This is the context:
There is an input event stream,
There are some methods to apply on
the stream, which applies different logic to evaluates each event,
saying it is a "good" or "bad" event.
An event can be a real "good" one only if it passes all the methods, otherwise it is a "bad" event.
There is an output event stream who has result of event and its eventID.
To solve this problem, I have two ideas:
We can apply each method sequentially to each event. But this is a kind of batch processing, and doesn't apply the advantages of stream processing, in the same time, it takes Time(M(ethod)1) + Time(M2) + Time(M3) + ....., which maybe not suitable to real-time processing.
We can pass the input stream to each method, and then we can run each method in parallel, each method saves the bad event into a permanent storage, then the Main method could query the permanent storage to get the result of each event. But this has some problems to solve:
how to execute methods in parallel in the programming language(e.g. Scala), how about the performance(network, CPUs, memory)
how to solve the synchronization problem? It's sure that those methods need sometime to calculate and save flag into the permanent storage, but the Main just need less time to query the flag， which a delay issue occurs.
etc.
This is not a kind of tech and design question, I would like to ask your guys' ideas, if you have some new ideas or ideas to solve the problem ? Looking forward to your opinions.
Parallel streams, each doing the full set of evaluations sequentially, is the more straightforward solution. But if that introduces too much latency, then you can fan out the evaluations to be done in parallel, and then bring the results back together again to make a decision.
To do the fan-out, look at the split operation on DataStream, or use side outputs. But before doing this n-way fan-out, make sure that each event has a unique ID. If necessary, add a field containing a random number to each event to use as the unique ID. Later we will use this unique ID as a key to gather back together all of the partial results for each event.
Once the event stream is split, each copy of the stream can use a MapFunction to compute one of evaluation methods.
Gathering all of these separate evaluations of a given event back together is a bit more complex. One reasonable approach here is to union all of the result streams together, and then key the unioned stream by the unique ID described above. This will bring together all of the individual results for each event. Then you can use a RichFlatMapFunction (using Flink's keyed, managed state) to gather the results for the separate evaluations in one place. Once the full set of evaluations for a given event has arrived at this stateful flatmap operator, it can compute and emit the final result.
Related
I have a stream of measurements keyed by an ID PCollection<KV<ID,Measurement>> and something like a changelog stream of additional information for that ID PCollection<KV<ID,SomeIDInfo>>. New data is added to the measurement stream quite regularly, say once per second for every ID. The stream with additional information on the other hand is only updated when a user performs manual re-configuration. We can't tell often this happens and, in particular, the update frequency may vary among IDs.
My goal is now to enrich each entry in the measurements stream by the additional information for its ID. That is, the output should be something like PCollection<KV<ID,Pair<Measurement,SomeIDInfo>>>. Or, in other words, I would like to do a left join of the measurements stream with the additional information stream.
I would expect this to be a quite common use case. Coming from Kafka Streams, this can be quite easily implemented with a KStream-KTable-Join. With Beam, however, all my approaches so far seem not to work. I already thought about the following ideas.
Idea 1: CoGroupByKey with fixed time windows
Applying a window to the measurements stream would not be an issue. However, as the additional information stream is updating irregularly and also significantly less frequently than the measurements stream, there is no reasonable common window size such that there is at least one updated information for each ID.
Idea 2: CoGroupByKey with global window and as non-default trigger
Refining the previous idea, I thought about using a processing-time trigger, which fires e.g. every 5 seconds. The issue with this idea is that I need to use accumulatingFiredPanes() for the additional information as there might be no new data for a key between two firings, but I have to use discardingFiredPanes() for the measurements stream as otherwise my panes would quickly become too large. This simply does not work. When I configure my pipeline that way, also the additional information stream discards changes. Setting both trigger to accumulating it works, but, as I said, this is not scalable.
Idea 3: Side inputs
Another idea would be to use side inputs, but also this solution is not really scalable - at least if I don't miss something. With side inputs, I would create a PCollectionView from the additional information stream, which is a map of IDs to the (latest) additional information. The "join" can than be done in a DoFn with a side input of that view. However, the view seems to be shared by all instances that perform the side input. (It's a bit hard to find any information regarding this.) We would like to not make any assumptions regarding the amount of IDs and the size of additional info. Thus, using a side input seems also not to work here.
