Hi Rahul, This pipeline should process millions of records per day with low latency. > I am avoiding Checkpointing, as the records in the Window operator and > in-flight records in the Async I/O operator are persisted along with the > Kafka offsets. But the records in Window and Async I/O operators can be > obtained just by replaying the records from Kafka Source, which is the > approach I want to take. > There is Deduplication logic in the pipeline. So, I am preferring to > replay records in case of failures rather than storing the incremental > snapshots. > Did you measure how much data is being checkpointed? A few million records per day should just be a few megabytes per checkpoint. It sounds to me as if you would rather want your approach because it's conceptually more sound but not because it is necessary.
I'm asking because delaying records and using async I/O is increasing your latency significantly anyways. So a couple of additional ms to the checkpoint don't sound like it will invalidate your use case to me. It will also be cheaper if you factor in your work time and will also work if you ever extend your pipeline to hold state. Recovery should also not be worse because you restore the records from blob storage instead of fetching it from the source system. Also let me repeat these questions > 1. Why do you need the window operator at all? Couldn't you just > backpressure on the async I/O by delaying the processing there? > 2. What's keeping you from attaching the offset of the Kafka records to A, > B, C and write the offset when writing the record into the sink? (Probably > need to wrap your sink function into a delegating sink function) > On Tue, Apr 13, 2021 at 12:33 PM Rahul Patwari <rahulpatwari8...@gmail.com> wrote: > Hi Arvid, > > Thanks for the reply. > > could you please help me to understand how the at least once guarantee >> would work without checkpointing in your case? >> > > This was the plan to maintain "at least once" guarantee: > Logic at Sink: > The DataStream on which Sink Function is applied, on the same DataStream, > apply a widowing operator and compute min of the timestamps of records and > persist the timestamp in Cassandra -> This will persist the record > timestamp below which all the records are processed, say, every 10 seconds. > (The delay created by the Windowing Operator used here makes sure that the > timestamp is persisted in Cassandra only after it is written to Sink) > Note: There is another Windowing Operator at the Source to delay the > processing of records. > > Logic at Source: > While creating the JobGraph, read the timestamp persisted in Cassandra for > each topic and configure the start position of Kafka Consumer. > > The problem is that the start positions are not respected when there are > Automatic restarts during failures. Basically, we wanted to read the > timestamp from Cassandra and start consuming from the timestamp even in > case of Automatic restarts during failures. > > This pipeline should process millions of records per day with low latency. > I am avoiding Checkpointing, as the records in the Window operator and > in-flight records in the Async I/O operator are persisted along with the > Kafka offsets. But the records in Window and Async I/O operators can be > obtained just by replaying the records from Kafka Source, which is the > approach I want to take. > There is Deduplication logic in the pipeline. So, I am preferring to > replay records in case of failures rather than storing the incremental > snapshots. > > Thanks, > Rahul > > On Tue, Apr 13, 2021 at 2:53 PM Arvid Heise <ar...@apache.org> wrote: > >> Hi Rahul, >> >> could you please help me to understand how the at least once guarantee >> would work without checkpointing in your case? >> >> Let's say you read records A, B, C. You use a window to delay processing, >> so let's say A passes and B, C are still in the window for the trigger. >> >> Now do you want to auto commit the offset of A after it being written in >> the sink? If so, what's keeping you from attaching the offset of the Kafka >> records to A, B, C and write the offset when writing the record into the >> sink? (Probably need to wrap your sink function into a delegating sink >> function) >> >> The way, Flink does the checkpointing is that it checkpoints the offset >> of C, and the state of the window (containing B, C) to avoid data loss. Why >> is that not working for you? Which state size do you expect? >> >> Why do you need the window operator at all? Couldn't you just >> backpressure on the async I/O by delaying the processing there? Then there >> would be no need to change anything. >> >> On Tue, Apr 13, 2021 at 10:00 AM Roman Khachatryan <ro...@apache.org> >> wrote: >> >>> Hi Rahul, >>> >>> Right. There are no workarounds as far as I know. >>> >>> Regards, >>> Roman >>> >>> On Mon, Apr 12, 2021 at 9:00 PM Rahul Patwari >>> <rahulpatwari8...