[
https://issues.apache.org/jira/browse/FLINK-4502?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15439445#comment-15439445
]
Elias Levy commented on FLINK-4502:
-----------------------------------
Chesnay, thanks for correcting me about the overwriting issue. I missed the
implication of setting the timestamp on the queries.
I still think the documentation is not clear that idempotent queries are a
prerequisite of exactly-once semantics when the WAL is enabled, as shown by the
portion of the documentation I quoted. The connector documentation page
mentions idempotent once, in the paragraph previous to the one describing the
WAL.
The WAL functionality description could also be clearer. For instance, is says
"The write-ahead log guarantees that the replayed checkpoint is identical to
the first attempt." But that does not seem accurate. The WAL doesn't appear
to guarantee that a replayed checkpoint is identical. Rather, it guarantees
that no output associated with a checkpoint is written to the sink until the
checkpoint is complete, which avoids possibly incompatible duplicate outputs
when data from non-completed checkpoints are replayed on recovery.
> Cassandra connector documentation has misleading consistency guarantees
> -----------------------------------------------------------------------
>
> Key: FLINK-4502
> URL: https://issues.apache.org/jira/browse/FLINK-4502
> Project: Flink
> Issue Type: Bug
> Components: Cassandra Connector, Documentation
> Affects Versions: 1.1.0
> Reporter: Elias Levy
>
> The Cassandra connector documentation states that "enableWriteAheadLog() is
> an optional method, that allows exactly-once processing for non-deterministic
> algorithms." This claim appears to be false.
> From what I gather, the write ahead log feature of the connector works as
> follows:
> - The sink is replaced with a stateful operator that writes incoming messages
> to the state backend based on checkpoint they belong in.
> - When the operator is notified that a Flink checkpoint has been completed
> it, for each set of checkpoints older than and including the committed one:
> * reads its messages from the state backend
> * writes them to Cassandra
> * records that it has committed them to Cassandra for the specific
> checkpoint and operator instance
> * and erases them from the state backend.
> This process attempts to avoid resubmitting queries to Cassandra that would
> otherwise occur when recovering a job from a checkpoint and having messages
> replayed.
> Alas, this does not guarantee exactly once semantics at the sink. The writes
> to Cassandra that occur when the operator is notified that checkpoint is
> completed are not atomic and they are potentially non-idempotent. If the job
> dies while writing to Cassandra or before committing the checkpoint via
> committer, queries will be replayed when the job recovers. Thus the
> documentation appear to be incorrect in stating this provides exactly-once
> semantics.
> There also seems to be an issue in GenericWriteAheadSink's
> notifyOfCompletedCheckpoint which may result in incorrect output. If
> sendValues returns false because a write failed, instead of bailing, it
> simply moves on to the next checkpoint to commit if there is one, keeping the
> previous one around to try again later. But that can result in newer data
> being overwritten with older data when the previous checkpoint is retried.
> Although given that CassandraCommitter implements isCheckpointCommitted as
> checkpointID <= this.lastCommittedCheckpointID, it actually means that when
> it goes back to try the uncommitted older checkpoint it will consider it
> committed, even though some of its data may not have been written out, and
> the data will be discarded.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
