Hi just my few thoughts
On 05.06.2017 11:44, Eno Thereska wrote:
there is no reasonable way to "skip" a crc error. How can you know the length you read was anything reasonable? you might be completely lost inside your response.Hi there, Sorry for the late reply, I was out this past week. Looks like good progress was made with the discussions either way. Let me recap a couple of points I saw into one big reply: 1. Jan mentioned CRC errors. I think this is a good point. As these happen in Kafka, before Kafka Streams gets a chance to inspect anything, I'd like to hear the opinion of more Kafka folks like Ismael or Jason on this one. Currently the documentation is not great with what to do once a CRC check has failed. From looking at the code, it looks like the client gets a KafkaException (bubbled up from the fetcher) and currently we in streams catch this as part of poll() and fail. It might be advantageous to treat CRC handling in a similar way to serialisation handling (e.g., have the option to fail/skip). Let's see what the other folks say. Worst-case we can do a separate KIP for that if it proved too hard to do in one go.
2. Damian has convinced me that the KIP should just be for deserialisation from the network, not from local state store DBs. For the latter we'll follow the current way of failing since the DB is likely corrupt. 3. Dead letter queue option. There was never any intention here to do anything super clever like attempt to re-inject the failed records from the dead letter queue back into the system. Reasoning about when that'd be useful in light of all sorts of semantic breakings would be hard (arguably impossible). The idea was to just have a place to have all these dead records to help with subsequent debugging. We could also just log a whole bunch of info for a poison pill record and not have a dead letter queue at all. Perhaps that's a better, simpler, starting point.
+1
Eos will not help the record might be 5,6 repartitions down into the topology. I haven't followed but I pray you made EoS optional! We don't need this and we don't want this and we will turn it off if it comes. So I wouldn't recommend relying on it. The option to turn it off is better than forcing it and still beeing unable to rollback badpills (as explained before)4. Agree with Jay on style, a DefaultHandler with some config options. Will add options to KIP. Also as part of this let's remove the threshold logger since it gets complex and arguably the ROI is low. 5. Jay's JSON example, where serialisation passes but the JSON message doesn't have the expected fields, is an interesting one. It's a bit complicated to handle this in the middle of processing. For example, some operators in the DAG might actually find the needed JSON fields and make progress, but other operators, for the same record, might not find their fields and will throw an exception. At a minimum, handling this type of exception will need to involve the exactly-once (EoS) logic. We'd still allow the option of failing or skipping, but EoS would need to clean up by rolling back all the side effects from the processing so far. Matthias, how does this sound?
6. Will add an end-to-end example as Michael suggested. Thanks EnoOn 4 Jun 2017, at 02:35, Matthias J. Sax <matth...@confluent.io> wrote: What I don't understand is this:From there on its the easiest way forward: fix, redeploy, start => doneIf you have many producers that work fine and a new "bad" producer starts up and writes bad data into your input topic, your Streams app dies but all your producers, including the bad one, keep writing. Thus, how would you fix this, as you cannot "remove" the corrupted date from the topic? It might take some time to identify the root cause and stop the bad producer. Up to this point you get good and bad data into your Streams input topic. If Streams app in not able to skip over those bad records, how would you get all the good data from the topic? Not saying it's not possible, but it's extra work copying the data with a new non-Streams consumer-producer-app into a new topic and than feed your Streams app from this new topic -- you also need to update all your upstream producers to write to the new topic. Thus, if you want to fail fast, you can still do this. And after you detected and fixed the bad producer you might just reconfigure your app to skip bad records until it reaches the good part of the data. Afterwards, you could redeploy with fail-fast again. Thus, for this pattern, I actually don't see any reason why to stop the Streams app at all. If you have a callback, and use the callback to raise an alert (and maybe get the bad data into a bad record queue), it will not take longer to identify and stop the "bad" producer. But for this case, you have zero downtime for your Streams app. This seems to be much simpler. Or do I miss anything? Having said this, I agree that the "threshold based callback" might be questionable. But as you argue for strict "fail-fast", I want to argue that this must not always be the best pattern to apply and that the overall KIP idea is super useful from my point of view. -Matthias On 6/3/17 11:57 AM, Jan Filipiak wrote:Could not agree more! But then I think the easiest is still: print exception and die. From there on its the easiest way forward: fix, redeploy, start => done All the other ways to recover a pipeline that was processing partially all the time and suddenly went over a "I cant take it anymore" threshold is not straight forward IMO. How to find the offset, when it became to