[IMPALA-9841] Table level writeIdList represent the filemetadata snapshot

• Type: Improvement
• Status: Open
• Priority: Major
• Resolution: Unresolved
• Component/s: Catalog
• Epic Color: ghx-label-11
• Assignee: Vihang Karajgaonkar
• Reporter: Vihang Karajgaonkar
• Created: 08/Jun/20 21:40
• Updated: 21/Dec/20 20:38

Description

Currently, in HdfsTable https://github.com/apache/impala/blob/master/fe/src/main/java/org/apache/impala/catalog/HdfsTable.java#L1017 we reload the ValidWriteIdList every time. However in case of incremental refresh, it is possible that the file-metadata is not loaded at all or only loaded for certain partitions. This would mean that the cached ValidWriteIdList of the table is out-of-sync with the actual file-metadata which is cached in catalog.

This is currently not a problem since we on rely on the ValidWriteIdList during the loading of the file-metadata. However, for a consistent reads of the transactional tables, we would need to make sure that this ValidWriteIdList is guaranteed to represent the cached files.

Issue Links

• is related to: IMPALA-9791 Support validWriteIdList in getPartialCatalogObject