Heron provides support for multiple delivery semantics, and you can select delivery semantics on a topology-by-topology basis. Thus, if you have topologies for which at-most-once semantics are perfectly acceptable, for example, you can run them alongside topologies with more stringent semantics (such as effectively once).
Heron supports three delivery semantics:
|Semantics||Description||When to use?|
|At most once||Heron processes tuples using a best-effort strategy. With at-most-once semantics, it's possible that some of the tuples delivered into the system may be lost due to some combination of processing, machine, and network failures. What sets at-most-once semantics apart from the others is that Heron will not attempt to retry a processing step upon failure, which means that the tuple may fail to be delivered.||When some amount of data loss is acceptable|
|At least once||Tuples injected into the Heron topology are guaranteed to be processed at least once; no tuple will fail to be processed. It's possible, however, that any given tuple is processed more than once in the presence of various failures, retries, or other contingencies.||When you need to guarantee no data loss|
|Effectively once||Heron ensures that the data it receives is processed effectively once---even in the presence of various failures---leading to accurate results. This applies only to stateful topologies. “Effectively” in this case means that there's a guarantee that tuples that cause state changes will be processed once (that is, they will have an effect on state once).||When you're using stateful topologies and need strong|
You can see a visual representation of these different delivery semantics in the figure below:
In this diagram, you see three Heron topologies, each of which is processing a series of tuples (
(1,5)tuple fails to be processed.
(7,8,11)tuple is processed more than once (perhaps due to a network glitch or a retry).
In order to use effectively-once semantics with a topology, that topology must satisfy two conditions:
It must be a stateful, idempotent topology.
The input stream into the topology must be strongly consistent. In order to provide effectively-once semantics, topologies need to be able to “rewind” state in case of failure. The state that it “rewinds” needs to be reliable state---preferably durably stored.
If the input to the topology is, for example, a messaging system that cannot ensure stream consistency, then effectively-once semantics cannot be applied, as the state “rewind” may return differing results. To put it somewhat differently, Heron can only provide delivery semantics as stringent as its data input sources can themselves provide.
There has been a lot of discussion recently surrounding so-called “exactly-once” processing semantics. We‘ll avoid this term in the Heron documentation because we feel that it’s misleading. “Exactly-once” semantics would mean that no processing step is ever performed more than once---and thus that no processing step is ever retried.
It‘s important to always keep in mind that no system can provide exactly-once semantics in the face of failures (as this article argues). But that’s okay because they don't really need to; the truly important thing is that a stream processing system be able to recover from failures by “rewinding” state to a previous, pre-failure point and to re-attempt to apply processing logic. We use the tern effectively once, following Victor Klang, for this style of semantics.
Heron can provide effectively-once guarantees if a topology meets the conditions outlined above, but it cannot provide “exactly-once” semantics.
The Heron topologies that you create can be either stateful or non stateful.
Stateful topologies, in turn, are of two types:
Idempotent stateful topologies are stateful topologies in which applying the processing graph to an input more than once, it will continue to return the same result. A basic example is multiplying a number by 0. The first time you do so, the number will change (always to 0), but if you apply that transformation again and again, it will not change.
For topologies to provide effectively-once semantics, they need to transform tuple inputs idempotently as well. If they don‘t, and applying the topology’s processing graph multiple times yields different results, then effectively-once semantics cannot be achieved.
If you'd like to create idempotent stateful topologies, make sure to write tests to ensure that idempotency requirements are being met.
Non-idempotent stateful topologies are stateful topologies that do not apply processing logic along the model of “multiply by zero” and thus cannot provide effectively-once semantics. An example of a non-idempotent
Remember: effectively-once semantics can only be applied to topologies that are: (a) stateful and (b) idempotent.