[
https://issues.apache.org/jira/browse/IGNITE-17627?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17606491#comment-17606491
]
Sergey Uttsel commented on IGNITE-17627:
----------------------------------------
As was discussed read(RowId rowId, HybridTimestamp timestamp) method must
return structure:
{
BinaryRow row
@Nullable UUID txId
@Nullable \{UUID, short} commitPartId
/** Last committed value timestamp. */
@Nullable HybridTimestamp timestamp
}
scan(Predicate<BinaryRow> keyFilter, HybridTimestamp timestamp) must return
cursor with this structure.
commitPartId contains \{ UUID tableId, short partId }
Need to pass commitPartId into MvPartitionStorage#addWrite to save it into a
write intent.
> Extend MvPartitionStorage API with write intent resolution capabilities
> -----------------------------------------------------------------------
>
> Key: IGNITE-17627
> URL: https://issues.apache.org/jira/browse/IGNITE-17627
> Project: Ignite
> Issue Type: Improvement
> Reporter: Ivan Bessonov
> Assignee: Semyon Danilov
> Priority: Major
> Labels: ignite-3
>
> Commit of RW transaction is not instantaneous. RO transaction might require
> reads of data that's in the process of being committed. Current API doesn't
> support such scenario.
> RO API in partition storage has only two methods: {{read}} and {{{}scan{}}}.
> h3. Read
> This one is pretty simple. It should return pair of {{binaryRow}} and
> {{{}txId{}}}. After that, caller can check the state of the transaction and
> either return the value or repeat the call.
> There must be a way to hint read method that uncommitted data must be skipped.
> An interesting way of reading data might be required: it there's a write
> intent, but we see a commit done after the timestamp, we can safely proceed
> with reading.
> Unfortunately, such optimization may be heavy on the storage read operations,
> because it requires a "deep" look-ahead request. So, whether or not we
> implement this depends on one thing - how often do we have write intent
> resolution in real RO transactions?
> API is to be defined.
> h3. Scan
> This one is tricky, we can't just return a cursor. Special type of cursor is
> required, and it must allow same read capabilities on each individual element.
> API is to be defined.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
