[
https://issues.apache.org/jira/browse/IMPALA-15117?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Michael Smith updated IMPALA-15117:
-----------------------------------
Labels: newbie (was: beginner-friendly)
> Optimize getFirstLevelAcidDirPath
> ---------------------------------
>
> Key: IMPALA-15117
> URL: https://issues.apache.org/jira/browse/IMPALA-15117
> Project: IMPALA
> Issue Type: Improvement
> Components: Frontend
> Affects Versions: Impala 4.5.0
> Reporter: Michael Smith
> Priority: Major
> Labels: newbie
>
> This could also apply to Hive.
> When loading ACID table metadata, {{AcidUtils.getFirstLevelAcidDirPath}}
> checks whether each candidate ACID subdirectory name (e.g. {{delta_*}},
> {{base_*}}, {{delete_delta_*}}) exists under a partition directory by calling
> {{FileSystem.isDirectory(path)}} for each candidate in a loop. On distributed
> filesystems each call is a separate RPC.
> Observed via async-profiler on a production Impala catalogd connected to
> Apache Ozone (OFS): 258 samples (94% blocked on RPC) in the stack
> {{AcidUtils.getFirstLevelAcidDirPath → FileSystem.isDirectory →
> OzoneBucket.getFileStatus → OM gRPC call}}. For a partition with N candidate
> paths, N separate {{GetFileStatus}} RPCs are issued where 1 {{listStatus}}
> would suffice.
> We could replace the N {{isDirectory(candidatePath)}} calls with a single
> {{listStatus(partitionDir)}} call and resolve membership locally:
> # Call {{fs.listStatus(partitionDir)}} once to get all immediate children
> with their type flags.
> # Build a local {{Set<String>}} of child directory names from the result.
> # Filter candidate ACID names against the set — no additional RPCs.
> This reduces the cost from O(N) RPCs to 1 regardless of filesystem
> implementation. On object-store-backed filesystems (OFS, S3, ABFS) where each
> {{isDirectory}} is a synchronous RPC, the reduction is especially significant.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
