Hi, This is slightly different for processing-time and event-time triggers.
First, event-time triggers: there are two data structures, a PriorityQueue (which is implemented as a heap) of timers that is sorted by timestamp, a set of registered timers that is used for deduplication. When adding a timer, we first check whether it already exists (using the set) and then add it to the queue. Whenever we receive a watermark we poll from the timer queue as long as the timestamp of the top timer is <= the watermark. We remote the timer from the set and call the user callback. For processing-time triggers it's very similar, except that we use a ProcessingTimeService instead of the watermark for advancing time. We always have one "physical" processing-time timer set at the ProcessingTimeService. When this fires we follow the same procedure as for event-time and then register a new "physical" timer for the next lowest processing-time timer. In you case this would mean 3 separate internal timers, but a timer is only a timestamp and a key (and a namespace). Best, Aljoscha > On 13. Oct 2017, at 13:56, Kien Truong <duckientru...@gmail.com> wrote: > > Hi Aljoscha, > > Could you clarify how the timer system works right now ? > > For example, let's say I have a function F, with 3 keys that are registered > to execute at processing time T. > Would Flink maintain a single internal timer at time T, then run the callback > on all 3 keys when it's triggered ? Or there'd be 3 internal timers that will > be triggered separately at time T ? > > Best regards, > > Kien > On 10/13/2017 6:43 PM, Aljoscha Krettek wrote: >> Hi, >> >> If you have multiple timers per key, then coalescing can make sense to >> reduce the burden on the timer system. Coalescing them across different keys >> would not be possible right now. >> >> Best, >> Aljoscha >> >> >>> On 13. Oct 2017, at 06:37, Kien Truong <duckientru...@gmail.com> >>> wrote: >>> >>> Hi, >>> >>> We are having a streaming job where we use timers to implement key timeout >>> for stateful functions. Should we implement coalescing logic to reduce the >>> number of timer trigger, or it is not necessary with Flink? >>> >>> Best regards, >>> Kien >>>
