[
https://issues.apache.org/jira/browse/CASSANDRA-14592?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16587593#comment-16587593
]
Benedict commented on CASSANDRA-14592:
--------------------------------------
Pushed an update that addresses (I think, it's been a while) Aleksey's offline
review comments.
We collaborated to modify the reconcile semantics a little further, so that
reconciliation is as consistent as possible. Now the only situations that
might arise with inconsistent reconciliation occur when one cell is expiring,
another is a tombstone, and only at the point where both are logically a
tombstone. Specifically, we now prefer:
# The most recent timestamp
# If either are a tombstone or expiring
## If one is regular, select the tombstone or expiring
## If one is expiring, select the tombstone
## The most recent deletion time
# The highest value (by raw ByteBuffer comparison)
> Reconcile should not be dependent on nowInSec
> ---------------------------------------------
>
> Key: CASSANDRA-14592
> URL: https://issues.apache.org/jira/browse/CASSANDRA-14592
> Project: Cassandra
> Issue Type: Bug
> Reporter: Benedict
> Assignee: Benedict
> Priority: Major
> Fix For: 4.0
>
>
> To have the arrival time of a mutation on a replica determine the
> reconciliation priority seems to provide for unintuitive database behaviour.
> It seems we should formalise our reconciliation logic in a manner that does
> not depend on this, and modify our internal APIs to prevent this dependency.
>
> Take the following example, where both writes have the same timestamp:
>
> Write X with a value A, TTL of 1s
> Write Y with a value B, no TTL
>
> If X and Y arrive on replicas in < 1s, X and Y are both live, so record Y
> wins the reconciliation. The value B appears in the database.
> However, if X and Y arrive on replicas in > 1s, X is now (effectively) a
> tombstone. This wins the reconciliation race, and NO value is the result.
>
> Note that the weirdness of this is more pronounced than it might first
> appear. If write X gets stuck in hints for a period on the coordinator to
> one replica, the value B appears in the database until the hint is replayed.
> So now we’re in a very uncertain state - will hints get replayed or not? If
> they do, the value B will disappear; if they don’t it won’t. This is despite
> a QUORUM of replicas ACKing both writes, and a QUORUM of readers being
> engaged on read; the database still changes state to the user suddenly at
> some arbitrary future point in time.
>
> It seems to me that a simple solution to this, is to permit TTL’d data to
> always win a reconciliation against non-TTL’d data (of same timestamp), so
> that we are consistent across TTLs being transformed into tombstones.
>
> 4.0 seems like a good opportunity to fix this behaviour, and mention in
> CHANGES.txt.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
