[
https://issues.apache.org/jira/browse/HUDI-7518?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Vinoth Chandar updated HUDI-7518:
---------------------------------
Description:
When there are repeated duplicate deletes to the partition file list in files
partition of the MDT, the current HoodieMetadataPayload merging logic drops
such "deletion", causing the file that is deleted from the file system and
supposed to be deleted from MDT file listing still left in MDT, because of the
following logic:
{code:java}
private Map<String, HoodieMetadataFileInfo>
combineFileSystemMetadata(HoodieMetadataPayload previousRecord) {
Map<String, HoodieMetadataFileInfo> combinedFileInfo = new HashMap<>();
// First, add all files listed in the previous record
if (previousRecord.filesystemMetadata != null) {
combinedFileInfo.putAll(previousRecord.filesystemMetadata);
}
// Second, merge in the files listed in the new record
if (filesystemMetadata != null) {
validatePayload(type, filesystemMetadata);
filesystemMetadata.forEach((key, fileInfo) -> {
combinedFileInfo.merge(key, fileInfo,
(oldFileInfo, newFileInfo) ->
newFileInfo.getIsDeleted()
? null
: new
HoodieMetadataFileInfo(Math.max(newFileInfo.getSize(), oldFileInfo.getSize()),
false));
});
} {code}
Here's a concrete example of how this bug causes the ingestion to fail:
(1) A data file and file group are replaced by clustering. The data file is
still on the file system and in MDT file listing.
(2) A cleaner plan is generated to delete the data file.
(3) The cleaner plan is executed the first time, and fails before commit due to
Spark job shutdown.
(4) The ingestion continues and succeeds, and another cleaner plan is generated
containing the same data file/file group to delete.
(5) The first cleaner plan is successfully executed, incurring deletion to the
file list with a metadata payload, and this is added to one log file in MDT,
e.g.,
{code:java}
HoodieMetadataPayload {key=partition, type=2, Files: {creations=[],
deletions=[7f6b146e-cd43-4fd3-9ce0-118232562569-0_63-29223-5579389_20240303214408245.parquet],
}}{code}
(6) The second cleaner plan is also successfully executed, incurring deletion
to the file list with a metadata payload containing the same data file to
delete, and this is added to a subsequent log file in the same file slice in
MDT, e.g.,
{code:java}
HoodieMetadataPayload {key=partition, type=2, Files: {creations=[],
deletions=[7f6b146e-cd43-4fd3-9ce0-118232562569-0_63-29223-5579389_20240303214408245.parquet],
}} {code}
(7) The replacecommit corresponds to the clustering is archived as the cleaner
has deleted the replaced file groups.
(8) When reading MDT or MDT compaction happens, the merging of these two
metadata payloads with identical deletes leads to empty deletion, so the data
file is not deleted from the partition file list in MDT. The expected behavior
is to keep the data file in the "deletions" field.
{code:java}
HoodieMetadataPayload {key=partition, type=2, Files: {creations=[],
deletions=[], }}{code}
(9) Next time, when doing upsert and indexing, the deleted data file is served
by the file system view based on MDT (e.g.,
"7f6b146e-cd43-4fd3-9ce0-118232562569-0_63-29223-5579389_20240303214408245.parquet"),
and the data file cannot be found on the file system, causing the ingestion to
fail.
was:
When there are repeated deletes to the partition file list in files partition
of the MDT, the current HoodieMetadataPayload merging logic drops such
"deletion", causing the file that is deleted from the file system and supposed
to be deleted from MDT file listing still left in MDT, because of the following
logic:
{code:java}
private Map<String, HoodieMetadataFileInfo>
combineFileSystemMetadata(HoodieMetadataPayload previousRecord) {
Map<String, HoodieMetadataFileInfo> combinedFileInfo = new HashMap<>();
// First, add all files listed in the previous record
if (previousRecord.filesystemMetadata != null) {
combinedFileInfo.putAll(previousRecord.filesystemMetadata);
}
// Second, merge in the files listed in the new record
if (filesystemMetadata != null) {
validatePayload(type, filesystemMetadata);
filesystemMetadata.forEach((key, fileInfo) -> {
combinedFileInfo.merge(key, fileInfo,
(oldFileInfo, newFileInfo) ->
newFileInfo.getIsDeleted()
? null
: new
HoodieMetadataFileInfo(Math.max(newFileInfo.getSize(), oldFileInfo.getSize()),
false));
});
} {code}
Here's a concrete example of how this bug causes the ingestion to fail:
(1) A data file and file group are replaced by clustering. The data file is
still on the file system and in MDT file listing.
(2) A cleaner plan is generated to delete the data file.
(3) The cleaner plan is executed the first time, and fails before commit due to
Spark job shutdown.
(4) The ingestion continues and succeeds, and another cleaner plan is generated
containing the same data file/file group to delete.
