• Complexity: Normal
• Discovered By: Fuzz Test
• Platform: All
• Since Version: 3.0.0

Source Control Link

https://github.com/apache/cassandra/commit/93496e826e7382adf52a99d4df38e73a43f892de

Test and Documentation Plan

a new time-bound multi-threaded unit test that "fuzzes" the problematic behavior described in this issue

• Assignee: Caleb Rackliffe
• Reporter: Caleb Rackliffe

Authors

Caleb Rackliffe

Reviewers

Andres de la Peña, Benedict Elliott Smith

• Created: 16/Jul/21 18:47
• Updated: 27/May/22 19:25
• Resolved: 21/Jul/21 17:59

Description

This problem could manifest on all versions, beginning on at least 3.0, but I’ll focus on the way it manifests in 4.0 here.

In what now seems like a wise move, CASSANDRA-16097 added an assertion to DigestResolver#responseMatch() that ensures the responses snapshot has at least one visible elements to compare (although of course only one element trivially cannot generate a mismatch and short-circuits immediately). However, at the point ReadCallback#onResponse() signals the waiting resolver, there is no guarantee that the size of the generated snapshot of the responses Accumulator is non-zero, or perhaps more worryingly, at least equal to the number of blocked-for responses. This seems to be a consequence of the documented weak visibility guarantees on Accumulator#add(). In short, if there are concurrent invocations of add(), is it not guaranteed that there is any visible size change after any one of them return, but only after all complete.

The particular exception looks something like this:

java.lang.AssertionError: Attempted response match comparison while no responses have been received.
	at org.apache.cassandra.service.reads.DigestResolver.responsesMatch(DigestResolver.java:110)
	at org.apache.cassandra.service.reads.AbstractReadExecutor.awaitResponses(AbstractReadExecutor.java:393)
	at org.apache.cassandra.service.StorageProxy.fetchRows(StorageProxy.java:2150)
	at org.apache.cassandra.service.StorageProxy.readRegular(StorageProxy.java:1979)
	at org.apache.cassandra.service.StorageProxy.read(StorageProxy.java:1882)
	at org.apache.cassandra.db.SinglePartitionReadCommand$Group.execute(SinglePartitionReadCommand.java:1121)
	at org.apache.cassandra.cql3.statements.SelectStatement.execute(SelectStatement.java:296)
	at org.apache.cassandra.cql3.statements.SelectStatement.execute(SelectStatement.java:248)
	at org.apache.cassandra.cql3.statements.SelectStatement.execute(SelectStatement.java:90)

It’s possible to reproduce this on simple single-partition reads without any short-read protection or replica filtering protection. I’ve also been able to reproduce this synthetically with a unit test on ReadCallback.

It seems like the most straightforward way to fix this would be to avoid signaling in ReadCallback#onResponse() until the visible size of the accumulator is at least the number of received responses. In most cases, this is trivially true, and our signaling behavior won’t change at all. In the very rare case that there are two (or more) concurrent calls to onResponse(), the second (or last) will signal, and having one more response than we strictly need should have no negative side effects. (We don’t seem to make any strict assertions about having exactly the number of required responses, only that we have enough.)

Issue Links

• relates to: CASSANDRA-16883 Weak visibility guarantees of Accumulator can lead to failure to recognize digest mismatches