@gmail.com> wrote: >>> > >>> > Hi Roman, Arvid, >>> > >>> > So, to achieve "at least once" guarantee, currently, automatic restart >>> of Flink should be disabled? >>> > Is there any workaround to get "at least once" semantics with Flink >>> Automatic restarts in this case? >>> > >>> > Regards, >>> > Rahul >>> > >>> > On Mon, Apr 12, 2021 at 9:57 PM Roman Khachatryan <ro...@apache.org> >>> wrote: >>> >> >>> >> Hi, >>> >> >>> >> Thanks for the clarification. >>> >> >>> >> > Other than managing offsets externally, Are there any other ways to >>> guarantee "at least once" processing without enabling checkpointing? >>> >> >>> >> That's currently not possible, at least with the default connector. >>> >> >>> >> Regards, >>> >> Roman >>> >> >>> >> On Mon, Apr 12, 2021 at 3:14 PM Rahul Patwari >>> >> <rahulpatwari8...@gmail.com> wrote: >>> >> > >>> >> > Hi Roman, >>> >> > >>> >> > Thanks for the reply. >>> >> > This is what I meant by Internal restarts - Automatic restore of >>> Flink Job from a failure. For example, pipeline restarts when Fixed delay >>> or Failure Rate restart strategies are configured. >>> >> > >>> >> > Quoting documentation in this link - Configuring Kafka Consumer >>> start position configuration >>> >> > >>> >> >> Note that these start position configuration methods do not affect >>> the start position when the job is automatically restored from a failure >>> >> > >>> >> > >>> >> > >>> >> > It seems that there will be data loss even when offsets are managed >>> externally when there are pipeline restarts due to a failure, say, an >>> exception. On the other hand, when the pipeline is stopped and >>> resubmitted(say, an upgrade), there won't be any data loss as offsets are >>> retrieved from an external store and configured while starting Kafka >>> Consumer. >>> >> > >>> >> > We do not want to enable checkpointing as the pipeline is >>> stateless. We have Deduplication logic in the pipeline and the processing >>> is idempotent. >>> >> > >>> >> > Other than managing offsets externally, Are there any other ways to >>> guarantee "at least once" processing without enabling checkpointing? >>> >> > >>> >> > Thanks, >>> >> > Rahul >>> >> > >>> >> > On Mon, Apr 12, 2021 at 6:00 PM Roman Khachatryan <ro...@apache.org> >>> wrote: >>> >> >> >>> >> >> Hi, >>> >> >> >>> >> >> Could you please explain what you mean by internal restarts? >>> >> >> >>> >> >> If you commit offsets or timestamps from sink after emitting >>> records >>> >> >> to the external system then there should be no data loss. >>> >> >> Otherwise (if you commit offsets earlier), you have to persist >>> >> >> in-flight records to avoid data loss (i.e. enable checkpointing). >>> >> >> >>> >> >> Regards, >>> >> >> Roman >>> >> >> >>> >> >> On Mon, Apr 12, 2021 at 12:11 PM Rahul Patwari >>> >> >> <rahulpatwari8...@gmail.com> wrote: >>> >> >> > >>> >> >> > Hello, >>> >> >> > >>> >> >> > Context: >>> >> >> > >>> >> >> > We have a stateless Flink Pipeline which reads from Kafka topics. >>> >> >> > The pipeline has a Windowing operator(Used only for introducing >>> a delay in processing records) and AsyncI/O operators (used for >>> Lookup/Enrichment). >>> >> >> > >>> >> >> > "At least Once" Processing semantics is needed for the pipeline >>> to avoid data loss. >>> >> >> > >>> >> >> > Checkpointing is disabled and we are dependent on the auto >>> offset commit of Kafka consumer for fault tolerance currently. >>> >> >> > >>> >> >> > As auto offset commit indicates that "the record is successfully >>> read", instead of "the record is successfully processed", there will be >>> data loss if there is a restart when the offset is committed to Kafka but >>> not successfully processed by the Flink Pipeline, as the record is NOT >>> replayed again when the pipeline is restarted. >>> >> >> > >>> >> >> > Checkpointing can solve this problem. But, since the pipeline is >>> stateless, we do not want to use checkpointing, which will persist all the >>> records in Windowing Operator and in-flight Async I/O calls. >>> >> >> > >>> >> >> > Question: >>> >> >> > >>> >> >> > We are looking for other ways to guarantee "at least once" >>> processing without checkpointing. One such way is to manage Kafka Offsets >>> Externally. >>> >> >> > >>> >> >> > We can maintain offsets of each partition of each topic in >>> Cassandra(or maintain timestamp, where all records with timestamps less >>> than this timestamp are successfully processed) and configure Kafka >>> consumer Start Position - setStartFromTimestamp() or >>> setStartFromSpecificOffsets() >>> >> >> > >>> >> >> > This will be helpful if the pipeline is manually restarted (say, >>> JobManager pod is restarted). But, how to avoid data loss in case of >>> internal restarts? >>> >> >> > >>> >> >> > Has anyone used this approach? >>> >> >> > What are other ways to guarantee "at least once" processing >>> without checkpointing for a stateless Flink pipeline? >>> >> >> > >>> >> >> > Thanks, >>> >> >> > Rahul >>> >>
