flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Kostas Kloudas <k.klou...@data-artisans.com>
Subject Re: Timer fault tolerance in Flink
Date Wed, 17 May 2017 11:42:33 GMT
Hi Rahul,

The timers are fault tolerant and their timestamp is the absolute value of when to fire.
This means that if you are at time t = 10 and you register a timer “10 ms from now”, the
timer will have a firing timestamp of 20.
This is checkpointed, so the “new machine” that takes over the failed task, will have
the timer with timestamp 20.

So the when the timer will fire depends on the “new machine” and it may differ from what
would happen in the previous machine in the 
following cases:
	For processing time, in case your new machine (the one that takes over the failed task) has
a clock that is out-of-sync with the 
previous machine that set the timer to 20.
	For event time, given that Flink does not checkpoint watermarks, the timer will fire when
the watermark on the new machine surpasses 
the timer's timestamp.

I hope this helps,

> On May 17, 2017, at 12:36 PM, Rahul Kavale <kavale.rahul@gmail.com> wrote:
> I am looking at timers in apache flink and wanted to confirm if the timers in flink are
fault tolerant.
> eg. when a timer registered with processFunction, of say 20 sec is running on a node
and after 15 seconds (since the timer started), the node failed for some reason. Does flink
guarantee that the timer resume on another node? if it does resume does it consider only the
remaining time for the timer ie 5 sec in this case?
> Thanks & Regards,
> Rahul

View raw message