Core: Fix skipped file counts in ManifestReader with deleted entries #8432
Conversation
| } | ||
|
|
||
| private CloseableIterable<ManifestEntry<F>> entries(boolean onlyLive) { | ||
| if (hasRowFilter() || hasPartitionFilter() || partitionSet != null) { |
There was a problem hiding this comment.
I had to add hasRowFilter() and hasPartitionFilter() to avoid cyclomatic complexity warnings. I also think it makes it more readable.
| ? scanMetrics.skippedDataFiles() | ||
| : scanMetrics.skippedDeleteFiles(), | ||
| open(projection(fileSchema, fileProjection, projectColumns, caseSensitive)), | ||
| onlyLive ? filterLiveEntries(entries) : entries, |
There was a problem hiding this comment.
The deleted entries must be discarded prior to evaluating stats primarily to have correct skipped counts. It should also help the performance.
| tableDir, tableName, SCHEMA, SPEC, SortOrder.unsorted(), formatVersion, reporter); | ||
| table.newAppend().appendFile(FILE_A).appendFile(FILE_D).commit(); | ||
| table.newAppend().appendFile(FILE_B).appendFile(FILE_C).commit(); | ||
| table.newAppend().appendFile(FILE_A).appendFile(FILE_B).appendFile(FILE_D).commit(); |
There was a problem hiding this comment.
I need to add FILE_B here so that we have a manifest with one skipped data file as the original manifest will be rewritten with FILE_A as deleted during the overwrite below.
| assertThat(result.scannedDataManifests().value()).isEqualTo(1); | ||
| assertThat(result.scannedDeleteManifests().value()).isEqualTo(0); | ||
| assertThat(result.skippedDataManifests().value()).isEqualTo(1); | ||
| assertThat(result.skippedDataManifests().value()).isEqualTo(2); |
There was a problem hiding this comment.
The overwrite rewrites the original manifest and adds a new one for the new data file.
| } | ||
|
|
||
| private CloseableIterable<ManifestEntry<F>> entries(boolean onlyLive) { | ||
| if (hasRowFilter() || hasPartitionFilter() || partitionSet != null) { |
There was a problem hiding this comment.
Do we need to consider the condition?
partitionSet.isEmpty()
If the partitionSet is empty, we seems to filter out all entries. Maybe we can return empty entries directly when the partitionSet is empty. I'm not sure whether we have guaranteed the partitionSet is not empty. Correct me if I'm wrong.
There was a problem hiding this comment.
I believe an empty partitionSet indicates there is no match and acts similar to alwaysFalse. We could add a branch for cases when the filters are always false or the matching partition set is empty, but we would have to analyze the usages first to make sure those are possible scenarios.
| return entries(true /* only live entries */); | ||
| } | ||
|
|
||
| private CloseableIterable<ManifestEntry<F>> filterLiveEntries( |
There was a problem hiding this comment.
Minor: these last two methods could be static.
There was a problem hiding this comment.
I did it this way so that I don't have to add F extends ContentFile<F> to these static methods and reuse the bound already defined in the class instead. I could use ? in the second one but not here.
This PR fixes skipped file counts in
ManifestReaderwith deleted entries as discussed here.