[
https://issues.apache.org/jira/browse/IGNITE-20395?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Kirill Sizov updated IGNITE-20395:
-----------------------------------
Description:
If a transaction was committed/aborted, but for any reason the cleanup
operation was not performed on a node, the write intent entries would still be
present in the storage.
For example, the primary node crashed before executing the cleanup. In this
case the storage will still have write intents even if the transaction is
finished.
When an RO transaction sees write intents, no matter on primary or on any other
node, it performs write intent resolution and returns the correct result.
When an RW transaction sees write intents (happens on primary only), it also
performs write intent resolution. But any following change of the rows having
write intents will result in a storage exception because the storage doesn't
support more than one write intent per row.
IGNITE-20041 added a way to trigger async write intent cleanup on the node that
executes write intent resolution. But RW transaction does not wait the
completion of the cleanup.
*Definition of done:*
If an RW transaction performs write intent resolution for a row and this is a
modifying action (not GET), we should synchronously cleanup the write intent
before proceeding with executing UpdateCommand.
*Implementation details*
We can either call cleanup directly (like in {{processTxCleanupAction}} or wait
for the cleanup future (the result of async cleanup execution method) to finish.
was:
If a transaction was committed/aborted, but for any reason the cleanup
operation was not performed on a node, the write intent entries would still be
present in the storage.
When an RO transaction sees such entries, no matter on primary or on any other
node, it performs write intent resolution and returns the correct result.
When an RW transaction sees write intents (happens on primary only), it also
performs write intent resolution. But any following change of the rows having
write intents will result in a storage exception because the storage doesn't
support more than one write intent per row.
IGNITE-20041 added a way to trigger async write intent cleanup on the node that
executes write intent resolution. But there are a few pieces left.
The cleanup operation after resolution is local, not distributed - if a primary
replica has been cleaned up, the other replicas will still have write intents
in their storages. Imagine the case when there are thee replicas, A, B and C. A
is primary.
A following RW transaction will successfully execute on primary since it had
its storage cleaned, but when replicating on B and C it will result in an
exception.
*We need to extend this functionality with the following(Definition of done):*
# If this is a primary replica and we are executing an RW transaction that has
already performed write intent resolution, we need to synchronously wait for a
pending cleanup of the affected rows to be able to safely change the storage.
# We need to perform the cleanup on all nodes of the replication group and
make sure RW transaction will synchronously wait for the corresponding cleanup
operation.
# Upsert operation does not read data, therefore will not perform resolution
and cleanup. Thus it may fail if there are unresolved intents in the storage.
Ensure upserts read data before inserting new.
> Make sure write intents are cleaned up for RW transaction
> ---------------------------------------------------------
>
> Key: IGNITE-20395
> URL: https://issues.apache.org/jira/browse/IGNITE-20395
> Project: Ignite
> Issue Type: Task
> Reporter: Kirill Sizov
> Priority: Major
> Labels: ignite-3
>
> If a transaction was committed/aborted, but for any reason the cleanup
> operation was not performed on a node, the write intent entries would still
> be present in the storage.
> For example, the primary node crashed before executing the cleanup. In this
> case the storage will still have write intents even if the transaction is
> finished.
> When an RO transaction sees write intents, no matter on primary or on any
> other node, it performs write intent resolution and returns the correct
> result.
> When an RW transaction sees write intents (happens on primary only), it also
> performs write intent resolution. But any following change of the rows having
> write intents will result in a storage exception because the storage doesn't
> support more than one write intent per row.
> IGNITE-20041 added a way to trigger async write intent cleanup on the node
> that executes write intent resolution. But RW transaction does not wait the
> completion of the cleanup.
> *Definition of done:*
> If an RW transaction performs write intent resolution for a row and this is
> a modifying action (not GET), we should synchronously cleanup the write
> intent before proceeding with executing UpdateCommand.
> *Implementation details*
> We can either call cleanup directly (like in {{processTxCleanupAction}} or
> wait for the cleanup future (the result of async cleanup execution method) to
> finish.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
