[
https://issues.apache.org/jira/browse/HIVE-30124?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Denys Kuzmenko updated HIVE-30124:
----------------------------------
Description: BasicLlapCacheHydration / LlapCacheMetadataSerializer
serialise and reload only ORC entries (LlapOrcCacheLoader). With HIVE-30059 the
LLAP data cache also holds Parquet column chunks (and the metadata cache holds
Parquet footers), but on daemon restart with hive.llap.io.cache.hydration
enabled only the ORC content comes back; every Parquet table starts cold.
fetchCachedContentInfo already walks both, so the gap is the loader: it needs a
Parquet counterpart that re-reads the recorded (fileKey, chunk ranges) through
ParquetEncodedDataReader's cache-put path, using the same ParquetCacheLayout
tiling so hydrated entries match what the readers expect.
> LLAP cache hydration should persist and reload native Parquet column-chunk
> entries
> ----------------------------------------------------------------------------------
>
> Key: HIVE-30124
> URL: https://issues.apache.org/jira/browse/HIVE-30124
> Project: Hive
> Issue Type: Sub-task
> Reporter: Denys Kuzmenko
> Priority: Major
>
> BasicLlapCacheHydration / LlapCacheMetadataSerializer serialise and reload
> only ORC entries (LlapOrcCacheLoader). With HIVE-30059 the LLAP data cache
> also holds Parquet column chunks (and the metadata cache holds Parquet
> footers), but on daemon restart with hive.llap.io.cache.hydration enabled
> only the ORC content comes back; every Parquet table starts cold.
> fetchCachedContentInfo already walks both, so the gap is the loader: it needs
> a Parquet counterpart that re-reads the recorded (fileKey, chunk ranges)
> through ParquetEncodedDataReader's cache-put path, using the same
> ParquetCacheLayout tiling so hydrated entries match what the readers expect.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
