flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Fabian Hueske <fhue...@gmail.com>
Subject Re: Flink recovery
Date Fri, 13 May 2016 21:13:02 GMT
Hi,

Flink's exactly-once semantics do not mean that events are processed
exactly-once but that events will contribute exactly-once to the state of
an operator such as a counter.
Roughly, the mechanism works as follows:
- Flink peridically injects checkpoint markers into the data stream. This
happens synchronously across all sources and markers.
- When an operator receives a checkpoint marker from all its sources, it
checkpoints its state and forwards the marker
- When the marker was received by all sinks, the distributed checkpoint is
noted as successful.

In case of a failure, the state of all operators is reset to the last
successful checkpoint and the sources are reset to the point when the
marker was injected.
Hence, some events are sent a second time to the operators but the state of
the operators was reset as well. So the repeated events contribute exactly
once to the state of an operator.

Note, you need a SinkFunction that supports Flink's checkpointing mechanism
to achieve exactly-once output. Otherwise, it might happen that results are
emitted multiple times.

Cheers, Fabian

2016-05-13 22:58 GMT+02:00 Madhire, Naveen <Naveen.Madhire@capitalone.com>:

> I checked the JIRA and looks like FLINK-2111 should address the issue
> which I am facing. I am canceling the job from dashboard.
>
> I am using kafka source and HDFS rolling sink.
>
> https://issues.apache.org/jira/browse/FLINK-2111
>
> Is this JIRA part of Flink 1.0.0?
>
>
>
> Thanks,
> Naveen
>
> From: "Madhire, Venkat Naveen Kumar Reddy" <Naveen.Madhire@capitalone.com>
> Reply-To: "user@flink.apache.org" <user@flink.apache.org>
> Date: Friday, May 13, 2016 at 10:58 AM
> To: "user@flink.apache.org" <user@flink.apache.org>
> Subject: Flink recovery
>
> Hi,
>
> We are trying to test the recovery mechanism of Flink with Kafka and HDFS
> sink during failures.
>
> I’ve killed the job after processing some messages and restarted the same
> job again. Some of the messages I am seeing are processed more than once
> and not following the exactly once semantics.
>
>
> Also, using the checkpointing mechanism and saving the state checkpoints
> into HDFS.
> Below is the checkpoint code,
>
> envStream.enableCheckpointing(11);
> envStream.getCheckpointConfig().setCheckpointingMode(CheckpointingMode.EXACTLY_ONCE);
> envStream.getCheckpointConfig().setCheckpointTimeout(60000);
> envStream.getCheckpointConfig().setMaxConcurrentCheckpoints(4);
>
> envStream.setStateBackend(new FsStateBackend("hdfs://ipaddr/mount/cp/checkpoint/"));
>
>
> One thing I’ve noticed is lowering the time to checkpointing is actually
> lowering the number of messages processed more than once and 11ms is the
> lowest I can use.
>
> Is there anything else I should try to have exactly once message
> processing functionality.
>
> I am using Flink 1.0.0 and kafka 0.8
>
>
> Thank you.
>
> ------------------------------
>
> The information contained in this e-mail is confidential and/or
> proprietary to Capital One and/or its affiliates and may only be used
> solely in performance of work or services for Capital One. The information
> transmitted herewith is intended only for use by the individual or entity
> to which it is addressed. If the reader of this message is not the intended
> recipient, you are hereby notified that any review, retransmission,
> dissemination, distribution, copying or other use of, or taking of any
> action in reliance upon this information is strictly prohibited. If you
> have received this communication in error, please contact the sender and
> delete the material from your computer.
>
> ------------------------------
>
> The information contained in this e-mail is confidential and/or
> proprietary to Capital One and/or its affiliates and may only be used
> solely in performance of work or services for Capital One. The information
> transmitted herewith is intended only for use by the individual or entity
> to which it is addressed. If the reader of this message is not the intended
> recipient, you are hereby notified that any review, retransmission,
> dissemination, distribution, copying or other use of, or taking of any
> action in reliance upon this information is strictly prohibited. If you
> have received this communication in error, please contact the sender and
> delete the material from your computer.
>

Mime
View raw message