The side input option you discuss is currently the best option, although you are correct about the scalability concern due to the side input being broadcast to all workers.
Alternatively, you can store the infrequently-updated side in an external key-value store and just do lookups from a DoFn. If you go this route, it's generally useful to do a GroupByKey first on the main input with ID as a key, which lets you cache the lookups with a good cache-hit ratio.
The Streams DSL documentation includes a caveat about using the aggregate method to transform a KGroupedTable → KTable, as follows (emphasis mine):
When subsequent non-null values are received for a key (e.g., UPDATE), then (1) the subtractor is called with the old value as stored in the table and (2) the adder is called with the new value of the input record that was just received. The order of execution for the subtractor and adder is not defined.
My interpretation of that last line implies that one of three things can happen:
subtractor can be called before adder
adder can be called before subtractor
adder and subtractor could be called at the same time
Here is the question I'm looking to get answered:
Are all 3 scenarios above actually possible when using the aggregate method on a KGroupedTable?
Or am I misinterpreting the documentation? For my use-case (detailed below), it would be ideal if the subtractor was always be called before the adder.
Why is this question important?
If the adder and subtractor are non-commutative operations and the order in which they are executed can vary, you can end up with different results depending on the order of execution of adder and subtractor. An example of a useful non-commutative operation would be something like if we’re aggregating records into a Set:
.aggregate[Set[Animal]](Set.empty)(
adder = (zooKey, animalValue, setOfAnimals) => setOfAnimals + animalValue,
subtractor = (zooKey, animalValue, setOfAnimals) => setOfAnimals - animalValue
)
In this example, for duplicated events, if the adder is called before the subtractor you would end up removing the value entirely from the set (which would be problematic for most use-cases I imagine).
Why am I doubting the documentation (assuming my interpretation of it is correct)?
Seems like an unusual design choice
When I've run unit tests (using TopologyTestDriver and
EmbeddedKafka), I always see the subtractor is called before the
adder. Unfortunately, if there is some kind of race condition
involved, it's entirely possible that I would never hit the other
scenarios.
I did try looking into the kafka-streams codebase as well. The KTableProcessorSupplier that calls the user-supplied adder/subtracter functions appears to be this one: https://github.com/apache/kafka/blob/18547633697a29b690a8fb0c24e2f0289ecf8eeb/streams/src/main/java/org/apache/kafka/streams/kstream/internals/KTableAggregate.java#L81 and on line 92, you can even see a comment saying "first try to remove the old value". Seems like this would answer my question definitively right? Unfortunately, in my own testing, what I saw was that the process function itself is called twice; first with a Change<V> value that includes only the old value and then the process function is called again with a Change<V> value that includes only the new value. Unfortunately, I haven't been able to dig deep enough to find the internal code that is generating the old value record and the new value record (upon receiving an update) to determine if it actually produces those records in that order.
The order is hard-coded (ie, no race condition), but there is no guarantee that the order won't change in future releases without notice (ie, it's not a public contract and no KIP is needed to change it). I guess there would be a Jira about it... But as a matter of fact, it does not really matter (detail below).
For the three scenarios you mentioned, the 3rd one cannot happen though: Aggregators are execute in a single thread (per shard) and thus either the adder or subtractor is called first.
first with a Change value that includes only
the old value and then the process function is called again with a Change
value that includes only the new value.
In general, both records might be processed by different threads and thus it's not possible to send only one record. It's just that the TTD simulates a single threaded execution thus both records always end up in the same processor.
Cf TopologyTestDriver sending incorrect message on KTable aggregations
However, the order actually only matters if both records really end up in the same processor (if the grouping key did not change during the upstream update).
Furthermore, the order actually depends not on the downstream aggregate implementation, but on the order of writes into the repartitions topic of the groupBy() and with multiple parallel upstream processor, those writes are interleaved anyway. Thus, in general, you should think of the "add" and "subtract" part as independent entities and not make any assumption about their order (also, even if the key did not change, both records might be interleaved by other records...)
