sivabalan narayanan created HUDI-7779:
-----------------------------------------
Summary: Guarding archival to not archive unintended commits
Key: HUDI-7779
URL: https://issues.apache.org/jira/browse/HUDI-7779
Project: Apache Hudi
Issue Type: Bug
Components: archiving
Reporter: sivabalan narayanan
Archiving commits from active timeline could lead to data consistency issues on
some rarest of occasions. We should come up with proper guards to ensure we do
not make such unintended archival.
Major gap which we wanted to guard is:
if someone disabled cleaner, archival should account for data consistency
issues and ensure it bails out.
We have a base guarding condition, where archival will stop at the earliest
commit to retain based on latest clean commit metadata. But there are few other
scenarios that needs to be accounted for.
a. Keeping aside replace commits, lets dive into specifics for regular commits
and delta commits.
Say user configured clean commits to 4 and archival configs to 5 and 6. after
t10, cleaner is supposed to clean up all file versions created at or before t6.
Say cleaner did not run(for whatever reason for next 5 commits).
Archival will certainly be guarded until earliest commit to retain based on
latest clean commits.
Corner case to consider:
A savepoint was added to say t3 and later removed. and still the cleaner was
never re-enabled. Even though archival would have been stopped at t3 (until
savepoint is present),but once savepoint is removed, if archival is executed,
it could archive commit t3. Which means, file versions tracked at t3 is still
not yet cleaned by the cleaner.
Reasoning:
We are good here wrt data consistency. Up until cleaner runs next time, this
older file versions might be exposed to the end-user. But time travel query is
not intended for on-clean commits and hence this is not an issue. None of
snapshot, time travel query or incremental query will run into issues as they
are not supposed to poll for t3.
At any later point, if cleaner is re-enabled, it will take care of cleaning up
file versions tracked at t3 commit.
b. The more tricky part is when replace commits are involved. Since replace
commit metadata in active timeline is what ensures the replaced file groups are
ignored for reads, before archiving the same, cleaner is expected to clean them
up fully. But are there chances that this could go wrong?
Corner case to consider. Lets add onto above scenario, where t3 has a
savepoint, and t4 is a replace commit which replaced file groups tracked in t3.
Cleaner will skip cleaning up files tracked by t3, but will clean up t4, t5 and
t6. So, earliest commit to retain will be pointing to t6. And say savepoint for
t3 is removed, but cleaner was disabled. In this state of the timeline, if
archival is executed, (since t3.savepoint is removed), archival might archive
t3 and t4.rc. This could lead to data duplicates as both replaced file groups
and new file groups from t4.rc would be exposed as valid file groups.
In other words, if we were to summarize the different scenarios:
i. replaced file group is never cleaned up.
- ECTR is less than this.rc and we are good.
ii. replaced file group is cleaned up.
- ECTR is > this.rc and is good to archive.
iii. tricky: ECTR moved ahead compared to this.rc, but due to savepoint, full
clean up did not happen. After savepoint is removed, and when archival is
executed, we should avoid archiving the rc of interest. This is the gap we
don't account for as of now.
We have 2 options to go about to solve this.
*Option A:*
Before archiving any replace commit by the archiver, lets explicitly check that
all replaced file groups are fully deleted.
Cons: Might need FileSystemView polling which might be costly.
*OptionB:*
Cleaner also tracks an additional metadata named, "fully cleaned up file
groups" at the end of clean planning and in completed clean commit metadata.
So, archival instead of polling FileSystemView (which might be costly), it can
check for clean commit metadata for the list of file groups and can deduce if
all file groups replaced by X.rc is fully deleted.
Pros:
Since clean planner anyways polls the file system view and has all file group
info already, no additional work might be required to deduce "fully cleaned up
file groups". Just that it needs to add an additional metadata.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