bad when it is not the latest commited one? How to reset there? with some reasonable stuff in your rockses? If one would do the following. The continuing Handler would measure for a threshold and would terminate after a certain threshold has passed (per task). Then one can use offset commit/ flush intervals to make reasonable assumption of how much is slipping by + you get an easy recovery when it gets to bad + you could also account for "in processing" records. Setting this threshold to zero would cover all cases with 1 implementation. It is still beneficial to have it pluggable Again CRC-Errors are the only bad pills we saw in production for now. Best Jan On 02.06.2017 17:37, Jay Kreps wrote:Jan, I agree with you philosophically. I think one practical challenge has to do with data formats. Many people use untyped events, so there is simply no guarantee on the form of the input. E.g. many companies use JSON without any kind of schema so it becomes very hard to assert anything about the input which makes these programs very fragile to the "one accidental message publication that creates an unsolvable problem. For that reason I do wonder if limiting to just serialization actually gets you a useful solution. For JSON it will help with the problem of non-parseable JSON, but sounds like it won't help in the case where the JSON is well-formed but does not have any of the fields you expect and depend on for your processing. I expect the reason for limiting the scope is it is pretty hard to reason about correctness for anything that stops in the middle of processing an operator DAG? -Jay On Fri, Jun 2, 2017 at 4:50 AM, Jan Filipiak <jan.filip...@trivago.com> wrote:IMHO your doing it wrong then. + building to much support into the kafka eco system is very counterproductive in fostering a happy userbase On 02.06.2017 13:15, Damian Guy wrote:Jan, you have a choice to Fail fast if you want. This is about giving people options and there are times when you don't want to fail fast. On Fri, 2 Jun 2017 at 11:00 Jan Filipiak <jan.filip...@trivago.com> wrote: Hi1. That greatly complicates monitoring. Fail Fast gives you that when you monitor only the lag of all your apps you are completely covered. With that sort of new application Monitoring is very much more complicated as you know need to monitor fail % of some special apps aswell. In my opinion that is a huge downside already. 2. using a schema regerstry like Avrostuff it might not even be the record that is broken, it might be just your app unable to fetch a schema it needs now know. Maybe you got partitioned away from that registry. 3. When you get alerted because of to high fail percentage. what are the steps you gonna do? shut it down to buy time. fix the problem. spend way to much time to find a good reprocess offset. Your timewindows are in bad shape anyways, and you pretty much lost. This routine is nonsense. Dead letter queues would be the worst possible addition to the kafka toolkit that I can think of. It just doesn't fit the architecture of having clients falling behind is a valid option. Further. I mentioned already the only bad pill ive seen so far is crc errors. any plans for those? Best Jan On 02.06.2017 11:34, Damian Guy wrote:I agree with what Matthias has said w.r.t failing fast. There are plentyoftimes when you don't want to fail-fast and must attempt to makeprogress.The dead-letter queue is exactly for these circumstances. Of course if every record is failing, then you probably do want to give up. On Fri, 2 Jun 2017 at 07:56 Matthias J. Sax <matth...@confluent.io>wrote:First a meta comment. KIP discussion should take place on the dev list-- if user list is cc'ed please make sure to reply to both lists.Thanks. Thanks for making the scope of the KIP clear. Makes a lot of sense tofocus on deserialization exceptions for now. With regard to corrupted state stores, would it make sense to fail a task and wipe out the store to repair it via recreation from the changelog? That's of course a quite advance pattern, but I want to bring it up to design the first step in a way such that we can get there (if we think it's a reasonable idea). I also want to comment about fail fast vs making progress. I think that fail-fast must not always be the best option. The scenario I have in mind is like this: you got a bunch of producers that feed the Streams input topic. Most producers work find, but maybe one producer miss behaves and the data it writes is corrupted. You might not even be able to recover this lost data at any point -- thus, there is no reason to stop processing but you just skip over those records. Of course, you need to fix the root cause, and thus you need to alert (either via logs of the exception handler directly) and you need to start to investigate to find the bad producer, shut it down and fix it. Here the dead letter queue comes into place. From my understanding, the purpose of this feature is solely enable post debugging. I don't think those record would be fed back at any point in time (so I don't see any ordering issue -- a skipped record, with this regard, is just "fully processed"). Thus, the dead letter queue should actually encode the original records metadata (topic, partition offset etc) to enable such debugging. I guess, this might also be possible if you just log the bad records, but it would be harder to access (you first must find the Streams instance that did write the log and extract the information from there). Reading it from topic is much simpler. I also want to mention the