The only guarantee provided is (given that you configured the producer correctly to avoid re-ordering during send()), that if the grouping key does not change, the send of the old and new value will not be re-ordered relative to each other. The order of the send is hard-coded in the upstream processor though:
https://github.com/apache/kafka/blob/trunk/streams/src/main/java/org/apache/kafka/streams/kstream/internals/KTableRepartitionMap.java#L93-L99
Thus, the order of the downstream aggregate processor is actually meaningless.
With Firestore Increment, what happens if you're using it in a Cloud Function and the Cloud Function is accidentally invoked twice?
To make sure that your function behaves correctly on retried execution attempts, you should make it idempotent by implementing it so that an event results in the desired results (and side effects) even if it is delivered multiple times.
E.g. the function is trying to increment a document field by 1
document("post/Post_ID_1").
updateData(["likes" : FieldValue.increment(1)])
So while Increment may be atomic it's not idempotent? If we want to make our counters idempotent we still need to use a transaction and keep track of who was the last person to like the post?
It will increment once for each invocation of the function. If that's not acceptable, you will need to write some code to figure out if any subsequent invocations are valid for your case.
There are many strategies to implement this, and it's up to you to choose one that suits your needs. The usual strategy is to use the event ID in the context object passed to your function to determine if that event has been successfully processed in the past. Maybe this involves storing that record in another document, in Redis, or somewhere that persists long enough for duplicates to be prevented (an hour should be OK).
For a long time now I am trying to wrap my head around RX. And, to be true, I am never sure if I got it - or not.
Today, I found an explanation on http://reactive-extensions.github.com/RxJS/ which - in my opinion - is horrible. It says:
RxJS is to events as promises are to async.
Great. This is a sentence so full of complexity that if you do not have the slightest idea of what RX is about, after that sentence you are quite as dumb as before.
And this is basically my problem: All the explanations in the usual places you find about RX make (at least me) feel dumb. They explain RX as a highly sophisticated concept with lots of highly complicated words and terms and whatsoever, and I am never quite sure what it is about.
So my question is: How would you explain RX to someone who is five years old? I'd like a clear, picturesque explanation of what it is, what it is good for, and what its main concepts are?
So, LINQ (in JavaScript, these are high-level array methods like map, filter, reduce, etc - if you're not a C# dev, just replace that whenever I mention 'LINQ') gives you a bunch of tools that you can apply to Sequences ("Lists" in a crude sense), in order to filter and transform an input into an output (aka "A list that's actually interesting to me"). But what is a list?
What is a List?
A List, is some elements, in a particular order. I can take any list and transform it into a better list with LINQ.
(Not necessarily sorted order, but an order).
An Event is a List
But what about an Event? Let's subscribe to an event:
OnKeyUp += (o,e) => Console.WriteLine(e.Key)
>>> 'H'
>>> 'e'
>>> 'l'
>>> 'l'
>>> 'o'
Hm. That looks like some things, in a particular order. It now suddenly dawns upon you, a list and an event are the same thing!
If Lists and Events are the Same....
...then why can't I transform and filter input events into more interesting events. That's what Rx is. It's taking everything you know about dealing with sequences, including all of the LINQ operators like Select and Where and Aggregate, and applies them to events.
Easy peasy.
A Callback is a Sequence Too
Isn't a Callback just basically an Event that only happens once? Isn't it basically just like a List with one item? Turns out it is, and one of the interesting things about Rx is that it lets us treat Events and Callbacks (and things like Geolocation requests) with the same language (i.e. we can combine the two, or wait for ether one or the other, etc etc).
Along with Paul's excellent answer I'd like to add the concept of pulling vs pushing data.
Pipeline
Lets take the example of some code that generates a series of numbers, and outputs the result. If you think of this as a stream on one end you have a producer that is creating new numbers for you, and on the other end you have a consumer that is doing something with those numbers.
Pull - Primes List
Lets say the producer is generating a list of prime numbers. Normally you would have some function that yields a list of numbers, and every time it returned it would push the next value it has calculated through the pipe to the consumer, which would output that number to the screen.
Prime Generator ---> Console.WriteLine
In this scenario it is easy to see that the producer is doing most of the work, and the consumer would be sitting around waiting for the producer to send the next value. The consumer is pulling on the pipeline, waiting for the producer to return the next value.
