[
https://issues.apache.org/jira/browse/HUDI-9030?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17932764#comment-17932764
]
sivabalan narayanan commented on HUDI-9030:
-------------------------------------------
[~ljain] : I had a chat w/ Ethan.
for file group reader, looks like we to enable it by default even w/ table
version 6.
so, lets plan to work on that.
Also, couple of comments:
- MarkerBasedRollbackStrategy#createRollbackRequestForCreateAndMerge removes
validation that log file should not have IOType as CREATE in table version 6.
CREATE is still used for log files with tbl version 6 in
LogFileCreationCallback#preFileCreation
Feedback:
Lets ensure we rollback all failed writes when someone moves from 0.x to 1.x
but still in table version 6.
If yes, it does not matter if we changed the IOtype in 1.0 even for table
version6. If not, we need to maintain parity with 0.x writer. So, we might
switch it to APPEND just for table version 6 in 1.0.
> Validate and certify log files and marker interplays in MOR using table
> version6
> ---------------------------------------------------------------------------------
>
> Key: HUDI-9030
> URL: https://issues.apache.org/jira/browse/HUDI-9030
> Project: Apache Hudi
> Issue Type: Sub-task
> Reporter: sivabalan narayanan
> Assignee: Lokesh Jain
> Priority: Major
> Labels: pull-request-available
> Fix For: 1.0.2
>
> Original Estimate: 6h
> Remaining Estimate: 6h
>
> Validate and certify log files and marker interplays in MOR using table
> version6 in comparison to using 0.x writer.
> Below table lists the behaviour related to log files and marker generation in
> 0.x and compares it with 1.x with table version 6 and 1.x with latest table
> version.
> || ||0.x||Table Version 6||1.x||
> |Log File name Instant|Base file instant|Base file instant|Deltacommit
> instant|
> |Log File name Write Token|Rollover log write token is always newly created
> Log write token is created from latest log file|Rollover log write token
> concept is removed. Log write token is used instead.|Rollover log write token
> concept is removed. Log write token is used instead.|
> |Log file version|Computed using latest log file|Computed using latest log
> file|Computed using latest log file|
> |Append to existing log file|Allowed|Not Supported. Writes happen to a new
> file.|Not Supported. Writes happen to a new file.|
> |Marker generation|Marker is created during append as well as when a new file
> is created (per log file marker)|Marker is created every time a new log file
> is create|Marker is created every time a new log file is created|
> |Rollback logic|*0.X :*
>
> For base files, we add it as part of "filesToDelete" list in Rollback plan.
> For log files, we added it as part of "logBlocksToDelete" in the rollback
> plan.
>
> during execution,
> any file in "filesToDelete" in the rollback plan will be deleted.
> for files in "logBlocksToDelete", a rollback command block will be added (new
> log file).
>
> per log file marker only works for Marker based rollback strategy. Its fair
> to say that Listing based is not supported anymore.
>
>
> |*Table Version 6 in 1.0:*
>
> For base files, we add it as part of "filesToDelete" list in Rollback plan.
> For log files, we added it as part of "logBlocksToDelete" in the rollback
> plan.
>
> during execution,
> any file in "filesToDelete" in the rollback plan will be deleted.
> for files in "logBlocksToDelete", a rollback command block will be added (new
> log file).
>
> per log file marker only works for Marker based rollback strategy. Its fair
> to say that Listing based is not supported anymore. |*1.X tbl v8:*
>
> there is no diff b/w a base file and a log file in 1.x. During rollback
> planning, any file from the marker, is added to "filesToDelete" list in the
> rollback plan.
>
> During execution, hudi deletes every file tracked in "filesToDelete" list in
> the rollback plan. |
> Other issues
> # For table version 6, AbstractHoodieLogRecordScanner ignores the log blocks
> which belong to inflight instants while scanning. PR fixes the logic so that
> such log blocks are not ignored. This is required for updating RLI which
> reads the deleted records from data table
> # hoodie.file.group.reader.enabled needs to be disabled in tbl version 6
> # The new rollback logic filters log files using the deltacommit timestamp
> and then marks them for deletion. This does not work for tbl version 6 since
> log files do not have deltacommit timestamp in the name. Therefore older
> rollback logic was brought back here.
> # PR removes the validation while scheduling compaction which validates that
> compaction instant should be greater than all completed deltacommit instants.
> This validation was added for table version 6 but is not really required.
> # KEY_GENERATOR_CLASS_NAME and KEY_GENERATOR_TYPE are new configs which are
> required in tbl version 6 as well. PR makes a change so that these configs
> are not ignored.
> # MarkerBasedRollbackStrategy#createRollbackRequestForCreateAndMerge removes
> validation that log file should not have IOType as CREATE in table version 6.
> CREATE is still used for log files with tbl version 6 in
> LogFileCreationCallback#preFileCreation
> # hoodie.datasource.read.incr.fallback.fulltablescan.enable needs to be
> disabled for tbl version 6
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
