vaultah commented on code in PR #13720:
URL: https://github.com/apache/iceberg/pull/13720#discussion_r2293737793
##########
spark/v4.0/spark/src/main/java/org/apache/iceberg/spark/actions/RewriteTablePathSparkAction.java:
##########
@@ -494,36 +483,60 @@ public RewriteContentFileResult
appendDeleteFile(RewriteResult<DeleteFile> r1) {
}
}
- /** Rewrite manifest files in a distributed manner and return rewritten data
files path pairs. */
- private RewriteContentFileResult rewriteManifests(
+ /**
+ * Rewrite manifest files in a distributed manner and return the resulting
manifests and content
+ * files selected for rewriting.
+ */
+ private Map<String, RewriteContentFileResult> rewriteManifests(
Set<Snapshot> deltaSnapshots, TableMetadata tableMetadata,
Set<ManifestFile> toRewrite) {
if (toRewrite.isEmpty()) {
- return new RewriteContentFileResult();
+ return Maps.newHashMap();
}
Encoder<ManifestFile> manifestFileEncoder =
Encoders.javaSerialization(ManifestFile.class);
+ Encoder<RewriteContentFileResult> manifestResultEncoder =
+ Encoders.javaSerialization(RewriteContentFileResult.class);
+ Encoder<Tuple2<String, RewriteContentFileResult>> tupleEncoder =
+ Encoders.tuple(Encoders.STRING(), manifestResultEncoder);
+
Dataset<ManifestFile> manifestDS =
spark().createDataset(Lists.newArrayList(toRewrite),
manifestFileEncoder);
Set<Long> deltaSnapshotIds =
deltaSnapshots.stream().map(Snapshot::snapshotId).collect(Collectors.toSet());
- return manifestDS
- .repartition(toRewrite.size())
- .map(
- toManifests(
- tableBroadcast(),
- sparkContext().broadcast(deltaSnapshotIds),
- stagingDir,
- tableMetadata.formatVersion(),
- sourcePrefix,
- targetPrefix),
- Encoders.bean(RewriteContentFileResult.class))
- // duplicates are expected here as the same data file can have
different statuses
- // (e.g. added and deleted)
- .reduce((ReduceFunction<RewriteContentFileResult>)
RewriteContentFileResult::append);
- }
-
- private static MapFunction<ManifestFile, RewriteContentFileResult>
toManifests(
+ Iterator<Tuple2<String, RewriteContentFileResult>> resultIterator =
+ manifestDS
+ .repartition(toRewrite.size())
+ .map(
+ toManifests(
+ tableBroadcast(),
+ sparkContext().broadcast(deltaSnapshotIds),
+ stagingDir,
+ tableMetadata.formatVersion(),
+ sourcePrefix,
+ targetPrefix),
+ tupleEncoder)
+ .toLocalIterator();
Review Comment:
I suggested pre-aggregating maps for discussion, but after thinking it
through, the final map still needs to be held by the driver, so the overall
memory footprint isn't reduced. Pre-aggregation only changes how results are
shipped, not how much memory is needed in the end. If anyone has a concrete
alternative or wants to discuss tradeoffs, I'm happy to follow up.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