Push - Progress percent events from a fast process (Reactive)
Ok, let's say you have a function that is processing 1,000,000 items. Each item takes milliseconds to process, and then the function yields out a percentage value of how far it has gotten. So lots of progress values, very fast.
At the other end of the pipeline you have a progress bar. Now if the progress bar was to handle every update the UI would block trying to keep up with the stream of values.
1-Million-Items-Processor ---> Progress Bar
In this scenario the data is being pushed through the pipeline by the producer and then the consumer is blocking because too much data is being pushed for it to handle.
Reactive allows you to put in delays, windows, or to sample the pipeline depending on how you wish to consume the data. In this case I would sample the data every second before updating the progress bar.
Lists vs Events
So lists and events are kinda the same. The difference is whether the data is pulled or pushed through the system. With lists the data is pulled. With events the data is pushed.
I'm trying to learn EventStore, I like the concept but when I try to apply in practice I'm getting stuck in same point.
Let's see the code:
foreach (var k in stream.CommittedEvents)
{
//handling events
}
Two question about that:
When an app start ups after some maintenance, how do we bookmark in a
safe way what events start to read? Is there a pattern to use?
as soon the events are all consumed, the cycle ends... what about the message arriving run time? I would expect the call blocking until some new message arrive ( of course need to be handled in a thread ) or having something like BeginRead EndRead.
Do I have to bind an ESB to handle run time event or does the EventSore provides some facility to do this?
I try to better explain with an example
Suppose the aggregate is a financial portfolio, and the application is an application showing that portfolio to a trader. Suppose the trader connect to the web app and he looks at his own portfolio. The current state will be the whole history, so I have to read potentially a lot of records to reproduce the status. I guess this could be done by a so called snapshot, but who's responsible for creating it? When one should choose to create an aggregate? How can one guess a snapshot for an aggregate exists ?
For the runtime part: as soon the user look at the reconstructed portfolio state, the real time part begin to run. The user can place an order and a new position can be created by succesfully execute that order in the market. How is the portfolio updated by the infrastructure? I would expect, but maybe I'm completely wrong, having the same event stream being the source of that new event new long position, otherwise I have two path handling the state of the same aggregate. I would like to know if this is how the strategy is supposed to work, even if I feel a little tricky having the two state agents, that can possibly overlap.
Just to clarify how I fear the overlapping:
I know events has to be idempotent, so I know it must not be a
problem anyway,
But let's consider the following:
I subscribe an event bus before streaming the event to update the state of the portfolio. some "open position event" appears on the bus: I must handle them, but maybe the portfolio is not in the correct state to handle it since is not yet actualized. Even if I'm able to handle such events I will find them again when I read the stream.
More insidious: I open the stream and I read all events and I create a state. Then I subscribe to the bus: some message on the bus happen in the middle between the end of the steram reading and the beggining of the subscription: those events are missing and the aggregate is not in the correct state.
Please be patient all, my English is poor and the argument is tricky, hope I managed to share my doubt :)
The current state will be the whole history, so I have to read
potentially a lot of records to reproduce the status. I guess this
could be done by a so called snapshot, but who's responsible for
creating it?
In CQRS and event sourcing, queries are served by projections which are generated from events emitted by aggregates. You don't use the aggregate instance as reconstituted from the event store to display information.
The term snapshot refers specifically to an optimization of the event store which allows rebuilding the aggregate without replaying all of the events.
Projections are essentially event handlers which maintain a denormalized view of aggregates. Events emitted from aggregates are published, possibly out of band, and the projection subscribes to and handles those events. A projection can combine multiple aggregates if a requirement exists to display summary information, for instance. In case of a trading application, each view will typically contain data from various aggregates. Projections are designed in a consumer-driven way - application requirements determine the different views of the underlying data that are needed.
With this type of workflow you have to embrace eventual consistency throughout your application. For instance, if an end user is viewing their portfolio and initiating new trades, the UI has to subscribe to updates to reflect updated projections in an asynchronous manner.
Take a look at here for an overview of CQRS and event sourcing.