[CASSANDRA-18670] Importer should build SSTable indexes successfully before making new SSTables readable

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 5.0-alpha1
• Component/s: Feature/SAI, Legacy/Tools, Tool/nodetool
• Epic Link: CEP-7 Storage Attached Index (SAI)
• Bug Category: Correctness - Transient Incorrect Response
• Severity: Critical
• Complexity: Normal
• Discovered By: Code Inspection
• Platform: All
• Since Version: 5.1-alpha1

Source Control Link

https://github.com/apache/cassandra/commit/83523ddd3d2aa5ddd9c9b455551a50cedcf8704a

Test and Documentation Plan

adding new tests around SAI and SSTableImporter

• Assignee: Caleb Rackliffe Assign to me
• Reporter: Caleb Rackliffe

Authors

Caleb Rackliffe

Reviewers

Andres de la Peña, Mike Adamson

• Created: 14/Jul/23 19:56
• Updated: 06/Dec/23 16:37
• Resolved: 24/Jul/23 22:53

Description

Importing SSTables w/ already built SAI components is pretty unproblematic. However, the case where we import SSTables without accompanying index components can be problematic in two ways, given the indexes are built after the SSTables themselves are added to the tracker:

1.) The new SSTables can be seen for reads on the base table before their attached indexes are built and ready to query in the index View.

2.) If a failure occurs in an index build, the import just leaves the index in a compromised state, and even though a failure to import will be reported to the user, the SSTables will be readable for queries on the base table.

If the index implementation allows it (and SAI does), SSTableImporter should make sure SSTable indexes are built before making SSTables visible as part of the base table. If an index build fails, the import should fail and it should appear to ongoing reads like it never happened, including the index continuing to be queryable.

Issue Links

• causes: CASSANDRA-19017 Ensure that empty SAI column indexes do not fail on validation after full-SSTable streaming
• causes: CASSANDRA-19084 Test Failure: IndexStreamingFailureTest.testAvailabilityAfterFailed*EntireFileStreaming
• is duplicated by: CASSANDRA-18653 Avoid unnecessary index builds and validations
• is related to: CASSANDRA-18653 Avoid unnecessary index builds and validations
• links to: GitHub Pull Request #2492