flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Paris Carbone <par...@kth.se>
Subject Re: flink snapshotting fault-tolerance
Date Thu, 19 May 2016 18:35:32 GMT
Invalidations are not necessarily exposed (I hope). Think of it as implementing TCP, you don’t
have to warn the user that packets are lost since eventually a packet will be received at
the other side in an eventually sunchronous system. Snapshots follow the same paradigm. Hope
that helps.

On 19 May 2016, at 20:33, Stavros Kontopoulos <st.kontopoulos@gmail.com<mailto:st.kontopoulos@gmail.com>>

Yes thats what i was thinking thnx. When people here exactly once they think are you sure,
there is something hidden there... because theory is theory :)
So if you keep getting invalidated snapshots but data passes through operators you issue a
warning or fail the pipeline and return an exception to the driver?

On Thu, May 19, 2016 at 9:30 PM, Paris Carbone <parisc@kth.se<mailto:parisc@kth.se>>
In that case, typically a timeout invalidates the whole snapshot (all states for the same
epoch) until eventually we have a full complete snapshot.

On 19 May 2016, at 20:26, Stavros Kontopoulos <st.kontopoulos@gmail.com<mailto:st.kontopoulos@gmail.com>>

"Checkpoints are only confirmed if all parallel subtasks successfully created a valid snapshot
of the state." as stated by Robert. So to rephrase my question... how confirmation that all
snapshots are finished is done and what happens if some task is very slow...or is blocked?
If you have N tasks confirmed and one missing what do you do? You start a new checkpoint for
that one? or a global new checkpoint for the rest of N tasks as well?

On Thu, May 19, 2016 at 9:21 PM, Paris Carbone <parisc@kth.se<mailto:parisc@kth.se>>

Regarding your last question,
If a checkpoint expires it just gets invalidated and a new complete checkpoint will eventually
occur that can be used for recovery. If I am wrong, or something has changed please correct


On 19 May 2016, at 20:14, Paris Carbone <parisc@kth.se<mailto:parisc@kth.se>>

Hi Stavros,

Currently, rollback failure recovery in Flink works in the pipeline level, not in the task
level (see Millwheel [1]). It further builds on repayable stream logs (i.e. Kafka), thus,
there is no need for 3pc or backup in the pipeline sources. You can also check this presentation
[2] which explains the basic concepts more in detail I hope. Mind that many upcoming optimisation
opportunities are going to be addressed in the not so long-term Flink roadmap.


[1] http://static.googleusercontent.com/media/research.google.com/en//pubs/archive/41378.pdf
[2] http://www.slideshare.net/ParisCarbone/tech-talk-google-on-flink-fault-tolerance-and-ha


On 19 May 2016, at 19:43, Stavros Kontopoulos <st.kontopoulos@gmail.com<mailto:st.kontopoulos@gmail.com>>

Cool thnx. So if a checkpoint expires the pipeline will block or fail in total or only the
specific task related to the operator (running along with the checkpoint task) or nothing

On Tue, May 17, 2016 at 3:49 PM, Robert Metzger <rmetzger@apache.org<mailto:rmetzger@apache.org>>
Hi Stravos,

I haven't implemented our checkpointing mechanism and I didn't participate in the design decisions
while implementing it, so I can not compare it in detail to other approaches.

From a "does it work perspective": Checkpoints are only confirmed if all parallel subtasks
successfully created a valid snapshot of the state. So if there is a failure in the checkpointing
mechanism, no valid checkpoint will be created. The system will recover from the last valid
There is a timeout for checkpoints. So if a barrier doesn't pass through the system for a
certain period of time, the checkpoint is cancelled. The default timeout is 10 minutes.


On Mon, May 16, 2016 at 1:22 PM, Stavros Kontopoulos <st.kontopoulos@gmail.com<mailto:st.kontopoulos@gmail.com>>

I was looking into the flink snapshotting algorithm details also mentioned here:

From other sources i understand that it assumes no failures to work for message delivery or
for example a process hanging for ever:

So my understanding (maybe wrong) is that this is a solution which seems not to address the
fault tolerance issue in a strong manner like for example if it was to use a 3pc protocol
for local state propagation and global agreement. I know the latter is not efficient just
mentioning it for comparison.

How the algorithm behaves in practical terms under the presence of its own failures (this
is a background process collecting partial states)? Are there timeouts for reaching a barrier?

PS. have not looked deep into the code details yet, planning to.


View raw message