Core: fix reading of split offsets in manifests #8834
Conversation
| private static final RecursiveComparisonConfiguration FILE_COMPARISON_CONFIG = | ||
| RecursiveComparisonConfiguration.builder() | ||
| .withIgnoredFields( | ||
| "dataSequenceNumber", "fileOrdinal", "fileSequenceNumber", "fromProjectionPos") |
There was a problem hiding this comment.
So this will compare split offsets now?
There was a problem hiding this comment.
yes, it will compare all fields from the input data files, except for these excluded fields which aren't expected to match
| @@ -127,6 +129,7 @@ public class TableTestBase { | |||
| .withFileSizeInBytes(10) | |||
| .withPartitionPath("data_bucket=3") // easy way to set partition data for now | |||
| .withRecordCount(1) | |||
| .withSplitOffsets(ImmutableList.of(4L, 10_000_000L, 20_000_000L)) | |||
There was a problem hiding this comment.
Minor comment but can we make the splits unique? Just want to make sure things will break in an obvious way if they are not matching
There was a problem hiding this comment.
Sure, I made this change
|
Other than to revert the optimize in #8336, is it better to invalidate the cached ...
case 12:
this.upperBounds = SerializableByteBufferMap.wrap((Map<Integer, ByteBuffer>) value);
return;
case 13:
this.keyMetadata = ByteBuffers.toByteArray((ByteBuffer) value);
return;
case 14:
this.splitOffsets = ArrayUtil.toLongArray((List<Long>) value);
this.splitOffsetList = null; // invalidate the cache
return;
case 15:
this.equalityIds = ArrayUtil.toIntArray((List<Integer>) value);
.... |
The goal of this PR is to revert the code to a known working state. There are a couple of optimizations that could be made (e.g. to equality IDs also) but felt that belongs in a separate PR. |
| "Should read the expected files", | ||
| Lists.newArrayList(FILE_A.path(), FILE_B.path(), FILE_C.path()), | ||
| files); | ||
| assertThat(files) |
There was a problem hiding this comment.
I believe this can be simplified to
assertThat(files)
.usingRecursiveComparison()
.ignoringFields(
"dataSequenceNumber", "fileOrdinal", "fileSequenceNumber", "fromProjectionPos")
.isEqualTo(Lists.newArrayList(FILE_A, FILE_B, FILE_C));
that way you don't need the static final field for the comparison config
There was a problem hiding this comment.
My thought was the config variable could be reused in other tests
If the current PR is intended to be included in the hot fix version 1.4.1, then I think it's the safe way to revert the code. |
| } | ||
|
|
||
| return splitOffsetList; | ||
| return ArrayUtil.toUnmodifiableLongList(splitOffsets); |
The list was not correctly invalidated when reusing the file.
This PR fixes a critical bug in reading split offsets from manifests. A change in #8336 added caching of the offsets collection in
BaseFileto avoid reallocation. However, the Avro reader will reuse the same BaseFile object when reading, thus only the offsets from the first entry are allocated, and then those are reused for all other entries.cc @aokolnychyi @RussellSpitzer @rdblue @danielcweeks This can result in corrupted metadata in cases the invalid offsets are read in then written back, e.g. when rewriting manifests.