(5) The first cleaner plan is successfully executed, incurring deletion to the
file list with a metadata payload, and this is added to one log file in MDT,
e.g.,
{code:java}
HoodieMetadataPayload {key=partition, type=2, Files: {creations=[],
deletions=[7f6b146e-cd43-4fd3-9ce0-118232562569-0_63-29223-5579389_20240303214408245.parquet],
}}{code}
(6) The second cleaner plan is also successfully executed, incurring deletion
to the file list with a metadata payload containing the same data file to
delete, and this is added to a subsequent log file in the same file slice in
MDT, e.g.,
{code:java}
HoodieMetadataPayload {key=partition, type=2, Files: {creations=[],
deletions=[7f6b146e-cd43-4fd3-9ce0-118232562569-0_63-29223-5579389_20240303214408245.parquet],
}} {code}
(7) The replacecommit corresponds to the clustering is archived as the cleaner
has deleted the replaced file groups.
(8) When reading MDT or MDT compaction happens, the merging of these two
metadata payloads with identical deletes leads to empty deletion, so the data
file is not deleted from the partition file list in MDT. The expected behavior
is to keep the data file in the "deletions" field.
{code:java}
HoodieMetadataPayload {key=partition, type=2, Files: {creations=[],
deletions=[], }}{code}
(9) Next time, when doing upsert and indexing, the deleted data file is served
by the file system view based on MDT (e.g.,
"7f6b146e-cd43-4fd3-9ce0-118232562569-0_63-29223-5579389_20240303214408245.parquet"),
and the data file cannot be found on the file system, causing the ingestion to
fail.
> Fix HoodieMetadataPayload merging logic around repeated deletes
> ---------------------------------------------------------------
>
> Key: HUDI-7518
> URL: https://issues.apache.org/jira/browse/HUDI-7518
> Project: Apache Hudi
> Issue Type: Bug
> Reporter: Ethan Guo
> Assignee: Ethan Guo
> Priority: Blocker
> Fix For: 0.15.0, 1.0.0
>
>
> When there are repeated duplicate deletes to the partition file list in files
> partition of the MDT, the current HoodieMetadataPayload merging logic drops
> such "deletion", causing the file that is deleted from the file system and
> supposed to be deleted from MDT file listing still left in MDT, because of
> the following logic:
> {code:java}
> private Map<String, HoodieMetadataFileInfo>
> combineFileSystemMetadata(HoodieMetadataPayload previousRecord) {
> Map<String, HoodieMetadataFileInfo> combinedFileInfo = new HashMap<>();
> // First, add all files listed in the previous record
> if (previousRecord.filesystemMetadata != null) {
> combinedFileInfo.putAll(previousRecord.filesystemMetadata);
> }
> // Second, merge in the files listed in the new record
> if (filesystemMetadata != null) {
> validatePayload(type, filesystemMetadata);
> filesystemMetadata.forEach((key, fileInfo) -> {
> combinedFileInfo.merge(key, fileInfo,
> (oldFileInfo, newFileInfo) ->
> newFileInfo.getIsDeleted()
> ? null
> : new
> HoodieMetadataFileInfo(Math.max(newFileInfo.getSize(),
> oldFileInfo.getSize()), false));
> });
> } {code}
> Here's a concrete example of how this bug causes the ingestion to fail:
> (1) A data file and file group are replaced by clustering. The data file is
> still on the file system and in MDT file listing.
> (2) A cleaner plan is generated to delete the data file.
> (3) The cleaner plan is executed the first time, and fails before commit due
> to Spark job shutdown.
> (4) The ingestion continues and succeeds, and another cleaner plan is
> generated containing the same data file/file group to delete.
> (5) The first cleaner plan is successfully executed, incurring deletion to
> the file list with a metadata payload, and this is added to one log file in
> MDT, e.g.,
> {code:java}
> HoodieMetadataPayload {key=partition, type=2, Files: {creations=[],
> deletions=[7f6b146e-cd43-4fd3-9ce0-118232562569-0_63-29223-5579389_20240303214408245.parquet],
> }}{code}
> (6) The second cleaner plan is also successfully executed, incurring deletion
> to the file list with a metadata payload containing the same data file to
> delete, and this is added to a subsequent log file in the same file slice in
> MDT, e.g.,
> {code:java}
> HoodieMetadataPayload {key=partition, type=2, Files: {creations=[],
> deletions=[7f6b146e-cd43-4fd3-9ce0-118232562569-0_63-29223-5579389_20240303214408245.parquet],
> }} {code}
> (7) The replacecommit corresponds to the clustering is archived as the
> cleaner has deleted the replaced file groups.
> (8) When reading MDT or MDT compaction happens, the merging of these two
> metadata payloads with identical deletes leads to empty deletion, so the data
> file is not deleted from the partition file list in MDT. The expected
> behavior is to keep the data file in the "deletions" field.
> {code:java}
> HoodieMetadataPayload {key=partition, type=2, Files: {creations=[],
> deletions=[], }}{code}
> (9) Next time, when doing upsert and indexing, the deleted data file is
> served by the file system view based on MDT (e.g.,
> "7f6b146e-cd43-4fd3-9ce0-118232562569-0_63-29223-5579389_20240303214408245.parquet"),
> and the data file cannot be found on the file system, causing the ingestion
> to fail.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
