Known issues v1.3

These are the currently known issues and limitations identified in this Analytics Accelerator release. Where applicable, we have included workarounds to help you mitigate the impact of these issues. These issues are actively tracked and are planned for resolution in a future release.

  • Spark Connect integration:
    • DirectScan is only supported if no tables in the query are schema-qualified (via search_path) or if all qualified tables originate from the same single catalog with matching Postgres and catalog names.
    • PGAA does not reuse Spark Connect sessions across queries, which can lead to an excessive number of parallel sessions during CompatScan operations.
    • Support is limited to Iceberg tables in Iceberg REST catalogs; AWS S3 Tables, PGFS, and Delta Tables are not currently supported.
    • OAuth2 authentication is not currently supported for Iceberg REST connections.
    • There is a known issue in the Iceberg Spark Library version 1.9.2 when using Spark to read from Iceberg tables. Under certain conditions, equality deletes may occasionally be skipped during concurrent executions. The current workaround is to disable the Spark application setting spark.sql.iceberg.executor-cache.enabled on your spark-defaults.conf file. Disabling this cache ensures data consistency by correctly processing all deletes, but it may have performance implications for high-volume read workloads.
  • Iceberg replication:

    • Compaction and table optimization require a configured Spark Connect endpoint.
    • Data loss or duplication issues may occur during PGD leader failovers or node crashes.
    • Comprehensive performance testing is ongoing; while improvements over Delta and copy-on-write are expected, final benchmarks are not yet available.
  • CTAS:

    • Queries are executed on the Postgres primary even if they are eligible for DirectScan mode.
    • The system does not automatically delete data in object storage or the catalog if it already exists at the target path.
    • Support for CTAS into a catalog-managed table has not been explicitly tested.