following. Assume you have such a topic with some bad records and some good records. If we always fail-fast, it's going to be super hard to process the good data. You would need to write an extra app that copied the data into a new topic filtering out the bad records (or apply the map() workaround withing stream). So I don't think that failing fast is most likely the best option in production is necessarily, true. Or do you think there are scenarios, for which you can recover the corrupted records successfully? And even if this is possible, it might be a case for reprocessing instead of failing the whole application? Also, if you think you can "repair" a corrupted record, should the handler allow to return a "fixed" record? This would solve the ordering problem. -Matthias On 5/30/17 1:47 AM, Michael Noll wrote:Thanks for your work on this KIP, Eno -- much appreciated! - I think it would help to improve the KIP by adding an end-to-end code example that demonstrates, with the DSL and with the Processor API, howtheuser would write a simple application that would then be augmented withtheproposed KIP changes to handle exceptions. It should also become much clearer then that e.g. the KIP would lead to different code paths forthehappy case and any failure scenarios.- Do we have sufficient information available to make informeddecisionsonwhat to do next? For example, do we know in which part of the topologytherecord failed? `ConsumerRecord` gives us access to topic, partition, offset, timestamp, etc., but what about topology-related information(e.g.what is the associated state store, if any)? - Only partly on-topic for the scope of this KIP, but this is about the bigger picture: This KIP would give users the option to send corrupted records to dead letter queue (quarantine topic). But, what patternwouldwe advocate to process such a dead letter queue then, e.g. how to allowforretries with backoff ("If the first record in the dead letter queuefailsagain, then try the second record for the time being and go back to thefirst record at a later time"). Jay and Jan already alluded toorderingproblems that will be caused by dead letter queues. As I said, retriesmight be out of scope but perhaps the implications should be consideredifpossible? Also, I wrote the text below before reaching the point in theconversationthat this KIP's scope will be limited to exceptions in the category of poison pills / deserialization errors. But since Jay brought up usercodeerrors again, I decided to include it again. ----------------------------snip---------------------------- A meta comment: I am not sure about this split between the code for the happy path (e.g. map/filter/... in the DSL) from the failure path(usingexception handlers). In Scala, for example, we can do:scala> val computation = scala.util.Try(1 / 0) computation: scala.util.Try[Int] = Failure(java.lang.ArithmeticException: / by zero) scala> computation.getOrElse(42) res2: Int = 42 Another example with Scala's pattern matching, which is similar to `KStream#branch()`: computation match { case scala.util.Success(x) => x * 5 case scala.util.Failure(_) => 42 } (The above isn't the most idiomatic way to handle this in Scala, butthat'snot the point I'm trying to make here.) Hence the question I'm raising here is: Do we want to have an API whereyoucode "the happy path", and then have a different code path for failures (using exceptions and handlers); or should we treat both Success and Failure in the same way? I think the failure/exception handling approach (as proposed in thisKIP)is well-suited for errors in the category of deserialization problemsakapoison pills, partly because the (default) serdes are defined throughconfiguration (explicit serdes however are defined through API calls). However, I'm not yet convinced that the failure/exception handlingapproachis the best idea for user code exceptions, e.g. if you fail to guard against NPE in your lambdas or divide a number by zero. scala> val stream = Seq(1, 2, 3, 4, 5) stream: Seq[Int] = List(1, 2, 3, 4, 5) // Here: Fallback to a sane default when encountering failedrecordsscala> stream.map(x => Try(1/(3 - x))).flatMap(t =>Seq(t.getOrElse(42))) res19: Seq[Int] = List(0, 1, 42, -1, 0) // Here: Skip over failed records scala> stream.map(x => Try(1/(3 - x))).collect{ case Success(s)=> s}res20: Seq[Int] = List(0, 1, -1, 0) The above is more natural to me than using error handlers to define howtodeal with failed records (here, the value `3` causes an arithmetic exception). Again, it might help the KIP if we added an end-to-endexamplefor such user code errors. ----------------------------snip---------------------------- On Tue, May 30, 2017 at 9:24 AM, Jan Filipiak <jan.filip...@trivago.com>wrote:Hi Jay,Eno mentioned that he will narrow down the scope to onlyConsumerRecorddeserialisation.I am working with Database Changelogs only. I would really not like tosee a dead letter queue or somethingsimilliar. how am I expected to get these back in order. Just grind to hold an call me on the weekend. I'll fix it then in a few minutes rather spend 2 weeks ordering dead letters.(wherereprocessing might be even the faster fix)Best Jan On 29.05.2017 20:23, Jay Kreps wrote: - I think we should hold off on retries unless we have worked outthefull usage pattern, people can always implement their own. Ithinkthe ideais that you send the message to some kind of dead letter queueandthenreplay these later. This obviously destroys all semanticguaranteeswe areworking hard to provide right now, which may be okay.